Studies in Computational Intelligence 682

Robert Koprowski

Processing of
Hyperspectral
Medical Images
Applications in Dermatology Using
MATLAB®
Studies in Computational Intelligence

Volume 682

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
About this Series

The series “Studies in Computational Intelligence” (SCI) publishes new develop-

ments and advances in the various areas of computational intelligence—quickly and
with a high quality. The intent is to cover the theory, applications, and design
methods of computational intelligence, as embedded in the fields of engineering,
computer science, physics and life sciences, as well as the methodologies behind
them. The series contains monographs, lecture notes and edited volumes in
computational intelligence spanning the areas of neural networks, connectionist
systems, genetic algorithms, evolutionary computation, artificial intelligence,
cellular automata, self-organizing systems, soft computing, fuzzy systems, and
hybrid intelligent systems. Of particular value to both the contributors and the
readership are the short publication timeframe and the worldwide distribution,
which enable both wide and rapid dissemination of research output.

More information about this series at http://www.springer.com/series/7092

Robert Koprowski

Processing of Hyperspectral
Medical Images
Applications in Dermatology Using
MATLAB®

123
Robert Koprowski
Department of Biomedical Computer
Systems, Institute of Computer Science
University of Silesia
Sosnowiec
Poland

Additional material to this book can be downloaded from http://extras.springer.com.

ISSN 1860-949X ISSN 1860-9503 (electronic)

Studies in Computational Intelligence
ISBN 978-3-319-50489-6 ISBN 978-3-319-50490-2 (eBook)
DOI 10.1007/978-3-319-50490-2

Library of Congress Control Number: 2016958714

® 2016 The MathWorks, Inc. MATLAB and Simulink are registered trade-marks of The MathWorks,
Inc. See www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand
names may be trademarks or registered trademarks of their respective handles.

© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publi-
cation does not imply, even in the absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword

Medical imaging is a field of knowledge dealing with the methods of acquisition

and analysis of images occurring in biological and medical research. The acquired
images are used for research, diagnostic, therapeutic or educational purposes. The
rapid development of diagnostic medical equipment and information technology
enables the growing interaction of these two areas of expertise for the benefit of
patients.
It is generally difficult to show the characteristics of real medical images pic-
torially and in a useful form for a physician. In addition to qualitative assessment,
the physician also needs quantification of medical images, which will illustrate the
various diagnostic parameters of medical objects. On their basis, the physician
makes decisions related to the course of treatment, the strategy and the selection of
appropriate drugs.
Quantitative assessment, achievable through the analysis of biomedical images,
involves profiling contemporary analysis methods and algorithms. Such methods
include not only image filtering but also its morphological and point transforma-
tions as well as their various classifications.
One of the rapidly developing techniques of image registration is the so-called
hyperspectral imaging, which is used, inter alia, in biology and medicine. Issues
related to the development of profiled software allowing for the hyperspectral
analysis of biological and medical images is the goal of this monograph.

Zygmunt Wróbel

v
Preface

Modern methods of infrared, visible light or UV-light imaging are used in many
fields of science, starting with astronomy through biophysics, physics, geography
and ending with medicine. One such method allowing for imaging in a wide
wavelength spectrum is hyperspectral imaging. The use of this type of imaging
provides ample opportunities not only in terms of the qualitative assessment of
acquired images but also in their quantification. The possibility of quantitative
assessment is the result of analysis performed in the software provided with
hyperspectral cameras. However, due to the large amount of data, this software has
numerous limitations and is user-friendly in a limited way. On the other hand, there
are well-known methods of 2D image analysis and processing. Their implemen-
tation in hyperspectral imaging is not an easy task. Apart from the need to extend
2D images into the third dimension (in which respect there are known methods of
image analysis and processing, but in visible light), there remains the issue of
optimization. It concerns optimization of computational complexity, optimization
of analysis time and performance of preliminary calculations commonly used by
users. The tasks that need to be solved by the users analysing hyperspectral medical
images are also specific by their very nature. The specificity of these images stems
directly from the inter-individual variability in patients and thus the images anal-
ysed. For this reason, for almost any task in question, object segmentation,
comparison, calculation of characteristics, individual profiling of an algorithm are
extremely important. Dedicated algorithms enable to conduct automated, repeatable
measurements of, for example, a specific disease entity. However, profiled
(dedicated) algorithms also have drawbacks—data overfitting. Therefore, these
methods must be tested on images acquired under different conditions, with dif-
ferent hardware settings and for different operators, for example, a hyperspectral
camera. Only in this case, it is certain that the proposed new algorithm will meet the
requirements of universality when it comes to the data source and manner of
acquisition and will be profiled for a particular application. Therefore, the key
element is not only to propose new dedicated methods of hyperspectral image
analysis and processing but also to try to implement them in practice and test their
properties.

vii
viii Preface

The presented methods of analysis and processing of hyperspectral medical

images have been tested in practice in the Matlab® environment. The applied source
code is attached to this monograph. The reader does not need to rewrite its frag-
ments from the text. The source code is also described in detail in the monograph.
The monograph is intended for computer scientists, bioengineers, doctoral stu-
dents and dermatologists interested in contemporary analysis methods. It can also
be used to teach senior students of engineering studies related to computer science
if the price of the book does not constitute a barrier. For the full understanding
of the issues discussed, it has been assumed that the reader knows the basic methods
and matrix operations in Matlab and knows the basic functions of Image
Processing, Signal Processing and Statistics Toolboxes. Finally, other group of
readers who want to know the way to solve the discussed problems in the field of
image analysis and processing in Matlab may become interested in this monograph.

Sosnowiec, Poland Robert Koprowski

Acknowledgments

First, I would like to thank Dr. Sławomir Wilczyński from the Medical University of
Silesia in Katowice for the inspiration and consultations in the area of dermatological
issues covered in the monograph. I would also like to thank Mr Raphael Stachiewicz
from Enformatic Sp. z. o. o., Poland, for providing the hyperspectral camera. It
allowed me to test the developed algorithms and perform measurements for objects
other than medical ones.
I also thank Prof. Zygmunt Wróbel from the University of Silesia in Katowice,
Poland, for his daily support and consultations.

ix
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Purpose and Scope of the Monograph . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.1 Coordinate System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.2 Evaluation of the Classifier Quality. . . . . . . . . . . . . . . . . . . 4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Image Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Image Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Affine Transformations of the Image . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Image Filtration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.1 Non-Adaptive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.2 Adaptive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Image Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Preliminary Version of the GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Block Diagram of the Discussed Transformations . . . . . . . . . . . . . 35
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4 Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1 Diagnostic Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Tracking Changes in the Shape of an Object . . . . . . . . . . . . . . . . . 40
4.3 Basic Analysis of Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 Block Diagram of the Discussed Transformations . . . . . . . . . . . . . 67
4.5 Measurement of Additional Features . . . . . . . . . . . . . . . . . . . . . . . 67
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

xi
xii Contents

5 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1 Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2 Naive Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.3 Discriminant Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.5 Receiver Operating Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.6 Pitfalls and Limitations of Classification . . . . . . . . . . . . . . . . . . . . . 98
5.7 Blok Diagram of the Discussed Transformations . . . . . . . . . . . . . . 105
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6 Sensitivity to Parameter Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.1 Respositioning the ROI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.2 Resizing the ROI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.3 Rotation of the ROI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.4 The Other Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Selected Symbols

m,u Row
n,v Column
i Number of the image in a sequence—wavelength
L(m,n,i) Point of the matrix of the image L
L(C) Image resulting from conditional operations
L(D) Image which complements the image L
LBIN Binary image
LGRAY Image in grey levels

Erosion symbol
 Dilation symbol
! Neighbourhood type
h Filter mask
pr Threshold
δg Measurement error
σ Standard deviation of the mean
θ Angle of inclination of the filter
A Amplitude
ACC Accuracy
B Number of bits of an image
FN False negative
FP False positive
I Number of images in a sequence
M Number of rows in a matrix
N Number of columns in a matrix
SE Structural element
SPC Specificity
TN True negative
TP True positive
TPR True positive rate

xiii
Chapter 1
Introduction

1.1 Purpose and Scope of the Monograph

The purpose of this monograph is to present new and known modified methods of
hyperspectral image analysis and processing and profile them in terms of their
usefulness in medical diagnostics and research, as well as develop quantitative
diagnostic tools that can be used in everyday medical practice. The algorithms
proposed in this monograph have the following characteristics:
• they are fully automatic—they do not require operator intervention, if it is
necessary to provide additional parameters of the algorithm operation, they are
selected automatically,
• the results obtained on their basis are fully reproducible,
• their operation was tested on a group of several thousands of hyperspectral
images,
• they were implemented in Matlab,
• they have an open and tested source code attached to this monograph (in the
form of an external link),
• they can be freely extended and modified—owing to the open source code.
The scope of the monograph includes medical images and, in particular, der-
matological ones. However, they are only used to test the discussed methods. The
scope of the monograph is divided into acquisition, image pre-processing, image
processing and their classification presented in the following chapters.

© Springer International Publishing AG 2017 1

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_1
2 1 Introduction

1.2 Material

Most of the images analysed in this monograph had a resolution

M × N × I = 696 × 520 × 128, where M—the number of rows, N—the number of
columns, I—the number of analysed wavelengths. Images of such or similar res-
olution (dependent on individual camera settings) were acquired with different
hyperspectral cameras. The overwhelming part (approximately 75%) of all 200,000
images was registered using the SOC710-VP Hyperspectral Imager with a colour
resolution B = 12 bits and spectral resolution from 400 to 1000 nm. This camera
enables to register 128 bands (I = 128) and is powered by 12 V. The analysed
images were obtained retrospectively and showed the skin of the hand, forearm, and
other areas of the body recorded for dozens of patients. The patients were subject to
exclusion criteria which were undisclosed skin diseases, fever, cardiac arrhythmias,
seizures, inflammation of the skin and pregnancy. The analysed areas were illu-
minated by sunlight or using 40 W halogen lamps of a constant radiation spectrum
ranging from 400 to 1000 nm. All the patients gave informed consent for the study
which was conducted in accordance with the Declaration of Helsinki. No tests,
measurements or experiments were performed on humans as part of this work. This
monograph only deals with the methods of analysis of their images and diagnostic
utility of the obtained results.

1.3 State of the Art

The subject of hyperspectral image analysis and the imaging method itself has been
known for many years. On the day of writing this monograph, the end of 2016, the
PubMed database contained 1922 publications containing the word “hyperspectral”
in the title or description. Slightly different numbers (the number of articles) were
given by the AuthorMapper database, namely 1825 publications, 18,643 authors
from 6105 institutions. A breakdown by countries, institutions, authors, journals
and subjects (the first 5 are listed) is presented in Table 1.1.
As shown above, the leaders in terms of publications on hyperspectral imaging
are the United States, author Chang Chein-I and the area of Computer Science with
1420, 56, 1299 publications respectively. Image Processing and Computer Vision is
a particularly exploited subject, which is extremely important from the point of
view of this monograph. This subject includes such areas as (the number of pub-
lications is given in parentheses): Signal, Image and Video Processing (39); Journal
of Real-Time Image Processing (37); Reference Recognition and Image Analysis
(27); Hyperspectral Imaging (26); Real-Time Progressive Hyperspectral Image
Processing (26); Neural Computing and Applications (25); Advances in Visual
Computing (22) Image Analysis and Recognition (22); Image and Signal
Processing (22); Multiple Classifier Systems (20); Machine Vision and
Applications (19); Hyperspectral Data Compression (17); Advanced Concepts for
1.3 State of the Art 3

Table 1.1 The first 5 countries, 5 institutions, 5 authors, 5 journals, 5 subjects related to the word
“hyperspectral”
Country United China Germany India France
States
Number of 1420 [1–5] 995 [6–10] 455 [11–15] 347 [16–20] 311 [21–25]
publications
Institution Chinese Zhejiang University of University of Wuhan
Academy of University California Maryland University
Sciences
Number of 162 [26–30] 67 [31–34] 64 [35–40] 47 [41–44] 43 [44–50]
publications
Author Chang, Graña, Manuel Sun, Da-Wen Goodacre, Wang, Liguo
Chein-I [51, [53, 54] [55–58] Royston [59, [61, 62]
52] 60]
Number of 56 29 21 18 18
publications
Journal Precision Journal of the Environmental Analytical Environmental
Agriculture Indian Society of Monitoring and and Earth Sciences
Remote Sensing Assessment Bioanalytical
Chemistry
Number of 144 [63, 64] 140 [65, 66] 100 [67, 68 91 [69, 70] 77 [71, 72]
publications
Subject Computer Life Sciences Artificial Earth Image
Science Intelligence Sciences Processing and
(incl. Robotics) Computer
Vision
Number of 1299 [73, 960 [75, 76] 908 [77, 78] 859 [79, 80] 830 [81, 82]
publications 74]

Intelligent Vision Systems (16); Journal of Signal Processing Systems (16);

Mathematical Morphology and Its Applications to Signal and Image Processing
(15); Remote Sensing Digital Image Analysis (15); Image Analysis (14);
Hyperspectral Image Fusion (13); Hyperspectral Image Processing (12); Journal of
Mathematical Imaging and Vision (11).
When reviewing publications [83–91] in terms of the described research prob-
lems and their solutions, several open issues in the field of hyperspectral image
analysis can be observed:
• the need for profiling methods of image analysis and processing to a particular
research problem,
• lack of universal methods of analysis and
• lack of or limited availability of source codes.
Therefore, this monograph describes a sample application for the analysis and
processing of hyperspectral images. The application was profiled to the area of
biomedical engineering, and includes both known and new algorithms for image
analysis and processing.
4 1 Introduction

The discussed scope of biomedical engineering involves the use of hyperspectral

cameras in dermatology. These issues have been partly presented in several pub-
lications [92–96]. Some of them are not profiled to solve a particular segmentation
issue and do not address the issue of the algorithm sensitivity to parameter changes
or the impact of different methods of acquisition on the results obtained.
Accordingly, the analysis of the impact of acquisition on the results obtained, at the
full automation of the proposed algorithm, constitutes another area (chapter) of this
monograph.

1.4 Basic Definitions

Basic definitions apply to two issues:

• orientation of coordinate systems and
• assessment of the classifier quality.
They are described in the following sections.

1.4.1 Coordinate System

The orientation of the coordinate system is strongly dependent on the individual

settings of the camera relative to the object, frame of reference. Regardless of the
individual camera setting, to which all the described algorithms cannot be sensitive,
it was assumed that the size of each image sequence would be defined by the
number of rows M numbered from one, the number of columns N and the number
of individual wavelengths I. The numbering from one and not zero, as in the case of
well-known programming languages C++, C#, results from the adopted nomen-
clature and numbering in Matlab, Scilab or Octave. As a result, it was adopted in
this monograph—Fig. 1.1 and Fig. 1.2.
The presented coordinate system (Fig. 1.1) will be used for all the presented
analyses and algorithms. When an image (matrix) is a single 2D matrix, dimension
I will be 1.

1.4.2 Evaluation of the Classifier Quality

Classifiers were usually induced by using the training data representing 2/3 of the
total number of data. The remaining 1/3 of the data was used to test the classifier
quality [96]. The training and test data were divided randomly. In the cases pre-
sented in this monograph, the division can be distorted. This is due to the fact that
1.4 Basic Definitions 5

Fig. 1.1 The following

symbols were adopted in the

coordinate system: M number
of rows, N number of
columns, I number of
wavelengths (random colours
of individual pixels were
adopted) 


 


 


 





Fig. 1.2 Orientation of the

adopted coordinate system for

a colour image—RGB (pixel
colours correspond to R, G
and B components)

1-R 


2-G 
 

3-B 

 

both the training and test vectors result from calculations for a sequence of images.
The number of images in the sequence to be analysed is affected by the operator.
Therefore, it does not have to be a strict division into 1/3 and 2/3. In each case,
evaluation of the classifier quality was based on determination of four values:
• False Positive (FP)—cases incorrectly classified as positive,
• False Negative (FN)—cases incorrectly classified as negative,
• True Positive (TP)—positive cases classified correctly,
• True Negative (FN)—negative cases classified correctly.
On their basis, sensitivity TPR and specificity SPC were defined as [97, 98]:
6 1 Introduction

TP
TPR ¼  100% ð1:1Þ
TP þ FN
TN
SPC ¼  100% ð1:2Þ
TN þ FP

and accuracy ACC:

TP þ TN
ACC ¼  100% ð1:3Þ
TP þ TN þ FP þ FN

The parameters SPC and TPR will be the basis for creating the receiver operating
characteristic curves (ROC), which are graphs of changes in TPR as a function of
100-SPC [99]. Additionally, the area under the curve (AUC) will be calculated
[100].

References

1. Wang, Liguo; Zhao, Chunhui, Introduction of Hyperspectral Remote Sensing Applications,

Hyperspectral Image Processing: 283–308, (2016).
2. Nguyen, Hien Van; Banerjee, Amit; Burlina, Philippe; Broadwater, Joshua; Chellappa,
Rama, Tracking and Identification via Object Reflectance Using a Hyperspectral Video
Camera, Machine Vision Beyond Visible Spectrum (2011-01-01) 1: 201–219, January 01,
2011.
3. Miguel, Agnieszka; Riskin, Eve; Ladner, Richard; Barney, Dane, Near-lossless and lossy
compression of imaging spectrometer data: comparison of information extraction perfor-
mance, Signal, Image and Video Processing (2012-11-01) 6: 597–611, November 01, 2012.
4. Le, Justin H.; Yazdanpanah, Ali Pour; Regentova, Emma E.; Muthukumar, Venkatesan A
Deep Belief Network for Classifying Remotely-Sensed Hyperspectral Data, Advances in
Visual Computing (2015-01-01): 9474, January 01, 2015.
5. Alam, Mohammad S.; Sakla, Adel Automatic Target Recognition in Multispectral and
Hyperspectral Imagery Via Joint Transform Correlation, Wide Area Surveillance
(2012-09-11) 6: 179–206, September 11, 2012.
6. Miller, Corey A.; Walls, Thomas J., Hyperspectral Scene Analysis via Structure from
Motion, Advances in Visual Computing (2015-01-01): 9474, January 01, 2015.
7. Zou, Xiaobo; Zhao, Jiewen, Hyperspectral Imaging Detection, Nondestructive Measurement
in Food and Agro-products (2015-01-01): 127–193, January 01, 2015.
8. Chen, Yu-Nan; Sun, Da-Wen; Cheng, Jun-Hu; Gao, Wen-Hong, Recent Advances for Rapid
Identification of Chemical Information of Muscle Foods by Hyperspectral Imaging Analysis,
Food Engineering Reviews (2016-04-04): 1–15, April 04, 2016.
9. Wang, Liguo; Zhao, Chunhui, Introduction of Hyperspectral Remote Sensing Applications,
Hyperspectral Image Processing (2016-01-01): 283–308, January 01, 2016.
10. Zhang, Baohua; Fan, Shuxiang; Li, Jiangbo; Huang, Wenqian; Zhao, Chunjiang; Qian, Man;
Zheng, Ling, Detection of Early Rottenness on Apples by Using Hyperspectral Imaging
Combined with Spectral Analysis and Image Processing, Food Analytical Methods
(2015-09-01) 8: 2075–2086, September 01, 2015.
References 7

11. Behmann, Jan; Mahlein, Anne-Katrin; Paulus, Stefan; Dupuis, Jan; Kuhlmann, Heiner;
Oerke, Erich-Christian; Plümer, Lutz, Generation and application of hyperspectral 3D plant
models: methods and challenges, Machine Vision and Applications (2015-10-03): 1–14,
October 03, 2015.
12. Lausch, Angela; Pause, Marion; Merbach, Ines; Zacharias, Steffen; Doktor, Daniel; Volk,
Martin; Seppelt, Ralf, A new multiscale approach for monitoring vegetation using remote
sensing-based indicators in laboratory, field, and landscape, Environmental Monitoring and
Assessment (2013-02-01) 185: 1215–1235, February 01, 2013.
13. Martin, Ron; Thies, Boris; Gerstner, Andreas OH, Hyperspectral hybrid method classifi-
cation for detecting altered mucosa of the human larynx, International Journal of Health
Geographics (2012-06-21) 11: 1–9, June 21, 2012.
14. Bigdeli, Behnaz; Samadzadegan, Farhad; Reinartz, Peter A Multiple SVM System for
Classification of Hyperspectral Remote Sensing Data, Journal of the Indian Society of
Remote Sensing (2013-12-01) 41: 763–776, December 01, 2013.
15. Wang, Ke; Gu, XingFa; Yu, Tao; Meng, QingYan; Zhao, LiMin; Feng, Li Classification of
hyperspectral remote sensing images using frequency spectrum similarity, Science China
Technological Sciences (2013-04-01) 56: 980–988, April 01, 2013.
16. Mookambiga, A.; Gomathi, V.Comprehensive review on fusion techniques for spatial
information enhancement in hyperspectral imagery, Multidimensional Systems and Signal
Processing (2016-04-27): 1–27, April 27, 2016.
17. Chutia, Dibyajyoti; Bhattacharyya, Dhruba Kumar; Kalita, Ranjan; Goswami, Jonali; Singh,
Puyam S.; Sudhakar, S., A model on achieving higher performance in the classification of
hyperspectral satellite data: a case study on Hyperion data, Applied Geomatics (2014-09-01)
6: 181–195, September 01, 2014.
18. Prabhu, N.; Arora, Manoj K.; Balasubramanian, R.Wavelet Based Feature Extraction
Techniques of Hyperspectral Data, Journal of the Indian Society of Remote Sensing
(2016-01-29): 1–12, January 29, 2016.
19. Prabhakar, M.; Prasad, Y. G.; Rao, Mahesh N., Remote Sensing of Biotic Stress in Crop
Plants and Its Applications for Pest Management, Crop Stress and its Management:
Perspectives and Strategies (2012-01-01): 517–545, January 01, 2012.
20. Vetrekar, N. T.; Gad, R. S.; Fernandes, I.; Parab, J. S.; Desai, A. R.; Pawar, J. D.; Naik, G.
M.; Umapathy, S., Non-invasive hyperspectral imaging approach for fruit quality control
application and classification: case study of apple, chikoo, guava fruits, Journal of Food
Science and Technology (2015-11-01) 52: 6978–6989, November 01, 2015.
21. Licciardi, Giorgio Antonino Hyperspectral Data in Urban Areas Encyclopedia of Earthquake
Engineering (2015-01-01): 1155–1164, January 01, 2015.
22. Luo, Bin; Chanussot, Jocelyn Supervised Hyperspectral Image Classification Based on
Spectral Unmixing and Geometrical Features, Journal of Signal Processing Systems
(2011-12-01) 65: 457–468, December 01, 2011.
23. Xia, Junshi; Chanussot, Jocelyn; Du, Peijun; He, Xiyan, Rotation-Based Ensemble
Classifiers for High-Dimensional Data, Fusion in Computer Vision (2014-03-26): 135–
160, March 26, 2014.
24. Hernández-Sánchez, Natalia; Moreda, Guillermo P.; Herre-ro-Langreo, Ana;
Melado-Herreros, Ángela, Assessment of Internal and External Quality of Fruits and
Vegetables, Imaging Technologies and Data Processing for Food Engineers (2016-01-01):
269–309, January 01, 2016.
25. Chen, Weishi; Guillaume, Mireille, HALS-based NMF with flexible constraints for
hyperspectral unmixing, EURASIP Journal on Advances in Signal Processing
(2012-03-05) 2012: 1–14, March 05, 2012.
26. Zhang, Bing; Yang, Wei; Gao, Lianru; Chen, Dongmei Real-time target detection in
hyperspectral images based on spatial-spectral information extraction EURASIP Journal on
Advances in Signal Processing (2012-07-13) 2012: 1–15, July 13, 2012.
8 1 Introduction

