View of Prediction of Melanoma Skin Cancer Using Veritable Support Vector Machine

(1)

Prediction of Melanoma Skin Cancer Using Veritable Support Vector Machine

M.Lingaraj^1*, A.Senthilkumar², J.Ramkumar³

1,2Sankara College of Science and Commerce, India

3VLB Janakiammal College of Arts and Science, India

ABSTRACT

Classification of Medical Image is a superior technique for computer-aided diagnostic systems (CAD) towards skin cancer where it grows in fast manner than other diseases. Skin diseases are tolerable at sometimes and intolerable in many times. The typical approaches used to detect skin diseases are mainly based on characteristics, color, texture and their combinations. Primary objective of this research is to analyze the issues and challenges present in detecting the skin cancer, and to propose a novel classifier namely veritable support vector machine (VSVM) that can assist doctor to diagnosis the presence of skin cancer at early stage. VSVM includes the following steps: (i) train VSVM under supervised learning methodology to classify the medical image’s raw pixels using feature vectors that have high-ranking, (ii) derive a set of typical features that are based on information available in the history of medical images, (iii) construction of effective model that can combine different feature groups. VSVM focuses to classify the images that have the presence of skin cancer.

VSVM provide an efficient way to build a classification model that can use raw pixels of medical image to construct the best classifier. However, since medical images are high in resolution and data sets are low, in- depth learning models are cost-effective and have minimum models. Matlab R2019b used to evaluate the performance of VSVM. The two datasets used for the evaluation of VSVM against SVM are HIS2828 and ISIC2017. Results indicate that VSVM has enhanced performance than SVM in terms of specificity, sensitivity, precision, recall, f-measure and classification accuracy. Comprehensive analysis of results indicates that the proposed classifier VSVM has increased performance in detecting the skin cancer and assist doctors than the previous classifier SVM.

Keywords

Skin Cancer; Medical; Image; Classification; SVM; HIS2828; ISIC2017.

Introduction

Classification of medical images is one of the main image recognition problems which help to classify medical images into various categories so that the doctors can recognize or examine diseases further. The classification of medical images can generally involve two steps where the first step is to obtain strong picture functionality and the second step is to construct templates that can be used to classify the image data collection. Several algorithms are being published every year for medical image classification’s research domain. Nevertheless, medical images taken from different sources can vary in the area of focus, contrast and white balance. Additionally, medical images typically have internal structures that involve different textures and pixel densities. Through traditional features medical images are difficult to identify such classes in an efficient manner; Support Vector Machine has grown in recent years among the top most research domain of computer applications and computing technologies researchers.

Support Vector Machine (SVM) is a traditional classification algorithm that is most commonly used in medical fields to predict diseases. Different classification algorithms (like extreme vector machine (EVM), relevance vector machine (RVM), etc) are available for predicting the diseases, but they are: (i) adaptable to specific types of diseases, (ii) limited to datasets, (iii) poor classification accuracy. SVM has better performance than EVM, RVM, etc, but still there exist a need to enhance the performance (i.e., the classification accuracy) of SVM.

(2)

Optimizations(Lingaraj M and Prakash A 2019; Ramkumar and Vadivel 2019, 2020) are used in classification for enhancing the accuracy.

This study focuses on novel cum efficient way to detect the presence of skin cancer using multiscale features. The proposed classifier veritable support vector machine (VSVM) focuses on extracting features from corresponding medical-images, so that the deep features are automatically extracted, while the traditional algorithms like SVM do it manually.

Simultaneously high and low-level representations could be used and representations of single features are also avoided.

