Journal of Digital Imaging

https://doi.org/10.1007/s10278-019-00269-1

ORIGINAL PAPER

Hybrid Transfer Learning for Classification of Uterine Cervix Images

for Cervical Cancer Screening
Vidya Kudva 1,2 & Keerthana Prasad 1 & Shyamala Guruvare 3

# Society for Imaging Informatics in Medicine 2019

Abstract
Transfer learning using deep pre-trained convolutional neural networks is increasingly used to solve a large number of problems in the
medical field. In spite of being trained using images with entirely different domain, these networks are flexible to adapt to solve a
problem in a different domain too. Transfer learning involves fine-tuning a pre-trained network with optimal values of hyperparameters
such as learning rate, batch size, and number of training epochs. The process of training the network identifies the relevant features for
solving a specific problem. Adapting the pre-trained network to solve a different problem requires fine-tuning until relevant features are
obtained. This is facilitated through the use of large number of filters present in the convolutional layers of pre-trained network. A very
few features out of these features are useful for solving the problem in a different domain, while others are irrelevant, use of which may
only reduce the efficacy of the network. However, by minimizing the number of filters required to solve the problem, the efficiency of
the training the network can be improved. In this study, we consider identification of relevant filters using the pre-trained networks
namely AlexNet and VGG-16 net to detect cervical cancer from cervix images. This paper presents a novel hybrid transfer learning
technique, in which a CNN is built and trained from scratch, with initial weights of only those filters which were identified as relevant
using AlexNet and VGG-16 net. This study used 2198 cervix images with 1090 belonging to negative class and 1108 to positive class.
Our experiment using hybrid transfer learning achieved an accuracy of 91.46%.

Keywords Cervical cancer screening . Deep learning . Transfer learning . Hybrid transfer learning . Machine learning . Medical
image classification . Artificial intelligence

Introduction cancers and is the second most common cause of death among
women worldwide [1]. According to WHO report [2], almost
Cervical cancer is one of the major reasons of illness and mor- 80% of deaths due to cervical cancer occur in developing coun-
tality among women worldwide [1]. It contributes to 12% of all tries. Pap smear is the most commonly used method for screen-
ing of cervical cancer. But it requires efficient networking be-
tween smear collection and cytology laboratories [3]. Also, it
suffers from low sensitivity of 52% [4]. In addition to being an
expensive test, results of this test are available approximately
* Keerthana Prasad after 2 weeks. Hence, this approach cannot facilitate the screen-
[email protected]
ing and treatment in a single visit. VIA is demonstrated as a
Vidya Kudva
simple, cost-effective test that facilitates screen and treat ap-
[email protected] proach [5, 6]. Application of 3–5% acetic acid turns the precan-
cerous lesions white, called acetowhite (AW) regions. Typical
Shyamala Guruvare
[email protected]
cervix images prior to and after the application of acetic acid
are shown in Fig. 1. Precancerous lesions, which have turned
1
Manipal School of Information Sciences, Manipal Academy of white, are marked as AW in Fig. 1b.
Higher Education, Manipal 576104, India These lesions can be destroyed instantly using cryotherapy
2
NMAM Institute of Technology, Nitte, Karkala 574110, India [7]. This screen and treat approach can be performed in a single
3
Department of Obstetrics and Gynecology, Kasturba Medical
visit to the hospitals, which suppresses further development of
College, Manipal 576104, India cervical cancer.
 J Digit Imaging

Fig. 1 Typical cervix images. a

Prior to the application of acetic
acid. b After the application of
acetic acid

Considerable training is required to discriminate between ac- the selection of hand-crafted features and classifying the im-
tual cancer lesions and lesions that are benign. The most impor- ages based on these features using a suitable classifier. These
tant deciding features for VIA-positive lesions are intensity of researchers tried to classify the images by extracting the fea-
acetowhitening, margin of AW lesions, and the rate at which AW tures related to intensity of acetowhitening, margin of AW
lesions appear and disappear. Identification of these features re- lesions, and the texture with in AW lesions. A few researchers
quires considerable amount of skill. Hence, accuracy of this test [10–14] considered acetowhite feature for the classification of
depends on the skill level of the person who performs the test [8]. cervix images into VIA-positive and VIA-negative classes.
Blumenthal et al. [9] reported that health workers need to be Another group [15, 16] used texture within the aceotowhite
trained well for successful implementation of VIA. This problem lesion as feature. Claude et al. [17] and Raad et al. [18] used
can be considered as a challenge in computer vision. Images of lesion margin feature. Identification of relevant features for
cervix during VIA examination can be acquired and analyzed detection of cervical cancer is a complex task. This problem
using image processing algorithms. Deploying these algorithms can be overcome by using deep learning.
in a cell phone or a comparable device with camera can be used Deep learning is a rapidly growing field in machine learn-
as decision support system for the health workers during cervical ing for the analysis of images. It can be used for wide variety
cancer screening. of image analysis tasks such as image classification, image
segmentation, object detection, and image registration. Deep
learning can be used for an image analysis task in two ways as
Related Work shown in Fig. 2.
CNN is a deep learning algorithm used extensively to
State-of-the-art algorithms developed by various researchers analyze images. They are mainly composed of different
used traditional machine learning method, which are based on layers, namely convolutional layer, activation layer,

