The Future of Computing: Ubiquitous Applications and Technologies

Automated Analysis of Medical Images in the Healthcare Domain

Parul Chhabra¹, *, Pradeep Kumar Bhatia¹, Vipin Babbar¹

¹ Department of Computer Science & Engineering, G. J. University of Science & Technology, Hisar, Haryana, India

Abstract

During lab tests, thousands of medical images are generated to trace the disease's symptoms. Manual interpretation of this data may consume excessive time and thus may delay diagnosis. Timely detection of critical diseases is very important as their stage can be changed over an interval. Automated analysis of medical data can reduce the gap between disease detection and its diagnosis and it also reduces the overall computational cost. In this paper, this goal will be achieved using different methods (Classification/ Segmentation/ Image Encoding/ Decoding/ Registration/ Restoration/ Morphology).

Keywords: Disease, Diagnosis, Healthcare, Medical image analysis, Prediction.

* Corresponding author Parul Chhabra: Department of Computer Science & Engineering, G. J. University of Science & Technology, Hisar, Haryana, India; E-mail: [email protected]

INTRODUCTION

Traditional healthcare services follow different steps i.e. disease detection, diagnosis, and keeping track of a patient’s history for clinical decision-making, as shown in Fig. (1). Medical data produced by each step must be examined by expert practitioners to avoid the incorrect diagnosis.

The disease detection phase may produce a large set of medical images and precise analysis of these medical images plays an important role in the identification of disease. It can also be used to track the progress of diagnosis as well as different stages of disease w.r.t. patients.

Fig. (1))

Health care services.

Machine learning can improve the efficiency of the analysis process and it can also be used to build a dataset/knowledgebase for healthcare services in such a way that patient/disease statistics can be shared worldwide. Medical images contain data in visual form and only expert practitioners can interpret that data [1-25].

To analyze this data automatically, machine learning offers the following ways as displayed in Fig. (2).

Fig. (2))

Medical image analysis.

Classification: Medical images can be classified w.r.t. disease types/features etc. and they can be used to detect disease and diagnostic purposes [26].

Segmentation: It is used to subdivide an image into multiple segments (i.e. objects/regions). It can be used for pathologies domain/object detection/ recognition, etc. [27].

Image Encoding/Decoding: It is used to compress the image whereas decoding follows the reverse operation to obtain the original image [28].

Registration: It can be used to align and stitch multiple images together for analysis purposes [29].

Restoration: It is used to filter noise level in an image, in order to produce clear and refined output [30].

Morphology: It deals with structural components, pixels, and shapes in a given image [31].

Following are the challenges and limitations of automated medical image analysis:

It requires a large volume of medical datasets, in order to build a training model for prediction.

Dataset validation is required to ensure the accuracy of the training model.

Quite complex to update the existing dataset.

Excessive computational resources are required to manage and process large-scale medical data.

Expert medical practitioners are still required to ensure the validity of outcomes.

The potential impact of automation of medical image analysis is given below:

It can reduce the processing time and computational cost for practitioners.

It can increase the accuracy of clinical decision-making.

It can optimize the errors in the diagnosis process.

Training model can be updated using the patient’s history, and health recovery with respect to recommended treatment.

Researchers have developed a few solutions for the analysis of medical imagery as discussed in the next section.

LITERATURE SURVEY

K. Rasheed et al. [6] investigated the various machine learning (ML) applications for the healthcare domain. Studies found that intelligent solutions can improve the diagnosis accuracy however, there are a few open issues i.e. lack of standards to generate the training models, dataset formats, incompatible interfaces for the data exchange, etc.

R. Buettner et al. [7] highlighted the various ML-based methods that can be utilized for medical image processing i.e. medical image encoding/decoding, segmentation, classification, image registration/restoration, morphological analysis, etc. Study shows that the accuracy of disease detection can be improved using these methods.

D. Tellez et al. [8] developed an image compression method that encodes the histopathology dataset and uses neural networks to compress noise level input. Outcomes show that it can produce refined images with optimal reconstruction error and these images can be easily interpreted by practitioners.