Literature Review

Spatial Adjacent Histogram (Liu et al. 2016) was proposed to enhance the confined binary patterns in medical images. Histogram strategy for spatial adjacent was also innovated to represent images in micro levels. While spatial relationships are ignored, the classification has faced the poor results. Discrete Bayesian Networks (Arias et al. 2016) was proposed as a pipeline for classifying the medical images. In this method, initially features are extracted and are combined for preprocessing with different methods. Deterministic classifiers were utilized to enhance its performance against the default image classification methodologies. Dimensionality Reduction (MingRu et al. 2019) was proposed to overcome minimum accuracy and low adaptability. Features are made to extract from the chrominance channel and it helped to extract multiple characteristic quantities. Support vector machine was utilized to perform the classification to which of those had poor results. Soft Set Classification (Lashari and Ibrahim 2013) proposed an algorithm for classification based on the ideas of (i) soft computing theory, (ii) advances in data mining algorithms, and (iii) specific soft computing applications. It provide technique for the classification of medical imaging that consists of 6 phases, which are: (i) acquisition, (ii) pre-processing, (iii) partitioning, (iv) classification, (v) analyzing and (vi) enhancement of performance.

Machine Learning Model (García-Floriano et al. 2019) suggested a method with the ensemble of a robust and efficient model of Machine Learning incorporating a digital image processing and mathematical morphology. A system of enhanced contrast and the implementation of simple morphological operations were used in the enclosed macular region. As the features of the rendered image, it uses invariant moments. The resulting vector for a drusen will be defined by support vector machine. Multiscale Representation Learning (Tang, Liu, and Liu 2017) proposed an imaging method to capture intrinsic scales for the classification task through sparse auto- encoder networks in medical images. Through sparse auto-encoders with different receptive field dimensions, the multiple function detectors are generated with the function maps through the convolution process. For medical images, this technique can represent better than single-scale structures of various sizes. Multi-Scale Non-Negative Classification (Zhang et al. 2017) proposed classification algorithm that was based on multi-scale and non-negative sparse coding of medical images. At first, medical images are made to decompose with multiple layers, so that specific visual information was derived from various layers of the scale. Secondly, non-negative sparse based coding model with fishing discriminatory analysis is done. The aim of the second step was to build scale layer to achieve a sparse that is discriminatory depiction of medical oriented images. Finally, the characteristics of multi-scale non-negative sparse are obtained by combining into a multiple-scale histogram for a medical image. Tensor-based Sparse Representations (Wang

(3)

et al. 2020) proposed a linear sparse coding extension for the extraction in Multi-phase Computed Tomography scan images of compact and efficient intermediate representations. It suggested the creation of three-layer volumes from the respective slices of multi-phase Computed Tomography scan images and the extraction from volumes of local spatiotemporal structures are portrayed as 3rd-order tensors.

Prototype System (Oliva et al. 2016) introduced as MIAS 3.0 to assist professionals in the research and analysis of colon tissue. Then machine-learning algorithm extracts the features for improving the classification. Texture descriptors Haralick, Amadasum, and Laws are provided.

Finally, medical images are then divided into regular or irregular. J48, the Closest Neighbor and Multilayer Perceptron, Back Propagation, Support Vector Machine and Naive Bayes are used for comparison. Synergic Deep Learning (Zhang et al. 2019) tackled the problem in Multiple Deep Convolutional-based Neural Networks (MDCNN) simultaneously and proposed a deep synergy- learning model that will make it possible to jointly learn from each other. The representations of images at each pair of MDCNN are concatenated as input of synergistic network with a completely connected structure, which assumes that the input pairs belong to the same class.

Deep Convolutional Neural Network (Yadav and Jadhav 2019), (Ashraf et al. 2020), Improved Support Vector Machine (Jiang et al. 2007), Explainable Deep Learning Model (Brunese et al.

2020), Optimized Convolutional Neural Network (An and Liu 2020), Augmented Deep Neural Networks (Pi et al. 2020), Machine learning Techniques (Seo et al. 2020) proposed to increase the classification accuracy of image classification.

Methods

SVM (i.e., Support Vector Machine) was proposed by Vapnik. It relies on empirical hypothesis of learning. They consider the classification scheme, which is an inductive framework for the learning from a restricted preparation of knowledge index.