Image analysis

Transfer
Full training
learning

Extract
Deep layer Shallow layer features from Fine-tune pre-
CNN CNN pre-trained trained CNN
CNN
Fig. 2 Deep learning approaches
J Digit Imaging

Table 1 Architecture of the

CNN-1 Layer type Input size Filter size Number of filters Stride Pad Output size

Data input 227X227X3 NA NA NA NA 227X227X3

Convolution 1 227X227X3 5X5X3 23 4 0 57X57X23
ReLU 1 57X57X23 NA NA NA NA 57X57X23
Max pool 1 57X57X23 3X3 1 2 0 28X28X23
Convolution 2 28X28X23 5X5X23 28 1 2 28X28X28
ReLU 2 28X28X28 NA NA NA NA 28X28X28
Max pool 2 28X28X28 3X3 1 2 0 13X13X28
Fully connected 1 13X13X28 NA NA NA NA 1024X1
Fully connected 2 1024X1 NA NA NA NA 512X1
Fully connected 3 512X1 NA NA NA NA 2X1

NA not applicable

pooling layer, and fully connected layer. CNNs can be is called full training. Depending on the number of layers
employed for medical image classification in two ways. in the architectures, CNNs can be classified as deep layer
The first approach is to train the CNN from scratch which CNN and shallow layer CNN. Full training of CNNs with

Fig. 3 Responses of a few filters in the first convolutional layer of AlexNet for VIA-positive and VIA-negative images

Fig. 4 Responses of a few filters in the first convolutional layer of VGG-16 net for VIA-positive and VIA-negative images
 J Digit Imaging

Input layer
227x227x3

Conv1 Filter weights are adapted

23, 5x5, S=1, from idenfied filters from
P=0 Conv1 of AlexNet and VGG-16
nets
ReLU1

Max pooling
3x3, S=2

Conv01 Conv02
48, 1x1, S=1, 64, 1x1, S=1,
P=0 P=0
Filter weights are adapted Filter weights are adapted
Conv2a Conv2b
from idenfied filters from from idenfied filters from
10, 5x5, S=1, 18, 5x5, S=1,
second Conv of AlexNet second Conv of VGG-16 net
P=2 P=2

Concatenaon layer S-Stride

P-Padding

ReLU2

Max pooling
3x3, S=2

Fully connected layer (1024)

ReLU3

Fully connected layer (512)

ReLU4

Fully connected layer (2)

Classificaon layer
Fig. 5 Architecture of CNN-2

95
deep architecture requires large amount of annotated data
90 which are usually not available in the medical domain.
85 Hence, the most popular deep CNNs [19–24] were trained
80
with images of non-medical domain. Deep CNNs are ca-
Accuracy (%)

75
pable of obtaining full representation of the training data.
70
65
60 Table 2 Performance of CNN-1 for LR = 0.001 and BS = 50
55
50 Accuracy (%) Sensitivity (%) Specificity (%) AUC
45
LR=0.0001 LR=0.001 LR=0.002 LR=0.003 Training 97.16 95.75 98.6 0.97
BS=50 BS=100 BS=150 BS=200 Testing 91.46 89.16 93.83 0.92
Validation 84.85 77.11 92.68 0.85
Fig. 6 Performance of CNN-1 for different learning rates and batch sizes
J Digit Imaging