27. Gao, Lianru; Zhang, Bing; Sun, Xu; Li, Shanshan; Du, Qian; Wu, Changshan Optimized
maximum noise fraction for dimensionality reduction of Chinese HJ-1A hyperspectral data
EURASIP Journal on Advances in Signal Processing (2013-04-02) 2013: 1–12, April 02,
2013.
28. Gao, Lianru; Zhuang, Lina; Wu, Yuanfeng; Sun, Xu; Zhang, Bing A quantitative and
comparative analysis of different preprocessing implementations of DPSO: a robust
endmember extraction algorithm Soft Computing (2014-11-06): 1–15, November 06, 2014.
29. Liu, Jun; Zhou, Xiran; Huang, Junyi; Liu, Shuguang; Li, Huali; Wen, Shan; Liu, Junchen
Semantic classification for hyperspectral image by integrating distance measurement and
relevance vector machine, Multimedia Systems (2015-03-18): 1–10, March 18, 2015.
30. Zhang, Hao; Hu, Hao; Zhang, Xiaobin; Wang, Kelin; Song, Tongqing; Zeng, Fuping
Detecting Suaeda salsa L. chlorophyll fluorescence response to salinity stress by using
hyperspectral reflectance, Acta Physiologiae Plantarum (2012-03-01) 34: 581–588, March
01, 2012.
31. Yu, Ke-Qiang; Zhao, Yan-Ru; Liu, Zi-Yi; Li, Xiao-Li; Liu, Fei; He, Yong Application of
Visible and Near-Infrared Hyperspectral Imaging for Detection of Defective Features in
Loquat Food and Bioprocess Technology (2014-11-01) 7: 3077–3087, November 01, 2014.
32. Li, Xiaorun; Cui, Jiantao; Zhao, Liaoying Blind nonlinear hyperspectral unmixing based on
constrained kernel nonnegative matrix factorization, Signal, Image and Video Processing
(2014-11-01) 8: 1555–1567, November 01, 2014.
33. Wu, Di; Wang, Songjing; Wang, Nanfei; Nie, Pengcheng; He, Yong; Sun, Da-Wen; Yao,
Jiansong Application of Time Series Hyperspectral Imaging (TS-HSI) for Determining
Water Distribution Within Beef and Spectral Kinetic Analysis During Dehydration Food and
Bioprocess Technology (2013-11-01) 6: 2943–2958, November 01, 2013.
34. Pan, Leiqing; Lu, Renfu; Zhu, Qibing; Tu, Kang; Cen, Haiyan Predict Compositions and
Mechanical Properties of Sugar Beet Using Hyperspectral Scattering, Food and Bioprocess
Technology (2016-03-07): 1–10, March 07, 2016.
35. Martinelli, Federico; Scalenghe, Riccardo; Davino, Salvatore; Panno, Stefano; Scuderi,
Giuseppe; Ruisi, Paolo; Villa, Paolo; Stroppiana, Daniela; Boschetti, Mirco; Goulart,
Luiz R.; Davis, Cristina E.; Dandekar, Abhaya M. Advanced methods of plant disease
detection. A review Agronomy for Sustainable Development (2015-01-01) 35: 1–25,
January 01, 2015.
36. Khorram, Siamak; Koch, Frank H.; Wiele, Cynthia F.; Nelson, Stacy A. C.Data Acquisition
Remote Sensing (2012-01-01): 17–37, January 01, 2012.
37. Pan, Zhihong; Healey, Glenn; Tromberg, Bruce Comparison of Spectral-Only and
Spectral/Spatial Face Recognition for Personal Identity Verification EURASIP Journal on
Advances in Signal Processing (2009-05-26) 2009: 1–6, May 26, 2009.
38. Underwood, E. C.; Mulitsch, M. J.; Greenberg, J. A.; Whiting, M. L.; Ustin, S. L.; Kefauver,
S. C.Mapping Invasive Aquatic Vegetation in the Sacramento-San Joaquin Delta using
Hyperspectral Imagery, Environmental Monitoring and Assessment (2006-10-01) 121: 47–
64, October 01, 2006.
39. Herold, Martin; Roberts, Dar A. The Spectral Dimension in Urban Remote Sensing, Remote
Sensing of Urban and Suburban Areas (2010-01-01) 10: 47–65, January 01, 2010.
40. Ramakrishna, Bharath; Plaza, Antonio J.; Chang, Chein-I; Ren, Hsuan; Du, Qian; Chang,
Chein-Chi Spectral/Spatial Hyperspectral Image Compression, Hyperspectral Data
Compression (2006-01-01): 309–346, January 01, 2006.
41. Benedetto, John J.; Czaja, Wojciech; Ehler, Martin Wavelet packets for time-frequency
analysis of multispectral imagery, GEM - International Journal on Geomathematics
(2013-11-01) 4: 137–154, November 01, 2013.
42. Yang, Chun-Chieh; Kim, Moon S.; Kang, Sukwon; Tao, Tao; Chao, Kuanglin; Lefcourt,
Alan M.; Chan, Diane E. The development of a simple multispectral algorithm for detection
of fecal contamination on apples using a hyperspectral line-scan imaging system, Sensing
and Instrumentation for Food Quality and Safety (2011-03-01) 5: 10–18, March 01, 2011.
References 9

43. Safavi, Haleh; Chang, Chein-I; Plaza, Antonio J.Projection Pursuit-Based Dimensionality
Reduction for Hyperspectral Analysis Satellite Data Compression (2011-01-01): 287–309,
January 01, 2011.
44. Benedetto, John J.; Czaja, Wojciech Dimension Reduction and Remote Sensing Using
Modern Harmonic Analysis, Handbook of Geomathematics (2015-01-01): 2609–2632,
January 01, 2015.
45. Dian, Yuanyong; Fang, Shenghui; Le, Yuan; Xu, Yongrong; Yao, Chonghuai Comparison
of the Different Classifiers in Vegetation Species Discrimination Using Hyperspectral
Reflectance Data Journal of the Indian Society of Remote Sensing (2014-03-01) 42: 61–72,
March 01,
46. Dian, Yuanyong; Le, Yuan; Fang, Shenghui; Xu, Yongrong; Yao, Chonghuai; Liu, Gang
Influence of Spectral Bandwidth and Position on Chlorophyll Content Retrieval at Leaf and
Canopy Levels Journal of the Indian Society of Remote Sensing (2016-02-13): 1–11,
February 13, 2016.
47. Zhang, Liangpei; Zhong, Yanfei Analysis of Hyperspectral Remote Sensing Images
Geospatial Technology for Earth Observation (2009-01-01): 235–269, January 01, 2009.
48. Du, Bo; Wang, Nan; Zhang, Liangpei; Tao, Dacheng Hyperspectral Medical Images
Unmixing for Cancer Screening Based on Rotational Independent Component Analysis
Intelligence Science and Big Data Engineering (2013-01-01) 8261: 336–343, January 01,
2013.
49. Zhang, Lefei; Zhang, Liangpei; Tao, Dacheng; Huang, Xin; Du, Bo Nonnegative
Discriminative Manifold Learning for Hyperspectral Data Dimension Reduction
Intelligence Science and Big Data Engineering (2013-01-01) 8261: 351–358, January 01,
2013.
50. Shen, Yingchun; Jin, Hai; Du, Bo An improved method to detect remote sensing image
targets captured by sensor network Wuhan University Journal of Natural Sciences
(2011-08-02) 16: 301–307, August 02, 2011.
51. Chang, Chein-I Hyperspectral Target Detection Real-Time Progressive Hyperspectral Image
Processing (2016-01-01): 131–172, January 01, 2016.
52. Ramakrishna, Bharath; Plaza, Antonio J.; Chang, Chein-I; Ren, Hsuan; Du, Qian; Chang,
Chein-Chi Spectral/Spatial Hyperspectral Image Compression, Hyperspectral Data
Compression (2006-01-01): 309–346, January 01, 2006.
53. Veganzones, Miguel A.; Graña, Manuel Hybrid Computational Methods for Hyperspectral
Image Analysis Hybrid Artificial Intelligent Systems (2012-01-01): 7209, January 01, 2012.
54. Moreno, Ramón; Graña, Manuel Segmentation of Hyperspectral Images by Tuned
Chromatic Watershed Recent Advances in Knowledge-based Paradigms and Applications
(2013-10-31) 234: 103–113, October 31, 2013.
55. Wu, Di; Sun, Da-Wen Hyperspectral Imaging Technology: A Nondestructive Tool for Food
Quality and Safety Evaluation and Inspection Advances in Food Process Engineering
Research and Applications (2013-09-13): 581–606, September 13, 2013.
56. Chen, Yu-Nan; Sun, Da-Wen; Cheng, Jun-Hu; Gao, Wen-Hong Recent Advances for Rapid
Identification of Chemical Information of Muscle Foods by Hyperspectral Imaging Analysis
Food Engineering Reviews (2016-04-04): 1–15, April 04, 2016.
57. Wu, Di; Sun, Da-Wen Hyperspectral Imaging Technology: A Nondestructive Tool for Food
Quality and Safety Evaluation and Inspection Advances in Food Process Engineering
Research and Applications (2013-09-13): 581–606, September 13, 2013.
58. Chen, Yu-Nan; Sun, Da-Wen; Cheng, Jun-Hu; Gao, Wen-Hong Recent Advances for Rapid
Identification of Chemical Information of Muscle Foods by Hyperspectral Imaging Analysis
Food Engineering Reviews (2016-04-04): 1–15, April 04, 2016.
59. Goodacre, Royston; Burton, Rebecca; Kaderbhai, Naheed; Timmins, Éadaoin M.;
Woodward, Andrew; Rooney, Paul J.; Kell, Douglas B. Intelligent Systems for the
Characterization of Microorganisms from Hyperspectral Data Rapid Methods for Analysis of
Biological Materials in the Environment (2000-01-01) 30: 111–136, January 01, 2000.
10 1 Introduction

60. Winder, Catherine L.; Cornmell, Robert; Schuler, Stephanie; Jarvis, Roger M.; Stephens,
Gill M.; Goodacre, Royston Metabolic fingerprinting as a tool to monitor whole-cell
biotransformations Analytical and Bioanalytical Chemistry (2011-01-01) 399: 387-401,
January 01, 2011.
61. Wang, Liguo; Zhao, Chunhui Introduction of Hyperspectral Remote Sensing Applications
Hyperspectral Image Processing (2016-01-01): 283-308, January 01, 2016.
62. Yang, Jinghui; Wang, Liguo; Qian, Jinxi Hyperspectral Imagery Classification Based on
Sparse Feature and Neighborhood Homogeneity, Journal of the Indian Society of Remote
Sensing (2015-09-01) 43: 445-457, September 01, 2015.
63. Yang, Chenghai; Everitt, James H. Using spectral distance, spectral angle and plant
abundance derived from hyperspectral imagery to characterize crop yield variation, Precision
Agriculture (2012-02-01) 13: 62-75, February 01, 2012.
64. Yang, Chenghai; Everitt, James H.; Bradford, Joe M. Airborne hyperspectral imagery and
linear spectral unmixing for mapping variation in crop yield, Precision Agriculture
(2007-12-01) 8: 279-296, December 01, 2007.
65. Prabhu, N.; Arora, Manoj K.; Balasubramanian, R. Wavelet Based Feature Extraction
Techniques of Hyperspectral Data, Journal of the Indian Society of Remote Sensing
(2016-01-29): 1–12, January 29, 2016.
66. Yang, Jinghui; Wang, Liguo; Qian, Jinxi Hyperspectral Imagery Classification Based on
Sparse Feature and Neighborhood Homogeneity, Journal of the Indian Society of Remote
Sensing (2015-09-01) 43: 445–457, September 01, 2015.
67. Lausch, Angela; Pause, Marion; Merbach, Ines; Zacharias, Steffen; Doktor, Daniel; Volk,
Martin; Seppelt, Ralf A new multiscale approach for monitoring vegetation using remote
sensing-based indicators in laboratory, field, and landscape Environmental Monitoring and
Assessment (2013-02-01) 185: 1215–1235, February 01, 2013.
68. Lausch, Angela; Pause, Marion; Doktor, Daniel; Preidl, Sebastian; Schulz, Karsten
Monitoring and assessing of landscape heterogeneity at different scales Environmental
Monitoring and Assessment (2013-11-01) 185: 9419–9434, November 01, 2013.
69. Vermeulen, Ph.; Fernández Pierna, J. A.; Egmond, H. P.; Zegers, J.; Dardenne, P.; Baeten,
V.Validation and transferability study of a method based on near-infrared hyperspectral
imaging for the detection and quantification of ergot bodies in cereals Analytical and
Bioanalytical Chemistry (2013-09-01) 405: 7765–7772, September 01, 2013.
70. Williams, Paul J.; Geladi, Paul; Britz, Trevor J.; Manley, Marena Near-infrared
(NIR) hyperspectral imaging and multivariate image analysis to study growth characteristics
and differences between species and strains of members of the genus Fusarium Analytical
and Bioanalytical Chemistry (2012-10-01) 404: 1759–1769, October 01, 2012.
71. Ebadi, Ladan; Shafri, Helmi Z. M.; Mansor, Shattri B.; Ashurov, Ravshan A review of
applying second-generation wavelets for noise removal from remote sensing data
Environmental Earth Sciences (2013-11-01) 70: 2679–2690, November 01, 2013.
72. Ebadi, Ladan; Shafri, Helmi Z. M.Compression of remote sensing data using
second-generation wavelets: a review Environmental Earth Sciences (2014-02-01) 71:
1379–1387, February 01, 2014.
73. Veganzones, Miguel A.; Graña, Manuel Hybrid Computational Methods for Hyperspectral
Image Analysis Hybrid Artificial Intelligent Systems (2012-01-01): 7209, January 01, 2012.
74. Qian, Shen-En Development of On-Board Data Compression Technology at Canadian Space
Agency Satellite Data Compression (2011-01-01): 1–28, January 01, 2011.
75. Zou, Xiaobo; Zhao, Jiewen Hyperspectral Imaging Detection Nondestructive Measurement
in Food and Agro-products (2015-01-01): 127–193, January 01, 2015.
76. Prabhakar, M.; Prasad, Y. G.; Rao, Mahesh N. Remote Sensing of Biotic Stress in Crop
Plants and Its Applications for Pest Management Crop Stress and its Management:
Perspectives and Strategies (2012-01-01): 517–545, January 01, 2012.
77. Mookambiga, A.; Gomathi, V. Comprehensive review on fusion techniques for spatial
information enhancement in hyperspectral imagery, Multidimensional Systems and Signal
Processing (2016-04-27): 1–27, April 27, 2016.
References 11

78. Appice, Annalisa; Guccione, Pietro; Malerba, Donato Transductive hyperspectral image
classification: toward integrating spectral and relational features via an iterative ensemble
system Machine Learning (2016-03-22): 1–33, March 22, 2016.
79. Prabhu, N.; Arora, Manoj K.; Balasubramanian, R. Wavelet Based Feature Extraction
Techniques of Hyperspectral Data, Journal of the Indian Society of Remote Sensing
(2016-01-29): 1–12, January 29, 2016.
80. Hasanlou, Mahdi; Samadzadegan, Farhad; Homayouni, Saeid SVM-based hyperspectral
image classification using intrinsic dimension Arabian Journal of Geosciences (2015-01-01)
8: 477–487, January 01, 2015.
81. Chang, Chein-I Back Matter - Real-Time Progressive Hyperspectral Image Processing
Real-Time Progressive Hyperspectral Image Processing (2016-01-01), January 01, 2016.
82. Wang, Liguo; Zhao, Chunhui Introduction of Hyperspectral Remote Sensing Applications
Hyperspectral Image Processing (2016-01-01): 283–308, January 01, 2016.
83. Chaudhuri, Subhasis; Kotwal, Ketan Introduction Hyperspectral Image Fusion
(2013-01-01): 1–18, January 01, 2013.
84. Wang, Liguo; Zhao, Chunhui Basic Theory and Main Processing Techniques of
Hyperspectral Remote Sensing Hyperspectral Image Processing (2016-01-01): 1–44,
January 01, 2016.
85. Ülkü, İrem; Töreyin, Behçet Uğur Sparse coding of hyperspectral imagery using online
learning Signal, Image and Video Processing (2015-05-01) 9: 959–966, May 01, 2015.
86. Sánchez, Sergio; Plaza, Antonio Fast determination of the number of endmembers for
real-time hyperspectral unmixing on GPUs Journal of Real-Time Image Processing
(2014-09-01) 9: 397–405, September 01, 2014.
87. Li, Xiaorun; Cui, Jiantao; Zhao, Liaoying Blind nonlinear hyperspectral unmixing based on
constrained kernel nonnegative matrix factorization Signal, Image and Video Processing
(2014-11-01) 8: 1555–1567, November 01, 2014.
88. Qin, Zhen; Shi, Zhenwei; Jiang, Zhiguo A quasi-Newton-based spatial multiple materials
detector for hyperspectral imagery Neural Computing and Applications (2013-08-01) 23:
403–409, August 01, 2013.
89. Chang, Chein-I Hyperspectral Target Detection Real-Time Progressive Hyperspectral Image
Processing (2016-01-01): 131–172, January 01, 2016.
90. Luo, Bin; Chanussot, Jocelyn Supervised Hyperspectral Image Classification Based on
Spectral Unmixing and Geometrical Features Journal of Signal Processing Systems
(2011-12-01) 65: 457–468, December 01, 2011.
91. Wang, Tao; Zhu, Zhigang; Krzaczek, Robert S.; Rhody, Harvey E.A System Approach to
Adaptive Multimodal Sensor Designs Machine Vision Beyond Visible Spectrum
(2011-01-01) 1: 159–176, January 01, 2011.
92. Shao, Ming; Wang, Yunhong; Liu, Peijiang Face Relighting Based on Multi-spectral
Quotient Image and Illumination Tensorfaces Computer Vision – ACCV 2009 (2010-01-01)
5996: 108–117, January 01, 2010.
93. Fadzil, M. H. Ahmad; Nugroho, Hermawan; Jolivot, Romuald; Marzani, Franck;
Shamsuddin, Norashikin; Baba, Roshidah Modelling of Reflectance Spectra of Skin
Phototypes III Visual Informatics: Sustaining Research and Innovations (2011-01-01) 7066:
352–360, January 01, 2011.
94. Lefèvre, Sébastien; Aptoula, Erchan; Perret, Benjamin; Weber, Jonathan Morphological
Template Matching in Color Images Advances in Low-Level Color Image Processing
(2013-12-17) 11: 241–277, December 17, 2013.
95. Galeano, July; Jolivot, Romuald; Marzani, Franck Analysis of Human Skin Hyper-Spectral
Images by Non-negative Matrix Factorization Advances in Soft Computing (2011-01-01)
7095: 431–442, January 01, 2011.
96. Jia, Hongjie; Ding, Shifei; Meng, Lingheng; Fan, Shuyan A density-adaptive affinity
propagation clustering algorithm based on spectral dimension reduction Neural Computing
and Applications (2014-12-01) 25: 1557–1567, December 01, 2014.
12 1 Introduction

97. Fisher R. A. The Use of Multiple Measurements in Taxonomic Problems, Annals of

Eugenics, 1936, 7, 179–88.
98. Hanley J.A., McNeil B. J., The meaning and use of the area under receiving operating
characteristic (ROC) curve, Radiology 1982, 43, 29–36.
99. Anscombe F. J., Graphs in Statistical Analysis, The American Statistician, 27 (1973), 17–21.
100. Bradley A. P., The use of the area under the ROC curve in the evaluation of machine
learning algorithms, Reference Recognition, 1997, 30 (7), 1145–59.
Chapter 2
Image Acquisition

2.1 Introduction

Image acquisition was carried out with the SOC710-VP hyperspectral camera. The
camera was positioned perpendicular to the table on which the object was placed.
The scanning area and the focal length were selected in such a way that the
analysed object filled, if possible, the entire stage—Fig. 2.1.
Two different types of illumination were adopted:
• natural light—sunlight,
• artificial lighting—40 W halogen lamps with a constant radiation spectrum in
the range from 400 to 1000 nm.
For each registered image LGRAY(m, n, i), a reference gray level in the full
spectrum sized 10 cm × 10 cm was used as a reference—Fig. 2.2. The reference
was an integral part of the camera equipment. In addition, for testing purposes, for
several cases, two additional series of images were recorded when there was no
light LDARK(m, n, i) and for the full illumination LWHITE(m, n, i)—in both cases
without an object. The images LDARK(m, n, i), LWHITE(m, n, i) are the basis for
normalization of the image LGRAY(m, n, i) which is described in the next chapter.
All images are saved in a *.cube format. They can also be saved in other formats,
*.raw and *.dat, which are further converted to Matlab in the form of a
three-dimensional matrix [1]. This conversion is specific to each of these types of
record (*.cube, *.raw or *.dat). The files with these extensions are saved by the
hyperspectral camera in the format shown in Fig. 2.3.
Saving data in the hyperspectral camera stems from the idea of its operation. The
first data saved to *.cube, *.raw or *.dat * files relate to the first row or column
(depending on the camera position relative to the object). The first row is stored for
all the wavelengths i 2 (1, I), then the next row etc. The number of rows, columns

© Springer International Publishing AG 2017 13

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_2
14 2 Image Acquisition

Fig. 2.1 Image acquisition—the position of the camera relative to the subject: 1 test object—
hand; 2 window; 3 hyperspectral camera; 4 table top illuminated by sunlight; 5 reference

Fig. 2.2 Zoom of the

acquisition area for artificial
lighting: 1 test object—
subject’s hand; 2 grey
reference; 3 stage; 4 halogen
bulbs
2.1 Introduction 15


m=2 m=4 m=6
m=1 m=3 m=5

i=1
i=2 

i=3 

 
i=4  

 
i=5 

i=6

i=1

raw file


Fig. 2.3 Data organization in *.raw, *.cube and *.dat files

and individual wavelengths is stored in a separate *.hdr file. An example of its

structure is shown below:
ENVI
Description = {}
samples = 520
lines = 696
bands = 128
header offset = 32768
major frame offsets = {0, 0}
file type = ENVI Standard
data type = 12
interleave = bil
sensor type = Unknown
byte order = 0
wavelength units = Unknown
wavelength = {
374.980011, 379.953130, 384.929945, 389.910456, 394.894663, 399.882566,
404.874165, 409.869460, 414.868451, 419.871138, 424.877521, 429.887600,
434.901375, 439.918846, 444.940014, 449.964877, 454.993436, 460.025691,
16 2 Image Acquisition

465.061642, 470.101289, 475.144632, 480.191671, 485.242406, 490.296837,

495.354964, 500.416787, 505.482306, 510.551521, 515.624432, 520.701039,
525.781342, 530.865341, 535.953036, 541.044427, 546.139514, 551.238297,
556.340776, 561.446951, 566.556822, 571.670389, 576.787652, 581.908612,
587.033267, 592.161618, 597.293665, 602.429408, 607.568847, 612.711982,
617.858813, 623.009340, 628.163563, 633.321482, 638.483097, 643.648408,
648.817415, 653.990118, 659.166517, 664.346612, 669.530403, 674.717890,
679.909073, 685.103952, 690.302527, 695.504798, 700.710765, 705.920428,
711.133787, 716.350842, 721.571594, 726.796041, 732.024184, 737.256023,
742.491558, 747.730789, 752.973716, 758.220339, 763.470658, 768.724673,
773.982384, 779.243791, 784.508894, 789.777693, 795.050188, 800.326379,
805.606266, 810.889849, 816.177128, 821.468103, 826.762774, 832.061141,
837.363204, 842.668963, 847.978418, 853.291569, 858.608416, 863.928959,
869.253199, 874.581134, 879.912765, 885.248092, 890.587115, 895.929834,
901.276249, 906.626360, 911.980167, 917.337670, 922.698869, 928.063764,
933.432355, 938.804642, 944.180625, 949.560304, 954.943679, 960.330750,
965.721517, 971.115980, 976.514139, 981.915994, 987.321545, 992.730792,
998.143735, 1003.560374, 1008.980709, 1014.404740, 1019.832467,
1025.263891, 1030.699010, 1036.137825}
The arrangement of individual elements is typical for almost all types of
hyperspectral cameras. The first elements of the header are designed to provide
information on the number of samples (samples = 520) or the number of columns
N, then the number of lines (lines = 696) or the number of rows M and the number
of different wavelengths (bands = 128) or I-th number of a matrix sized
M × N. Then there are two more important elements: header offset = 32,768
relating to the transfer of data in bytes (in this case 32,768 bytes), and data type
data type = 12 meaning that there is 16-bit unsigned integer per one pixel [2]. For
other values of variable ‘data type’ per one pixel there is:
• 8-bit unsigned integer (data type = 1),
• 16-bit signed integer (data type = 2),
• 32-bit signed integer (data type = 3),
• 32-bit single-precision (data type = 4),
• 64-bit double-precision floating-point (data type = 5),
• real-imaginary pair of single-precision floating-point (data type = 6),
• 16-bit unsigned integer (data type = 12),
• 32-bit unsigned long integer (data type = 13),
• 64-bit long integer (data type = 14),
• 64-bit unsigned long integer (data type = 15).
The last element in the *.hdr file is the variable wavelength. It means the
wavelengths in nanometres for which the individual images were acquired. In the
2.1 Introduction 17

present case, it is 374.980011 nm (i = 1), 379.953130 nm (i = 2), 384.929945 nm

(i = 3) etc.
Therefore, the *.hdr file is useful as it downloads the parameters (number of
rows, columns, wavelengths) necessary to read data in the *.cube, *.raw or *.dat
files. The following are excerpts (separated by ‘…’) of the source code of the file
read_envi_header enabling to read and interpret the file *.hdr consisting of
three blocks: the search for the sign '{}= ', the search for the numerical value of
the variable lines and the values of the variable 'Wavelength' i.e.:

function
[lines,bands,samples,Wavelength,data_type,header_offse
t]=read_envi_header(src)

...

fid = fopen(src);
[drep] = textscan(fid,'%s','delimiter','{}=
','MultipleDelimsAsOne', 1);
fclose(fid); drep_=drep{:,1};

Dlines=strcmpi(drep{1,:},'lines');
DlinesY=[Dlines,(1:size(Dlines,1))'];
DlinesY(DlinesY(:,1)~=1,:)=[];
nrl=DlinesY(1,2);
lines=str2num(str2mat(drep_{nrl+1}));

...

Dwavelength=strcmpi(drep{1,:},'Wavelength');
DwavelengthY=[Dwavelength,(1:size(Dwavelength,1))'];
DwavelengthY(DwavelengthY(:,1)~=1,:)=[];
nrsw=DwavelengthY(end,2);
Wavelength=str2num(str2mat(drep_{(nrsw+1) :
(nrsw+1+bands-1)}));

...

