Computer aided diagnosis of melanoma using Computer Vision and Machine Learning

This paper presents a computer-aided analysis of pigmented skin lesions following the ABCD guidelines used by Dermatologists. There are 3 steps to this process: image segmentation, feature analysis and machine learning. The most important and foundational step is image segmentation. In this paper we present 3 lesion segmentation algorithms: Radial Search, Region Growing Search using Flood Fill technique and Statistical Region Growing algorithm. We also present detailed description of the algorithms used to extract asymmetry, border irregularity, color variegation and diametric features. The aforementioned segmentation and feature extraction algorithms were tested on 350 test cases consisting of malignant melanocytic lesions, atypical nevus, superficial spreading, seborrhoeic keratosis, non-melanocytic lesions and benign lesions. An Artificial Neural Network (ANN) was trained using the lesion features with 80% accuracy.