Fig. 7 Plot of training accuracy

and validation accuracy for LR =
0.001 and BS = 50 for CNN-1

But requirement of large number of annotated data hin- trained from scratch, with initial weights of only those
ders its applicability in the analysis of images in medical filters which were identified as relevant using AlexNet
domain. Hence, full training using shallow layered CNNs and VGG-16 net.
[25–29] was considered for image analysis in medical
domain. To exploit the advantages of deep architecture
of CNN on the limited dataset, another approach known
as transfer learning was utilized. This approach facilitates Method
the use of the knowledge acquired by pre-trained CNNs
which were trained using large number of non-medical Data Collection and Ground Truth
data. Transfer learning can be performed in two ways.
First approach extracts the features from different layers The cervix images were collected from the Kasturba Medical
after retraining them with small dataset [30–33]. The fea- College, Manipal, India. Ground truth for the study was pro-
tures extracted are classified using different classifiers. vided by a Gynaecologic Oncology expert. A total of 231
Second approach of transfer learning involves fine- images were collected of which 31 images were VIA positive.
tuning various layers of pre-trained CNNs to perform a We also collected 1413 cervix images from the National
different task [34–37]. It was demonstrated by these re- Cancer Institute (NCI) and the National Institute of Health
searchers that transfer learning is the best choice for the (NIH) archive. This consisted of 890 VIA-negative and 523
analysis of medical images, where large amount of anno- VIA-positive images. Combining both the datasets, we had
tated data is not available. A large number of filters pres- 1644 images in total with 1090 VIA negative and 554 VIA
ent in the convolutional layers of pre-trained network fa- positive. To increase the size of the dataset in VIA-positive
cilitates extraction of wide variety of features. A very few class, we used data augmentation. We rotated the VIA-
features out of these features are useful for solving the positive images by an angle of 90°. With data augmentation,
problem in a different domain, while others are irrelevant, we have 1108 VIA-positive images in this dataset consisting
use of which reduces the efficacy of the network. of 2198 images.
However, by minimizing the number of filters required
to solve the problem, the efficiency of the training the
network can be improved. In this study, we consider iden- Hybrid Transfer Learning
tification of relevant filters using the pre-trained networks,
namely AlexNet and VGG-16 net to detect cervical cancer We designed two shallow layer CNN architectures from
from cervix images. This paper presents a novel hybrid scratch. We name them as CNN-1 and CNN-2 in this
transfer learning technique, in which a CNN is built and manuscript.

Fig. 8 Plot of training loss and

validation loss for CNN-1 for
LR = 0.001 and BS = 50
 J Digit Imaging

85
53 distinctly highlighted the acetowhite feature in the image.
80 We identified 23 filters which distinctly highlighted the
75 acetowhite information in the image. These 23 filters were
Accuracy (%)

70 adapted into first convolutional layer of the proposed CNN-

65 1. Size of filters in the first convolutional layer of AlexNet and
60 VGG-16 net is 11X11X3 and 3X3X3 respectively. We resized
55 the size of selected filters to 5X5X3, which is the size of the
50 filters in the first convolutional layer of the proposed CNN.
45 We also extracted the filter bias weights of the visualized
LR=0.01 LR=0.02 LR=0.03 LR=0.04 LR=0.001 LR=0.002 LR=0.003 LR=0.004
filters and used them as the initial filter bias weights of the
BS=50 BS=100 BS=150 BS=200 BS=250 proposed CNN-1.
In the similar manner, we visualized responses of the filter
Fig. 9 Performance of CNN-1 by initializing the filter weights to random
values in the second convolutional layer of AlexNet and VGG-16
net. Filters 15, 30, 48, 69, 112, 120, 140, 184, 232, and 248
CNN-1 of AlexNet and filters 2, 3, 9, 10, 11, 13, 16, 17, 19, 20, 23, 25,
26, 30, 42, 53, 55, and 60 of VGG-16 net distinctly highlight-
Architecture of CNN-1 is listed in Table 1. Number of filters ed the acetowhite features. Hence, 28 filters from second
in the convolutional layers of the CNN-1 was selected based convolutional layers of pre-trained networks highlighted
on how many filters in the pre-trained networks highlight the acetowhite feature. Thus, the second convolutional layer of
specific features required. the proposed CNN-1 has 28 filters. Filters were resized to
Two pre-trained networks AlexNet and VGG-16 net were 5X5X23. Since depth of the filters in the second convolutional
considered for visualizing the responses of different filters in layer is required to be 23, we randomly selected 23 slices of 64
convolutional layers. The main feature used for identification slices from VGG-16 net. Similarly, from AlexNet, we ran-
of cervical cancer is acetowhite feature. Further, margin of domly picked up 23 slices out of 48 slices. Bias weights of
acetowhite region and texture within the acetowhite region relevant filters in the second convolutional layer of pre-trained
are used to distinguish between various kinds of possible networks were used as the initial bias weights of the filters in
acetowhite regions. Filter responses were visualized for a the second convolutional layer of the proposed CNN-1. Two
few sample cervix images including VIA-positive and VIA- max pooling layers were used to reduce the dimensionality of
negative category. Responses of filter in the first and second the features. Finally, three fully connected layers were added
convolutional layers of AlexNet and VGG-16 net were visu- whose weights were initialized randomly.
alized. Responses of a few filters in the first convolutional The CNN-1 was trained using cervix images. Dataset was
layer of AlexNet for VIA-positive and VIA-negative images divided as 85% for training, 7.5% for validation, and 7.5% for
is shown in Fig. 3. testing. Two hyperparameters, namely learning rate (LR) and
Out of the 96 filters in the first convolutional layer of batch size (BS), were varied and performance of CNN was
AlexNet, filters 10, 23, 32, 39, 40, 41, 77, 84, and 95 observed. The parameter which provided the best perfor-
highlighted the whiteness features distinctly. Similarly, re- mance was selected. The CNN was trained for 100 epochs.
sponses of a few filters in the first convolutional layer of We compared the performance of the proposed hybrid
VGG-16 net are shown in Fig. 4. CNN-1 with the performance of proposed CNN-1 architecture
Out of the 64 filters in the first convolutional layer of VGG- by initializing the filter weights in the convolutional layers to
16 net, filters 2, 3, 9, 10, 13, 14, 16, 17, 20, 23, 24, 26, 30, and random values.

