Robust deep learning framework for the detection of melanoma in images

Abstract

Melanoma, a type of skin cancer, occurs when melanocytes become cancerous and is a common cause of death in adults. The presence of melanoma can be conclusively proved through biopsies, but these lap reports often take time. Early detection of melanoma could improve mortality rates and reduce costs. AI-based assistive tools can aid early detection. Most studies focus on detection either in dermoscopic images or in non-dermoscopic images, not both. In this paper, we propose a novel generalised framework which can detect melanoma in both dermoscopic and non-dermoscopic images. The framework includes a preprocessing pipeline, data augmentation and resolving class imbalances, followed by a VGG-16 model. The model gives a sensitivity (for melanoma cases) of 87% on non-dermoscopic images and 91 % on dermoscopic images.