Set with 𝑛 emphasis (𝑥1, 𝑦1), . . . , (𝑥𝑛, 𝑦) in which 𝑥𝑗 ∈ 𝑅𝑁 and 𝑦𝑗 ∈ {−1, 1}, 𝑗 = 1, . . . , 𝑛.

Assume that certain hyper-plane will separate the positive and negative specimens. This indicates that the positive and the negative specimens with the type are specifically capable of:

𝑑 𝑥 = 𝑤. 𝑥 + 𝑏 (1)

𝐷(𝑥𝑗) = 1, if 𝑦𝑗 = 1 for each 𝑥𝑗 represents the test preparation; 𝑑(𝑥𝑗) < −1, generally. This skill is also known as the choice capacity. A function 𝑥 can be transformed as:

𝑦 = 𝑠𝑖𝑔𝑛 𝑑 𝑥 (2)

Many possible hyper planes may be identified for a given preparation dataset to separate both classes effectively.(i.e., positive and negative). By that, the edges around the hyperplane are enclosed by SVM with the hope to find an optimal arrangement. The solution is an ideal isolating line for a case in two-dimensional space.

The hyper planes are the supporting vectors that focus towards classification:

𝑦𝑗 𝑤 ⋅ 𝑥𝑗 + 𝑏 = 1 (3)

As an example, if {𝑥𝑖, 𝑦𝑖} is not found on the hyperplane of support vectors, then mathematically it can be said as:

(4)

𝑦𝑗 𝑤 ⋅ 𝑥𝑗 = (1 − 𝑏) (4) Scientifically, the edge 𝑀 is finally reached between two vectors of support.

𝑀 = 2

𝑤 (5)

Where, 𝑤 is the standard of 𝑤.

Subsequently, the enhancement of the edge 𝑀 is equal to the limitation of the entire vector with the drawback (i.e., error). Its helps VSVM to train under supervised learning methodology for classification. This constraint can be seen as:

𝑦𝑗 𝑤 ⋅ 𝑥𝑗 + 𝑏 ≥ 1 (6)

For the situation, the first specimens couldn't be isolated by any hyper plane, SVM will change the first examples into a higher dimensional space by utilizing a nonlinear mapping. Here, 𝛷(𝑥) indicates the mapping from RN to a higher dimensional space𝑍. It helps to derive new typical features. A hyper plane should be found in the higher dimensional space with the most extreme edge as:

𝑤 ⋅ 𝑧 + 𝑏 = 0 (7)

For. point(𝑧𝑗, 𝑦𝑗), with the end goal ,where 𝑧𝑗 = 𝑡𝑜𝑡𝑎𝑙(𝑥𝑗):

𝑦𝑗 𝑤 ⋅ 𝑧 + 𝑏 ≥ = 𝑗 1, 𝐾, 𝑛 (8)

When the dataset is not directly divisible, 𝑛 non-negative factors are utilized and it can be addressed to delicate end of the results, with the ultimate aim that the requirement for each specimen in Eq. (8) is updated as follows:

𝑦𝑗(𝑤 ⋅ 𝑧𝑗 + 𝑏) ≥ 1 − 𝜉 = 𝑗 1, 𝐾, 𝑛 (9) The solution to the problem is the ideal hyper plane problem:

𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 1

2 𝑤 . 𝑤 + 𝐶 ^𝑘 ξ

𝑗 =1

(10) With subject to

𝑦𝑗 𝑤 ⋅ 𝑧𝑗 + 𝑏 ≥ 1 − 𝜉 = 𝑗 1, 𝐾, 𝑛 (11) Where, the primary term in Eq.(11), is the calculated edge between support vectors, with which the different errors are calculated. ′𝐶′ is an ongoing parameter that harmonizes the extreme edge with mix-up base grouping. It leads a way to construct an effective model that can combine the different features. At this point, the grouping result 𝑦 is given as: for a test point 𝑥, which is mapped to 𝑧 in the element space is given as:

𝑦 = 𝑠𝑖𝑔𝑛 𝑤 ⋅ 𝑧 𝑏 (12)

(5)

Dataset

This research has developed the Matlab coding network for the extracting high-level features from the toolbox “MatConvnet”, which makes the neural network implementation and traditional features extraction based on the moment of color and texture. Our studies on two benchmark medical imaging datasets are designed to check the efficiency of our process, which are (i) HIS2828 dataset, (ii) ISIC2017 dataset. All our evaluations were performed on desktop computing with configuration i3-4130 3.4 GHz CPU, 8GB Main Memory.