Fig. 10 A plot of training and

validation accuracy for LR =
0.004 and BS = 250 for CNN-1
when filter weights are initialized
with random values
J Digit Imaging

Fig. 11 Plot of training and

validation loss for CNN-1 when
filter weights are initialized with
random values

CNN-2 Initial learning rate was set to 0.001 and dropped by a factor of
0.1 in every 50 epochs to avoid overfitting.
Architecture of CNN-2 is shown in Fig. 5. As in the CNN-1 We compared the performance of the proposed hybrid
architecture, first convolutional layer (conv1 of Fig. 5) con- transfer learning approach with traditional transfer learning
tains 23 filters weights of which were adapted from identified approach. We fine-tuned pre-trained AlexNet and VGG-16
filters from AlexNet and VGG-16 net which highlighted net using cervix images after replacing the final fully connect-
acetowhite features. In the second convolutional layer of ed layers with a new fully connected layer with two neurons to
CNN-1, depth of the adapted filters was scaled to 23 by ran- suit the proposed two class problem.
dom selection of 23 slices. But in the second convolutional
layer of CNN-2 (conv2a and conv2b of Fig. 5), the entire set
of depths (48 in AlexNet and 64 inVGG-16 net) of adapted
filters from the second convolutional layers of pre-trained net- Results
works was used. This is facilitated through the use of 1X1
convolutional layers (conv01 and conv02 of Fig. 5). Outputs All the experiments were performed using system with Tesla
of conv2a and conv2b layers are concatenated along the K20Xm graphics card of 6GB memory. MATLAB R2019a
depth. The rest of the layers in CNN-2 after concatenation was used to implement the network.
layer is the same as that of layers after conv2 layer of CNN- Performance of CNN-1 for different learning rates and
1. CNN-2 is trained in a similar manner as that of CNN-1. batch sizes is shown in Fig. 6.

Fig. 12 Responses of filters 10,

23, 32, 39, 40, 41, 77, 84, and 95
in the first convolutional layer of
AlexNet before training
 J Digit Imaging

Fig. 13 Responses of
corresponding filters in CNN-1
after training

Fig. 14 Reponses of filters 2, 3, 9,

10. 13, 14, 16, 17, 20, 23, 24, 26,
30, and 53 in the first
convolutional layer of VGG-16
net before training
J Digit Imaging

Fig. 15 Responses of
corresponding filters of VGG-16
net after training

It can be observed from Fig. 6 that for learning rate of 0.001 From Figs. 7 and 8, it can be observed that training and
and batch size of 50, CNN-1 provides the best accuracy of validation accuracy and loss curves are close to each other
91.46%. Hence, we selected the learning rate as 0.001 and which indicate that the proposed CNN-1 is trained properly.
batch size as 50. We also observed the performance of CNN-1 by initializ-
Accuracy, sensitivity, specificity, and AUC values for train- ing the weights of the filters in convolutional layers to random
ing, testing, and validation steps of CNN-1 for LR = 0.001 and values. Figure 9 depicts the performance of CNN-1 by initial-
batch size = 50 are listed in Table 2. izing the filter weights to random values.
A plot of training accuracy and validation accuracy for From Fig. 9, it can be observed that the best accuracy of
LR = 0.001 and BS = 50 is shown in Fig. 7. 84.15% was achieved for LR = 0.004 and BS = 250.
A plot of training loss and validation loss for CNN-1 for A plot of training and validation accuracy for LR = 0.004
LR = 0.001 and BS = 50 is shown in Fig. 8. and BS = 250 is shown in Fig. 10.

