Intelligent prognostics model for disease prediction and classification from dermoscopy images using a convolutional neural network

Abstract

Skin cancer has been reported to be one of the most leading forms of cancer diseases, especially amongst Caucasian descendant and light-skinned people. Basal Cell Carcinoma (BCC) is the malignant types of skin cancer and their classification in earlier stage is biggest issue. While remediable with primary classification is useful, only extremely trained specialists are capable of accurately recognizing the disease from skin lesions dermoscopy images. As expertise is in limited contribute, an automated systems capable of classifying disease could save human lives, and also help to reduce unnecessary biopsies, and reduce extra costs. On the way to achieve this goal, we proposed a disease classification system that conglomerates current developments in deep learning with Convolutional Neural Network (CNN) structure, creating hybrid algorithm of segmentation with Particle Swarm Optimization (PSO) that are capable of segmenting accurate skin lesions region from dermoscopy images, along with analyzing the detected area and surrounding tissue for BCC. Using k-means segmentation technique, the foreground and background component is separated into two regions. To improve the segmentation results, PSO is used with the novel concept of hair removal from lesion region. The proposed system is evaluated using the largest publicly accessible standard skin lesions dataset of dermoscopic images, containing 600 training and 400 testing images. When the evaluation parameters of proposed work is associated with not many other state-of-art methods, the proposed technique attains the best performance of 98.5%in terms of area under the curve (AUC) in distinguishing BCC from benign lesions utilizing only the extracted vascular Speed Up Robust Features (SURF). These concerns have propelled the need to provide automated systems for medical diagnosis of skin cancer diseases within a strict time window towards reducing the unnecessary biopsy, increasing the speed of diagnosis and providing reproducibility of diagnostic results.