Melanoma Detection by Meta-Heuristically-Optimized MLP Parameters Using Non-Dermatoscopy Images

Abstract

This paper represents a scheme of melanoma detection using handcrafted feature set with meta-heuristically optimized multilayer perceptron (MLP) parameters. Features including shape, color, and texture are extracted from camera images of skin lesion collected from University of Waterloo database. The features are used in two different ways for binary classification of the data into benign and malignant class. 1) The extracted features are ranked on their relevance using ReleifF ranking algorithm and also converted into PCA components and ranked according to their variance. Best result is obtained with 50 best ranked raw features with accuracy of 87.1%. 2) All 1,888 features are fed to an MLP with two hidden layers, with number of neurons optimized by two different metaheuristic algorithms, namely particle swarm optimization (PSO) and simulated annealing (SA) separately. The latter method is found to be more efficient, and an accuracy of 88.38%, sensitivity of 92.22%, and specificity of 83.07% are achieved by PSO, which is better in comparison with the latest research on this dataset.