The dots ‘…’ (as mentioned above) mean that some part of the source code has
been removed. It should be noted that they play a different role than the dotted line
‘…’ in Matlab which indicates that a further part of the code will be continued in
the next line.
18 2 Image Acquisition

The read values of lines, bands, samples, Wavelength, data_type,

header_offset are further used for reading the image data contained in the files *.
cube, *.raw or *.dat. The function designed for this purpose called read_envi_-
data is as follows:

function
[LGRAY]=read_envi_data(src,lines,bands,samples,data_ty
pe,band_no,header_offset)

...

if (data_type~=12) && (data_type~=4)

disp('Unsupported file type')
else
fid = fopen(src);
frewind(fid)
if data_type==12
fseek(fid,band_no*samples*2+header_offset,
'bof');
[LGRAY, COUNT] =
fread(fid,[samples,lines],[mat2str(samples),'*integer*
2'],samples*bands*2-samples*2);
else % data_type==4
fseek(fid,band_no*samples*4+header_offset,
'bof');
[LGRAY, COUNT] =
fread(fid,[samples,lines],[mat2str(samples),'*float'],
samples*bands*4-samples*4);
end
fclose(fid);
LGRAY=LGRAY';
end

In its first part, the data type is checked. Two types of data mentioned above
numbered ‘12’ and ‘4’are handled [3]. If the data type is different, the message
‘Unsupported file type’ will be displayed. Then the data will be read from the
header_offset. The reading for the data type ‘12’ takes place every sam-
ples*bands*2-samples*2 while in the case of data type ‘4’ every sam-
ples*bands*4-samples*4, 16 and 32 bits per pixel respectively. The result is
the matrix LGRAY used for image pre-processing.
References 19

References

1. Robert Koprowski, Sławomir Wilczyński, Zygmunt Wróbel, Barbara Błońska-Fajfrowska,

Calibration and segmentation of skin areas in hyperspectral imaging for the needs of
dermatology, BioMedical Engineering OnLine 2014 13:113.
2. Koprowski R, Wilczyński S, Wróbel Z, Kasperczyk S, Błońska-Fajfrowska B. Automatic
method for the dermatological diagnosis of selected hand skin features in hyperspectral
imaging. Biomed Eng Online. 2014 13:47.
3. Koprowski R. Hyperspectral imaging in medicine: image pre-processing problems and
solutions in Matlab. J Biophotonics. 2015 Nov;8(11–12):935–43. doi: 10.1002/jbio.
201400133. Epub 2015 Feb 9.
Chapter 3
Image Pre-Processing

Preliminary analysis and processing of images is associated with three main

elements:
• affine transformation of the image,
• image filtering and
• image calibration.
These three elements are presented in the following subchapters. The source
code of these three elements was implemented in three Matlab files: GUI_hyper-
spectral_trans, GUI_hyperspectral and GUI_hyperspe-
ctral_fun. The first one concerns affine transformations, the second one relates
to the graphical user interface and the third one concerns the function associated
with the response to specific user’s actions. The content, source code, of these
functions (mainly GUI_hyperspectral_fun) will be presented in fragments in
the order of its description in the text.
It is possible to read the image owing to the functions read_envi_data and
read_envi_header described above. In the file GUI_hyperspectral_fun,
each 2D image read correctly is saved to disk with the same name as the input file *.
cube, *.raw or *.dat with the extension *.mat. When re-reading the file *.cube, *.raw
or *.dat, it is checked whether the file *.mat exists. If it exists, it is loaded. Saving
individual 2D files with the extension *.mat means that the data are read at least 2
times faster. A fragment of the source code of the file GUI_hyperspe-
ctral_fun is shown below:

© Springer International Publishing AG 2017 21

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_3
22 3 Image Pre-Processing

try
load([src,mat2str(band_no),'.mat'],'L1');
catch
[L1]=read_envi_data(src,lines,bands,samples,data_type,
band_no,header_offset);
save([src,mat2str(band_no),'.mat'],'L1');
L_(:,band_no,[1 3])=0;L_(:,band_no,2)=1;
end

The menu allowing for the selection of the file *.cube, *.raw or *.dat is invoked
at the beginning of the function GUI_hyperspectral_fun, i.e.:

[FileName,PathName,FilterIndex] = uiget-
file({'*.cube';'*.dat';'*.raw'},'Select file');
if FilterIndex~=0
src=[PathName,FileName];
if strcmp(src(end-2:end),'ube')
[lines,bands,samples,Wavelength,data_type,header_offse
t]=read_envi_header([src(1:end-4),'hdr']);
else
[lines,bands,samples,Wavelength,data_type,header_offse
t]=read_envi_header([src(1:end-3),'hdr']);
end
band_no=round(bands/2);
[L1]=read_envi_data(src,lines,bands,samples,data_type,
band_no,header_offset);

According to the source code shown above, the middle 2D image from the image
sequence, when properly loaded, is displayed by default. Since the number of
individual *.mat files initially converted by Matlab (with prior reading of the same
file) is not known, its status is read. Reading and showing the status in the visual
form involves a sequential attempt to read all the files *.mat, i.e.:

L_=cat(3,ones([20 bands]),zeros([20 bands]),zeros([20

bands]));
for band_no=1:bands
if ex-
ist([src,mat2str(band_no),'.mat'])==2
L_(:,band_no,[1
3])=0;L_(:,band_no,2)=1;
end
end
Lvar=imresize(L_,[20 800]);
Lvar(Lvar<0)=0;Lvar(Lvar>1)=1;
set(hObj(20),'CData',Lvar);
3 Image Pre-Processing 23

*.mat exist *.mat no exist

Fig. 3.1 Exemplary image indicating the distribution: the existence of the file *.mat is shown in
green, its absence in red

The displayed image is 20 × 800 pixels. The red stripes represent the absence of
the *.mat file, while the green ones mean that it is located on the disk. An exem-
plary image is shown in Fig. 3.1.

3.1 Affine Transformations of the Image

The range of affine transformations applied in hyperspectral imaging is much wider

than in the case of classic 2D images. Only those which are most often applied in
hyperspectral imaging were selected. These are:
• rotation by any angle in the angular range of α 2 (0,360] degrees every 10
degrees—the new coordinates of pixels in this case (mA, nA) are as follows:



N N M
nA ¼ þ n  cosðaÞ  m   sinðaÞ ð3:1Þ
2 2 2



M N M
mA ¼ þ n   sinðaÞ þ m   cosðaÞ ð3:2Þ
2 2 2

• reordering of rows—the image LGRAYM(m, n, i)—mirrored around the x-axis,

• reordering of columns—the image LGRAYN(m, n, i)—mirrored around the y-axis,
i.e.:

LGRAYM ðm; n; iÞ ¼ LGRAY ðM  m; n; iÞ ð3:3Þ

LGRAYN ðm; n; iÞ ¼ LGRAY ðm; N  n; iÞ ð3:4Þ

24 3 Image Pre-Processing

• Cropping—cutting a portion of the image, i.e.:

LGRAYROI ¼ LGRAY ðm; n; iÞ ð3:5Þ

where: m, n 2 ROI.
Since in practice, the user manually selects the option of rotation, shift or
cropping, the variable names in the function were standardized to ‘L1’ for sim-
plicity. In practice, a record in the source code overwrites the value in variable L1,
but it is consistent and clear:
if get(hObj(7),'Value')==1 % ROI
if min(r(2:3))>1
L1=L1(r(2):(r(2)+r(4)),r(1):(r(1)+r(3)));
end
end

if get(hObj(8),'Value')==1
L1=L1(:,end:-1:1);
end
if get(hObj(9),'Value')==1
L1=L1(end:-1:1,:);
end
if get(hObj(10),'Value')==1
L1=mat2gray(L1);
end
if get(hObj(12),'Value')~=1
L1=imrotate(L1,(get(hObj(12),'Value')-
1)*10,'crop');
end
For each condition if, the value set by the user is taken—from the handle
hObj. In this case, these are the handles to checkbox (7,8,9,10) and the
pull-down menu (popup), the value of 12.

3.2 Image Filtration

3.2.1 Non-Adaptive

The read image LGRAY(m, n, i) and the calibrated images LDARK(m, n, i) and
LWHITE(m, n, i) are subjected to noise removal. The noise is removed using a
median filter with a mask hw sized Mw × Nw = 3 × 3 pixels or more set manually
using the graphical user interface (GUI). Each 2D image is filtered individually. The
minimum size of the mask hw was selected based on the maximum size of a single
3.2 Image Filtration 25

distortion whose area of concentration did not exceed 4 pixels. The size of dis-
tortions in hyperspectral images may be different and therefore the size of the filter
is set manually. The specific size is set in the menu checkbox with the handle
hObj(14), i.e.:

if get(hObj(14),'Value')==2
L1=medfilt2(L1,[3 3],'symmetric');
end
if get(hObj(14),'Value')==3
L1=medfilt2(L1,[5 5],'symmetric');
end
if get(hObj(14),'Value')==4
L1=medfilt2(L1,[7 7],'symmetric');
end
if get(hObj(14),'Value')==5
L1=medfilt2(L1,[9 9],'symmetric');
end
if get(hObj(14),'Value')==6
L1=medfilt2(L1,[11 11],'symmetric');
end
The above source code shows that for the value ‘1’, the image L1 is not filtered
in any way, and the filtration itself ranges from Mw × Nw = 3 × 3 pixels to
Mw × Nw = 11 × 11 pixels. This is enough to remove noise from most hyper-
spectral images.

3.2.2 Adaptive

The second type of noise removal from a sequence of hyperspectral images is

adaptive adjustment of the size of the filter [1–7], for example, the median filter.
There are three options here:
• adaptation of the filter size to the 2D image content independently for each i-th
image,
• adaptation to the i-th 2D image—depending on the wavelength,
• adaptation of the filter size to both the 2D image content and the i-th image.
Choosing the right solution for hyperspectral images should be preceded by the
analysis of changes in the Peak Signal-to-Noise Ratio (PSNR) for individual i
images in a series of measurements. The values of PSNR, the vector LPSNR(i), are
defined as:

2
!
ð 2B  1Þ
LPSNR ðiÞ ¼ 10  log10 ð3:6Þ
LMSE ðm; n; iÞ
26 3 Image Pre-Processing

where: B is the number of bits per one image pixel, LMSE(m, n, i) is the mean
squared error. i.e.:

1 X M X N
LMSE ðm; n; iÞ ¼ ðLMEAN ðiÞ  LGRAY ðm; n; iÞÞ2 ð3:7Þ
M  N m¼1 n¼1

1 X M X N
LMEAN ðiÞ ¼ LGRAY ðm; n; iÞ ð3:8Þ
M  N m¼1 n¼1

The problem of selecting the size of the filter hw (its size Mw × Nw) and making
it dependent on LPSNR(i) is directly related to the content of images. This content
may be different in each case—especially when it comes to diagnosis of the skin. In
this regard, the selected ROIs shown in Fig. 3.2 were analysed.
The sample results shown in Fig. 3.2 confirm an increase in noise for hyper-
spectral cameras for extreme wavelength values. In addition, it should be noted that
median filtering with a mask hw sized 3 × 3 pixels increases the value of PSNR to
the greatest extent (almost 10 dB). The source code of m-file GUI_hyperspe-
ctral_filter_test providing the above graphs is as follows:

(a)

(b) (c)

Fig. 3.2 Results of analysis of PSNR values for two selected regions: a the input image LGRAY(m,
n,i = 80), b a graph of changes in LPSNR(i) for i images without median filtering and with median
filtering using masks sized Mw × Nw = 3 × 3, 5 × 5, 9 × 9 and 11 × 11 pixels; c an analogous
graph for another ROI
3.2 Image Filtration 27

L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(80),'.mat']);
Lgrayi=mat2gray(L1.L1);
figure;
[X,Y,I2,RECT] = IMCROP(Lgrayi);
hObj=waitbar(0,'Please wait...');
LPSNR=[];
for i=1:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgrayi=mat2gray(L1.L1);
Lgrayi=IMCROP(Lgrayi,RECT);
LPSNR(i,1)=20*log10(1)-10*log10( sum(sum(
(mean(Lgrayi(:)) - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );

Lgrayi=mat2gray(L1.L1);
Lgrayi=medfilt2(Lgrayi,[3 3]);
Lgrayi=IMCROP(Lgrayi,RECT);
LPSNR(i,2)=20*log10(1)-10*log10( sum(sum(
(mean(Lgrayi(:)) - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );

Lgrayi=mat2gray(L1.L1);
Lgrayi=medfilt2(Lgrayi,[5 5]);
Lgrayi=IMCROP(Lgrayi,RECT);
LPSNR(i,3)=20*log10(1)-10*log10( sum(sum(
(mean(Lgrayi(:)) - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );
Lgrayi=mat2gray(L1.L1);
Lgrayi=medfilt2(Lgrayi,[9 9]);
Lgrayi=IMCROP(Lgrayi,RECT);
LPSNR(i,4)=20*log10(1)-10*log10( sum(sum(
(mean(Lgrayi(:)) - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );
Lgrayi=mat2gray(L1.L1);
Lgrayi=medfilt2(Lgrayi,[11 11]);
Lgrayi=IMCROP(Lgrayi,RECT);
LPSNR(i,5)=20*log10(1)-10*log10( sum(sum(
(mean(Lgrayi(:)) - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );
waitbar(i/128)
end
28 3 Image Pre-Processing

close(hObj)
figure
plot(LPSNR(:,1),'-r*'); grid on; hold on
plot(LPSNR(:,2),'-g*');
plot(LPSNR(:,3),'-b*');
plot(LPSNR(:,4),'-m*');
plot(LPSNR(:,5),'-k*');
xlabel('i [pixel]','FontSize',14,'FontAngle','Italic')
ylabel('L_{PSNR}
[dB]','FontSize',14,'FontAngle','Italic')
legend('None','3x3','5x5','9x9','11x11')

The first part of the code allows to identify the ROI in the image i = 80. Then,
the images from i = 1 to 128 are loaded sequentially from the disk and the ROI is
separated. Then, the value of PSNR after median filtering with different mask sizes
is calculated.
The results shown in Fig. 3.2 could suggest that increasing the size of the mask
hw of the median filter to the value of Mw × Nw = 11 × 11 pixels and more is the
right approach. The attentive reader probably drew attention to the formulas (3.7)
and (3.8), where due to the lack of the source image (devoid of noise), the mean
value of LMEAN(i) is taken into account. These calculations are only justified when a
homogeneous ROI is analysed and there is no source image free from noise. In
other cases, the formulas (3.7) and (3.8) must be modified by replacing
LMEAN(i) with an image devoid of noise. Since there is no noise-free image, it will
be artificially added to the existing i images in gray levels. The function imnoise
enables to add Gaussian or salt and pepper noise to the image LGRAY(m, n, i). The
resulting image LNOISE(m, n, i) will be further used to test changes in PSNR but for
the entire image (without the need to manually select the ROI). Filtration efficiency
is here compared with the adaptive median filter, median filter and image without
filtration. The source code of the m-file GUI_hyperspectral_filter_test2
allowing for this type of calculations is shown below:
3.2 Image Filtration 29

hObj=waitbar(0,'Please wait...');
LPSNR=[]; LSEU=[];
for i=1:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgrayi=mat2gray(L1.L1);Lgrayi_=Lgrayi;
Lnoise=imnoise(Lgrayi,'salt & pepper',0.5);
LPSNR(i,1)=20*log10(1)-10*log10( sum(sum( (Lnoise
- Lgrayi).^2 )) ./ (size(Lgrayi,1) * size(Lgrayi,2))
);
[Lnoise_a,LSE]=GUI_hyperspectral_adaptive_filter(Lnois
e);
LSEU(i,1:4)=hist(LSE(:),[0 3 5 7]);
LPSNR(i,2)=20*log10(1)-10*log10( sum(sum(
(Lnoise_a - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );

Lnoise_m=medfilt2(Lnoise,[7 7]);
LPSNR(i,3)=20*log10(1)-10*log10( sum(sum(
(Lnoise_m - Lgrayi).^2 )) ./ (size(Lgrayi,1) *
size(Lgrayi,2)) );
if i==80
figure; imshow([Lgrayi,Lnoise,Lnoise_a]);
figure; imshow([Lgrayi,Lnoise,Lnoise_m]);
end
waitbar(i/128)
end
close(hObj)
figure
plot(LPSNR(:,1),'-r*'); grid on; hold on
plot(LPSNR(:,2),'-g*');
plot(LPSNR(:,3),'-b*');
xlabel('i [pixel]','FontSize',14,'FontAngle','Italic')
ylabel('L_{PSNR}
[dB]','FontSize',14,'FontAngle','Italic')
legend('None','Adaptive','Median 3x3')
LSEU=LSEU./(size(Lgrayi,1) * size(Lgrayi,2))*100;
figure
plot(LSEU(:,2),'-r*'); grid on; hold on
plot(LSEU(:,3),'-g*');
plot(LSEU(:,4),'-b*');
xlabel('i [pixel]','FontSize',14,'FontAngle','Italic')
ylabel('L_{SEU}(i)
[%]','FontSize',14,'FontAngle','Italic')
legend('3x3','5x5','7x7')
30 3 Image Pre-Processing

The function GUI_hyperspectral_adaptive_filter is implemented

with adaptive median filtering with a mask hw whose size ranges from
Mw × Nw = 3 × 3 pixels to Mw × Nw = 7 × 7 pixels. The main idea of the pro-
posed adaptive filtering is to calculate erosion (image LGRAYE(m, n, i)), dilation
(image LGRAYD(m, n, i)) and perform median filtering (image LMED(m, n, i)) with a
structural element SEw (in the case of erosion and dilation) and a mask hw (in the
case of filtration) sized 3 × 3, 5 × 5 and 7 × 7 pixels, i.e.:
 
LGRAYE ðm; n; iÞ ¼ min LGRAY ðm þ mSEw ; n þ nSEw ; iÞ ð3:9Þ
mSEw ;nSEw 2SEw

 
LGRAYD ðm; n; iÞ ¼ max LGRAY ðm þ mSEw ; n þ nSEw ; iÞ ð3:10Þ
mSEw ;nSEw 2SEw

In order to calculate the resulting image L(c)MED(m, n, i) after filtration with

the adaptive median filter, auxiliary variables (binary images) Lcw(m, n, i),
Lgw(m, n, i) and the image in gray levels Low(m, n, i) need to be introduced:
8  
<1 if LGRAYE ðm; n; iÞ\LMED ðm; n; iÞ ^ 
Lcw ðm; n; iÞ ¼ LGRAYD ðm; n; iÞ [ LMED ðm; n; iÞ ð3:11Þ
:
0 other
8  
>
<1 LGRAYE ðm; n; iÞ\LGRAY ðm; n; iÞ
if  
Lgw ðm; n; iÞ ¼ ^ LGRAYD ðm; n; iÞ [ LGRAY ðm; n; iÞ ð3:12Þ
>
:
0 other
8    
< LMED ðm; n; iÞ if Lcw ðm; n; iÞ ¼ 1 ^ Lgw ðm; n; iÞ ¼ 0
Low ðm; n; iÞ ¼
:
0 other
ð3:13Þ

where LMED(m, n, i) is the result of filtration of the image LGRAY(m, n, i) for the
mask sized Mw × Nw. In order to simplify the notation of the results of filtration,
erosion and dilation carried out for a specific mask size, the size Mw = Nw was
given as one number as a subscript, for example, LMED,3(m, n, i) is the result of
median filtering with a mask sized Mw × Nw = 3 × 3 pixels. Thus, the resulting
image L(c)
MED(m, n, i) is equal to:

8    
>
> L ðm; n; iÞ if Lcw;3 ðm; n; iÞ ¼ 0 ^ Lcw;5 ðm; n; iÞ ¼ 1
>
< ow;5
ðcÞ    
LMED ðm; n; iÞ ¼
>
> Low;7 ðm; n; iÞ if Lcw;3 ðm; n; iÞ ¼ 0 ^ Lcw;5 ðm; n; iÞ ¼ 0
>
:
Low;3 ðm; n; iÞ other
ð3:14Þ
3.2 Image Filtration 31

The individual subscripts, e.g. 3, (as mentioned above) are directly related to the
size of the mask (structural element) amounting to, for example, 3 × 3 pixels.
Adaptive filtration is performed by the afore-mentioned and already used function
GUI_hyperspectral_adaptive_filter with the following source code:

function
[Lout,LSE]=GUI_hyperspectral_adaptive_filter(Lgrayi)
Lout3=Lgrayi;
Lout5=Lgrayi;
Lout7=Lgrayi;
LSE=zeros(size(Lgrayi));
SE3=ones(3);
SE5=ones(5);
SE7=ones(7);
Lgrayei3=imerode(Lgrayi,SE3);
Lgrayei5=imerode(Lgrayi,SE5);
Lgrayei7=imerode(Lgrayi,SE7);
Lgraydi3=imdilate(Lgrayi,SE3);
Lgraydi5=imdilate(Lgrayi,SE5);
Lgraydi7=imdilate(Lgrayi,SE7);
Lmedi3=medfilt2(Lgrayi,size(SE3));
Lmedi5=medfilt2(Lgrayi,size(SE5));
Lmedi7=medfilt2(Lgrayi,size(SE7));
Lc3=(Lgrayei3<Lmedi3) & (Lgraydi3>Lmedi3);
Lc5=(Lgrayei5<Lmedi5) & (Lgraydi3>Lmedi5);
Lc7=(Lgrayei7<Lmedi7) & (Lgraydi3>Lmedi7);
Lg3=(Lgrayei3<Lgrayi) & (Lgraydi3>Lgrayi);
Lg5=(Lgrayei5<Lgrayi) & (Lgraydi5>Lgrayi);
Lg7=(Lgrayei7<Lgrayi) & (Lgraydi7>Lgrayi);
Lout3( (Lc3==1) & (Lg3==0) )=Lmedi3( (Lc3==1) &
(Lg3==0) );
Lout5( (Lc5==1) & (Lg5==0) )=Lmedi5( (Lc5==1) &
(Lg5==0) );
Lout7( (Lc7==1) & (Lg7==0) )=Lmedi7( (Lc7==1) &
(Lg7==0) );
Lout=Lout3;
Lout((Lc3==0)&(Lc5==1))=Lout5((Lc3==0)&(Lc5==1));
Lout((Lc3==0)&(Lc5==0))=Lout7((Lc3==0)&(Lc5==0));
LSE((Lc3==1) & (Lg3==0))=3;
LSE((Lc3==0)&(Lc5==1))=5;
LSE((Lc3==0)&(Lc5==0))=7;
32 3 Image Pre-Processing

Fig. 3.3 Results of analysis of LPSNR(i) and LSEU(i) for all i images LGRAY(m,n,i). The graph
a shows the measurement results of LPSNR(i) without filtration, with adaptive filtration and with
median filtration. The graph b shows the values of LSEU(i), the percentage share in filtration of the
masks sized Mw × Nw = 3 × 3, 5 × 5, 7 × 7 pixels

The above source code has a block structure associated with conducting triple
calculations of individual variables for three different mask sizes, i.e.: 3 × 3, 5 × 5,
7 × 7 pixels. These calculations are necessary to determine the final form of the
image L(c)
MED(m, n, i) in accordance with the formula (3.14). This function, with the
source code mentioned above, provides practically relevant results—Fig. 3.3.
The results of analysis of LPSNR(i) values presented in Fig. 3.3(a) clearly indicate
the advantage of applying an adaptive median filter over the conventional median
filter with a mask sized 7 × 7 pixels (a difference of about 3 dB for i 2 (40,60)) and
compared to an image without any interference (filtration). Figure 3.3 b) shows the
percentage share with respect to all pixels in the image LGRAY(m, n, i) of individual
masks Mw × Nw = 3 × 3, 5 × 5, 7 × 7 pixels. As can be seen in Fig. 3.3(b), the
share of the mask sized 3 × 3 pixels is the largest, about 45%. Additionally, the
percentage share of the mask sized 5 × 5 pixels is similar to the distribution shown
in Fig. 3.2. For extreme images (extreme values of i), more filtration is required,
while the middle ones require less filtration. In each case of filtration, sample
images and their visual assessment are much more convincing than PSNR.
Therefore, Fig. 3.4 shows the images LNOISE(m, n, i) and LMED,7(m, n, i) as well as
L(c)
MED(m, n, i).
The problem presented at the beginning of this subchapter, i.e.: adaptation of the
filter size to the 2D image content independently for each i-th image and adaptation
to the i-th 2D image depending on the wavelength, is solved by the above adaptive
approach. Therefore, there is no need to develop two separate algorithms.
The presented adaptive filtration was not deliberately included in the GUI or the
m-files of the program. At this point, I encourage the reader to make the appropriate
changes in the files GUI_hyperspectral and GUI_hyperspectral_fun
so that adaptive filtering will be available in the main application menu.
3.3 Image Calibration 33

(a) (b) (c)

Fig. 3.4 Results of analysis for a sample image i = 80 a input image LGRAY(m, n, i = 80); b result
of median filtering LMED(m, n, i = 80) for the mask sized Mw × Nw = 7 × 7 pixels; c result of
adaptive median filtering L(c)
MED(m, n, i = 80)

3.3 Image Calibration

Calibration of hyperspectral images is a very important element, because it is

affected by many factors that can introduce significant errors to the interpretation of
results (the read intensity). In practice, assuming constant light intensity in the full
spectral range of the camera, there are two methods of calibration:
• using a reference visible in the analysed image or
• using calibrating images.
In the first case, calibration is related to image normalization from the value of
minimum brightness occurring in the image to the mean value read from the area
visible in the reference image. Therefore, the file after calibration LCAL(m,n,i) is
calculated as:

L ðm; n; iÞ
LCAL ðm; n; iÞ ¼  CA2  ð3:15Þ
max max LCA2 ðm; n; iÞ
n m
34 3 Image Pre-Processing

where:
8
< LGRAY ðm; n; iÞ if LGRAY ðm; n; iÞ\Calw ðiÞ
LCA2 ðm; n; iÞ ¼  
: max max LGRAY ðm; n; iÞ other
n m
ð3:16Þ

and:
1 X
Calw ðiÞ ¼ LGRAY ðm; n; iÞ ð3:17Þ
Mc  Nc m;n2ROIc

Mc and Nc—are the numbers of rows and columns of the ROIc being the reference—
Fig. 2.2.
The size of the ROIc is most often Mc × Nc = 40 × 40 pixels. The calibrated
image LCAL(m, n, i) has values (for bright pixels) limited from the top by the mean
brightness from the area of the reference Calw(i). Implementation of this fragment
in Matlab is simple:

if get(hObj(11),'Value')==1
if min(rc(2:3))>1

Cal_w=mean(mean(L1(rc(2):(rc(2)+rc(4)),rc(1):(rc(1)+rc
(3)))));
L1(L1>Cal_w)=Cal_w;
L1=mat2gray(L1);
end
end

Four values stored in the variable rc come from manual selection of the ROIc. It
must be made clear that this calibration method can be fully automated with a
constant position of the reference—e.g. always in the upper left corner of the stage.
In this case, it is enough to assign the variable rc to 4 constants—x and y coor-
dinates and the size of the ROIc in x- and y-axis.
In the second case, calibration is related to the performance of 2 additional
registrations of images LDARK(m, n, i) and LWHITE(m, n, i). The idea of this cali-
bration is shown in Fig. 3.5.
These images (Fig. 3.5) are the basis for calibration. The calibrated image
L(2)
CAL(m, n, i) is calculated as:

ð2Þ
LCAL ðm; n; iÞ ¼  LGRAY ðm;n;iÞLDARK ðm;n;iÞ :
max maxðLGRAY ðm;n;iÞLDARK ðm;n;iÞÞ
n m
ð3:18Þ
LWHITE ðm; n; iÞ
3.3 Image Calibration 35

LWHITE(m,n,i)

LGRAY(m,n,i)
intensity


LCAL(m,n,i)
LDARK(m,n,i)

n
Fig. 3.5 Schematic graph of calibration results for L(2)
CAL(m, n, i) of brightness changes in the
image LGRAY(m, n, i) using the images LDARK(m, n, i) and LWHITE(m, n, i) when m = const

I encourage the reader to implement this second calibration method in practice.

