Image-based Classification of Skin Cancer using Convolution Neural Network

Abstract

Skin cancer is a category or collection of cancer affecting the tissues and layers of skin. Skin cancer is classified into several types depending on the type of cell it affects. These types include melanoma, melanocytic nevus, basal cell carcinoma, benign keratosis, actinic keratosis, dermatofibroma, vascular lesion, and squamous cell carcinoma. Melanoma which affects the melanocytes and is considered to be the most fatal and deadly cancer, is growing at an alarming rate, especially in the western hemisphere and the Pacific region. The proposed system contained a web-based application where the image of an affected skin area can be uploaded and the likelihood of skin cancer is displayed. This system has used a convoluted neural network (CNN) based binary and multi-classification model making efficient use of image processing, computer vision, OpenCV, and Python to classify dermatoscopic lesion images into cancerous and non-cancerous along with their types. The implemented binary classifier achieves an accuracy of 92%. Further, the multi-class classification model is implemented based on CNN to classify dermatoscopic cancerous lesion images into nine types which achieved an accuracy of 97%. Among nine classes one of the classes is non-cancerous. The models aim to provide a means of diagnostic tool that will help in the preliminary diagnosis of skin lesions. Early detection and diagnosis are appropriate measures to combat the spread and lethality of skin cancer.