Diagnosis of skin cancer using novel computer vision and deep learning techniques

Recent years have noticed an increase in the total number of skin cancer cases and it is projected to grow exponentially, however mortality rate of malignant melanoma can be decreased if it is diagnosed and treated in its early stage. Notwithstanding the fact that visual similarity between benign and malignant lesions makes the task of diagnosis difficult even for an expert dermatologist, thereby increasing the chances of false prediction. This dissertation proposes two novel methods of computer-aided diagnosis for the classification of malignant lesion. The first method pre-processes the acquired image by the Dull razor method (for digital hair removal) and histogram equalisation. Henceforth the image is segmented by the proposed method using LR-fuzzy logic and it achieves an accuracy, sensitivity and specificity of 96.50%, 97.50% and 96.25% for the PH2 dataset; 96.16%, 91.88% and 98.26% for the ISIC 2017 dataset; 95.91%, 91.62% and 97.37% for ISIC 2018 dataset respectively. Furthermore, the image is classified by the modified You Only Look Once (YOLO v3) classifier and it yields an accuracy, sensitivity and specificity of 98.16%, 95.43%, and 99.50% respectively. The second method enhances the images by removing digital artefacts and histogram equalisation. Thereafter, triangular neutrosophic number (TNN) is used for segmentation of lesion, which achieves an accuracy, sensitivity, and specificity of 99.00%, 97.50%, 99.38% for PH2; 98.83%, 98.48%, 99.01% for ISIC 2017; 98.56%, 98.50%, 98.58% for ISIC 2018; and 97.86%, 97.56%, 97.97% for ISIC 2019 dataset respectively. Furthermore, data augmentation is performed by the addition of artefacts and noise to the training dataset and rotating the images at an angle of 650, 1350, and 2150 such that the training dataset is increased to 92838 from 30946 images. Additionally, a novel classifier based on inception and residual module is trained over augmented dataset and it is able to achieve an accuracy, sensitivity and specificity of 99.50%, 100%, 99.38% for PH2; 99.33%, 98.48%, 99.75% for ISIC 2017; 98.56%, 97.61%, 98.88% for ISIC 2018 and 98.04%, 96.67%, 98.52% for ISIC 2019 dataset respectively. Later in our dissertation, the proposed methods are deployed into real-time mobile applications, therefore enabling the users to diagnose the suspected lesion with ease and accuracy.