Technology-assisted platforms provide reliable solutions in almost every field these days. One such important application in the medical field is the skin cancer classification in preliminary stages that need sensitive and precise data analysis. For the proposed study the Kaggle skin cancer dataset is utilized. The proposed study consists of two main phases. In the first phase, the images are preprocessed to remove the clutters thus producing a refined version of training images. To achieve that, a sharpening filter is applied followed by a hair removal algorithm. Different image quality measurement metrics including Peak Signal to Noise (PSNR), Mean Square Error (MSE), Maximum Absolute Squared Deviation (MXERR) and Energy Ratio/ Ratio of Squared Norms (L2RAT) are used to compare the overall image quality before and after applying preprocessing operations. The results from the aforementioned image quality metrics prove that image quality is not compromised however it is upgraded by applying the preprocessing operations. The second phase of the proposed research work incorporates deep learning methodologies that play an imperative role in accurate, precise and robust classification of the lesion mole. This has been reflected by using two state of the art deep learning models: Inception-v3 and MobileNet. The experimental results demonstrate notable improvement in train and validation accuracy by using the refined version of images of both the networks, however, the Inception-v3 network was able to achieve better validation accuracy thus it was finally selected to evaluate it on test data. The final test accuracy using state of art Inception-v3 network was 86%.
Cancer nowadays is one of the greatest growing groups of diseases throughout the
world, among which skin cancer is most common of them. According to stats and
figures, the annual rate of skin cancer is increasing at an alarming rate each year [
Dermoscopic diagnosis refers to a non-invasive skin imaging method, which has
become a core tool in the diagnosis of melanoma and other pigmented skin lesions.
However, performing dermoscopy using conventional methods may lower down the
diagnostic accuracy which can lead to more chances of errors. These errors are
generally caused by the complexity of lesion structures and the subjectivity of visual
interpretations [
Computer-Aided Diagnosis (CAD) system is a type of digitized platform based on advanced computer vision, deep learning, and pattern recognition techniques for skin cancer classification. For the proposed study we have designed a CAD system for skin cancer classification by utilizing advanced deep neural networks. The system consists of the following steps: Firstly, a preprocessing of the digital images which includes removing clutter such as hair from that part of the skin where the pigmented mole is present and applying a sharpening filter to make that area more clear and visible thus minimizing the chances of error. The next essential step includes the feature extraction and classification process to extract the results for the cases under consideration by utilizing deep learning techniques. Section 2 presents the background and related study and highlights the medical aspects regarding skin cancer. Section 3 describes the detailed methodology of the proposed system whereas Section 4 presents the implementation and experimental results of the proposed study. Section 5 draws the overall conclusion of the paper. 2
The human skin is the largest organ of the overall human body. It covers all other organs
of the body. It guards the entire body from microbes, bacterium, ultraviolet radiation,
helps to regulate body temperature and permits the sensations of touch, heat, and cold
[
Mole or nevus on human skin can be described as a dark, erected spot comprised of
skin cells that are grown in a group rather than individually. These cells are generally
known as melanocytes which are responsible for producing melanin, the pigment color
in our skin. The main reason behind mole development on human skin is predominantly
because of direct sun exposure and any kind of extreme injury. The fair skin population
has a greater ratio of skin moles due to the lower quantity of melanin (natural pigments)
in their skins [
The study proposed by Simon Kalouche utilizes [
In the proposed study, an efficient skin cancer diagnosis system has been implemented for precise classification between malignant melanoma and benign cases. The complete algorithm consists of several steps starting from the input phase of applying image preprocessing ranging to the analysis of the case under consideration in the form of the probability of lesion Malignancy. Fig. 1 shows the complete workflow of the proposed algorithm. 3.1
For the proposed study, the Kaggle skin cancer dataset [
It is noteworthy that image quality is refined after applying the image preprocessing operations. This is shown in Section 4 of the paper where the results from four different image quality metrics Peak Signal to Noise Ratio (PSNR), Mean Squared Error (MSE), Maximum Absolute Squared Deviation (MXERR) (MXERR) and Ratio of Squared Norms (L2RAT) on both ground truth images, and preprocessed images are presented. 4 Fig 1. Workflow diagram of the proposed method a b c
In the next step, the processed images are fed to state-of-the-art deep neural networks
in order to perform the feature extraction and classification steps. In this work, the
Inception-v3 and MobileNet deep learning architectures are utilized. These architectures
play a vital role by extracting feature values from raw pixel images. The Inception-v3
has state of the art performance in the classification task. It is made up of 48 layers
stacked on top of each other [
The overall algorithm was implemented using Matlab R2018a for computing image
quality metrics and TensorFlow [