In this case, the reader should duplicate the reading of files in the GUI and add the
relevant fragment in the file GUI_hyperspectral_fun. The further course of
action and the algorithm fragment remain unchanged.

3.4 Preliminary Version of the GUI

The issues of data reading and image pre-processing presented in the previous
chapters have been linked with the preparation of a preliminary version of the GUI.
The GUI has been divided into several areas—Fig. 3.6.
The GUI presented in Fig. 3.6 allows for opening *.cube, *.raw or *.dat files,
automatic conversion to *.mat files, reordering of image rows and columns, nor-
malization, image rotation, artificial colouring of images, median filtering, visual-
ization of the number of image columns and rows as well as the number of images
for individual wavelengths, viewing and analysis of individual images, viewing the
analysed image, displaying changes in the mean, minimum and maximum bright-
ness for the entire area or the selected ROI for individual images, selecting the ROI,
image calibration, displaying text data on the wavelength and the file name.
This GUI will be further expanded and its functionality will be increased.

3.5 Block Diagram of the Discussed Transformations

The discussed transformations along with the source code excerpts and the corre-
sponding m-files can be presented in the form of a block diagram. This diagram is
shown in Fig. 3.7.
36 3 Image Pre-Processing

16 17 18 19
1

2
3
4
5
6
7
8

9
10
11

12 15
13 14

Fig. 3.6 Main menu of the application: 1 default window menu; 2 open button; 3 conversion
button; 4 reordering image rows; 5 normalization; 6 image rotation; 7 artificial colour palette; 8
median filter size; 9 number of samples (number of columns); 10 number of lines (number of
rows); 11 number of bands (number of images for each wavelength); 12 slider for viewing and
analysis of individual images; 13 image showing the amount of converted images *.mat; 14
viewing the analysed image; 15 graph of the mean, minimum and maximum brightness for the
entire area or the selected ROI; 16 reordering of image columns; 17 option of selecting the ROI; 18
calibration; 19 text data on the wavelength and the file name

The algorithm discussed so far has been divided into three blocks: image
acquisition, image pre-processing and image processing discussed later in this
monograph.
The m-files containing the discussed functions and methods are available in this
book as an attachment. It should be borne in mind that the files will be further
expanded to add new functionality. For this reason, a container has been developed
for readers interested in testing the discussed scope of functionality of the proposed
algorithms. The container includes the discussed functions in GUI_ver_pre.zip
attached to the book.
3.5 Block Diagram of the Discussed Transformations 37

Fig. 3.7 Block diagram of

the initial version of the

Image acquisition
algorithm. The block diagram
has been divided into three Camera
main parts: image acquisition,
image pre-processing and
image processing discussed
later in this monograph. This *.raw, *.dat,
diagram includes one of the *.cube, *.hdr
blocks highlighted in blue
whose functionality has not
been deliberately included in
the main application Rotate Angle

x,y translation Yes/No

Image pre-processing

Crop image ROI

mask
Median filter
h

Adaptive Median Filter

Calibration image LDARK

LWHITE

Image processing
38 3 Image Pre-Processing

References

1. J. Astola and P. Kuosmanen, “Fundamentals of Nonlinear Digital Filtering”, CRC Press, 1997.
2. T. Sun, M. Gabbouj and Y. Neuvo, “Center weighted median filters: Some properties and their
applications in image processing”, Signal Processing, vol. 35, Issue 3, pp 213–229, February
1994.
3. T. Chen, K. K. Ma, L.H. Chen, “Tri-State Median Filter for Image Denoising”, IEEE
Transactions on Image Processing, vol. 8, Issue 12, pp 1834–1838, 1999.
4. Z. Wang, D. Zhang, “Progressive switching median filter for the removal of impulse noise from
highly corrupted images”, IEEE Transactions on Circuits and Systems, vol. 46, Issue 1, pp 78–80,
1999.
5. Y. Zhao, D. Li, Z. Li, “Performance enhancement and analysis of an adaptive median filter”,
International Conference on Communications and Networking, pp. 651–653, 2007.
6. V. Backman, R. Gurjar, K. Badizadegan, I. Itzkan, R. R. Dasari, L.T. Perelman and M.S. Feld,
“A New Fast and Efficient Decision-Based Algorithm for Removal of High-Density Impulse
Noises”, Signal Processing Letters, IEEE, Vol. 14, Issue 3, pp 189–192, 2007.
7. V.R. Vijaykumar, P.T. Vanathi, P. Kanagasabapathy and D. Ebenezer, “High Density Impulse
Noise Removal Using Robust Estimation Based Filter”, IAENG International Journal of
Computer Science, August 2008.
Chapter 4
Image Processing

4.1 Diagnostic Expectations

One of the key elements in the construction of the image processing algorithm is the
diagnostic usefulness of the results. The literature review presented in the intro-
duction and the publications from [1–4] show that there is a wide range of seg-
mentation methods. Due to the large amount of information extracted from
hyperspectral images, there are virtually no restrictions (relating to the minimum
amount of data) to use any method of image analysis and segmentation. Therefore,
many authors use segmentation methods (also used in classification) such as sup-
port vector machines (SVM) [1, 2], the nearest neighbours [3] and others [4]. These
segmentation methods are based primarily on a set of data obtained from the
manually selected for all acquired wavelengths. From a practical, dermatological,
point of view, these methods used in hyperspectral imaging should:
• allow for segmentation of objects,
• allow for spectral analysis of any image portion,
• enable to compare the spectral characteristics of any two areas,
• allow for building a classifier based on binary decision trees, discriminant
analysis and others,
• test the created classifier for different images.

© Springer International Publishing AG 2017 39

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_4
40 4 Image Processing

In addition, the methods used in dermatology and biomedical engineering in the

field of hyperspectral imaging should be fully automatic, provide reproducible
results, allow for batch data analysis (selecting only a catalogue), and be resistant to
individual variability of patients. Given these expectations, both diagnostic and
functional, image processing presented in the following sections has been proposed.

4.2 Tracking Changes in the Shape of an Object

An extremely important element in hyperspectral imaging is the analysis of the

object/objects. This analysis is determined by the method of selecting the object.
Besides simple selection of the ROI, automatic selection is often used in practice.
Automatic selection is simply binarization or another more advanced method of
segmentation. However, regardless of the segmentation method, a hyperspectral
image sequence, by definition, does not provide the same image for different
wavelengths. Accordingly, the segmentation process typically occurs for one of the
images and its result (binary image) is used for the subsequent images. The object
shape hardly ever remains the same in successive images. Therefore the only
solution is to perform 3D segmentation. However, this type of segmentation
requires the analysis of the entire sequence of cube images sized, for example,
M
N
I = 696
520
128 pixels. In most types of computer programs, it is
not possible due to the large amount of data for analysis (100 MB for data
type = 12). Therefore, one possible solution to such problems is to track changes in
the shape of an object for the successive images. One possibility is to use the
methods of conditional erosion and dilation. These methods are typically used to
improve the quality of binary images obtained (most often) by binarization.
Tracking changes in the shape of an object requires:
• indicating the image that will be the basis for segmentation (binarization),
• indicating the algorithm enabling the correction of the object in the binarized
image relative to the other images in gray levels, for subsequent wavelengths.
The method of conditional erosion and dilation involves designation of one of
the images LGRAY(m, n, i), most often for i = 1, and then segmentation, for example,
using binarization with a threshold pbg, providing the image LBIN(m, n, i), i.e.:


1 if LGRAY ðm; n; iÞ [ pbg
LBIN ðm; n; iÞ ¼ ð4:1Þ
0 other
4.2 Tracking Changes in the Shape of an Object 41

As a consequence, two images are obtained (e.g. for i = 1), namely LGRAY(m, n,
i = 1) and LBIN(m, n, i = 1). Let us assume that there is only one object in the binary
image LBIN(m, n, i = 1). Conditional erosion and dilation for the adopted symmetric
structural element SE2(mSE2, nSE2) sized MSE2
NSE2 are shown below:

ðC Þ
LBINE ðm; n; iÞ
(
LBIN ðm; n; 1Þ if pc ðm; n; iÞ\pdc ð4:2Þ
¼ min ðLBIN ðm þ mSE2 ; n þ nSE2 ; iÞÞ other
mSE2 ;nSE2 2SE2

( ðm;n;1Þ
ðCÞ LBIN if pc ðm; n; iÞ [ pec
LBIND ðm; n; iÞ ¼ max ðLBIN ðm þ mSE2 ; n þ nSE2 ; iÞÞ other
mSE2 ;nSE2 2SE2

ð4:3Þ

where

1 X
M SE2 X
NSE2
pc ðm; n; iÞ ¼ LGRAY ðm þ mSE2 ; n þ nSE2 ; iÞ ð4:4Þ
MSE2  NSE2 m ¼1 n ¼1
SE2 SE2

as well as pec and pdc—the thresholds set by the user.

The constants pec and pdc determining the effectiveness of erosion and dilation
respectively take values dependent on the type of the variable in which the image
LGRAY(m, n, i) is stored. In the case of the variable type double, these are values
from the range pe = pd 2 (0, 1), whereas for the variable type uint8—pe = pd
(0, 255). The corresponding source codes for erosion and dilation are shown below
—these are functions GUI_hyperspectral_dilate_c and GUI_hypers-
pectral_erode_c:
42 4 Image Processing

function
[Lbind]=GUI_hyperspectral_dilate_c(Lgray,Lbin,se2,pdc,
vi)
...
[sm,sn]=size(se2);
[m,n]=size(Lbin);
ppp=floor(([sm,sn]+1)/2); pm=ppp(1); pn=ppp(2);
if vi==1
hg = waitbar(0,'Working.... ');
end
Lbind=Lbin;
for mm=1:(m-sm)
for nn=1:(n-sn)
Lsu=sum(sum(Lbin( mm : (mm+sm-1) , nn :
(nn+sn-1) ).*se2 ));
if Lsu<sum(se2(:))
if Lsu>0
if Lbin(mm+pm-1,nn+pn-1)==1;
wyy=Lgray( mm : (mm+sm-1) , nn : (nn+sn-1)
);
if sum(sum( wyy .*se2
))/sum(sum(se2>0))>pdc;
Lbind( mm : (mm+sm-1) , nn : (nn+sn-1)
)=se2;
end
end
end
end
end
if vi==1
waitbar(mm/(m-sm),hg)
end
end
if vi==1
close(hg)
end
4.2 Tracking Changes in the Shape of an Object 43

Analogously for dilation, i.e.:

function
[Lbine]=GUI_hyperspectral_erode_c(Lgray,Lbin,se2,pec,v
i)
...
[sm,sn]=size(se2);
[m,n]=size(Lbin);
if vi==1
hg = waitbar(0,'Working....');
end
Lbine=Lbin;
for mm=1:(m-sm)
for nn=1:(n-sn)
Lsu=sum(sum(Lbin( mm : (mm+sm-1) , nn :
(nn+sn-1) ).*se2 ));
if Lsu<sum(se2(:))
if Lsu>0
if Lbin( round((mm+sm-1+mm)/2) , round((nn+sn-
1+nn)/2) )==1;
wyy=Lgray( mm : (mm+sm-1) , nn : (nn+sn-1)
);
if sum(sum( wyy .*se2
))/sum(sum(se2>0))<pec;
Lbine( mm : (mm+sm-1) , nn : (nn+sn-1)
)=0;
end
end
end
end
end
if vi==1
waitbar(mm/(m-sm),hg)
end
end
if vi==1
close(hg);
end

The described functions of conditional opening and closing do not fulfil, in

comparison with the classical approach, the following relationships:
• subsequent operations of opening or closing may cause further changes in the
size of the object in the image, i.e.:
44 4 Image Processing

ðLBIN ðm; n; iÞ  SE2Þ  SE2

ð4:5Þ
6¼ ðððLBIN ðm; n; iÞ  SE2Þ  SE2Þ  SE2Þ  SE2

ðLBIN ðm; n; iÞ  SE2Þ  SE2

ð4:6Þ
6¼ ðððLBIN ðm; n; iÞ  SE2Þ  SE2Þ  SE2Þ  SE2

where symbols ;  refer to erosion and dilation respectively.

• opening the image completeness is not equal to the completeness of its closing
and vice versa, i.e.:

    ðDÞ
ðDÞ ðDÞ
LBIN ðm; n; iÞ  SE2  SE2 6¼ LBIN ðm; n; iÞ  SE2  SE2 ð4:7Þ

  ðDÞ  
ðDÞ ðDÞ
LBIN ðm; n; iÞ  SE2  SE2 6¼ LBIN ðm; n; iÞ  SE2  SE2 ð4:8Þ

where
BIN—completeness (superscript D) of the image LBIN,
L(D)
• the images resulting from opening for the included structural elements ‘a’ and
‘b’ do not need to be included, i.e.:

   
ðaÞ ðbÞ
LBIN ðm; n; iÞ  SE2  SE2 6 LBIN ðm; n; iÞ  SE2  SE2 ð4:9Þ
   
ðaÞ ðbÞ
LBIN ðm; n; iÞ  SE2  SE2 6 LBIN ðm; n; iÞ  SE2  SE2 ð4:10Þ

BIN  LBIN.
where L(a) (b)

Specificity of the described conditional erosion and dilation operations is based

on the sequential performance of conditional erosion and dilation, i.e.:

ðF Þ
LBIN ðm; n; iÞ
ð4:11Þ
¼ ððððLBIN ðm; n; iÞ  SE2Þ  SE2Þ  SE2Þ  SE2Þ  SE2

The number of performed sequential operations of conditional erosion and

dilation strictly depends on the size of the structural element SE2 and the shape of
the object in the binary image LBIN(m, n, i). Figure 4.1 shows changes in the surface
area for successive iterations and various values of the threshold pec = pdc
4.2 Tracking Changes in the Shape of an Object 45

LGRAY (m,n,i=1) LBIN (m,n,i=1)

m [pixel]
m [pixel]

n [pixel] n [pixel]

Fig. 4.1 Changes in the surface area for successive iterations and various values of the threshold
pec = pdc

2 {0.2, 0.3, …, 0.7, 0.8} for a sample object with the total surface area of
3667 pixels. It also shows the first images LGRAY(m, n, i = 1) and LBIN(m, n, i = 1)
for which the calculations were made.
Figure 4.1 shows three different situations. The first one is the complete removal
of the object from the image obtained for pec = pdc = 0.8. The second one is the
adjustment of the position and shape of the object visible when pec = pdc 2 {0.4,
0.5, 0.6, 0.7}. The third one is zooming the object to the full size of the image for
pec = pdc2{0.2, 0.3}. The source code providing the above graph is shown below:
46 4 Image Processing

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube50.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.2);
Lbin=Lgray>0.4;
figure; imshow(Lbin)

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube80.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.2);
figure; imshow(Lgray,[])

Lorg=Lbin;
se2=ones(3); pam=[];
for pec=0.2:0.1:0.8;
vi=1; Lbin=Lorg;
pami=[];
pami=[pami;[0, sum(sum(Lbin))]];
for it=1:16

[Lbin]=GUI_hyperspectral_erode_c(Lgray,Lbin,se2,pec,vi
);
% imshow(Lbin)
% pause(0.1)
pami=[pami;[it, sum(sum(Lbin))]];

[Lbin]=GUI_hyperspectral_dilate_c(Lgray,Lbin,se2,pec,v
i);
% imshow(Lbin)
% pause(0.1)
pami=[pami;[it, sum(sum(Lbin))]];
end
pam=[pam,pami(:,2)];
end
figure; plot(pam,'-*'); grid on
xlabel('it [/]','FontSize',14)
ylabel('area [pixel]','FontSize',14)
leg-
end('p_{ec}=p_{de}=0.2','p_{ec}=p_{de}=0.3','p_{ec}=p_
{de}=0.4','p_{ec}=p_{de}=0.5','p_{ec}=p_{de}=0.6','p_{
ec}=p_{de}=0.7','p_{ec}=p_{de}=0.8')

In the above source code, a change in the value of pec and pdc ranging from 0.2
to 0.8 in each loop circulation is noteworthy. Then, according to the idea presented
above, conditional erosion and dilation, functions GUI_hyper-
spectral_dilate_c and GUI_hyperspectral_erode_c, are calculated
alternately.
4.2 Tracking Changes in the Shape of an Object 47

The assessment of the convergence of the algorithm should be carried out also
for other sizes of the structural element SE2 (in the present case it was
3
3 pixels). A more detailed analysis of the various types of collected images
and various sizes of objects confirmed that typically the convergence of the algo-
rithm can be reached after approximately 15 iterations, when fluctuations around
the correct value of the surface area are in the range of ±10% (quasi steady state).
An increase in the size of the structural element SE2 increases the rate of con-
vergence but also the error of approximately ±40% in relation to the object sep-
arated by an expert. The accuracy of 10% is usually obtained if the size of the
structural element SE2 constitutes ≅3% of the object surface area. This relationship
is clearly visible in Fig. 4.2a which shows a graph of changes in the surface area of
the object for subsequent iterations and resizing the structural element SE2 from
3
3 pixels to 11
11 pixels.
The other graphs in Fig. 4.2b–d show the results for the input image resolution
M
N = 279
208 pixels, M
N = 557
416 pixels as well as 32 and 64

Fig. 4.2 Graph of changes in the surface area of the object for subsequent iterations and resizing
the structural element SE2 from 3
3 pixels to 11
11 pixels: a for M
N = 279
208 pixels
and 32 iterations; b for M
N = 557
416 pixels and 32 iterations; c for
M
N = 279
208 pixels and 64 iterations; d for M
N = 557
416 pixels and 64 iterations
48 4 Image Processing

iterations. The source code allowing for the calculations for the first graph presented
in Fig. 4.2a is shown below:

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube50.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.4);
Lbin=Lgray>0.4;
figure; imshow(Lbin)

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube80.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.4);
figure; imshow(Lgray,[])

Lorg=Lbin;
pam=[];
for MSE2NSE2=3:2:11
SE2=ones(MSE2NSE2);
pec=0.6;
vi=1; Lbin=Lorg;
pami=[];
pami=[pami;[0, sum(sum(Lbin))]];
for it=1:16

[Lbin]=GUI_hyperspectral_erode_c(Lgray,Lbin,SE2,pec,vi
);
pami=[pami;[it, sum(sum(Lbin))]];

[Lbin]=GUI_hyperspectral_dilate_c(Lgray,Lbin,SE2,pec,v
i);
pami=[pami;[it, sum(sum(Lbin))]];
end
pam=[pam,pami(:,2)];
end
figure; plot(pam,'-*'); grid on
xlabel('it [/]','FontSize',14)
ylabel('area [pixel]','FontSize',14)
leg-
end('M_{SE2}=N_{SE2}=3','M_{SE2}=N_{SE2}=5','M_{SE2}=N
_{SE2}=7','M_{SE2}=N_{SE2}=9','M_{SE2}=N_{SE2}=11')

To better understand and illustrate the transformations in the above source code,
the parts responsible for reading the image (for i = 50 and i = 80) are separated.
The presented loop enables to resize the mask SE2 = ones(MSE2NSE2) in the
range from 3 to 11 every 2 pixels.
4.2 Tracking Changes in the Shape of an Object 49

For zero iteration (it = 0), which is the initial state, the total surface area shown
in Fig. 4.2 is greater than the surface area for the same iteration shown in Fig. 4.1.
This is due to the adopted change in the resolution of the input image. In the first
case, the resolution is reduced to 20% of the original size, while in the second case
it is 40% of the original image resolution.
Due to the nature of conditional erosion and dilation, two-dimensional image
convolution, the time necessary to obtain the results depends on the image reso-
lution, the size and shape of the object and the size of the structural element SE2.
Some selected times of analysis are shown in Table 4.1. IT indicates the maximum
number of iterations, and it* the number of iterations after which the quasi steady
state is achieved.
The time of analysis shown in Table 4.1 is affected to the greatest extent by the
image resolution. According to the intuition, doubling the image resolution results
in an almost fourfold increase in computation time. The structural element SE2
influences the calculation time to the least extent.
The results of the proposed algorithm for tracking the object based on condi-
tional erosion and dilation are presented in Figs. 4.3 and 4.4.

Table 4.1 Some selected Time M
N (pixel) MSE2
NSE2 IT [/] it* [/]
times of analysis for two (s) (pixel)
different sizes of the object,
five different sizes of the 7.7 279
208 3
3 16.2 21
structural element and 7.5 279
208 5
5 16.2 11
different numbers of iterations 7.7 279
208 7
7 16.2 7
(for Intel® Xenon® CPU 7.9 279
208 9
9 16.2 6
[email protected] GHz)
8.4 279
208 11
11 16.2 6
30.1 557
416 3
3 16.2 32
30.4 557
416 5
5 16.2 20
31.6 557
416 7
7 16.2 13
32.6 557
416 9
9 16.2 12
34 557
416 11
11 16.2 8
15.5 279
208 3
3 32.2 28
15.1 279
208 5
5 32.2 18
15.6 279
208 7
7 32.2 11
16.0 279
208 9
9 32.2 9
16.7 279
208 11
11 32.2 7
59.7 557
416 3
3 32.2 40
60.9 557
416 5
5 32.2 19
62.9 557
416 7
7 32.2 13
65.4 557
416 9
9 32.2 11
69.6 557
416 11
11 32.2 8
50 4 Image Processing

(a)
m [pixel]
(c)

n [pixel]
(b)
m [pixel]

m n

n [pixel]

Fig. 4.3 Results of conditional erosion and dilation for an artificial binary image representing a
rectangle. Part a shows the output image LBINE(m, n, i) as a binary image superimposed on the
image in gray levels LGRAY(m, n, i). Part b shows the input image LBINE(m, n, i) as a binary image
superimposed on the input image in gray levels LGRAY(m, n, i). Subsequent stages of erosion and
dilation are shown in part (c) for pec = pdc = 0.8

The results of conditional erosion and dilation presented in Figs. 4.3 and 4.4
were obtained for an artificial binary image representing a rectangle. Part a shows
the output image as a binary image superimposed on the image in gray levels.
Figures 4.3b and 4.4 show the input image as a binary image superimposed on the
input image in gray levels. The subsequent stages of erosion and dilation are shown
in Figs. 4.3c and 4.4c. These are the successive stages of conditional erosion and
dilation for successive conditional erosions and dilations of the images LBINE(m, n,
i) and LBIND(m, n, i). Figure 4.3 shows the results for pec = pdc = 0.8 and Fig. 4.4
for pec = pdc = 0.4. In both cases the size of the structural element SE2 was the
same, namely 3
3 pixels. Therefore, Figs. 4.3 and 4.4 show how conditional
erosion and dilation, which enable to change the shape of the object present in the
image from a rectangle to the shape corresponding to the content of the image
LGRAY(m, n, i), work in practice.
4.2 Tracking Changes in the Shape of an Object 51

(a)
(c)
m [pixel]

n [pixel]
(b)
m [pixel]

m n

n [pixel]

Fig. 4.4 Results of conditional erosion and dilation for an artificial binary image representing a
rectangle. Part a shows the output image LBINE(m, n, i) as a binary image superimposed on the
image in gray levels LGRAY(m, n, i). Part b shows the input image LBINE(m, n, i) as a binary image
superimposed on the input image in gray levels LGRAY(m, n, i). Subsequent stages of erosion and
dilation are shown in part (c) for pec = pdc = 0.4

In this way, the above algorithm was implemented to track an object whose
shape changes for successive i images in a series. To this end, setting the threshold
manually or automatically, the first image LGRAY(m, n, i = 1) can be subjected to
binarization providing the image LBIN(m, n, i = 1) and then, conditional erosion and
dilation of the images LBINE(m, n, i 6¼ 1) and LBIND(m, n, i 6¼ 1) can be performed
alternately. In practice, however, the first image (i = 1) is rarely used as a basis for
binarization and then determination of the starting object whose shape is further
corrected. This is due to the large amount of noise in the image. The middle image
in a series is most commonly adopted as the value i, i.e.: i = I/2 = 64 (assuming an
even number of I). The results are shown in Fig. 4.5.
The results of conditional erosion and dilation shown in Fig. 4.5 were obtained
for successive images in a sequence for i 2 (1, I) and IT = 11. The analysis was
started from the binary image LBIN(m, n, i = I/2). The next images in Fig. 4.5a–d
52 4 Image Processing

Fig. 4.5 Results of conditional erosion and dilation performed for subsequent images in a
sequence for i 2 (1, I) and IT = 11 starting with the binary image LBIN(m, n, i = I/2) when:
a pec = pdc = 0.5 and MSE2
N SE2 = 3
3; b pec = pdc = 0.3 and MSE2
N SE2 = 3
3;
c pec = pdc = 0.5 and MSE2
N SE2 = 5
5; d pec = pdc = 0.3 and MSE2
NSE2 = 5
5

were obtained for pec = pdc 2 {0.3, 0.5} and MSE2
NSE2 2 {3
3, 5
5}.
Depending on the selected parameters of the algorithm, the shape of the tracked
object changes significantly. These changes are due to the different amount of noise
in the image, the individual changes in the size of the object for the adjacent 2D
images etc. In each case, these parameters (pec, pdc, MSE2, NSE2) are selected
individually.
The source code for displaying the results from Fig. 4.5a is shown below:
4.2 Tracking Changes in the Shape of an Object 53

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube64.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.4);
Lbin=Lgray>0.4;
Lbini=[];
Lbini(1:size(Lbin,1),1:size(Lbin,2),64)=Lbin;
SE2=ones(5);
pec=0.5;
vi=0;
for i=65:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.4);
for it=1:2:10

[Lbin]=GUI_hyperspectral_erode_c(Lgray,Lbin,SE2,pec,vi
);