HIS2828 dataset

The HIS2828 dataset consists of four groups of fundamental images of tissue, which represent various types of tissue. It is made available at www.challenge.kitware.com. Every RGB image has the resolution of 720*480. This dataset comprises 2828 pictures that are classified as follows:

1026 medical images of nervous tissue, 484 medical images of Connective, 804 medical images of

Epithelial tissue and 514 medical images of Muscular tissue, which are used as the labels with 1, 2, 3, and 4. Table 1 indicates the HIS2828 dataset structure.

Table 1. HIS2828 Dataset Structure

𝐈𝐦𝐚𝐠𝐞 𝐜𝐚𝐭𝐞𝐠𝐨𝐫𝐲 𝐍𝐮𝐦𝐛𝐞𝐫 𝐨𝐟 𝐢𝐦𝐚𝐠𝐞𝐬 𝐋𝐚𝐛𝐞𝐥

𝑁𝑒𝑟𝑣𝑜𝑢𝑠 𝑡𝑖𝑠𝑠𝑢𝑒 1026 1

𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑣𝑒 𝑡𝑖𝑠𝑠𝑢𝑒 484 2

𝐸𝑝𝑖𝑡𝑕𝑒𝑙𝑖𝑎𝑙 𝑡𝑖𝑠𝑠𝑢𝑒 804 3

𝑀𝑢𝑠𝑐𝑢𝑙𝑎𝑟 𝑡𝑖𝑠𝑠𝑢𝑒 514 4

ISIC2017 dataset

The International Skin Imaging Collaboration (ISIC) offers ISIC2017 dataset of skin lesions. It is available at www.informed.unal.edu.co. This includes 2000 images, of which 374 are malignant tumors of the skin known as "melanoma" and 1626 are benign tumors of the skin keratization.

This then distinguishes between (a) Melanoma, (b) Nevus of Seborrhean Keratosis, as a binary image classification function. Every image has a different resolution in this dataset and must be handled. The ISIC2017 dataset information is shown in Table 2.

Table 2. ISIC2017 Dataset Structure

𝐈𝐦𝐚𝐠𝐞 𝐜𝐚𝐭𝐞𝐠𝐨𝐫𝐲 𝐍𝐮𝐦𝐛𝐞𝐫 𝐨𝐟 𝐢𝐦𝐚𝐠𝐞𝐬 𝐋𝐚𝐛𝐞𝐥

𝑀𝑒𝑙𝑎𝑛𝑜𝑚𝑎 374 1

𝑁𝑒𝑣𝑢𝑠 𝑜𝑓 𝑠𝑒𝑏𝑜𝑟𝑟𝑕𝑒𝑖𝑐 𝑘𝑒𝑟𝑎𝑡𝑜𝑠𝑖𝑠 1626 2

PERFORMANCE METRICS

The default output metrics used in data mining for validation are sensitivity, specificity, precision, recall, accuracy, and f-measure. In the medical sector, the positive level of true and false values are mostly important. The efficiency of classifiers is also to be tested by accuracy, recall, precision, and f-measure. Confusion matrix is viewed as a specific type of table for showing the output of algorithms. While making a consideration of dual-class problems

(6)

(i.e. Class 1 and Class 2) then it is necessary to count the true positives (TP), false positive (FP), true negatives (TN) and false negative (FN). The following are the description of the measures:

 𝑇𝑃 - Class 1 samples correctly identified by statistical method of calculation.

 𝑇𝑁 - Class 2 samples based on statistical measurement methods are specifically graded.