P. Seeböck [9] introduced an ML-based method that uses supervised learning for the analysis of retina images. It enforces binary classification, noise filtering, and clustering over input data to detect anomalies. Experimental results show that it has an average accuracy/ROC curve.

K. Gong et al. [10] introduced an image reconstruction method that builds a learning model using neural networks. It estimates the energy levels (low/high) in a given input and uses multipliers to reconstruct the images. Experiments show that it is more efficient as compared to traditional denoising methods.

Y. Qi et al. [11] developed a neural network-based method to improve the quality of images. It estimates different parameters (contrast/coherence/signal-to-noise- ratio) to reproduce the high-resolution images. Analysis indicates that it is more efficient as compared to existing solutions.

Q. Abbas et al. [12] developed an ML model to analyze medical images. It builds metrics using various processes (segmentation/regression/regeneration/ augmentation/loss function/data loading). Experiments indicate that these metrics can be used to enhance the diagnosis accuracy as well as reduce operational costs.

X. Zhou et al. [13] developed a neural network to classify histopathological data. It uses segmentation to produce outcomes and tests have shown that it is more accurate and efficient as compared to traditional deep-learning neural networks.

H. Guan et al. [14] conducted a survey to analyze the impact of different factors associated with medical images i.e. data type/volume/quality etc. Analysis shows that the accuracy of disease detection and decision-making is affected by learning methods (supervised/unsupervised/semi-supervised) and computational costs may vary due to heterogeneous data types.

X. Wang et al. [15] explored the relationship between image analysis and diagnostic accuracy and developed a classification model using supervised learning. It builds labeled data for each input to perform classification. Outcomes show that it has an optimal computational cost, higher accuracy, and efficiency in contrast with traditional methods.

H. Pinckaers et al. [16] used neural networks to improve the quality of image data. It extracts the Metadata of images and forms a correlation metric to produce high-resolution data. The analysis has shown that it can manage variations in the input size, pixel size, etc. and it has an optimal ROC value.

B. M. Rashed et al. [17] investigated different ML algorithms that can be used for data mining of medical images and these are Support Vector Machine (SVM)/ k-Nearest Neighbors (KNN)/Decision Trees/Random Forest/Logistic Regression. The common processes found in the study are prediction, data mining, classification, regression, clustering, dimension reduction, etc. that can be used for image analysis.

K. Naveen et al. [18] studied the role of machine learning/deep learning algorithms for the medical image analysis. It has been found that learning methods can affect the accuracy of analysis and the computational cost may also vary. This study states that the optimal selection of a learning method is necessary to ensure good outcomes in the classification process.

S. P. Shayesteh et al. [19] developed a method to extract features from ultrasound samples. It uses logistic regression classifier to perform selective feature selection. The analysis has shown that it offers optimal sensitivity with higher accuracy of disease detection in contrast to existing solutions.

T. Zhang et al. [20] identified that noise in images can degrade the outcomes of a classifier and introduced a noise adaption solution that enforces noise patterns over ultrasound samples. Outcomes show that it has higher accuracy under the constraints of noise level variations.

Geetha et al. [21] compared the performance of supervised/unsupervised learning approaches with respect to healthcare data. The analysis has shown that the accuracy of a prediction model may vary with respect to learning methods. Also, the complexity of medical data types may reduce its efficiency, which may also affect the decision making process and increase the operational cost.

N. Nahar et al. [22] explored the association between disease detection and the analysis of X-ray images. The study found that deep learning algorithms can efficiently predict the presence of diseases in given input samples and improve the diagnostic accuracy. However, the study also indicates that there is no single solution to analyze different types of medical imagery.

U. Khan et al. [23] found that ML algorithms can be used to extract the clinical data from medical imagery efficiently by using classification and segmentation

techniques. The outcomes of the analysis can be further used to improve the accuracy of diagnosis and clinical