[Lbin]=GUI_hyperspectral_dilate_c(Lgray,Lbin,SE2,pec,v
i);
[i it]
end
Lbini(1:size(Lbin,1),1:size(Lbin,2),i)=Lbin;
if (i==80)|(i==120)
Lbeg=Lgray; Lbeg(Lbin==1)=max(Lbeg(:));
figure; imshow(Lbeg,[]); ti-
tle(['i=',mat2str(i)])
end
end
Lbin=Lbini(1:size(Lbin,1),1:size(Lbin,2),64);
for i=63:-1:1
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,0.4);
for it=1:2:10

[Lbin]=GUI_hyperspectral_erode_c(Lgray,Lbin,SE2,pec,vi
);

[Lbin]=GUI_hyperspectral_dilate_c(Lgray,Lbin,SE2,pec,v
i);
[i it]
end
54 4 Image Processing

Lbini(1:size(Lbin,1),1:size(Lbin,2),i)=Lbin;
if (i==10)|(i==40)
Lbeg=Lgray; Lbeg(Lbin==1)=max(Lbeg(:));
figure; imshow(Lbeg,[]); ti-
tle(['i=',mat2str(i)])
end
end

figure
[x,y,z]=meshgrid( 1:size(Lbini,2) , 1:size(Lbini,1) ,
1:size(Lbini,3));
p1 = patch(isosurface(x,y,z,Lbini,0.5),'FaceColor',[0
0 1 ],'EdgeColor','none');
view(33,12);
camlight; camlight(-80,-10); lighting phong;
grid on
alpha(p1,0.3)
hold on
axis square
axis([0 220 20 290 0 128])
xlabel('m [pixel]','FontSize',14)
ylabel('n [pixel]','FontSize',14)
zlabel('i [/]','FontSize',14)

The above source code consists of two parts. In the first one, there is a loop
designed for the analysis of images from i equal to 65 to 128. This analysis involves
a sequence of conditional erosions and dilations for it = 1:2:10. The other part
of the source code concerns the analysis for decreasing values of i, i.e. from 63 to 1.
Thus, the 2D image for i = 64 is the beginning of the analysis running in both
directions (decreasing and increasing i).
The presented method for tracking the shape of an object was not deliberately
included in the overall GUI of the described program to encourage readers to its
independent implementation. The full source code shown above is attached to this
monograph in the form of the following m-files:
GUI_hyperspectral_erode_dilate_test,
GUI_hyperspectral_erode_dilate_test2,
GUI_hyperspectral_erode_dilate_test3
and
GUI_hyperspectral_erode_dilate_test4.
4.3 Basic Analysis of Features 55

4.3 Basic Analysis of Features

The basic element of hyperspectral image analysis is a comparative analysis of

features such as the mean value or contrast both of the whole image and the ROI. In
this case, two groups of data are compared.
The first group of data is derived directly from the analysed image. This may
be, as previously mentioned, the mean value of brightness LS(i) of the selected
ROIS, for example [similarly to (3.17)]:

1 X
LS ð i Þ ¼ LGRAY ðm; n; iÞ ð4:12Þ
MS
NS m;n2ROIS

It may be also the value of minimum or maximum brightness. The first group of
data can be also created as a result of texture analysis. These may be, for example,
the results of analysis of gray-level co-occurrence matrix (GLCM), i.e.:

N X
X M
LGLCM ðu; v; iÞ ¼ LGB ðm; n; i; u; vÞ ð4:13Þ
n¼1 m¼1

where LGB can be calculated for the horizontal neighbourhood (an arrow as a
superscript) L!
GB :

L!
GB ðm; n; i; u; vÞ


1 if ðLGRAY ðm; n; iÞ ¼ uÞ ^ ðLGRAY ðm þ 1; n; iÞ ¼ vÞ ð4:14Þ
¼
0 other

for m 2 (1, M − 1) and n 2 (1, N), u 2 (1, U) and v 2 (1, V) where U and V are
equal to the number of brightness levels, i.e. 2B where B is the number of bits per
pixel. The above notation concerns the comparison of the horizontal neighbourhood
of pixels. For the vertical arrangement, the formula (4.14) is converted to the
following relationship:

L#GB ðm; n; i; u; vÞ


1 if ðLGRAY ðm; n; iÞ ¼ uÞ ^ ðLGRAY ðm þ 1; n; iÞ ¼ vÞ ð4:15Þ
¼
0 other

for m 2 (1, M) and n 2 (1, N − 1).

56 4 Image Processing

On this basis (LGLCM(u, v, i)), the parameters such as contrast LCON(i), energy
LENE(i) or homogeneity LHOM(i) are calculated, i.e.:

V X
X U
LCON ðiÞ ¼ ðv uÞ2 LGLCM ðu; v; iÞ ð4:16Þ
v¼1 u¼1

V X
X U
LENE ðiÞ ¼ LGLCM ðu; v; iÞ2 ð4:17Þ
v¼1 u¼1

V X
X U
1
LHOM ðiÞ ¼ LGLCM ðu; v; iÞ ð4:18Þ
v¼1 u¼1 1 þ ð v uÞ 2

Apart from analysis of GLCM, other texture features such as the surface area of
the recognized references can be also analysed. In the simplest form, this is the sum
LDET(i) of the image LWZ(m, n, i) after binarization using the threshold prw:

M X
X N
LDET ðiÞ ¼ LD ðm; n; iÞ ð4:19Þ
m¼1 n¼1

where


1 if LOC ðm; n; iÞ [ prw
LD ðm; n; iÞ ¼ ð4:20Þ
0 other

LOC ðm; n; iÞ ¼ LGRAY ðm; n; iÞ

     
ð4:21Þ
min min max max minðLGRAY ðm; n; iÞÞ ; LGRAY ðm; n; iÞ
m;n SE SE SE SE

SE—is a structural element whose shape corresponds to the shape of the rec-
ognized reference.
All the above new features are calculated separately for each ith wavelength and
will be further used.
The other group of data is derived from another portion of the same image. It
can be also acquired from another image or it may be a data vector (loaded outside).
To distinguish between these two groups of data, upper indexes were introduced
—‘W’ for the second group of data and ‘E’ for the first group of data. A basic
comparison involves calculating the difference between the data vectors, for
example, the calculated mean value of brightness, i.e.:
 W 
L ðiÞ LE ðiÞ
dS ð i Þ ¼ S S
ð4:22Þ
S ðiÞ
max LW
i2ð1;I Þ
4.3 Basic Analysis of Features 57

Fig. 4.6 The idea of (a) W

calculating FN, FP, TP and  LS(i)
TN for a sample graph of the E
variable dS(i): a a graph of the LS(i)

intensity
mean brightness for
comparable areas and their
difference dS(i), b the areas prg
marked in red and green by
S(i)
measurement
the user meet or do not meet
the condition of compliance; 
c the results of comparisons of i
the red and green areas from (b)
user
parts (a) and (b)
(c)
TP FN TP TN FP result

These differences are then binarized with respect to the threshold pq that is set
manually and expressed as a percentage of the value of the variable dS(i).
The analysis of the two discussed groups of data and the values they provide is
related to the areas selected manually. On this basis, the values of FN, FP, TP and
TN are calculated. The idea of these calculations is shown in Fig. 4.6.
The values of the variable dS(i) below the assumed threshold prg satisfy the
condition of allowable differences between comparable features - in this case the
mean brightness values. These wavelengths (values i) that are different or exceed
the threshold prg are marked in Fig. 4.6a with a red background. Similar colours
(red and green) were used to mark in Fig. 4.6b the areas that must comply with the
condition of miscellaneous brightness levels below the set threshold. The following
excerpt in the function GUI_hyperspectral_fun is to enable the user to
manually select these areas on a graph:

ax=axis(hObj(21));

set(hObj(70),'XData',[ax(1)],'YData',[ax(4)],'FaceColo
r','g');

set(hObj(71),'XData',[ax(1)],'YData',[ax(4)],'FaceColo
r','r');
re=getrect(hObj(21));
rect=[re(1),re(1)+re(3)];
set(hObj(70),'XData',[ax(1) rect(1) rect(1)
rect(2) rect(2) ax(2)],'YData',[ax(4) ax(4) 0 0 ax(4)
ax(4) ],'FaceColor','r');
set(hObj(71),'XData',[rect(1)
rect(2)],'YData',[ax(4) ax(4)],'FaceColor','g');
alpha(0.1)

A comparison of the two results calculated automatically and set manually is

shown in Fig. 4.6c. On this basis, the aforementioned values of FN, FP, TP and TN
58 4 Image Processing

as well as ACC are calculated. The usefulness of this type of analysis is very high in
practical applications and diagnostics, for which the degree of compliance of the
mean brightness level with the actual values is analysed.
A fragment of the source code responsible for this part of calculations is located in
four m-files GUI_hyperspectral_diff, GUI_hyperspectral_class,
GUI_hyperspectral and GUI_hyperspectral_fun. The source code of
the function GUI_hyperspectral_diff is shown below:

function
[diff_test_reference,wavelength_reference_test]=GUI_hy
perspectral_diff(reference,test)

tw=test.Wavelength_;
td=test.anal;

pw=reference.Wavelength_;
pd=reference.anal;
twd=[tw,td];
pwd=[pw,pd];

twd(min(pwd(:,1))>twd(:,1),:)=[];
twd(max(pwd(:,1))<twd(:,1),:)=[];
pwd(min(twd(:,1))>pwd(:,1),:)=[];
pwd(max(twd(:,1))<pwd(:,1),:)=[];

if (size(twd,1)>0) && (size(pwd,1)>0)

y1 = interp1(twd(:,1),twd(:,4),pwd(:,1));
y2 = interp1(twd(:,1),twd(:,5),pwd(:,1));
y3 = interp1(twd(:,1),twd(:,6),pwd(:,1));
y4 = interp1(twd(:,1),twd(:,7),pwd(:,1));
diff_test_reference=[twd(:,1),abs(pwd(:,4)-
y1)./max(pwd(:,4)),abs(pwd(:,5)-
y2)./max(pwd(:,5)),abs(pwd(:,6)-
y3)./max(pwd(:,6)),abs(pwd(:,7)-y4)./max(pwd(:,7))];
wave-
length_reference_test=[twd(:,1),pwd(:,4:7),y1,y2,y3,y4
];
else
diff_test_reference=[];
wavelength_reference_test=[];
end

The presented source code can be divided into 2 elements:

• standardization of the variables test and reference so that they cover the
same range of wavelengths (of the measured i images). This is due to the
4.3 Basic Analysis of Features 59

versatility of the approach for which, in any case, it is possible to perform

calculations in the proposed application for any range of wavelengths.
• calculation of the values of dS(i) as well as dCON(i), dENE(i) and dHOM(i) – stored
in the variable diff_test_reference.
The values included in the variables test and reference are obtained using
two possibilities:
• the first one is the aforementioned manual selection of the ROI. Manual
selection of the ROI was realized in the fragment of the source code in the file
GUI_hyperspectral_fun, i.e.:

if sw==14
ax=axis(hObj(21));
set(hObj(70),'XData',[ax(1)],'YData',[ax(4)],'Fac
eColor','g');

set(hObj(71),'XData',[ax(1)],'YData',[ax(4)],'Fac
eColor','r');
re=getrect(hObj(21));
rect=[re(1),re(1)+re(3)];
set(hObj(70),'XData',[ax(1) rect(1) rect(1)
rect(2) rect(2) ax(2)],'YData',[ax(4) ax(4) 0 0
ax(4) ax(4) ],'FaceColor','r');
set(hObj(71),'XData',[rect(1)
rect(2)],'YData',[ax(4) ax(4)],'FaceColor','g');
alpha(0.1)
GUI_hyperspectral_fun(8)
end

It (the above code fragment) is invoked in the function GUI_hyperspectral

in the fragment:

hObj(34) =uicontrol('Style', 'pushbut-

ton','units','normalized', 'String',
'SELECT','Position', [0.69 0.95 0.3 0.03],...
'Callback',
'GUI_hyperspectral_fun(14)','BackgroundColor',col
or2);

• the other one is reading from the external file test.mat and/or refer-
ence.mat. The code fragment of the function GUI_hyperspectral_fun
is responsible for reading, i.e.:
60 4 Image Processing

if sw==6 % GET FILE TEST

[filen_, pathn_] = uigetfile('test.mat',
'Select test.mat file');
if isequal(filen_,0)==0
test=load([pathn_,filen_]);
set(hObj(35),'ForegroundColor','g')
GUI_hyperspectral_fun(8)
else
set(hObj(35),'ForegroundColor','r')
errordlg('File test.mat not found','File
Error');
end
end
if sw==7 % GET FILE REFERENCE
[filen_, pathn_] = uiget-
file('reference.mat', 'Select reference.mat
file');
if isequal(filen_,0)==0
reference=load([pathn_,filen_]);
set(hObj(36),'ForegroundColor','g')
GUI_hyperspectral_fun(8)
else
set(hObj(36),'ForegroundColor','r')
errordlg('File reference.mat not
found','File Error');
end
end

When a wrong file is indicated or no file is indicated, there appears errordlg

saying 'File reference.mat not found' and/or
'File test.mat not found'. Reading is possible by placing the two buttons
in the main window, in the file GUI_hyperspectral, i.e.:

hObj(37) =uicontrol('Style', 'pushbut-

ton','units','normalized', 'String',
'OPEN','Position', [0.06 0.48 0.05 0.04],...
'Callback',
'GUI_hyperspectral_fun(6)','BackgroundColor',colo
r2);
hObj(38) =uicontrol('Style', 'pushbut-
ton','units','normalized', 'String',
'OPEN','Position', [0.06 0.44 0.05 0.04],...
'Callback',
'GUI_hyperspectral_fun(7)','BackgroundColor',colo
r2);
4.3 Basic Analysis of Features 61

Fig. 4.7 Graphs of dS(i) as

well as dCON(i), dENE(i) and
dHOM(i) as a function of
wavelength and (black line)
the manually set threshold pq

Fig. 4.8 Graph of changes in

the mean, minimum and
maximum brightness for the
S ðiÞ
reference LW

Figure 4.7 shows the graphs of dS(i) as well as dCON(i), dENE(i) and dHOM(i) as a
function of wavelength, whereas Fig. 4.8 shows a graph of changes in the mean,
minimum and maximum brightness values for the reference LW S (i).
62 4 Image Processing

Fig. 4.9 The main menu of the application and additional windows: 1 main menu; 2 the results of
ACC, TPR, SPC, TN, TP, FN, FP; 3 GLCM; 4 graphs of LCON(i), LENE(i) and LHOM(i); 5 graphs of
dS(i) as well as dCON(i), dENE(i) and dHOM(i); 6 changes in the brightness for the mth row and nth
column selected manually by moving the slider

The first graph (Fig. 4.7) is shown in a separate application window. The second
one (Fig. 4.8) constitutes the right part of the main window—Fig. 4.9.
The values of ACC, TPR, SPC, TN, TP, FN, FP are calculated using the function
GUI_hyperspectral_class, i.e.:
4.3 Basic Analysis of Features 63

function
[ACC,TP,TN,FP,FN,TPR,SPC,meas,measT,nam,TRR]=GUI_hyper
spectral_class(diff_test_reference,wavelength_referenc
e_test,rect,prg,type)
...
global hObj

err=[diff_test_reference(:,1)];
meas=[];
measT=[];
nam=[];
TRR=(err(:,1)>rect(1))&(err(:,1)<rect(2));

if get(hObj(30),'Value')==1
err=[err,diff_test_reference(:,2)];
meas=[meas,wavelength_reference_test(:,2)];
measT=[measT,wavelength_reference_test(:,6)];
nam{length(nam)+1}='L_{INT}(i)';
end
if get(hObj(31),'Value')==1
err=[err,diff_test_reference(:,3)];
meas=[meas,wavelength_reference_test(:,3)];
measT=[measT,wavelength_reference_test(:,7)];
nam{length(nam)+1}='L_{CON}(i)';
end
if get(hObj(32),'Value')==1
err=[err,diff_test_reference(:,4)];
meas=[meas,wavelength_reference_test(:,4)];
measT=[measT,wavelength_reference_test(:,8)];
nam{length(nam)+1}='L_{ENE}(i)';
end
if get(hObj(33),'Value')==1
err=[err,diff_test_reference(:,5)];
meas=[meas,wavelength_reference_test(:,5)];
measT=[measT,wavelength_reference_test(:,9)];
nam{length(nam)+1}='L_{HOM}(i)';
end
if size(err,2)>=2;
if type==1
err=[err,sum(err(:,2:end)>prg,2)>0,(err(:,1)>rect(1))&
(err(:,1)<rect(2))];
TP=sum( (err(:,end-1)==1).*(err(:,end)==1) );
TN=sum( (err(:,end-1)==0).*(err(:,end)==0) );
FP=sum( (err(:,end-1)==1).*(err(:,end)==0) );
FN=sum( (err(:,end-1)==0).*(err(:,end)==1) );
ACC=round( (TP+TN)/(TP+TN+FP+FN).*100);
TPR=round(TP/(TP+FN)*100);
SPC=round(TN/(TN+FP)*100);
64 4 Image Processing

When analysing the next fragments of the proposed source code, it is divided
into two areas:
• in the first area the values are gathered in the variable err constituting the basis
for further analysis. Gathering is directly related to manual (by the user)
determination which features (dS(i) as well as dCON(i), dENE(i) and dHOM(i)) are
taken into account in the analysis. This option is provided by a suitable code
fragment in the GUI_hyperspectral, i.e.:

hObj(30) =uicontrol('Style',
'checkbox','units','normalized', 'String',
'INT','Position', [0.01 0.57 0.05 0.05],...

'Callback','GUI_hyperspectral_fun(8)','Background
Color',color2,'Value',1);
hObj(31)=uicontrol('Style',
'checkbox','units','normalized', 'String',
'CON','Position', [0.051 0.57 0.059 0.05],...

'Callback','GUI_hyperspectral_fun(8)','Background
Color',color2);
hObj(32)=uicontrol('Style',
'checkbox','units','normalized', 'String',
'ENE','Position', [0.01 0.52 0.05 0.05],...

'Callback','GUI_hyperspectral_fun(8)','Background
Color',color2);
hObj(33)=uicontrol('Style',
'checkbox','units','normalized', 'String',
'HOM','Position', [0.051 0.52 0.059 0.05],...

'Callback','GUI_hyperspectral_fun(8)','Background
Color',color2);

When appropriate values (as selected) are added to the variable err, there
follows its binarization with the threshold prg sum(err(:,2:end) >
prg,2) > 0. Then the range of TP and TN located in the manually selected ROI is
calculated, i.e.: (err(:,1) > rect(1))&(err(:,1) < rect(2)).
• In the other area, the values of ACC, TPR, SPC, TN, TP, FN, FP are calculated.
The obtained results are shown in a separate window—Fig. 4.9 (2).
Depending on the option INT, ENE, VAR, HOM (Fig. 4.10) chosen to calculate
ACC, TPR, SPC, TN, TP, FN, FP, different configurations of features are taken into
account. Therefore, for the set threshold prg (in this case equal to 15%), different
results are obtained. For example, for the test and reference ROIs shown in
Fig. 4.11, the results for different configurations of features are presented in
4.3 Basic Analysis of Features 65

Fig. 4.10 Fragment of the

main menu with the elements 10
responsible for image
pre-processing—green and
image processing—red: 1
calculation of intensity; 2
calculation of energy; 3 types
of analysis ERR/DEC
TREE/BAYES/DISC/SVM; 4
threshold prg; 5 on/off
window of classification
results; 6 on/off window of 1 11
results of ACC, TPR, SPC,
TN, TP, FN, FP; 7 on/off 2 12
window of texture analysis; 8
on/off window of GLCM; 9 13
on/off window of graphs of
14
feature errors; 10 choice of
test/pattern analysis; 11 3
calculation of contrast; 12
calculation of homogeneity; 4
13 loading of the external file
5
test.mat; 14 loading of the
external file reference.mat 6
7
8
9

Table 4.2. The task is to verify the quality of the recognition of the skin area for a
selected characteristic spectral region—Fig. 4.11 (right). In this case, 86 mea-
surements of TN and 42 measurements of TP are marked. The number of mea-
surements is equivalent to the number I of images for individual wavelengths.
Table 4.2 shows that the presented simple method of analysis of features does
not work in every case. First, the results of TPR, SPC obtained for individual
features (INT, ENE, VAR or HOM) are at the level of 0 and 100% or 0 and 87% (for
TPR and SPC respectively). Secondly, in the present case, the feature which is
brightness (INT) improves the results regardless of the presence of the other features
66 4 Image Processing

pattern test TN TP
m [pixel]

n [pixel]

Fig. 4.11 Reference and test ROIs and the marked range of TN and TP for the tested case

Table 4.2 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
different configurations of features—on/off position of checkbox objects (INT, ENE, VAR, HOM)a
INT CON ENE HOM ACC TPR SPC TN TP FN FP
0 0 0 1 67 0 100 86 0 42 0
0 0 1 0 67 0 100 86 0 42 0
0 0 1 1 67 0 100 86 0 42 0
0 1 0 0 59 0 87 75 0 42 11
0 1 0 1 59 0 87 75 0 42 11
0 1 1 0 59 0 87 75 0 42 11
0 1 1 1 59 0 87 75 0 42 11
1 0 0 0 90 69 100 86 29 13 0
1 0 0 1 90 69 100 86 29 13 0
1 0 1 0 90 69 100 86 29 13 0
1 0 1 1 90 69 100 86 29 13 0
1 1 0 0 81 69 87 75 29 13 11
1 1 0 1 81 69 87 75 29 13 11
1 1 1 0 81 69 87 75 29 13 11
1 1 1 1 81 69 87 75 29 13 11
a
The value of ‘0’ means that the feature does not occur, ‘1’ that it occurs in calculations
4.3 Basic Analysis of Features 67

(see the last rows in Table 4.2), i.e.: TPR = 69% and SPC = 100%. To sum up, this
simple method of analysis of features cannot use the full potential and results
obtained from texture analysis. Consequently, the selected types of classifiers
described in the following sections were implemented.

4.4 Block Diagram of the Discussed Transformations

The discussed transformations, extraction of features, including the parts of the

source code are presented in the form of a block diagram in Fig. 4.12. This diagram
applies to parts of the algorithm responsible for fundamental analysis of features.
Additionally, it was subdivided into the part responsible for the extraction of fea-
tures. The blocks of image acquisition and pre-processing have been discussed in
earlier chapters of the monograph.
The block diagram will be further supplemented by the blocks associated with
classification. The diagram intentionally does not include the module, the portion of
the source code and the corresponding functionality of tracking changes in the
contour, which, as mentioned above, readers can implement themselves.
The collected features are the basis for the construction of classifiers.

4.5 Measurement of Additional Features

The previous chapter discusses the features obtained from the analysis and pro-
cessing of hyperspectral images, and to be more specific, derived from GLCM
analysis of the ROI selected by the user. This analysis has been implemented in the
described software both in terms of the source codes and the GUI. However, this is
not the only possible implementation as well as not the only possible set of features.
Typical analyses of textures used for hyperspectral images include: quadtree
decomposition, Hough transform, entropy and not discussed above—correlation—
extracted from the GLCM. Acquisition of these new features from the ROI is almost
intuitive, and requires only the knowledge and the correct use of the function
hough, entropy and the parameter of the function graycoprops which is
‘Correlation’. Acquisition of quantitative scalar features from quadtree
decomposition requires the use of qtdecomp and the following code implemented
for:
68 4 Image Processing

Image acquisition

Image pre-processing

Conditional opening
and closing

Min
Feature extraction

Intensity Max

Mean
GLCM
Image processing

Contrast Energy Homo.

