Melanoma Diagnosis from Dermoscopy Images Using Artificial Neural Network

Abstract

Melanoma is the deadliest and unpredictable type of skin cancer. Fortunately, if it is diagnosed and treated at its early stage, the survival rate is very high. To avoid invasive skin biopsy, melanoma diagnosis from dermoscopy images has been introduced for last few decades. But it is very challenging due to low interclass variance between melanoma and non-melanoma images, and high intraclass variance in melanoma images. This paper presents a new approach for diagnosing melanoma skin cancer from dermoscopy images based on fundamental ABCD (Asymmetry, Border, Color, and Diameter) rule which are associated with shape, size and color properties of the images. Two new features related to area and perimeter of the skin lesion are proposed in this paper along with the other existing features which are distinguishing between melanoma and benign images. Dull razor algorithm is applied for black hair removal from the input images and Chan-Vese method is employed for segmentation. The extracted features are applied to an Artificial Neural Network (ANN) model for training and finally detecting melanoma images from the input images. This proposed approach achieves overall accuracy of 98%. This promising result would be able to assist dermatologist for making decision clinically.