Vision-Based Classification of Skin Cancer using Deep Learning

This study proposes the use of deep learning algorithms to detect the presence of skin cancer, specifically melanoma, from images of skin lesions taken by a standard camera. Skin cancer is the most prevalent form of cancer in the US where 3.3 million people get treated each year. The 5-year survival rate of melanoma is 98% when detected and treated early yet over 10,000 people are lost each year due mostly to late-stage diagnoses [2]. Thus, there is a need to make melanoma screening and diagnoses methods cheaper, quicker, simpler, and more accessible. This study aims to produce an inexpensive and fast computer-vision based machine learning tool that can be used by doctors and patients to track and classify suspicious skin lesions as benign or malignant with adequate accuracy using only a cell phone camera. The data set was trained on 3 separate learning models with increasingly improved classification accuracy. The 3 models included logistic regression, a deep neural network, and a fine-tuned, pre-trained, VGG-16 Convolutional Neural Network (CNN) [7]. Preliminary results show the developed algorithm’s ability to segment moles from images with 70% accuracy and classify skin lesions as melanoma with 78% balanced accuracy using a fine-tuned VGG-16 CNN.