Yes/No Yes/No Yes/No Yes/No

Classification

Fig. 4.12 Block diagram of the algorithm part responsible for fundamental analysis of features.
The block diagram is subdivided into a part responsible for the extraction of the features. Blocks of
image acquisition and pre-processing have been discussed in earlier chapters of the monograph.
This diagram includes one of the blocks highlighted in blue whose functionality has not been
included in the main application
4.5 Measurement of Additional Features 69

L1=load('D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube50.mat');
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,[512 512]);
figure; imshow(Lgray)
L2 = qtdecomp(Lgray,.27);
L3 = zeros(size(L2));
pam=[];
for q=0:9
pam=[pam;[2^q length(find(L2==2^q))]];
if pam(end,2)>0
L4=ones([2^q 2^q pam(end,2)]);
L4(2:2^q,2:2^q,:)=0;
L3=qtsetblk(L3,L2,2^q,L4);
end
end
L3(end,1:end)=1;
L3(1:end,end)=1;
L5=Lgray; L5(L3==1)=1;
figure, imshow(L5,[]);
xlabel('n
[pixel]','FontSize',14,'FontAngle','Italic');
ylabel('m [pixel]','FontSize',14,'FontAngle','Italic')
figure; plot(pam(:,1),pam(:,2),'-r*'); hold on; grid
on
xlabel('N_q=M_q
[pixel]','FontSize',14,'FontAngle','Italic');
ylabel('number of blocks [/]','FontSize',14);

The presented source code contains a loop that enables to change the size of the
sought areas (q=0:9 for a code fragment find(L2==2^q)). If at least one area of
this size (if pam(end,2)>0) is found, it is filled with blocks:
L4=ones([2^q 2^q pam(end,2)]);
L4(2:2^q,2:2^q,:)=0;
The results obtained for the threshold above which the division into smaller
blocks was performed, i.e.: pqt = 0.27, is shown in Figs. 4.13 and 4.14.
Figure 4.13 shows the image LGRAY(m, n, i) with a superimposed division into
individual blocks sized from Mq
Nq = 1
1 pixel to Mq
Nq =
512
512 pixels. The size of each block is a power of 2, i.e.: it is equal to 2q for
q 2 (0, 10) for the analysed case. Attention should be paid here to the need to resize
the image LGRAY(m, n, i) to the size of rows and columns that are a power of two.
Figure 4.14 shows a graph of the total number of blocks Mq
Nq as a function of
70 4 Image Processing

Fig. 4.13 Image LGRAY(m, n, i) with superimposed division into individual blocks sized Mq
Nq

Fig. 4.14 Graph of the total

number of blocks as a
function of their size
Mq
Nq
4.5 Measurement of Additional Features 71

their size. This type of analysis can complement the existing analysis presented in
the previous chapters. It may also be carried out for all i images of the sequence (for
different wavelengths). In this case (analysis of an image sequence for i 2 (1, 128)),
the source code has been modified to the following form:

pam=[];hObj=waitbar(0,'Please wait...');
for i=1:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgray=mat2gray(L1.L1);
Lgray=imresize(Lgray,[512 512]);
L2 = qtdecomp(Lgray,.27);
L3 = zeros(size(L2));
for d=0:9
pam(i,d+1)=length(find(L2==2^d));
end
waitbar(i/128)
end
close(hObj)
figure, mesh(pam);
xlabel('q+1
[pixel]','FontSize',14,'FontAngle','Italic');
ylabel('i [/]','FontSize',14,'FontAngle','Italic');
zlabel('number of blocks
[/]','FontSize',14,'FontAngle','Italic');
figure, plot(pam(:,1),'-r*'); hold on; grid on
xlabel('i [/]','FontSize',14,'FontAngle','Italic');
ylabel('number of blocks
[/]','FontSize',14,'FontAngle','Italic')

As is apparent from the above source code, the said modification involves the
introduction of automatic analysis of all images of the sequence (i 2 (1, 128)), and
plotting, at the end of the algorithm, a graph. The results obtained are shown in
Figs. 4.15 and 4.16.
Figure 4.15 shows a graph of changes in the number of blocks sized Mq
Nq as
a function of their size for subsequent i images. As is apparent from the presented
graph, the number of the smallest areas sized Mq
Nq = 1
1 pixel is the greatest
for each i image. Figure 4.16 shows a graph of the total number of blocks sized
Mq
Nq = 1
1 pixel as a function of subsequent i images.
The above are excerpts of the code of m-files
GUI_hyperspectral_qtdecomp_test and
GUI_hyperspectral_qtdecomp_test2 that are available to the reader in
the form of supporting materials attached to this monograph.
Apart from the discussed features that can be analysed, other calculations
can also be performed. New features can be obtained from image analysis using
72 4 Image Processing

Fig. 4.15 Graph of changes

in the number of blocks sized
Mq
Nq as a function of their
size (2q) for subsequent
i images

Fig. 4.16 Graph of the total

number of blocks sized
Mq
Nq = 1
1 pixel as a
function of subsequent
i images

Riesz transform [5–7] or Gabor filtration. The basis will be a Gaussian function for
three dimensions [8–10], i.e.:

hGA ðmGA ; nGA ; iGA ; rm ; rn ; ri Þ

 
m2 n2 i2 ð4:23Þ
¼ AGA
exp GA 2 GA 2 GA 2
2
rm 2
r n 2
ri
4.5 Measurement of Additional Features 73

where

1
AGA ¼ 3 ð4:24Þ
r m
r n
r i
ð 2
pÞ 2

and rm, rn, ri—standard deviation of the mean for three dimensions m, n, i.
mGA, nGA, iGA—values m, n, i normalized to the range from −0.5 to 0.5, for
example for mGA:
m
mGA ¼ 0:5 ð4:25Þ
M

Due to the nature of hyperspectral images of the skin, in practice, it is often

necessary to rotate the mask hGA but only in two dimensions, i.e. the new coor-
dinates (mGAh, nGAh) after rotation are equal (similarly to Sect. 3.1. Affine trans-
formations of the image):

mGAh ¼ mGA
sinðhÞ þ nGA
cosðhÞ ð4:26Þ

nGAh ¼ nGA
cosðhÞ þ mGA
sinðhÞ ð4:27Þ

Also in practical applications [11–16], it is often necessary to use a derivative in

each of the three possible dimensions. Therefore, it was assumed, for simplicity of
calculations, that the variable hdevGA will be the result of calculating the derivative
in three dimensions. The superscript will mean the wth degree of the derivative for
devGA is the first derivative in the third
three consecutive dimensions. For example, h(0,0,1)
dimension, i.e.:

ð0;0;1Þ
hdevGA ðmGAh ; nGAh ; iGAh ; rm ; rn ; ri Þ
@hGA ðmGAh ; nGAh ; iGAh ; rm ; rn ; ri Þ ð4:28Þ
¼
@iGA

On this basis, it is possible to create a pyramid of masks hdevGA for changes in

various arguments. Changes in the angle h in the range h 2 (0, 2p), e.g. every value
of 0.1, are most often used. The results obtained in the form of a pyramid of masks
hdevGA are shown in Fig. 4.17.
74 4 Image Processing

 




 

Fig. 4.17 Pyramid of masks hdevGA for h2(0, 2p) every 0.1 and various degrees of derivatives
hdevGA ð0;0;0Þ ; hdevGA ð1;0;0Þ ; hdevGA ð1;1;0Þ ; hdevGA ð1;1;1Þ ; hdevGA ð2;0;0Þ and hdevGA ð0;2;1Þ . Negative values of
the mask are marked in red, and positive values in blue. In each case, one of the masks for h = 0°
has been placed in the top left corner
4.5 Measurement of Additional Features 75

The masks hdevGA shown in Fig. 4.17 were calculated for h 2 (0,2p) every 0.1
and various degrees of derivatives hdevGA ð0;0;0Þ ; hdevGA ð1;0;0Þ ; hdevGA ð1;1;0Þ ; hdevGA ð1;1;1Þ ;
hdevGA ð2;0;0Þ and hdevGA ð0;2;1Þ . Negative values of the mask are marked in red
(Fig. 4.17), and positive values in blue. These results were obtained using two
functions. The first one is dergauss which calculates the derivative of the
Gaussian function for the row w 2 (0, 4), i.e.:

function y = dergauss(x,sigma,w)
if w==0
y = exp(-x.^2/(2*sigma^2)) / (sigma*sqrt(2*pi));
elseif w==1
y =(-x./(sigma.^2)).*exp(-x.^2/(2.*sigma.^2)) ./
(sigma.*sqrt(2.*pi));
elseif w==2
y =((x.^2-sigma.^2)./(sigma.^4)).*exp(-
x.^2/(2.*sigma.^2)) ./ (sigma.*sqrt(2.*pi));
elseif w==3
y =( (x.^3-3.*x.*sigma.^2) ./(sigma.^6)).*exp(-
x.^2/(2.*sigma.^2)) ./ (sigma.*sqrt(2.*pi));
elseif w==4
y =( (x.^4-
6.*x.^2.*sigma.^2+3.*sigma.^4)./(sigma.^8)).*exp(-
x.^2/(2.*sigma.^2)) ./ (sigma.*sqrt(2.*pi));
else
end

As is apparent from the presented function, for each condition (the degree of
derivative—the value of the variable w), the value of y is calculated using a
different formula.
The second function is Gauss_test with the following source code:
76 4 Image Processing

hdevGA=[];
sigman=0.1;
sigmam=0.1;
M=200;N=200;
devm=3;
devn=0;
hObj = waitbar(0,'Please wait...');
for theta=0:0.1:(2*pi);
[nGA,mGA]=meshgrid(linspace(-0.5,0.5,M),linspace(-
0.5,0.5,N));
nGAtheta=nGA.*cos(theta)+mGA.*sin(theta);
mGAtheta=-nGA.*sin(theta)+mGA.*cos(theta);
hdevGA(1:M,1:N,round(theta*10+1))=dergauss(nGAtheta,si
gman,devn).*dergauss(mGAtheta,sigmam,devm);
if theta ==0
figure; mesh(hdevGA)
xlabel('m
[pixel]','FontSize',14,'FontAngle','Italic')
ylabel('n
[pixel]','FontSize',14,'FontAngle','Italic')
zlabel('h_{devGA}
[rad]','FontSize',14,'FontAngle','Italic')
end

waitbar(theta/(2*pi),hObj)
end
close(hObj)
figure
[n,m,i]=meshgrid( 1:size(hdevGA,2) , 1:size(hdevGA,1)
, 0:0.1:(2*pi));
p1 =
patch(isosurface(n,m,i,hdevGA>1,0.1),'FaceColor',[0 0
1 ],'EdgeColor','none');
alpha(p1,0.2)
p2 = patch(isosurface(n,m,i,hdevGA<-
1,0.1),'FaceColor',[1 0 0 ],'EdgeColor','none');
alpha(p2,0.9)
view(41,24);
camlight; camlight(-80,-10); lighting phong;
grid on
hold on
axis square
axis([1 N 1 M 0 2*pi])
xlabel('m [pixel]','FontSize',14,'FontAngle','Italic')
ylabel('n [pixel]','FontSize',14,'FontAngle','Italic')
zlabel('\theta [deg]','FontSize',14)

The first part of this source code relates to the declaration of variables and
determination of their values. Then the values of matrices nGA,mGA are deter-
mined, which are the basis for calculating the mask hdevGA. Rotation by the angle
4.5 Measurement of Additional Features 77

theta is initially implemented. In the next step, the previously discussed function
dergauss is used. In the final stage, a three-dimensional graph (Fig. 4.17) is
shown using the functions patch and isosurface.
The results of the presented functions (dergauss,Gauss_test) are shown
in Fig. 4.17, but this is only one of many cases of the pyramid. In hyperspectral
imaging the derivative in the third axis is also often used for analysis and acqui-
sition of features (the results shown earlier involve only two axes and rotation). The
function dergauss will be still used as well as the following new command
sequence Gauss_test2:

hdevGA=[];
sigman=0.08;
sigmam=0.08;
sigmai=0.08;
devm=2;
devn=0;
devi=0;
M=100;N=100;I=100;
Ii=linspace(-0.5,0.5,I);
Iid=dergauss(Ii,sigmai,devi);
hObj = waitbar(0,'Please wait...');
for i=1:length(Iid);
[nGA,mGA]=meshgrid(linspace(-0.5,0.5,M),linspace(-
0.5,0.5,N));

hdevGA(1:M,1:N,i)=dergauss(nGA,sigman,devn).*dergauss(
mGA,sigmam,devm).*Iid(i);
waitbar(i/length(Iid),hObj)
end

close(hObj)
figure
[n,m,i]=meshgrid( 1:size(hdevGA,2) , 1:size(hdevGA,1)
, 1:size(hdevGA,3));
p1 =
patch(isosurface(n,m,i,hdevGA>1,0.1),'FaceColor',[0 0
1 ],'EdgeColor','none');
alpha(p1,0.9)
p2 = patch(isosurface(n,m,i,hdevGA<-
1,0.1),'FaceColor',[1 0 0 ],'EdgeColor','none');
alpha(p2,0.9)
view(41,24);
camlight; camlight(-80,-10); lighting phong;
grid on
hold on
axis square
axis([1 N 1 M 0 length(Iid)])
xlabel('m [pixel]','FontSize',14,'FontAngle','Italic')
ylabel('n [pixel]','FontSize',14,'FontAngle','Italic')
zlabel('i [pixel]','FontSize',14,'FontAngle','Italic')
78 4 Image Processing

As in the previously discussed source code, the values of constants are initially
declared along with 3D spaces of variable parameters ([nGA,mGA]=mesh-
grid…). Next, for subsequent i, values of hdevGA(1:M,1:N,i)are declared.
In the last stage, the results in three dimensional space are shown. They are pre-
sented in Fig. 4.18.
The results in Fig. 4.18 indicate the range of variation in the masks hdevGA with
respect to different degrees of derivatives for three dimensions. Other parameters
that can be changed for individual dimensions are h and r. Figure 4.19 shows
different variants of a sequence of masks for different values of the degree of
derivatives, and various values of standard deviations of the mean r. The following
is a portion of the source code that was used to create Fig. 4.19, i.e.:

hdevGA=[];
sigman=0.01;
sigmam=0.18;
sigmai=0.08;
devm=0;
devn=1;
devi=2;
M=100;N=100;I=100;
Ii=linspace(-0.5,0.5,I);
Iid=dergauss(Ii,sigmai,devi);
hObj = waitbar(0,'Please wait...');
for i=1:length(Iid);
[nGA,mGA]=meshgrid(linspace(-0.5,0.5,M),linspace(-
0.5,0.5,N));
theta=i/I*pi;
nGAtheta=nGA.*cos(theta)+mGA.*sin(theta);
mGAtheta=-nGA.*sin(theta)+mGA.*cos(theta);

hdevGA(1:M,1:N,i)=dergauss(nGAtheta,sigman,devn).*derg
auss(mGAtheta,sigmam,devm).*Iid(i);
waitbar(i/length(Iid),hObj)
end

Masks hdevGA defined for different values of parameters enable to acquire fea-
tures that are not available for typical methods of analysis and typical known mask
filters (Sobel, Roberts or Canny). This problem is visible for simple binarization of
a sequence of images LGRAY(m, n, i) for two thresholds equal to 0.5 (Fig. 4.20—red)
and 0.4 (Fig. 4.20—blue).
The results of the convolution of the image LGRAY(m, n, i) with the pyramid of
masks hdevGA ð1;0;0Þ , h = 0° and r = 0.08 (see Fig. 4.18), are shown in Fig. 4.21. The
colours (Fig. 4.21) indicate the results of binarization for the negative areas (blue)
and positive ones (red). The acquisition of a feature from each ith image, necessary
4.5 Measurement of Additional Features 79

 

 

 

Fig. 4.18 Pyramid of masks hdevGA for three dimensions and different degrees of derivatives
hdevGA ð0;0;0Þ ; hdevGA ð1;0;0Þ ; hdevGA ð1;1;0Þ ; hdevGA ð1;1;1Þ ; hdevGA ð2;0;0Þ and hdevGA ð0;2;1Þ (respectively) when
h = 0° and r = 0.08. Negative values of the mask are marked in red, and the positive ones in blue
80 4 Image Processing

m m
 n 
n
i i

m m 


n n
i i

m m 

 n
n
i i

Fig. 4.19 Pyramid of masks hdevGA for three dimensions, different degrees of derivatives and
various values of r. Negative values of the mask are marked in red, and the positive ones in blue
4.5 Measurement of Additional Features 81

Fig. 4.20 Results of

binarization of the image
LGRAY(m, n, i) for two
thresholds equal to 0.5 (red)
and 0.4 (blue)

Fig. 4.21 Result of the

convolution of the image
LGRAY(m, n, i) with the
pyramid of masks hdevGA ð1;0;0Þ ,
h = 0° and r = 0.08 (see
Fig. 4.18). The colours
indicate the results of
binarization for the negative
areas (blue) and positive ones
(red)

for further use in the classification, is dependent on the type of the processing task.
This can be either the maximum or minimum value of the image convolution
LCONV(m, n, i) (Fig. 4.21), as well as the surface area of the areas above or below
the predetermined threshold. As mentioned earlier, pyramids of masks hdevGA are
generally profiled to a specific type of images and their nature. This applies to both
hyperspectral images and other types of medical images.
The final parts of the source code are in the m-files Gauss_test3 and
Gauss_test4 and are available as supplementary material attached to this
monograph.
82 4 Image Processing

References

1. Quinzán I, Sotoca JM, Latorre-Carmona P, Pla F, García-Sevilla P, Boldó E., Band selection
in spectral imaging for non-invasive melanoma diagnosis. Biomed Opt Express. 2013 Apr 1;4
(4):514–9
2. Qi B, Zhao C, Youn E, Nansen C. Use of weighting algorithms to improve traditional support
vector machine based classifications of reflectance data. Opt Express. 2011 Dec 19;19
(27):26816–26
3. Hennessy R, Bish S, Tunnell JW, Markey MK. Segmentation of diffuse reflectance
hyperspectral datasets with noise for detection of Melanoma. Conf Proc IEEE Eng Med Biol
Soc. 2012;2012:1482–5.
4. Chen TF, Baranoski GV. Effective compression and reconstruction of human skin
hyperspectral reflectance databases. Conf Proc IEEE Eng Med Biol Soc.
2015 Aug;2015:7027–30.
5. D. Van De Ville, N. Chenouard and M. Unser, “Steerable Pyramids and Tight Wavelet
Frames in L2(ℝd)”, IEEE Transactions on Image Processing, vol. 20, no. 10, pp. 2705–2721,
October 2011.
6. E. Simoncelli and W. Freeman, “The steerable pyramid: A flexible architecture for multi-scale
derivative computation,” in Proceedings of the International Conference on Image Processing,
Oct. 23–26, 1995, vol 3, pp. 444-447.
7. M. Unser, D. Sage, D. Van De Ville,”Multiresolution Monogenic Signal Analysis Using the
Riesz-Laplace Wavelet Transform”, IEEE Transactions on Image Processing, vol. 18, no. 11,
pp. 2402–2418, November 2009.
8. M. Michaelis and G. Sommer, “A Lie group approach to steerable filters,” Reference
Recognit. Lett., vol. 16, no. 11, pp. 1165–1174, 1995.
9. Y. Hel-Or and P. C. Teo, “Canonical decomposition of steerable functions,” J. Math. Imag.
Vis., vol. 9, no. 1, pp. 83–95, 1998.
10. P. C. Teo and Y. Hel-Or, “Lie generators for computing steerable functions,” Reference
Recognit. Lett., vol. 19, no. 1, pp. 7–17, 1998.
11. E. P. Simoncelli, W. T. Freeman, E. H. Adelson, and D. J. Heeger, “Shiftable multiscale
transforms,” IEEE Trans. Inf. Theory, vol. 38, no. 2, pp. 587–607, 1992.
12. A. Karasaridis and E. Simoncelli, “A filter design technique for steerable pyramid image
transforms,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Processing, May 7–10,
1996, vol. 4, pp. 2387–2390.
13. E. Simoncelli and W. Freeman, “The steerable pyramid: a flexible architecture for multi-scale
derivative computation,” in Proc. Int. Conf. Image Processing, Oct. 23–26, 1995, vol. 3,
pp. 444–447.
14. Y. Wiaux, L. Jacques, and P. Vandergheynst, “Correspondence principle between spherical
and Euclidean wavelets,” Astrophys. J., vol. 632, p. 15, 2005.
15. J. G. Daugman, “Complete discrete 2-D Gabor transforms by neural networks for
image-analysis and compression,” IEEE Trans. Acoust. Speech Signal Process., vol. 36,
no. 7, pp. 1169–1179, Jul. 1988.
16. E. J. Candès and D. L. Donoho, “Ridgelets: A key to higher-dimensional intermittency?,”
Phil. Trans. Roy. Soc. Lond. A., pp. 2495–2509, 1999.
Chapter 5
Classification

The acquired image features such as mean brightness, contrast, energy and
homogeneity can be used for machine learning and classification. Of the many
types of classifiers, decision trees, the naive Bayes classifier, discriminant analysis
and support vector machine were selected. The training mode for all classifiers is
carried out in the same way. The group of data is a set of four features (brightness,
contrast, energy and homogeneity) calculated for two ROIs selected by the user.
The first region, whose data may be also loaded as a reference.mat file, relates to
the training area. The second ROI concerns the test area, and the data on individual
features can also be loaded from an external file test.mat. The length of the test and
training data vectors is dependent on the user and the number of analysed i images.
It is also dependent on the number of common, for both the training and test group,
wavelengths. Each ith image complying with these conditions creates a new record
in the training vector entering four subsequent scalar values. The idea of selecting
the values reference and test and the ROI is shown schematically in Fig. 5.1.
Figure 5.1 shows a schematic diagram of an exemplary pattern and test vector
created from the variables LS(i), LCON(i), LENE(i) and LHOM(i). The results that are
deleted are marked in red, whereas the results participating in training and testing
the classifier in blue. The green area results from harmonisation, for each analysis,
of common wavelengths. In this case, these are the values 900, 901, 902, 903 and
904 nm. The vectors (training and test) thus prepared are used in the construction
and testing of the following classifiers (mentioned above):
• decision trees,
• naive Bayes classifier,
• discriminant analysis and
• support vector machine.

© Springer International Publishing AG 2017 83

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_5
84 5 Classification

ROI ROI

pattern test

LS(i), LCON(i), LENE(i), LHOM(i) LS(i), LCON(i), LENE(i), LHOM(i)

wavelength, intensity, contrast, energy, homogeneity wavelength, intensity, contrast, energy, homogeneity
 
 
 
 
SELECT

SELECT
 
 
 
 
 


Fig. 5.1 Schematic diagram of an exemplary pattern and test vector created from the variables
LS(i), LCON(i), LENE(i) and LHOM(i). The results which are deleted are marked in red, and the results
participating in training and testing the classifier in blue

These classifiers and their implementation are described in detail in the next
subchapters (for this purpose Statistics Toolbox is additionally required). For the
classifiers described in the following subchapters, the same test and pattern data
were used. They were obtained from the manually selected areas of the image
shown in Fig. 4.11. In total, 92 cases of wavelengths for which there should be
compliance with the pattern and 36 cases where such compliance cannot exist were
obtained for the test and training vectors.
The length of both vectors is due to the results of harmonisation of wavelengths
(see Fig. 4.6). In this case, there was complete compatibility of wavelengths—the
analysed ROIs come from the same image LGRAY(m, n, i).

5.1 Decision Trees

Decision trees have been used in machine learning [1] for many years [2–17]. They
have been also used and implemented in Matlab for a few years. Several functions
are designed for this purpose:
• classregtree—function responsible for the tree induction,
• test—function responsible for testing the tree,
• prune—function responsible for pruning the tree.
5.1 Decision Trees 85

The Gini index is the criterion for assessing the split point of decision trees used
in Matlab. Decision trees were induced using the CART algorithm. Classification
with the use of decision trees was implemented in the function
GUI_hyperspectral_class_dec_tree. This function can be divided into
several areas. In the first area, the true and false cases were divided into and
strings, i.e.:

In the next area, the tree is induced and tested for the training data, i.e.:

In the next area, the tree is pruned.

The last part of the function refers to the visualization of both cross-validation
and resubstitution and the site of tree pruning. 2D and 3D graphs are also shown
depending on the number of features selected for analysis, i.e.:
86 5 Classification
5.1 Decision Trees 87

This function (GUI_hyperspectral_class_dec_tree) is activated by

the function GUI_hyperspectral_class in one of the lines verifying the
user’s choice of a decision tree as a classifier, i.e.:
88 5 Classification

Fig. 5.2 Graphs of a the complete decision tree; b dependence of cross-validation errors on the
number of tree nodes; c the pruned decision tree; d the results of classification

The decision trees thus implemented were tested for three features (brightness,
energy, homogeneity) acquired from the reference and test ROIs shown in Fig. 4.11.
The obtained results are presented in Fig. 5.2.
The graph presented in Fig. 5.2d and the pruned decision tree provide in this
case: ACC = 88%, TPR = 56%, SPC = 100%, TN = 92, TP = 20, FN = 16 and
FP = 0. Table 5.1 shows the results of ACC, TPR, SPC, TN, TP, FN, FP (expressed
as a percentage) for various combinations of the features.
As is apparent from Table 5.1, brightness (INT) improves sensitivity to 56%
compared to other combinations of features. In the case of the combinations of some
of the features such as.: HOM, ENE, ENE and HOM, CON and ENE, CON ENE
and HOM, TPR = 0%. Much better results are obtained in the case of the same data
for another type of a classifier.
5.2 Naive Bayes Classifier 89

Table 5.1 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
the pruned decision tree for different combinations of features—on/off position of checkbox
objects (INT, ENE, VAR, HOM)a
INT CON ENE HOM ACC TPR SPC TN TP FN FP
0 0 0 1 66 0 92 85 0 36 7
0 0 1 0 72 0 100 92 0 36 0
0 0 1 1 72 0 100 92 0 36 0
0 1 0 0 75 11 100 92 4 32 0
0 1 0 1 75 11 100 92 4 32 0
0 1 1 0 72 0 100 92 0 36 0
0 1 1 1 72 0 100 92 0 36 0
1 0 0 0 88 56 100 92 20 16 0
1 0 0 1 88 56 100 92 20 16 0
1 0 1 0 88 56 100 92 20 16 0
1 0 1 1 88 56 100 92 20 16 0
1 1 0 0 88 56 100 92 20 16 0
1 1 0 1 88 56 100 92 20 16 0
1 1 1 0 88 56 100 92 20 16 0
1 1 1 1 88 56 100 92 20 16 0
a
The value of ‘0’ means that the feature does not occur, ‘1’ that it occurs in calculations

5.2 Naive Bayes Classifier

One possible classification method implemented in MATLAB is a naive Bayes

classifier [18–31]. It is based on the assumption of the mutual independence of the
independent variables [19–21]. This simple probabilistic classifier provides good
results in hyperspectral image classification. The implementation of the naive Bayes
classifier [23–25] was carried out in the function GUI_hyperspec-
tral_class_naiv_bayes. The greater part of the source code is the same as
in the function GUI_hyperspectral_class_dec_tree for decision trees.
The following shows only a portion of the source code for the most important
differences between these functions, i.e.:
90 5 Classification

Fig. 5.3 Results of classification with a naive Bayes classifier. Negative and positive cases
(wavelengths) are marked in green and red and the classification function in blue: a for features:
intensity (LINT(i)), homogeneity (LHOM(i)) and energy (LENE(i)); b for features: contrast (LCON(i)),
homogeneity (LHOM(i)) and energy (LENE(i))

The results of classification, for the same training and test data as in the case of
decision trees described in the previous subchapter (for the test and training vectors,
92 cases of wavelengths for which there should be compliance with the pattern and
36 cases where such compliance cannot exist), are shown in Fig. 5.3 and Table 5.2.
As follows from the numerical values in Table 5.2, the highest value of
ACC = 94% was obtained for the combination of features INT and ENE. The
smallest value of ACC = 66% was obtained for a single feature HOM.