Image 99 99 99 99 99 99 99 99 19.3975 847.0221 0.9747 224 x 224 The experimental results show clearly that image quality is not comprised however it is upgraded which is evident from high PSNR values and other metrics after applying image preprocessing operations especially the hair removal filter. The image quality metrics were carried out on more than fifty images and the same observations were 23.2953 measured. The second part of the experiments includes the training of the classifiers using the two state of the art deep learning networks i.e. Inception-v3 and MobileNet. For Inception-V3 the data was resized to 299 x 299 since the network has an image input size of 299 by 299. The classifiers were trained on both sets of images i.e. original (ground truth) images and images after applying the preprocessing operations to them. Both the networks were trained using the same hyperparameters. The learning rate was set to 0.005 with a batch size of 32 and total iterations were set to 5000. The training data was split in the ratio of 75% and 25% for training and validations images respectively. Fig. 3 and Fig. 4 display the training and validation accuracy graphs along with the error rate (cross-entropy) graph of MobileNet and Inception-v3 networks. The accuracy graphs in Fig. 3 show that training and validation accuracy before applying the image preprocessing was 86% and 79.8% and it was increased to 89% and 85.9% by using a refined version of images obtained after applying the image preprocessing operations. Similarly, the validation error rate was also decreased from 61% to 32% by using the refined version of images. Training loss Validation loss
Iterations The accuracy graphs in Fig. 4 show that training and validation accuracy before applying the image preprocessing was 88.3% and 84.2%. By using the refined version of images training accuracy tends to remain the same thought the validation accuracy was increased to 86.1%. Similarly, the validation error rate was also decreased from 36% to 32.3% by using the refined version of images.
Overall, in both networks, significant improvements were measured after using the refined version of images. The experimental results show that the Inception-v3 network was able to achieve better validation accuracy using a refined version of training data i.e. 86.1 % thus we will be using the Inception-v3 network for evaluating it on the test data. For evaluating the classifiers on the test data, we have picked numerous cases from the test set from both classes, benign and malignant melanoma among which visually complex and challenging test cases were selected for the proposed research work. It is pertinent to mention that the network was tested using the original images (unrefined version) to test the overall effectiveness of the classifier. Fig. 5 shows some of the results predicted correctly on test images. Table 2 illustrates the complete results on visually complex test cases selected for the proposed study which will be further used to evaluate overall testing accuracy, sensitivity (true positive rate), specificity (true negative rate) and precision metrics. The rows highlighted with red color indicates the misclassified test cases when compared with ground truth results. b
Predicted results using Inception-v3 Network trained
on original images Benign – Low risk – 97.8%
Malignant–High risk – 91.8%
Benign – Low risk – 98.4%
Benign – Low risk – 98.4%
Malignant – High risk – 96.4%
Malignant – High risk – 98.7%
Malignant – High risk – 98.8%
Malignant – High risk – 99.4%
Benign – Low risk – 71.2%
Malignant – High risk – 85.9%
Malignant – High risk – 99.5%
Malignant – High risk – 98.5%
Malignant – High risk – 70.9%
Benign– Low risk – 74.8%
Low risk
Malignant – High risk – 97.5 %
The overall performance of the Inception-v3 network on test data has been evaluated
using five quantitative measures: Accuracy, sensitivity, specificity, precision and F1
score [
Table 3 illustrates the results of all the four quantitative measures: Accuracy, sensitivity, specificity, and precision of the Inception-v3 network before and after using the image preprocessing operations on test data. It can be observed that testing accuracy is increased to 86% by training the classifier using the refined version of images. ( ( ) = The main purpose of the proposed study was to improve the overall accuracy level of two state of art deep learning networks which include Inception-v3 and MobileNet by using the refined version of skin cancer images obtained after applying image preprocessing operations. The experiments were conducted using the Kaggle Skin Cancer Dataset by applying initial sharpening filter and hair removal algorithms. Initially, we applied these algorithms as image pre-processing mechanisms to remove the clutters thus producing the refined version of images. Different image quality metrics including Peak Signal to Noise (PSNR), Mean Square Error (MSE), Maximum Absolute Squared Deviation (MXERR) and Energy Ratio/ Ratio of Squared Norms (L2RAT) were used to compare the image quality before and after applying the pre-processing techniques. These metrics prove that image quality was upgraded after applying sharpening filter and hair removal algorithms. In the next phase of experimental results, we have seen substantial improvement in training, validation and test accuracy after applying image pre-processing operation. Thus, we have achieved an overall test accuracy of 86% using state of the art Inception-v3 network by fine-tuning the last layer of the network with a refined version of kaggle skin cancer training dataset.
For future work, more image pre-processing techniques like neural networks based
super image algorithms and other such techniques could be used to improve the image
quality to a better extent. Moreover, other state of the art deep neural networks such as
ResNet-101 [