Fig. 16 Performance of CNN-2 87

for different batch sizes
86
85
Sensivity(%)

84
Specificity(%)
Accuracy(%)

83
82
81
80
79
78
BS=50 BS=100 BS=150 BS=200 BS=250
Batch size

Accuracy Sensivity Specificity

 J Digit Imaging

Table 3 Performance of CNN-2 for BS = 150 Performance of CNN-2 for different batch sizes is shown in
Accuracy (%) Sensitivity (%) Specificity (%) AUC Fig. 16.
At batch size of 150 CNN performance was better in terms
Training 86.41 79.72 93.20 0.86 of accuracy and sensitivity. Accuracy, sensitivity, specificity,
Testing 85.37 86.75 83.95 0.85 and AUC values for training, testing, and validation steps of
Validation 78.18 75.90 80.49 0.78 CNN-2 for batch size = 150 are listed in Table 3.
A plot of training and validation accuracy and loss for
BS = 150 for CNN-2 is shown in Fig. 17.
A plot of training and validation loss for CNN-1 when filter Performance of traditional transfer learning approach using
weights are initialized with random values is shown in Fig. 11. fine-tuning pre-trained AlexNet, VGG-16 net, and perfor-
By initializing the filter weights to random values, we ob- mance of proposed CNNs is listed in Table 4.
tained an accuracy of 84.15%, sensitivity of 83.13%, specific- From Table 4, it can be observed that proposed CNN-1 and
ity of 85.19%, and AUC of 0.84. CNN-2 perform equally well when compared to the perfor-
To understand if the initialized filters from AlexNet and mance of traditional transfer learning with lesser training time
VGG-16 have improved after training, we visualized the filter and learnable parameters.
outputs of CNN-1 and compared with the corresponding filter
responses of AlexNet and VGG-16 net which were identified
as relevant. Reponses of filters 10, 23, 32, 39, 40, 41, 77, 84,
and 95 in the first convolutional layer of AlexNet for VIA- Discussions
positive image (original image) depicted in Fig. 3 before train-
ing is shown in Fig. 12. One of the principal challenges in deep learning is overfitting,
Responses of corresponding filters in CNN-1 after training as an overfit model is not generalizable to unseen data.
are shown in Fig. 13. Overfitting of a CNN on training data is usually identified
It can be observed from Figs. 12 and 13 that filters 10, 23, and by monitoring the loss and accuracy on the training and vali-
32 do not highlight the acetowhite region distinctly after training, dation datasets. In the case if the model performs well only on
while in other filters, the acetowhite regions are more distinct the training set compared to the validation set, then the model
after training when compared to their responses before training. has likely been overfit to the training data. As per statistical
Reponses of filters 2, 3, 9, 10. 13, 14. 16, 17, 20, 23, 24, 26, learning theory [38], large number of learnable parameters in a
30, and 53 in the first convolutional layer of VGG-16 net for predictor make the model hard to train in the case of limited
VIA-positive image before training is shown in Fig. 14. data samples. As per Zhang et al. [39], with 2n + d number of
Responses of corresponding filters in CNN-1 after training parameters, a simple neural network with two-layers is capa-
are shown in Fig. 15. ble of perfectly fitting any dataset of n samples of dimension
From Figs. 14 and 15, it can be observed that filters 3, 16, d. However, the pre-trained networks used in transfer learning
20, 30, and 53 do not highlight the acetowhite region distinctly have much more than 2n + d parameters which may lead to
after training, while in other filters, the acetowhite regions are overfitting. Also, it requires more computational resources
more distinct or remained same after training when compared and training time as shown in Table 4. There have been a
to their responses before training. In total, 5 filters out of 23 few techniques proposed to limit overfitting, for example, reg-
filters did not highlight the acetowhite feature after training. ularization with dropout, batch normalization, data

Fig. 17 Plot of training and

validation accuracy and loss for
BS = 150 for CNN-2
J Digit Imaging

Table 4 Performance of fine-

tuning AlexNet and VGG-16 net Accuracy (%) Sensitivity (%) Specificity (%) Training time (s) Learnable
parameters (millions)