Table 5.2 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
the nave Bayes classifier for different combinations of features—on/off position of checkbox
objects (INT, ENE, VAR, HOM)a
INT CON ENE HOM ACC TPR SPC TN TP FN FP
0 0 0 1 66 25 83 76 9 27 16
0 0 1 0 92 89 93 86 32 4 6
0 0 1 1 84 69 90 83 25 11 9
0 1 0 0 85 61 95 87 22 14 5
0 1 0 1 82 58 91 84 21 15 8
0 1 1 0 90 83 92 85 30 6 7
0 1 1 1 86 75 90 83 27 9 9
1 0 0 0 89 61 100 92 22 14 0
1 0 0 1 90 64 100 92 23 13 0
1 0 1 0 94 81 99 91 29 7 1
1 0 1 1 91 78 96 88 28 8 4
1 1 0 0 91 67 100 92 24 12 0
1 1 0 1 91 72 99 91 26 10 1
1 1 1 0 91 81 96 88 29 7 4
1 1 1 1 91 81 95 87 29 7 5
a
The value of ‘0’ means that the feature does not occur, ‘1’ that it occurs in calculations
5.3 Discriminant Analysis 91

5.3 Discriminant Analysis

Discriminant analysis [32–38] feasible in Matlab is associated with the function

classify. It enabled to implement one possible type of discriminant function—
quadratic (fits multivariate normal densities with covariance estimates stratified
by group) in the function GUI_hyperspectral_class_disc. Besides the
above type, it is possible to use the following discriminant functions [35–37]:
• linear—fits a multivariate normal density to each group,
• diaglinear—with a diagonal covariance matrix estimate.
• diagquadratic—with a diagonal covariance matrix estimate.
• mahalanobis—uses Mahalanobis distances with stratified covariance
estimates.
Implementation of a discriminant function in the function GUI_hyperspec-
tral_class_disc is as follows:

The above excerpt provides the results presented in Fig. 5.4.

Similarly to the previously discussed classifiers, in Fig. 5.4 negative and positive
cases (wavelengths) are marked in green and red and the classification function in
blue: in Fig. 5.4a for features: intensity (LINT(i)), homogeneity (LHOM(i)) and
energy (LENE(i)); in Fig. 5.4b for features: contrast (LCON(i)), homogeneity
(LHOM(i)) and energy (LENE(i)). For these and other combinations of features, the
numerical results of ACC, TPR, SPC, TN, TP, FN, FP are shown in Table 5.3.
In Table 5.3, there is only one case when ACC = 95%. This is the accuracy
value for the combination of features INT and CON. The minimum value of
accuracy is 65% and occurs for a single feature HOM.
As mentioned above, discriminant analysis enables to use different types of
discriminant function. Figure 5.5 and Table 5.4 show the results obtained for three
features (homogeneity, intensity and energy) for various types of discriminant
analysis: ‘linear’, ‘diaglinear’, ‘diagquadratic’ and ‘mahalanobis’.
92 5 Classification

Fig. 5.4 Results of classification, discriminant analysis—’quadratic’. Negative and positive

cases (wavelengths) are marked in green and red and the classification function in blue: a for
features: intensity (LINT(i)), homogeneity (LHOM(i)) and energy (LENE(i)); b for features: contrast
(LCON(i)), homogeneity (LHOM(i)) and energy (LENE(i))

Table 5.3 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
discriminant analysis for different combinations of features—on/off position of checkbox objects
(INT, ENE, VAR, HOM)a
INT CON ENE HOM ACC TPR SPC TN TP FN FP
0 0 0 1 65 50 71 65 18 18 27
0 0 1 0 91 92 90 83 33 3 9
0 0 1 1 84 75 87 80 27 9 12
0 1 0 0 88 83 90 83 30 6 9
0 1 0 1 84 75 88 81 27 9 11
0 1 1 0 87 92 85 78 33 3 14
0 1 1 1 88 92 86 79 33 3 13
1 0 0 0 92 72 100 92 26 10 0
1 0 0 1 91 75 97 89 27 9 3
1 0 1 0 94 94 93 86 34 2 6
1 0 1 1 94 94 93 86 34 2 6
1 1 0 0 95 86 98 90 31 5 2
1 1 0 1 94 94 93 86 34 2 6
1 1 1 0 90 94 88 81 34 2 11
1 1 1 1 90 97 87 80 35 1 12
a
The value of ‘0’ means that the feature does not occur, ‘1’ that it occurs in calculations

Different results, presented in Table 5.4, are obtained depending on the selected
type of discriminant analysis.
The best results (Table 5.4) are for ‘mahalanobis’ type of discriminant analysis,
i.e. ACC = 97%. The worst results are for ‘diagquadratic’ type, i.e. ACC = 92%.
The differences arise directly from the type of analysis and the distribution of values
of individual features.
5.4 Support Vector Machine 93

Fig. 5.5 Results of classification for three features—homogeneity LHOM(i)), intensity (LINT(i))
and energy (LENE(i)) for different types of discriminant analysis: a ‘linear’; b ‘diaglinear’;
c ‘diagquadratic’; d ‘mahalanobis’. Negative and positive cases (wavelengths) are
marked in green and red and the classification function in blue

Table 5.4 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
different types of discriminant analysis and the features INT, ENE and HOM
Type ACC TPR SPC TN TP FN FP
Linear 93 75 100 92 27 9 0
Diaglinear 94 92 95 87 33 3 5
Quadratic 94 94 93 86 34 2 6
Diagquadratic 92 86 95 87 31 5 5
Mahalanobis 97 92 99 91 33 3 1

5.4 Support Vector Machine

Support vector machine has been one of the most popular classifiers in recent years
[38–51]. It allows for the appointment of a hyperplane that enables to separate two
classes [41, 42, 44] with the greatest possible margin [45–48]. The implementation
of the SVM classifier was carried out in the function GUI_hyperspectral_
94 5 Classification

class_svm which, in turn, uses the function svmtrain. A fragment of the

source code of the function GUI_hyperspectral_class_svm responsible for
SVM classification is shown below:

The results of SVM classification are shown in Fig. 5.6 and Table 5.5. Negative
and positive cases (wavelengths) (Fig. 5.6) are marked in green and red and the
classification function in blue for features: intensity (LINT(i)), homogeneity
(LHOM(i)) and energy (LENE(i))—Fig. 5.6a); for features: contrast (LCON(i)),
homogeneity (LHOM(i)) and energy (LENE(i))—Fig. 5.6b).
Numerical results of individual combinations of features in the training and test
vectors are shown in Table 5.5.
The greatest values of ACC = 91% were obtained for a few combinations of
features ENE, ENE and HOM, INT and HOM, INT and ENE and HOM, INT and
CON and HOM, INT and CON and ENE and HOM. In almost each of these

Fig. 5.6 Results of SVM classification. Negative and positive cases (wavelengths) are marked in
green and red and the classification function in blue: a for features: intensity (LINT(i)),
homogeneity (LHOM(i)) and energy (LENE(i)); b for features: contrast (LCON(i)), homogeneity
(LHOM(i)) and energy (LENE(i))
5.4 Support Vector Machine 95

Table 5.5 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage)
for SVM for different combinations of features—on/off position of checkbox objects (INT, ENE,
VAR, HOM)a
INT CON ENE HOM ACC TPR SPC TN TP FN FP
0 0 0 1 72 0 100 92 0 36 0
0 0 1 0 91 89 91 84 32 4 8
0 0 1 1 91 89 91 84 32 4 8
0 1 0 0 84 50 98 90 18 18 2
0 1 0 1 88 81 91 84 29 7 8
0 1 1 0 87 53 100 92 19 17 0
0 1 1 1 87 69 93 86 25 11 6
1 0 0 0 88 58 100 92 21 15 0
1 0 0 1 91 69 100 92 25 11 0
1 0 1 0 88 58 100 92 21 15 0
1 0 1 1 91 69 100 92 25 11 0
1 1 0 0 89 61 100 92 22 14 0
1 1 0 1 91 69 100 92 25 11 0
1 1 1 0 89 61 100 92 22 14 0
1 1 1 1 91 69 100 92 25 11 0
a
The value of ‘0’ means that the feature does not occur, ‘1’ that it occurs in calculations

combinations, there is the feature HOM. The feature ENE is noteworthy as it is able
(as a single feature) to provide the best results of accuracy for the considered SVM
classifier.
Summing up, the best results for the analysed case are obtained for ‘maha-
lanobis’ type of discriminant analysis, i.e. ACC = 97%. The results should be
treated only illustratively as they present possible problems and methods of analysis
of the results obtained from the implemented classifiers.

5.5 Receiver Operating Characteristics

The receiver operating characteristics (ROC) curve is obtained on the basis of

classification results. Changes in sensitivity as a function of specificity are analysed.
The individual measurement points can, in the general case, be the result of a
classifier operation for different types of changes. A classic example is a change in
the position of the cut-off threshold during data classification. For the described
types of classifiers and the presented data, the ROC curve enables to show extre-
mely important information, which is sensitive to changing parameters. The
assessment of sensitivity relates here to resizing the averaging filter hw (from the
default parameters Mw × Nw = 3 × 3 pixels), choosing the size of the ROIc
96 5 Classification

Classification
decision trees

iv
TPR




1-SPC

Fig. 5.7 Block diagram of calculating the values for the ROC curve and changes in the location of
the TP area in the range of ±iv

(typically Mc × Nc = 40 × 40 pixels), choosing the size of the structural element

SE2 for conditional erosion and dilation (the default size MSE2 × NSE2), and the
values of the thresholds pec and pdc, classifier type (decision trees, naive Bayes
classifier, SVM), choosing the threshold prg, choosing True and False areas
(Fig. 4.8). The potential impact of the settings (selection) of the above variables on
the results obtained must be verified in practice. The impact of changes in the last
parameter—manual selection of True and False areas—has been further described
in the monograph (Fig. 4.8). Figure 5.7 shows the idea of measuring the ROC
curve.
A change in the range of ±iv (Fig. 5.7) of the location of the TP area will affect
the division into TP and TN of the training vector, a different structure of the
decision tree and thus different values of SPC and TPR representing a point on the
ROC plot. The source code responsible for the calculation of the various values
needed to plot the ROC curve is in the function GUI_hyperspectral_class.
It has been deliberately marked (symbol ‘%’) as a comment to encourage, at this
point, the readers to introduce their own element (e.g. a button) on the menu
5.5 Receiver Operating Characteristics 97

Fig. 5.8 ROC curve for

changes in the location of the
TP area in the range of ±iv for
iv = 30

iv =-30
iv =30

that would run the ROC analysis for changes in the range of ±iv. A fragment
of the source code responsible for changes in the range of ±iv for iv = 30 is as
follows:

The result is the ROC curve presented in Fig. 5.8.

98 5 Classification

The above example is only one possible application of this approach for plotting
ROC curves. I encourage the reader to further test and analyse changes in the value
of sensitivity and specificity for changes in other aforementioned variables, for
example, the impact of resizing the averaging filter hw from Mw × Nw = 3 × 3
(default parameters) to, for example, Mw × Nw = 33 × 33 pixels.

5.6 Pitfalls and Limitations of Classification

Classification in most cases is the crowning stage of the tedious process of image
analysis. Proper preparation of the data vector(s) is extremely important from a
practical point of view [52]. An increase in the length of the training vector and/or a
reduction in the length of the test vector improve the results obtained, but the
created classifier is also less universal [53, 54]. In addition, the pressure to improve
the results obtained is high, especially if the area of the publication of results in
scientific journals and (almost) the need to confirm the superiority of the developed
method are taken into account. Apart from changing the length of the training and
test vector, also other errors can be made during classification [52, 55, 56]. The
most common errors are:
• providing the results of classification for the training vector as those for the test
vector,
• reducing the length of the test vector—increasing the length of the training
vector (as mentioned above),
• overfitting,
• an excessive number of features at a too small vector length,
• leakage of data between the training and test data,
• artificial reproduction of data,
• failure to provide an appropriate range of variation.
Ignoring the mathematical relations and moving on to the practical implemen-
tation, two files are taken into account: test.mat and reference.mat. These
files are the result of previously conducted analysis and were previously stored on
the disk. After reading them and standardizing common wavelengths, an SVM
classifier will be built and tested for both the training and test data. The corre-
sponding source code (the part concerning graphs is similar to the previously
described one—in previous subchapters) is shown below:
5.6 Pitfalls and Limitations of Classification 99
100 5 Classification

The results obtained for the training and test data (see %measT = meas;) are
shown in Figs. 5.9 and 5.10.
The results shown in Figs. 5.9 and 5.10 are clearly better (a change in ACC from
41 to 89% for the training vector). However, the classifier due to its characteristics
does not fit to the data (there is no problem of overfitting).
The second discussed issue is reducing the length of the test vector. Let us
assume the length u of the test vector is changed from 1 (one positive or negative
case) to 100 cases (the number of positive and negative cases is random), i.e.
u 2 (1, 100). The results obtained are shown in the graph in Fig. 5.11.

Fig. 5.9 Results of the SVM

classification for exemplary
features of the training data.
TP = 85, TN = 24, FN = 6,
FP = 7, ACC = 89%,
TPR = 93%, SPC = 77%
5.6 Pitfalls and Limitations of Classification 101

Fig. 5.10 Results of the

SVM classification for
exemplary features of the test
data. TP = 19, TN = 31,
FN = 72, FP = 0,
ACC = 41%, TPR = 21%,
SPC = 100%

Fig. 5.11 Graph of changes

in ACC for different lengths of
the test vector

The results shown in the graph in Fig. 5.11 relate to changes in ACC for different
lengths of the test vector. Each value of u is the result of 1000 random data.
Accordingly, the maximum values are marked in green (Fig. 5.11), whereas the
minimum ones in blue. The presented graph indicates almost complete dependence
of the result (ACC) on the appropriate selection of data if there are no more than 10
of them (for the considered case). The more data are drawn, the narrower the range
of variability of ACC. In an extreme case, ACC = 40 ± 20% in the graph in
Fig. 5.11. To sum up, a reduction in the length of the test vector allows for almost
any change in the results of, for example, accuracy.
The third discussed issue is overfitting. Overfitting the data is typical for most
classifiers. For example, this is the induction of binary decision trees without
pruning. This problem is mentioned in the subchapter “Decision trees”. Figure 5.2b
shows a graph of dependence of cross-validation errors on the number of tree
102 5 Classification

nodes. The problem of overfitting produces very good results of accuracy, sensi-
tivity and specificity. Unfortunately, an induced decision tree (classifier) is not able
to generalize data, so its practical clinical usefulness is limited. Overfitting the data
applies not only to the classifier construction. It is also related to the type of data
source.
The induction of the decision tree for data from a single source, one hospital
and/or one imaging device, causes excessive fit to the data despite, for example,
pruning the tree. A similar problem occurs in the analysis of data from a single
operator (physician) operating the imaging device.
The discussed issues are well illustrated by the example of the induction of a
decision tree for the training data and showing the results obtained for the training
and test data of both the complete and pruned decision tree—Table 5.6.
The results presented in Table 5.6 show small differences in accuracy for the
pruned trees and test and training data (5%). There are also big differences between
the test and training data obtained for the complete decision tree—the difference in
accuracy of 35%. Pruning the decision tree produces worse results but the created
decision tree is less sensitive to the data. In addition, transparency increases and the
computational complexity of the created tree decreases.
The fourth issue is an excessive number of features in relation to the vector
length (number of analysed cases). To some extent this is justified. In the case of the
tedious process of designing the algorithm for analysis and processing, the cul-
mination of this work is the data vector—the vector of features. Adding a new
feature is the result of a small amount of work compared to the said process of
algorithm designing. Therefore, the authors of various works induce a classifier and
test it for dozens and even hundreds of features which are acquired from only a few
cases of data (patients). The question to which the answer will be given below
refers to the maximum number of features that can be used for classification so as to
obtain reliable and diagnostically useful results. For this purpose, the SVM clas-
sifier was used and a different number of features k 2 (1, 50) and different lengths
of the vector u 2 (4, 50) were chosen at random (uniform distribution in the set {0,
1}). The results of ACC are shown in Fig. 5.11 and Fig. 5.12.
The source code for performing these calculations and creating the graph shown
in Fig. 5.11 is as follows:

Table 5.6 Results of TN, TP, FN, FP as well as ACC, TPR, SPC (expressed as a percentage) for
the complete and pruned decision tree for the training and test data
Decision tree Data type ACC TPR SPC TN TP FN FP
Complete Traininga 92 91 92 52 60 5 5
Complete Test 57 26 85 15 55 42 10
Pruned Training 73 75 71 43 46 14 19
Pruned Test 68 79 58 45 38 12 27
a
The lack of specificity, sensitivity and accuracy equal to 100% for the complete decision tree and
the test data is due to 10-fold cross-validation of data which is determined by default
5.6 Pitfalls and Limitations of Classification 103

The result of changes in ACC as a function of changes in the number of features

and vector length shown in Fig. 5.11 is of great practical importance. It enables to
determine the minimum u/k ratio which provides diagnostically useful results.
Figure 5.13 shows the dependence of ACC as a function of u/k which shows that
for u/k = 10 and less, the results exceed the value of 50% of accuracy. This means
that the classification accuracy of the classifier, regardless of the data (they may be
even random), is well above 50%, and for u/k = 5 and less accuracy is approxi-
mately 100%. Therefore, the practical utility of such a classifier is questionable. It is
therefore necessary to provide in each case at least several-fold increase in the
training vector length in relation to the number of features.
The fifth problem is the data leakage between the training and test data. The
leakage is further understood as the overlap of data from the training and test
vectors. If we assume that s will mean the percentage overlap of the two data
104 5 Classification

Fig. 5.12 Graph of changes

in ACC for different lengths of
the test vector u and a
different number of features k

vectors, the result of accuracy (ACC) for different values of s 2 (0, 100) will run for
the SVM classifier in accordance with the graph shown in Fig. 5.14.
According to the graph shown in Fig. 5.14 and as expected, the larger the
percentage share of the test vector in the training vector, the better the results
obtained. For example, for a few per cent (from 0 to about 20) of common data, the
value of ACC does not change. For s ranging from 40 to 85%, the value of ACC
increased by 9%. It should be noted here that data leakage is mainly associated with
the wrong (intended or unintended) implementation of the classifier.
The last mentioned problems, i.e. artificial duplication of data and failure to
provide an adequate range of variability of data, have already been partially dis-
cussed in the previous problems.
In the presented application related to hyperspectral images, the number of
features is disproportionately smaller than the length of the data vector (i = 128 for
4 features). Lack of data leakage ensures proper implementation of the transfor-
mations discussed in earlier chapters. Only deliberate action, such as reducing the

Fig. 5.13 Graph of changes

in ACC for different values of
the ratio of the test vector
length u to the number of
features k
5.6 Pitfalls and Limitations of Classification 105

Fig. 5.14 Graph of changes

in ACC for different s degrees
of overlap of the test vector
and the training vector

training vector length to several 2D images can cause the discussed problems. At
this point, I encourage the reader to create appropriate warnings for the application
user about the occurrence of one of the discussed problems, for example using the
function warndlg .
The source codes of all the discussed problems have been saved in separate m-
files in the materials attached to this monograph, namely Class_test,
Class_test2, Class_test3, Class_test4 and Class_test5.

5.7 Blok Diagram of the Discussed Transformations

The block diagram of the discussed issues is presented in Fig. 5.15.

The block diagram presented in Fig. 5.15 concerns the algorithm part respon-
sible for the selection of a classifier. Depending on the operator’s individual choice,
it is possible to analyse in the proposed application the results obtained for the
following classifiers: decision trees, naive Bayes classifier, discriminant analysis
and support vector machine. In Matlab in Statistics Toolbox, implementation of
other types of classifications, e.g., k-means, is also possible. I encourage the
readers at this point to perform their own implementation of the above functions for
one of the possible definitions of distance, i.e.: ‘sqEuclidean’ (squared eucli-
dean distance); ‘cityblock’ (sum of absolute differences); ‘cosine’ (one minus
the cosine of the included angle between points); ‘correlation’ (one minus the
sample correlation between points); ‘Hamming’—(percentage of bits that differ).
106 5 Classification

Fig. 5.15 Block diagram of

the algorithm part responsible Image acquisition
for the selection of a classifier
and for the ROC curves Image pre-processing

Image processing

Receiver Operating Characteristics

Classification

discriminant analysis
decision trees

Naive Bayes

SVM

References

1. Ye F, Chen ZH, Chen J, Liu F, Zhang Y, Fan QY, Wang L Chi-squared Automatic Interaction
Detection Decision Tree Analysis of Risk Factors for Infant Anemia in Beijing, China. Chin
Med J (Engl). 2016 20th May;129(10):1193–1199. doi:10.4103/0366-6999.181955.
2. Heikkilä P, Forma L, Korppi M. High-flow oxygen therapy is more cost-effective for
bronchiolitis than standard treatment-a decision-tree analysis. Pediatr Pulmonol. 2016 May 5
3. Kadi I, Idri A. Cardiovascular Dysautonomias Diagnosis Using Crisp and Fuzzy Decision
Tree: A Comparative Study. Stud Health Technol Inform. 2016;223:1–8.
4. Kraszewska-Głomba B, Szymańska-Toczek Z, Szenborn L. Procalcitonin and C-reactive
protein-based decision tree model for distinguishing PFAPA flares from acute infections.
Bosn J Basic Med Sci. 2016 Mar 10;16(2):157–61. doi:10.17305/bjbms.2016.974.
References 107

5. Fernández L, Mediano P, García R, Rodríguez JM, Marín M. Risk Factors Predicting

Infectious Lactational Mastitis: Decision Tree Approach versus Logistic Regression Analysis.
Matern Child Health J. 2016 Apr 11. [Epub ahead of print]
6. Garcia RJ, von Winterfeldt D. Defender-Attacker Decision Tree Analysis to Combat
Terrorism. Risk Anal. 2016 Apr 1. doi:10.1111/risa.12574. [Epub ahead of print]
7. Bamber JH, Evans SA. The value of decision tree analysis in planning anaesthetic care in
obstetrics. Int J Obstet Anesth. 2016 Feb 27. pii: S0959-289X(16)00036-4. doi:10.1016/j.ijoa.
2016.02.007. [Epub ahead of print] Review.
8. Leach HJ, O’Connor DP, Simpson RJ, Rifai HS, Mama SK, Lee RE. An exploratory decision
tree analysis to predict cardiovascular disease risk in African American women. Health
Psychol. 2016 Apr;35(4):397–402. doi:10.1037/hea0000267.
9. Wylie CE, Shaw DJ, Verheyen KL, Newton JR. Decision-tree analysis of clinical data to aid
diagnostic reasoning for equine laminitis: a cross-sectional study. Vet Rec. 2016 Apr 23;178
(17):420. doi:10.1136/vr.103588. Epub 2016 Mar 11.
10. Oksel C, Winkler DA, Ma CY, Wilkins T, Wang XZ. Accurate and interpretable nanoSAR
models from genetic programming-based decision tree construction approaches.
Nanotoxicology. 2016 Apr 6:1–12. [Epub ahead of print]
11. Montorsi F, Oelke M, Henneges C, Brock G, Salonia A, d’Anzeo G, Rossi A, Mulhall JP,
Büttner H. Exploratory Decision-Tree Modeling of Data from the Randomized REACTT
Trial of Tadalafil Versus Placebo to Predict Recovery of Erectile Function After Bilateral
Nerve-Sparing Radical Prostatectomy. Eur Urol. 2016 Mar 3. pii: S0302-2838(16)00214-1.
doi:10.1016/j.eururo.2016.02.036. [Epub ahead of print]
12. Baneshi MR, Haghdoost AA, Zolala F, Nakhaee N, Jalali M, Tabrizi R, Akbari M. Can
Religious Beliefs be a Protective Factor for Suicidal Behavior? A Decision Tree Analysis in a
Mid-Sized City in Iran, 2013. J Relig Health. 2016 Feb 29. [Epub ahead of print]
13. Medeiros LB, Trigueiro DR, Silva DM, Nascimento JA, Monroe AA, Nogueira JA,
Leadebal OD. Integration of health services in the care of people living with aids: an approach
using a decision tree Cien Saude Colet. 2016 Feb;21(2):543–552. English, Portuguese.
14. Hashem S, Esmat G, Elakel W, Habashy S, Abdel Raouf S, Darweesh S, Soliman M,
Elhefnawi M, El-Adawy M, ElHefnawi M. Accurate Prediction of Advanced Liver Fibrosis
Using the Decision Tree Learning Algorithm in Chronic Hepatitis C Egyptian Patients.
Gastroenterol Res Pract. 2016;2016:2636390. doi:10.1155/2016/2636390. Epub 2016 Jan 6.
15. Collet JF, Lacave R, Hugonin S, Poulot V, Tassart M, Fajac A. BRAF V600E detection in
cytological thyroid samples: A key component of the decision tree for surgical treatment of
papillary thyroid carcinoma. Head Neck. 2016 Feb 8. doi:10.1002/hed.24393. [Epub ahead of
print]
16. Kang Y, McHugh MD, Chittams J, Bowles KH. Utilizing Home Healthcare Electronic Health
Records for Telehomecare Patients With Heart Failure: A Decision Tree Approach to Detect
Associations With Rehospitalizations. Comput Inform Nurs. 2016 Apr;34(4):175–82. doi:10.
1097/CIN.0000000000000223.
17. Owen EB, Woods CR, O’Flynn JA, Boone MC, Calhoun AW, Montgomery VL. A Bedside
Decision Tree for Use of Saline With Endotracheal Tube Suctioning in Children. Crit Care
Nurse. 2016 Feb;36(1):e1–e10. doi:10.4037/ccn2016358.
18. Asafu-Adjei JK, Betensky RA.A Pairwise Naïve Bayes Approach to Bayesian Classification.
Intern J Reference Recognit Artif Intell. 2015 Oct 1;29(7). pii: 1550023.
19. Chen C, Zhang G, Yang J, Milton JC, Alcántara AD. An explanatory analysis of driver injury
severity in rear-end crashes using a decision table/Naïve Bayes (DTNB) hybrid classifier.
Accid Anal Prev. 2016 May;90:95–107.
20. Geng H, Lu T, Lin X, Liu Y, Yan F. Prediction of Protein-Protein Interaction Sites Based on
Naive Bayes Classifier. Biochem Res Int. 2015;2015:978193.
21. Miasnikof P, Giannakeas V, Gomes M, Aleksandrowicz L, Shestopaloff AY, Alam D,
Tollman S, Samarikhalaj A, Jha P. Naive Bayes classifiers for verbal autopsies: comparison to
physician-based classification for 21,000 child and adult deaths. BMC Med. 2015 Nov
25;13:286.
108 5 Classification