 𝐹𝑃 - Class 1 was mistakenly listed as class 2 samples based on statistical calculation process.

 𝐹𝑁 - Class 2 was mistakenly listed as class 1 samples based on the statistical calculation process.

This research work performs the calculation of results using benchmarking performance metrics, with the aforementioned steps, and is defined as:

𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑃

𝑇𝑃 + 𝐹𝑁 (13)

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑇𝑁

𝑇𝑁 + 𝐹𝑃 (14)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃

𝑇𝑃 + 𝐹𝑃 (15)

𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃

𝑇𝑃 + 𝐹𝑁 (16)

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁

𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 (17)

𝐹 − 𝑀𝑒𝑎𝑠𝑢𝑟𝑒 = 2 ×(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙)

(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙) (18)

RESULTS AND DISCUSSION

Proposed classifier VSVM is evaluated against SVM algorithm on two datasets namely HIS2828 dataset and ISIC2017. The experimental result’s performance are discussed in the below sections.

From Figure.1 to Figure. 10, x-axis is plotted with datasets. From Figure. 1 to Figure. 4, y-axis is plotted with count of medical images, and from Figure. 5 to Figure. 10 y-axis is plotted with results in percentages.

(7)

True Positive and True Negative Analysis

Figure 1 and Figure 2 contrasts VSVM's success against SVM with TP and TN. The output of VSVM is obviously better than that of SVM. VSVM classifies according to selected features. But only in the sequence of medical images, SVM identifies the medical images which lead to poor result. Numerical values of Figure 1 shown in Table 3 and Figure 2 shown in Table 4.

Figure 1. True Positive Analysis

Figure 2. True Negative Analysis

(8)

Table 3. Numerical Values of True Positive Analysis Result

Classifier Name HIS2828 dataset ISIC2017 dataset

SVM 810 613

VSVM 1303 968

Table 4. Numerical Values of True Negative Analysis Result

SVM 753 526

VSVM 1019 794

False Positive and False Negative Analysis

Figure 3 and Figure 4 indicate that VSVM performs better SVM. Obviously, SVM delivers more FP and FN in worse results. VSVM also provides FP and FN, which is very small compared to SVM. VSVM has decreased FP and FN because of consideration of the hidden state values.

Numerical values of Figure 3 shown in Table 5 and Figure 4 shown in Table 6.

Table 5. Numerical Values of False Positive Analysis Result

SVM 629 516

VSVM 208 101

Table 6. Numerical Values of False Negative Analysis Result

SVM 636 345

VSVM 298 137

Figure 3. False Positive Analysis

(9)

Figure 4. False Negative Analysis Sensitivity and Specificity Analysis

The sensitivity and specificity results of VSVM which is compared with SVM are contrasted in Figure 5 and Figure 6. VSVM obviously has improved results and showed its enhanced performance. In VSVM, the medical images of two datasets are considered in Eqn. (10) and Eqn.

(11), in order to minimize the falsify classification. Numerical values of Figure 5 shown in Table 7 and Figure 6 shown in Table 8.

Figure 5. Sensitivity Analysis

(10)

Figure 6. Specificity Analysis

Table 7. Numerical Values of Sensitivity Analysis Result

SVM 56.02 63.99

VSVM 81.39 87.60

Table 8. Numerical Values of Specificity Analysis Result

SVM 54.49 50.48

VSVM 83.05 88.72

Precision and Recall Analysis

The precise results of VSVM and SVM are compared in Figure 7 and Figure 8. VSVM helps to provide improved results than SVM by using the results obtained from newly derived cum selected features. SVM received poor accuracy and reminder results, as it did not prioritize the expected results and evaluation of the classification. Numerical values of Figure 7 shown in Table 9 and Figure 8 shown in Table 10.