AlexNet 84.31 93.50 75.00 5079.32 61

VGG-16 84.15 83.13 85.18 62,479.52 138
CNN-1 91.46 89.16 93.83 1506.08 5.4
CNN-2 85.37 86.75 83.95 661.84 5.4

augmentation, and early stopping. In spite of these efforts, proposed method with that of state-of-the-art traditional
there is still a concern of overfitting. The best solution for methods for classification of cervix images as listed in
reducing overfitting is to obtain more training data. A model Table 5 reveals that proposed method performs equally well
trained on a larger dataset typically generalizes better, though and in some cases even better than state-of-the-art methods in
that is not always attainable in medical imaging. Wu et al. [40] spite of larger data compared to the number of data used by
in their experimentation of classifying Tiny ImageNet data state-of-the-art methods.
using deep and shallow pre-trained CNNs concluded that ben- Considering only acetowhite feature of images in the fil-
efits of going deeper in the case of small data is limited. ters, we were able to achieve a classification accuracy of
Hence, solution is reducing architectural complexity. This is 91.46%. Performance of the network may be improved by
achieved in the proposed hybrid transfer learning approach. having more filters in the convolutional layers which highlight
Along with having simple architecture, since weights of filters the margin of acetowhite region and texture within the
in the convolutional layers have been adapted from learnt acetowhite region.
filters, this provides comparable performance to the state-of-
the-art transfer learning method with lesser training time,
learning parameters, and memory requirements. This type of Conclusion
transfer learning can be used for classification when number
of available data are small as in healthcare. Drawback of this A novel hybrid transfer learning approach for cervix image
type of hybrid transfer learning is that it can be performed only classification was reported in this paper. Two shallow layer
on color images as the ImageNet dataset consists of color CNNs were designed and initial weights of the filters were
images having red, green, and blue channels. It cannot be adapted from pre-trained AlexNet and VGG-16 net which
applied to healthcare aspects which has gray scale images as highlight the acetowhite features. These networks were
in the case of radiology. A comparison of performance of trained to classify cervix images. It provided a classification

Table 5 Comparison of performance of proposed method with that of state-of-the-art traditional methods

Authors Features Classifier Number of images Performance

Park et al. [10] RGB intensity values, ratios, KNN, SVM 29 Sensitivity, 79%
gray scale values Specificity, 88%
Li et al. [11] G component of RGB color space, K-means 99 Sensitivity, 94%
contrast feature Specificity,87%
Kim et al. [12] Pyramid color histogram in CIE Lab SVM 2000 Sensitivity, 73%
space (PLAB), pyramid histogram Specificity, 77%
of oriented gradients (PHOG)
Xu et al. [13] PHOG, PLAB, and pyramid histogram AdaBoost 345 Sensitivity, 86.4%
of local binary pattern (PLBP) Specificity,74.2%
Accuracy, 80.3%
Kudva et al. [14] Color, SGLDM, NGTDM SVM 102 Sensitivity, 99.05%
NGLDM, wavelet, LBP Specificity, 97.16%
Accuracy, 97.94%
Ji et al. [15] 24 texture features such as joint Minimum-distance 300 Accuracy, 87.03%
entropy, angle entropy, energy, classifier
contrast, etc.
Claude et al. [17] Contour features Multilayer perceptron 30 Accuracy, 95.8%
Proposed method Hybrid transfer learning CNN 2198 Sensitivity, 91.46%
Specificity, 89.16%
Accuracy, 93.83%
 J Digit Imaging