22. Griffis JC, Allendorfer JB, Szaflarski JP. Voxel-based Gaussian naïve Bayes classification of
ischemic stroke lesions in individual T1-weighted MRI scans. J Neurosci Methods. 2016 Jan
15;257:97–108.
23. Wang M, Zuo W, Wang Y. A Multilayer Naïve Bayes Model for Analyzing User’s
Retweeting Sentiment Tendency. Comput Intell Neurosci. 2015;2015:510281.
24. Marucci-Wellman HR, Lehto MR, Corns HL.A practical tool for public health surveillance:
Semi-automated coding of short injury narratives from large administrative databases using
Naïve Bayes algorithms. Accid Anal Prev. 2015 Nov;84:165–76. doi:10.1016/j.aap.2015.06.
014. Epub 2015 Sep 26.
25. Carvajal G, Roser DJ, Sisson SA, Keegan A, Khan SJ. Modelling pathogen log10 reduction
values achieved by activated sludge treatment using naïve and semi naïve Bayes network
models. Water Res. 2015 Nov 15;85:304–15. doi:10.1016/j.watres.2015.08.035. Epub 2015
Aug 21.
26. Dou Y, Guo X, Yuan L, Holding DR, Zhang C. Differential Expression Analysis in RNA-Seq
by a Naive Bayes Classifier with Local Normalization. Biomed Res Int. 2015;2015:789516.
27. Pogodin PV, Lagunin AA, Filimonov DA, Poroikov VV. PASS Targets: Ligand-based
multi-target computational system based on a public data and naïve Bayes approach.
SAR QSAR Environ Res. 2015;26(10):783–93.
28. Waleska Simões P, Mazzuchello LL, Toniazzo de Abreu LL, Garcia D, dos Passos MG,
Venson R, Bisognin Ceretta L, Veiga Silva AC, da Rosa MI, Martins PJ. A Comparative
Study of Bayes Net, Naive Bayes and Averaged One-Dependence Estimators for
Osteoporosis Analysis. Stud Health Technol Inform. 2015;216:1075.
29. Minnier J, Yuan M, Liu JS, Cai T.Risk Classification with an Adaptive Naive Bayes Kernel
Machine Model. J Am Stat Assoc. 2015 Apr 22;110(509):393–404.
30. Zhang H, Yu P, Zhang TG, Kang YL, Zhao X, Li YY, He JH, Zhang J. In silico prediction of
drug-induced myelotoxicity by using Naïve Bayes method. Mol Divers. 2015 Nov;19(4):
945–53.
31. Prinyakupt J, Pluempitiwiriyawej C. Segmentation of white blood cells and comparison of
cell morphology by linear and naïve Bayes classifiers. Biomed Eng Online. 2015 Jun
30;14:63. doi:10.1186/s12938-015-0037-1.
32. Close ME, Abraham P, Humphries B, Lilburne L, Cuthill T, Wilson S. Predicting
groundwater redox status on a regional scale using linear discriminant analysis. J Contam
Hydrol. 2016 Apr 29;191:19–32.
33. Wu D, Fu X, Wen Y, Liu B, Deng Z, Dai L, Tan D. High-resolution melting combines with
Bayes discriminant analysis: a novel hepatitis C virus genotyping method. Clin Exp Med.
2016 May 13.
34. McDonald LS, Panozzo JF, Salisbury PA, Ford R. Discriminant Analysis of Defective and
Non-Defective Field Pea (Pisum sativum L.) into Broad Market Grades Based on Digital
Image Features. PLoS One. 2016 May 13;11(5):e0155523.
35. Han B, Peng H, Yan H. Identification of Medicinal Mugua Origin by Near Infrared
Spectroscopy Combined with Partial Least-squares Discriminant Analysis. Pharmacogn Mag.
2016 Apr-Jun;12(46):93–7.
36. Mandelkow H, de Zwart JA, Duyn JH. Linear Discriminant Analysis Achieves High
Classification Accuracy for the BOLD fMRI Response to Naturalistic Movie Stimuli. Front
Hum Neurosci. 2016 Mar 31;10:128.
37. Amores-Ampuero A, Alemán I. Comparison of cranial sex determination by discriminant
analysis and logistic regression. Anthropol Anz. 2016 Apr 5:1–8.
38. Abbruzzo A, Tamburo E, Varrica D, Dongarrà G, Mineo A. Penalized linear discriminant
analysis and Discrete AdaBoost to distinguish human hair metal profiles: The case of
adolescents residing near Mt. Etna. Chemosphere. 2016 Jun;153:100–6.
39. Utkin LV, Chekh AI, Zhuk YA. Binary classification SVM-based algorithms with
interval-valued training data using triangular and Epanechnikov kernels. Neural Netw.
2016 Apr 27;80:53–66.
References 109

40. Ghandi M, Mohammad-Noori M, Ghareghani N, Lee D, Garraway L, Beer MA. gkmSVM:

an R package for gapped-kmer SVM. Bioinformatics. 2016 Apr 19. pii: btw203.
41. Lee D. LS-GKM: a new gkm-SVM for large-scale datasets. Bioinformatics. 2016 Mar 15. pii:
btw142.
42. Wang X, Du H, Tan JOnline Fault Diagnosis for Biochemical Process Based on FCM and
SVM. Interdiscip Sci. 2016 Apr 29.
43. Wang MY, Li P, Qiao PL. The Virtual Screening of the Drug Protein with a Few Crystal
Structures Based on the Adaboost-SVM. Comput Math Methods Med. 2016;2016:4809831.
44. Sun H, Pan P, Tian S, Xu L, Kong X, Li Y, Dan Li, Hou T. Constructing and Validating
High-Performance MIEC-SVM Models in Virtual Screening for Kinases: A Better Way for
Actives Discovery. Sci Rep. 2016 Apr 22;6:24817.
45. Astorino A, Fuduli A. The Proximal Trajectory Algorithm in SVM Cross Validation. IEEE
Trans Neural Netw Learn Syst. 2016 May;27(5):966–77.
46. Kong Y, Qu D, Chen X, Gong YN, Yan A. Self-Organizing Map (SOM) and Support Vector
Machine (SVM) Models for the Prediction of Human Epidermal Growth Factor Receptor
(EGFR/ ErbB-1) Inhibitors. Comb Chem High Throughput Screen. 2016;19(5):400–11.
47. Huang X, Shi L, Suykens JA. Solution Path for Pin-SVM Classifiers With Positive and
Negative τ Values. IEEE Trans Neural Netw Learn Syst. 2016 Apr 8.
48. Bogaarts JG, Gommer ED, Hilkman DM, van Kranen-Mastenbroek VH, Reulen JP. Optimal
training dataset composition for SVM-based, age-independent, automated epileptic seizure
detection. Med Biol Eng Comput. 2016 Mar 31.
49. Peker M. A decision support system to improve medical diagnosis using a combination of
k-medoids clustering based attribute weighting and SVM. J Med Syst. 2016 May;40(5):116.
50. Kieslich CA, Smadbeck J, Khoury GA, Floudas CA. conSSert: Consensus SVM Model for
Accurate Prediction of Ordered Secondary Structure. J Chem Inf Model. 2016 Mar 28;56
(3):455–61.
51. Ju Z, Cao JZ, Gu H. Predicting lysine phosphoglycerylation with fuzzy SVM by incorporating
k-spaced amino acid pairs into Chou‫׳‬s general PseAAC. J Theor Biol. 2016 May 21;397:
145–50.
52. Foster, K.R., Koprowski, R., Skufca, J.D.: Machine learning, medical diagnosis, and
biomedical engineering research—commentary. Biomed. Eng. Online 13, 94 (2014).
53. Smialowski P, Frishman D, Kramer S. Pitfalls of supervised feature selection. Bioinformatics.
2010;26(3):440–443.
54. Ennett CM, Frize M. Selective Sampling to Overcome Skewed a priori Probabilities.
Proceed AMIA Symposium. 2000. pp. 225–229
55. Steyerberg EW, Bleeker SE, Moll HA, Grobbee DE, Moons KG. Internal and external
validation of predictive models: a simulation study of bias and precision in small samples.
J Clin Epidemiol. 2003;56(5):441–447
56. Cyran KA, Kawulok J, Kawulok M, Stawarz M, Michalak M, Pietrowska M, Polańska
J. Support Vector Machines in Biomedical and Biometrical Applications. In Emerging
Paradigms in Machine Learning. Springer Berlin Heidelberg. 2013;13:379–417.
Chapter 6
Sensitivity to Parameter Changes

In every algorithm and software, especially those designed for the needs of med-
icine, it is important to assess the algorithm sensitivity to parameter changes [1].
This evaluation should be a standard item for each algorithm. Unfortunately, this is
rarely encountered in practice. The authors of the new software solutions do not
mention these parameters in fear of both lack of interest of buyers of the created
software and the possibility of rejection of the scientific article for that reason.
Sensitivity to parameter changes of any algorithm is usually strongly related to its
internal structure (e.g. setting the parameters of its operation automatically) and the
test method (selection of a method for changing parameters) [2, 3]. It should be
emphasized here that each algorithm allows for errors at the level of 100% in
extreme cases of its application [4, 5]. Therefore, it is very important to link the
range of variability of parameters with the error value.
In the case of medical hyperspectral imaging, the issue of sensitivity to
parameter changes is not easier. In practical terms, there are numerous different
combinations of measured features and parameter changes. It is difficult to separate
those combinations that are not only the most interesting but also the most diag-
nostically important. Certainly the most interesting element is the analysis of the
sensitivity of patient positioning and the whole image acquisition on the classifi-
cation results. This area, however, due to its specificity (very high dependence on
the affected place and the type and severity of the disease), is left to the reader.
Below there is a narrower analysis of the algorithm sensitivity to parameter

© Springer International Publishing AG 2017 111

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_6
112 6 Sensitivity to Parameter Changes

changes. The selected parameter is the mean brightness of the manually selected
ROI. The size of the ROI and its position relative to the original one (specified by
the operator) will be changed.
The evaluation of the algorithm sensitivity in the evaluation of the mean
brightness value for each i frame of the image LGRAY(m, n, i) was performed for
repositioning the ROI, resizing the ROI and its rotation around its axis. The results
obtained are presented in the following subchapters. The evaluation criterion JM in
each case is defined as:

LMROIT ðiÞ  LMROIP ðiÞ

JM ðiÞ ¼  100% ð6:1Þ
1

where
LMROIT(i) and LMROIP(i) are the mean brightness values for subsequent ROIs.
The value of ‘1’ results from the adopted range of the brightness level (from 0 to 1
—variable type double).
For example LMROIT(i) is equal to:

1 X
LMROIT ðiÞ ¼ LGRAY ðm; n; iÞ ð6:2Þ
MROI  NROI m;n2ROI

where MROI and NROI are the number of rows and columns of the ROI.

6.1 Respositioning the ROI

Repositioning the ROI involves changing its position in the row axis by Δm and in
the column axis by Δn. The position of the ROI, in accordance with the operator’s
selection, shown in Fig. 4.11 was adopted by default. Its position was changed in
the range Δm = ±10 pixels and Δn = ±10 pixels. The resolution of the ROI was
MROI × NROI = 100 × 100 pixels. The results obtained are shown in Figs. 6.1
and 6.2.
The presented graphs of sensitivity to repositioning the ROI show that the
brightness changes are smaller than ±2%. The source code providing the graph
shown in Fig. 6.1 is as follows:
6.1 Respositioning the ROI 113

m

Fig. 6.1 Graph of changes in JM for different values of Δm and subsequent i images

n

Fig. 6.2 Graph of changes in JM for different values of Δn and subsequent i images
114 6 Sensitivity to Parameter Changes

L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(80),'.mat']);
Lgrayi=mat2gray(L1.L1);
figure;
[X,Y,I2,RECTP] = IMCROP(Lgrayi);
Lgrayi=IMCROP(Lgrayi,RECTP);
hObj=waitbar(0,'Please wait...');
JM=[];
for i=1:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgrayi=mat2gray(L1.L1);
LgrayiP=IMCROP(Lgrayi,RECTP);
LMROIP=mean(LgrayiP(:));
for deltam=-10:10
RECTT=RECTP; RECTT(2)=RECTT(2)+deltam;
LgrayiT=IMCROP(Lgrayi,RECTT);
LMROIT=mean(LgrayiT(:));
JM(i,deltam+11)=(LMROIT-LMROIP)*100;
end
waitbar(i/128)
end
close(hObj)
[i,deltam]=meshgrid((1:size(JM,2))-10,1:size(JM,1));
figure
mesh(i,deltam,JM); grid on; hold on
ylabel('i [pixel]','FontSize',14,'FontAngle','Italic')
xlabel('\Delta m
[pixel]','FontSize',14,'FontAngle','Italic')
zlabel('J_M [%]','FontSize',14,'FontAngle','Italic')
view(-25,32)

Similar results (Figs. 6.1 and 6.2) are obtained for resizing the ROI, which is
presented in the next subchapter.

6.2 Resizing the ROI

The impact of resizing the ROI (as specified above) on the percentage change in the
mean brightness was determined in the same way as in the previous subchapter. In
this case, the size M × N was changed in the range ΔM == ±10 pixels and
ΔN == ±10 pixels. The results are shown in Figs. 6.3 and 6.4.
Similarly to the results obtained in the previous subchapter, the sensitivity of the
brightness change to resizing the ROI is less than ±2%. The next subchapter
presents the effect of rotation on the change in the mean brightness in the ROI.
6.2 Resizing the ROI 115

M

Fig. 6.3 Graph of changes in JM for different values of ΔM and subsequent i images

N

Fig. 6.4 Graph of changes in JM for different values of ΔN and subsequent i images
116 6 Sensitivity to Parameter Changes

6.3 Rotation of the ROI

Similarly to the previous subchapters, the algorithm sensitivity (the mean brightness
value) to rotation of the ROI (as specified above) was verified. In this case, the
rotation λ ranged from 0° to 360° (Fig. 6.5).
In this case, slightly worse results were obtained. Sensitivity to rotation of the
ROI is the greatest in comparison with its repositioning and resizing. It is related not
only to the participation of new pixels resulting from the rotation of the ROI itself
but also from interpolation problems and the method of filling the missing pixels in
the corners (see the function imrotate with the parameter ‘crop’). The sensi-
tivity values in this case are not greater than 15% compared with the absolute value.
The full source code for the examples discussed above can be found in the m–files
GUI_hyperspectral_para_changes, GUI_hyperspectral_para_
changes2, GUI_hyperspectral_para_changes3 and GUI_hypers-
pectral_para_changes4. Once again I encourage the readers to create their
own m-files designed to assess the algorithm sensitivity to changes in other
parameters or to include classification in the analysis. Extension of this analysis will
provide a lot of useful and new information on the nature of the algorithm operation
and its weaknesses. Especially the latter makes the operator more attentive to their
skilful and careful selection. The analysis can also be based on the ROC curves
presented in one of the earlier chapters.

Fig. 6.5 Graph of changes in JM for different values of λ and subsequent i images
6.4 The Other Parameters 117

m

Fig. 6.6 Graph of changes in ACC for different values of Δm

6.4 The Other Parameters

The analysis of the algorithm sensitivity to parameter changes is a broad issue. Any
change of any parameter located in the acceptable range should be examined in
terms of its impact on the results obtained—brightness in the simplest form, clas-
sification results in an advanced form. Therefore, in addition to the above-discussed
impact of the size, position and rotation of the ROI on the obtained results of the
mean brightness, it is also interesting to analyse the impact of lighting or posi-
tioning the pattern relative to the analysed object. It is possible to analyse inde-
pendently the sensitivity of the classification results to repositioning and resizing
the ROI and the impact of noise in the image (due to both the properties of the
camera and the type and brightness of lighting). Below there are the results of
measurement of ACC for the manually selected ROIs, with the test area being
moved in the row axis by Δm = ±20 pixels and in the column axis by Δn = ±20
pixels. Two features are analysed: the mean brightness in the ROI and the standard
deviation of the mean. The results obtained are shown in Figs. 6.6 and 6.7.
The source code in this case is a bit different than that presented in the previous
subchapters, i.e.:
118 6 Sensitivity to Parameter Changes

L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(80),'.mat']);
Lgrayi=mat2gray(L1.L1);
figure; res=[]; hObj=waitbar(0,'Please wait...');
[X,Y,I,RECTP] = IMCROP(Lgrayi);
for deltam=-20:20
RECTT=RECTP;
RECTT(2)=RECTT(2)+deltam;
meas=[];
measT=[];
for i=1:128
L1=load(['D:/k/_I20_L0-511_13-1-
2016_13.5.59.cube',mat2str(i),'.mat']);
Lgrayi=mat2gray(L1.L1);
LgrayiP=IMCROP(Lgrayi,RECTP);
LgrayiT=IMCROP(Lgrayi,RECTT);
meas(i,1)=mean(LgrayiP(:));
meas(i,2)=std(LgrayiP(:));
measT(i,1)=mean(LgrayiT(:));
measT(i,2)=std(LgrayiT(:));
end
TRR=zeros([size(meas,1), 1]);
TRR(20:110)=1;
species=[];
for ijj=1:length(TRR)
if TRR(ijj)==1
species{ijj}='Yes';
else
species{ijj}='No';
end
end
svmStruct = svmtrain(meas,species');
grpname = svmclassify(svmStruct,measT);
TP=sum( strcmp(grpname,'Yes').*
strcmp(species','Yes') );
TN=sum( strcmp(grpname,'No') .*
strcmp(species','No') );
FN=sum( strcmp(grpname,'No') .*
strcmp(species','Yes') );
FP=sum( strcmp(grpname,'Yes').*
strcmp(species','No') );
ACC= round((TP+TN)/(FN+FP+TN+TP).*100);
TPR= round(TP/(TP+FN).*100);
SPC= round(TN/(TN+FP).*100);
res=[res;[deltam, TP, TN, FN, FP, ACC, TPR, SPC]];
waitbar((deltam+20)/40)
end
6.4 The Other Parameters 119

close(hObj)
figure; plot(res(:,1),res(:,6),'-r*'); hold on; grid
on;
xlabel('\Delta m
[pixel]','FontSize',14,'FontAngle','Italic')
ylabel('ACC [%]','FontSize',14,'FontAngle','Italic')

The source code in its first part enables to manually identify the ROI in the image
for i = 80. Then the test ROI is artificially moved, i.e.: RECTT = RECTP; RECTT
(2) = RECTT(2) + deltam; and two features are calculated: the mean
brightness mean(LgrayiP(:)) and the standard deviation of the mean std
(LgrayiP(:). The length of the training and test vectors is the same and is equal
to the total number of frames, i.e. I = 128. In the next stage, the SVM classifier is
trained (variable meas) and tested, and ACC is calculated (variable measT). The
results are shown in the last part of the presented source code figure; plot…).
Noise can be introduced artificially to the ith sequence of images using the
previously applied function imnoise. In this case, the value of ACC for the SVM
classifier (as in the previous example) was initially analysed for different values of
variance v 2 (0, 1) and a zero mean value. The results obtained are shown in
Fig. 6.8.
The form of the source code is almost identical to the previous example. The only
significant difference is the change in the value of v in each loop circulation in the

n

Fig. 6.7 Graph of changes in ACC for different values of Δn

120 6 Sensitivity to Parameter Changes

Fig. 6.8 Graph of changes in ACC for different values of v

range from 0 to 1 every 0.01 and the notation LgraiTm = imnoise(IMCROP

(Lgrayi,RECTT), 'gaussian',0,v) allowing for the addition of noise to the
image. Only one type of noise ('gaussian') is analysed. By a slight modification of
the source code, similar tests can be performed for the following noise types:
'localvar', 'poisson', 'salt & pepper' or 'speckle'.
It is apparent from the graph presented in Fig. 6.8 that the value of accuracy is
reduced for successive values of variance from 92 to 72%—by exactly 20%. This is
a significant change when compared, for example, with the graphs shown in
Figs. 6.6 and 6.7. For repositioning the test ROI, the change in ACC was only a few
per cent. This information gives the picture of the algorithm sensitivity (in this case
the results of SVM classification) to the degree of noise in the image.
At the end it should be emphasized that the presented results of the evaluation of
the algorithm sensitivity to parameter changes are exemplary. They do not exhaust
in any way the full diversity of hyperspectral images dependent on individual
variability of patients, the imaging area, lighting and many others.

References

1. KOPROWSKI R.: Quantitative assessment of the impact of biomedical image acquisition on the
results obtained from image analysis and processing. BioMedical Engineering OnLine 2014,
13:93.
References 121

2. Gonzalez R, Woods R: Digital Image Processing. New York: Addison-Wesley Publishing

Company; 1992.
3. Sonka M, Michael Fitzpatrick J: Medical Image Processing and Analysis. In Handbook of
Medical, Imaging. Belligham: SPIE; 2000.
4. Dash M, Liu H: Feature Selection for Classification, in Intelligent Data Analysis. New York:
Elsevier; 1997:131–156.
5. Duda R, Hart P, Stork D: Reference Classification. 2nd edition. New York: John Wiley &
Sons, Inc.; 2001.
Chapter 7
Conclusions

This monograph presents both new and known methods of analysis and processing
of hyperspectral medical images. The developed GUI allows for easy and intuitive
performance of basic operations both on a single image and a sequence of hyper-
spectral images. These are operations such as filtration, separation of an object,
measurements of basic and complex texture features as well as classification.
Therefore the developed GUI may be useful both diagnostically in the analysis, for
example, of dermatological images and may also serve as a foundation for software
development. In addition, the monograph presents new approaches to analysis and
processing of hyperspectral images. After minor modifications, they can be used for
other purposes and image analyses. The algorithm sensitivity to changes in the
selected parameters has also been evaluated. The presented source code can be used
without licensing restrictions provided this monograph is cited. It should also be
emphasized here that the author is not responsible for the consequences of wrong
use and operation of this software. Despite the author’s best efforts, errors may
occur in the presented source code. The presented software is deliberately free of
restrictions, which should encourage the reader to its subsequent modifications and
improvements. An equally open issue is the time optimization of the described
methods of image analysis and processing, which has been deliberately omitted in
almost the entire monograph (except Table 4.1). Thus the presented software does
not close the interesting subject of analysis and processing of hyperspectral medical
images.

© Springer International Publishing AG 2017 123

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2_7
Appendix

A set of Matlab m-files is attached to this monograph so that the reader does not
have to rewrite each selected part of the source code from the text. According to the
information given in the monograph, the files have been divided into two
containers:
• GUI_ver_pre.zip—containing 5 m-files enabling to test the initial version of the
application;
• GUI_ver_full.zip—containing 21 test m-files and 15 GUI files.
The container GUI_ver_pre.zip includes 5 m-files with the following names and
functionalities:
• read_envi_header—reading the header from *.hdr file,
• read_envi_data—reading data from *.cube, *.raw or *.dat files,
• GUI_hyperspectral_trans—affine transformations of the image,
• GUI_hyperspectral—main GUI file (run first),
• GUI_hyperspectral_fun—function responsible for the functionality of
individual menu elements.
The container GUI_ver_full.zip includes 36 m-files (including m-files for tests)
with the following names and functionalities:
• read_envi_header—reading the header from *.hdr file (the same as in the
container GUI_ver_pre.zip)
• read_envi_data—reading data from *.cube, *.raw or *.dat files (the same
as in the container GUI_ver_pre.zip)
• GUI_hyperspectral_trans—affine transformations of the image (the
same as in the container GUI_ver_pre.zip)
• GUI_hyperspectral—main GUI file extended with respect to the file from
the container GUI_ver_pre.zip (run first),

© Springer International Publishing AG 2017 125

R. Koprowski, Processing of Hyperspectral Medical Images,
Studies in Computational Intelligence 682,
DOI 10.1007/978-3-319-50490-2
126 Appendix

• GUI_hyperspectral_fun—function responsible for the functionality of

individual menu elements extended with respect to the file from the container
GUI_ver_pre.zip,
• Class_test—m-file for testing different classification variants,
• Class_test2—m-file for testing different classification variants,
• Class_test3—m-file for testing different classification variants,
• Class_test4—m-file for testing different classification variants,
• Class_test5—m-file for testing different classification variants,
• Gauss_test—m-file for testing different variants of the Gaussian function,
• Gauss_test2—m-file for testing different variants of the Gaussian function,
• Gauss_test3—m-file for testing different variants of the Gaussian function,
• Gauss_test4—m-file for testing different variants of the Gaussian function,
• Dergauss—function of the Gaussian function derivatives,
• GUI_hyperspectral_adaptive_filter—function of the adaptive filter,
• GUI_hyperspectral_class—function of selecting classification type,
• GUI_hyperspectral_class_dec_tree—classifier function—decision
trees,
• GUI_hyperspectral_class_disc—classifier function—discriminant
analysis,
• GUI_hyperspectral_naive_bayes—classifier function—naive Bayes
classifier,
• GUI_hyperspectral_class_svm—classifier function—SVM,
• GUI_hyperspectral_diff—function responsible for calculating bright-
ness differences,
• GUI_hyperspectral_dilate_c—function of conditional dilation,
• GUI_hyperspectral_erode_c—function of conditional erosion,
• GUI_hyperspectral_erode_dilate_test—m-file for testing the
properties of conditional erosion and dilation,
• GUI_hyperspectral_erode_dilate_test2—m-file for testing the
properties of conditional erosion and dilation,
• GUI_hyperspectral_erode_dilate_test3—m-file for testing the
properties of conditional erosion and dilation,
• GUI_hyperspectral_erode_dilate_test4—m-file for testing the
properties of conditional erosion and dilation,
• GUI_hyperspectral_filter_test—m-file for testing dedicated
filtration,
• GUI_hyperspectral_filter_test2—m-file for testing dedicated
filtration,
Appendix 127

• GUI_hyperspectral_qtdecomp_test—m-file for testing square-tree

decomposition,
• GUI_hyperspectral_qtdecomp_test2—m-file for testing square-tree
decomposition,
• GUI_hyperspectral_para_change—m-file for testing the effect of
parameter changes on the results obtained,
• GUI_hyperspectral_para_changes2—m-file for testing the effect of
parameter changes on the results obtained,
• GUI_hyperspectral_para_changes3—m-file for testing the effect of
parameter changes on the results obtained,
• GUI_hyperspectral_para_changes4—m-file for testing the effect of
parameter changes on the results obtained.
The two described containers containing all the m-files discussed in this
monograph along with the source codes are available at http://extras.springer.com/.