Table 9. Numerical Values of Precision Analysis Result

SVM 56.29 54.30

VSVM 86.23 90.55

(11)

Table 10. Numerical Values of Recall Analysis Result

SVM 56.02 63.99

VSVM 81.39 87.60

Figure 7. Precision Analysis

Figure 8. Recall Analysis

(12)

Accuracy and F-Measure Analysis

Figure 9 and Figure 10 shows that VSVM has out performing results than SVM, which clearly makes a indication that VSVM has better performance towards predicting presence of skin cancer. VSVM harmonizes the extreme edges of medical images in an efficient view by effectively considering the features. However, SVM lacks due to the fact that edges are not taken into account. Numerical values of Figure 9 shown in Table 11 and Figure 10 shown in Table 12.

Figure 9. Classification Accuracy Analysis

Figure 10. F-Measure Analysis

(13)

Table 11. Numerical Values of Classification Accuracy Analysis Result Classifier Name HIS2828 dataset ISIC2017 dataset

SVM 55.27 56.95

VSVM 82.11 88.10

Table 12. Numerical Values of F-Measure Analysis Result

SVM 56.15 58.74

VSVM 83.74 89.05

CONCLUSION

A novel algorithm namely VSVM is proposed for predicting the presence of Melanoma Skin Cancer. It blends high-quality features with conventional images. It involves training, deriving of typical features and construction of an effective model that incorporates derived and available features. The results of experiments show that VSVM has achieved accuracy of 82.11% and 88.10% on HIS2828 and ISIC2017 medical image datasets, and it outperforms SVM. VSVM is limited to prediction of skin cancer and can give best performance in HIS2828 and ISIC2017 medical image datasets. Future scope of this research can be focused t utilize the optimization techniques to attain better classification accuracy even more.

References

1. An, Feng Ping, and Jun E. Liu. 2020. “Medical Image Segmentation Algorithm Based on Optimized Convolutional Neural Network-Adaptive Dropout Depth Calculation.” Complexity 2020. doi: 10.1155/2020/1645479.

2. Arias, Jacinto, Jesus Martínez-Gómez, Jose A. Gámez, Alba G. Seco de Herrera, and Henning Müller. 2016. “Medical Image Modality Classification Using Discrete Bayesian Networks.”

Computer Vision and Image Understanding 151:61–71. doi: 10.1016/j.cviu.2016.04.002.

3. Ashraf, Rehan, Muhammad Asif Habib, Muhammad Akram, Muhammad Ahsan Latif, Muhammad Sheraz Arshad Malik, Muhammad Awais, Saadat Hanif Dar, Toqeer Mahmood, Muhammad Yasir, and Zahoor Abbas. 2020. “Deep Convolution Neural Network for Big Data Medical Image Classification.” IEEE Access 8:105659–70. doi: 10.1109/ACCESS.2020.2998808.

4. Brunese, Luca, Fabio Martinelli, Francesco Mercaldo, and Antonella Santone. 2020. “Deep Learning for Heart Disease Detection through Cardiac Sounds.” Procedia Computer Science 176:2202–11. doi: https://doi.org/10.1016/j.procs.2020.09.257.

5. García-Floriano, Andrés, Ángel Ferreira-Santiago, Oscar Camacho-Nieto, and Cornelio Yáñez- Márquez. 2019. “A Machine Learning Approach to Medical Image Classification: Detecting Age- Related Macular Degeneration in Fundus Images.” Computers and Electrical Engineering 75:218–29. doi: 10.1016/j.compeleceng.2017.11.008.

6. Jiang, Yun, Zhanhuai Li, Longbo Zhang, and Peng Sun. 2007. “An Improved SVM Classifier for Medical Image Classification.” Pp. 764–73 in Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 4585 LNAI. Springer Verlag.

(14)

7. Lashari, Saima Anwar, and Rosziati Ibrahim. 2013. “A Framework for Medical Images Classification Using Soft Set.” Procedia Technology 11:548–56. doi:

10.1016/j.protcy.2013.12.227.

8. Lingaraj M, and Prakash A. 2019. “Power Aware Routing Protocol (PARP) to Reduce Energy Consumption in Wireless Sensor Networks.” International Journal of Recent Technology and Engineering 7(5):380–85.