accuracy of 91.46%. Performance of the network may be im- settings: Focus on visual inspection with acetic acid and cryother-
apy. International Journal of Gynecology & Obstetrics 89(2):S30–
proved by having more filters in the convolutional layers
S37, 2005
which highlight the margin of acetowhite regions and texture 10. Park SY, Follen M, Milbourne A, Rhodes H, Malpica A,
within the acetowhite region. MacKinnon N, MacAulay C, Markey MK, Richards-Kortum R:
Automated image analysis of digital colposcopy for the detection
Acknowledgments This publication is made possible by a sub-agreement of cervical neoplasia. Journal of Biomedical Optics 13(1):014029,
from the Consortium for Affordable Medical Technologies (CAMTech) 2008
at Massachusetts General Hospital with funds provided by the generous 11. Li W, Venkataraman S, Gustafsson U, Oyama JC, Ferris DG,
support of the American people through the United States Agency for Lieberman RW: Using acetowhite opacity index for detecting cer-
International Development (USAID grant number 224581). We would vical intraepithelial neoplasia. Journal of Biomedical Optics 14(1):
like to acknowledge the support of Mark Schiffman, M.D, M.P.H., 014020, 2009
Division of Cancer Epidemiology and Genetics, National Cancer 12. Kim E, Huang X: A Data Driven Approach to Cervigram Image
Institute, USA, for providing us with cervix images. We would like to Analysis and Classification. In: Celebi M, Schaefer G Eds. Color
acknowledge the support of Dr. Suma Nair, Associate Professor, Medical Image Analysis. Lecture Notes in Computational Vision
Community Medicine Department, Kasturba Medical College, Manipal, and Biomechanics, Vol. 6. Dordrecht: Springer, 2013
for facilitating the acquisition of images during the screening programs 13. T. Xu, E. Kim and X. Huang 2015 Adjustable AdaBoost classifier
conducted. and pyramid features for image-based cervical cancer diagnosis. In
Proc. IEEE 12th International Symposium on Biomedical Imaging,
pages 281–285, New York, NY, USA
Compliance with Ethical Standards 14. Kudva V, Prasad K, Guruvare S: Andriod Device-Based Cervical
Cancer Screening for Resource-Poor Settings. Journal of Digital
Disclaimer The contents are the responsibility of Manipal Academy of Imaging 31(5):646–654, 2018
Higher Education and do not necessarily reflect the views of 15. Ji Q, Engel J, Craine E: Texture analysis for classification of cervix
Massachusetts General Hospital, USAID or the United States lesions. IEEE Transactions on Medical Imaging 19(11):1144–1149,
Government. 2000
16. Song D, Kim E, Huang X, Patruno J, Munoz-Avila H, Hein J:
Multimodal entity coreference for cervical dysplasia diagnosis.
IEEE Transactions on Medical Imaging 34(1):229–245, 2015
References 17. I. Claude, R. Winzenrieth, P. Pouletaut and J. C. Boulanger.
Contour features for colposcopic images classification by articial
1. J. Ferlay, I. Soerjomataram, M. Ervik, R. Dikshit, S. Eser, C. neural networks. In Proc. 16th International Conference on Pattern
Mathers, M. Rebelo, D. M. Parkin, D. Forman and F. Bray 2013 Recognition, pages 771–774, Quebec City, Quebec, Canada, 2002.
GLOBOCAN 2012 v1.0, Cancer incidence and mortality world- 18. V. V. Raad, Z. Xue and H. Lange. Lesion margin analysis for auto-
wide: IARC CancerBase No. 11, Lyon, France: IARC [internet]. mated classification of cervicalcancer lesions. In Proc. SPIE
Available from:http://globocan.iarc.fr, . Last accessed on 28-07- Medical Imaging: Image Processing, volume 6144, pages 1–13,
2018. San Diego, California, United States, 2006.
2. World Health Organization 2002 Cervical cancer screening in de- 19. Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
veloping countries: Report of a WHO consultation. Program on S. Guadarrama and T. Darrell 2014 CAFFE: Convolutional
Cancer Control. Last accessed on 28-03-2018. Architecture for Fast Feature Embedding. arXiv:1408.5093. Last
3. Fahey MT, Irwiq L, Macaskill P: Meta-analysis of pap test accura- accessed on 28-03-2018.
cy. American Journal of Epidemiology 141(7):680–689, 1995 20. A. Krizhevsky, I. Sutskever and G. E. Hinton. ImageNet classifica-
4. Bhattacharyya AK, Nath JD, Deka H: Comparative study between tion with deep convolutional neural networks. In Proc. 25th
pap smear and Visual Inspection with Acetic acid (VIA) in screen- International Conference on Neural Information Processing
ing of CIN and early cervical cancer. Journal of Mid-life Health Systems, volume 1, page 1097–1105, Lake Tahoe, Nevada, 2012.
6(2):53–58, 2015 21. LeCun Y, Bottou L, Bengio Y, Haner P: Gradient-based learning
5. Sankaranarayanan R, Wesley R, Thara S, Dhakad N, Chandralekha applied to document recognition. Proceedings of the IEEE 86(11):
B, Sebastian P, Chithrathara K, Parkin DM, Nair MK: Test charac- 2278–2324, 1998
teristics of visual inspection with 4% acetic acid and Lugol’s Iodine 22. P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus and Y.
in cervical cancer screening in Kerala. India. International Journal LeCun. Over-Feat: Integrated recognition, localization and detec-
of Cancer 106(3):404–408, 2003 tion using convolutional networks. In Proc. International
6. Belinson J, Pretorius R, Zhang W, Wu LY, Qiao YL, Elson P: Conference on Learning Representations (ICLR2014), arXiv:
Cervical cancer screening by simple visual inspection after acetic 1312.6229, 2014. Last accessed on 28-03-2018.
acid. Obstetrics & Gynecology 98(3):441–444, 2001 23. K. Simonyan and A. Zisserman 2014 Very deep convolutional net-
7. Chumworathayi B, Blumenthal PD, Limpaphayom KK, Kamsa- works for large-scale image recognition. arXiv:1409.1556. Last
Ard S, Wongsena M, Supaatakorn P: Effect of single-visit VIA accessed on 28-03-2018.
and cryotherapy cervical cancer prevention program in Roi Et, 24. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D.
Thailand: a preliminary report. Journal of Obstetrics and Erhan, V. Vanhoucke and A. Rabinovich 2015 Going deeper with
Gynaecology Research 36(1):79–85, 2010 convolutions. In Proc. IEEE Conference on Computer Vision and
8. Sangwa-Lugoma G, Mahmud S, Nasr SH, Liaras J, Patrick KK, Pattern Recognition, pages 1–9, Boston, MA, USA.
Tozin RR, Drouin P, Lorincz A, Ferenczy A, Franco EL: Visual 25. Zhanga W, Lia R, Dengb H, Wangc L, Lind W, Jia S, Shenc D:
inspection as a cervical cancer screening method in a primary Deep convolutional neural networks for multi-modality isointense
health-care setting in Africa. International Journal of Cancer infant brain image segmentation. NeuroImage 108:214–224, 2015
119(6):1389–1395, 2006 26. Gao Z, Wang L, Zhou L, Zhang J: HEp-2 cell image classification
9. Blumenthal P, Lauterbach M, Sellors J, Sankaranarayanan R: with deep convolutional neural networks. IEEE Journal of
Training for cervical cancer prevention programs in low-resource Biomedical and Health Informatics 21(2):416–428, 2017
J Digit Imaging

