Diagnoses of Melanoma Lesion Using YOLOv3

The most modifiable risk factor for skin cancer is ultraviolet radiation (UVR) exposure. Melanoma or malignant melanoma is the rarest but at the same time deadliest form of skin cancer. While prevention of melanoma is possible to some extent by educating masses to involve in safe sun practices as avoiding sun exposure during peak radiation hours, using protective clothing, applying sunscreen and distancing oneself from artificial sources of UV light, early detection and accurate treatment of the disease may curtail the fatality of the deadly disease. If statistics are to be believed, the lifetime risk of developing melanoma in the year 1935 was 1 in 1500 as compared to 1 in 50 in 2010, indicating its dramatic increase in the last century. While effective and timely treatment of melanoma has been a subject of prime importance for researchers and the medical fraternity alike, several invasive and non-invasive techniques have come to the fore from time to time for diagnosis of melanoma. Analysis of the several methods developed during the years suggests that easier access to skin examinations increase the chances of accurate and well-timed detection of melanoma and computer-aided diagnosis (CAD) has played a major role in fulfilling the same. This work proposes a novel CAD approach which includes preprocessing of the dermoscopic images by Dull Razor algorithm followed by classification by deep learning-based algorithm ‘You Only Look Once’ (YOLO) and finally segmentation of the identified image by a self-designed algorithm. The experiments have been conducted on three publicly available datasets—PH2, ISBI 2017 and ISIC 2016. The combination of the total methodology offers a Jac score of 86.12% and Dic of 92.55% which is way superior to results of contemporary works in the area.