9. Liu, Dong, Shengsheng Wang, Dezhi Huang, Gang Deng, Fantao Zeng, and Huiling Chen. 2016.

“Medical Image Classification Using Spatial Adjacent Histogram Based on Adaptive Local Binary Patterns.” Computers in Biology and Medicine 72:185–200. doi:

10.1016/j.compbiomed.2016.03.010.

10. MingRu, Kong, Qin Zheng, Song Kui Yan, and N. Arunkumar. 2019. “Medical Image Classification Algorithm Based on Principal Component Feature Dimensionality Reduction.”

Future Generation Computer Systems 98:627–34. doi: 10.1016/j.future.2018.11.056.

11. Oliva, Jefferson Tales, Huei Diana Lee, Newton Spolaôr, Claudio Saddy Rodrigues Coy, and Feng Chung Wu. 2016. “Prototype System for Feature Extraction, Classification and Study of Medical Images.” Expert Systems with Applications 63:267–83. doi: 10.1016/j.eswa.2016.07.008.

12. Pi, Yong, Zhen Zhao, Yongzhao Xiang, Yuhao Li, Huawei Cai, and Zhang Yi. 2020. “Automated Diagnosis of Bone Metastasis Based on Multi-View Bone Scans Using Attention-Augmented Deep Neural Networks.” Medical Image Analysis 65:101784. doi: 10.1016/j.media.2020.101784.

13. Ramkumar, J., and R. Vadivel. 2019. “Performance Modeling of Bio-Inspired Routing Protocols in Cognitive Radio Ad Hoc Network to Reduce End-to-End Delay.” International Journal of Intelligent Engineering and Systems 12(1):221–31. doi: 10.22266/ijies2019.0228.22.

14. Ramkumar, J., and R. Vadivel. 2020. “Bee Inspired Secured Protocol for Routing in Cognitive Radio Ad Hoc Networks.” Indian Journal of Science And Technology 13(30):3059–69. doi:

10.17485/IJST/v13i30.1152.

15. Seo, Hyunseok, Masoud Badiei Khuzani, Varun Vasudevan, Charles Huang, Hongyi Ren, Ruoxiu Xiao, Xiao Jia, and Lei Xing. 2020. “Machine Learning Techniques for Biomedical Image Segmentation: An Overview of Technical Aspects and Introduction to State‐of‐art Applications.”

Medical Physics 47(5):e148–67. doi: 10.1002/mp.13649.

16. Tang, Qiling, Yangyang Liu, and Haihua Liu. 2017. “Medical Image Classification via Multiscale Representation Learning.” Artificial Intelligence in Medicine 79:71–78. doi:

10.1016/j.artmed.2017.06.009.

17. Wang, Jian, Jing Li, Xian Hua Han, Lanfen Lin, Hongjie Hu, Yingying Xu, Qingqing Chen, Yutaro Iwamoto, and Yen Wei Chen. 2020. “Tensor-Based Sparse Representations of Multi- Phase Medical Images for Classification of Focal Liver Lesions.” Pattern Recognition Letters 130:207–15. doi: 10.1016/j.patrec.2019.01.001.

18. Yadav, Samir S., and Shivajirao M. Jadhav. 2019. “Deep Convolutional Neural Network Based Medical Image Classification for Disease Diagnosis.” Journal of Big Data 6(1):113. doi:

10.1186/s40537-019-0276-2.

19. Zhang, Jianpeng, Yutong Xie, Qi Wu, and Yong Xia. 2019. “Medical Image Classification Using Synergic Deep Learning.” Medical Image Analysis 54:10–19. doi: 10.1016/j.media.2019.02.010.

20. Zhang, Ruijie, Jian Shen, Fushan Wei, Xiong Li, and Arun Kumar Sangaiah. 2017. “Medical Image Classification Based on Multi-Scale Non-Negative Sparse Coding.” Artificial Intelligence in Medicine 83:44–51. doi: 10.1016/j.artmed.2017.05.006.