27. Lequan Y, Hao C, Dou Q, Qin J, Heng PA: Automated melanoma 34. Chen H, Ni D, Qin J, Li S, Yang X, Wang T, Heng PA: Standard
recognition in dermoscopy images via very deep residual networks. plane localization in fetal ultrasound via domain transferred deep
IEEE Transactions on Medical Imaging 36(4):994–1004, 2017 neural networks. IEEE Journal of Biomedical and Health
28. Sharma H, Zerbe N, Klempert I, Hellwich O, Hufnagl P: Deep Informatics 19(5):1627–1636, 2015
convolutional neural networks for automatic classification of gas- 35. Tajbakhsh N, Shin JY, Gurudu SR, Hurst RT, Kendall CB, Gotway
tric carcinoma using whole slide images in digital histopathology. MB, Liang J: Convolutional neural networks for medical image
Computerized Medical Imaging and Graphics 61:2–13, 2017 analysis: Full training or fine tuning? IEEE Transactions on
29. Kudva V, Prasad K, Guruvare S: Automation of Detection of Medical Imaging 35(5):1299–1312, 2016
Cervical Cancer Using Convolutional Neural Networks. Critical 36. Paras L: Deep convolutional neural networks for endotracheal tube
Reviews in Biomedical Engineering 46(2):135–145, 2018 position and X-ray image classification: Challenges and opportuni-
30. Ribeiro E, Andreas U, Wimmer G, Hafner M: Exploring deep learning ties. Journal of Digital Imaging 30(4):460–468, 2017
and transfer learning for colonic polyp classification. Computational 37. Phillip MC, Malhi HS: Transfer learning with convolutional neural
and Mathematical Methods in Medicine 2016:1–16, 2016 networks for classification of abdominal ultrasound images. Journal
31. Abdolmanafi A, Duong L, Dahdah N, Cheriet F: Deep feature of Digital Imaging 30(2):234–243, 2017
learning for automatic tissue classification of coronary artery using 38. V. Vapnik 2013 The nature of statistical learning theory. Springer
optical coherence tomography. Biomedical Optics Express 8(2): science & business media.
1203–1220, 2017 39. C. Zhang, S. Bengio and M. Hardt. Understanding deep learning
requires rethinking generalization. arXiv:1611.03530v2 [cs.LG] 26
32. Abbas Q, Fondon I, Sarmiento A, Jiménez S, Alemany P:
Feb 2017. Available online at https://arxiv.org/pdf/1611.03530.pdf.
Automatic recognition of severity level for diagnosis of diabetic
Last accessed on 15.06.2019
retinopathy using deep visual features. Medical & Biological
40. J. Wu, Q. Zhang and G. Xu. Tiny ImageNet challenge. Technical
Engineering & Computing 55(11):1959–1974, 2017
report, Stanford University, 2017. Available online at http://cs231n.
33. Beevi KS, Madhu SN, Bindu GR: Automatic mitosis detection in stanford.edu/reports/2017/pdfs/930.pdf. Last accessed on 14-06-
breast histopathology images using Convolutional Neural Network 2019.
based deep transfer learning. Biocybernetics and Biomedical
Engineering 39:214–223, 2019
Publisher’s Note Springer Nature remains neutral with regard to jurisdic-
tional claims in published maps and institutional affiliations.