Classification of Melanoma Lesions Using Wavelet-Based Texture Analysis

Abstract

This paper presents a wavelet-based texture analysis method for classification of melanoma. The method applies tree-structured wavelet transform on different color channels of red, green, blue and luminance of dermoscopy images, and employs various statistical measures and ratios on wavelet coefficients. Feature extraction and a two-stage feature selection method, based on entropy and correlation, were applied to a train set of 103 images. The resultant feature subsets were then fed into four different classifiers: support vector machine, random forest, logistic model tree and hidden naive bayes to classify melanoma in a test set of 102 images, which resulted in an accuracy of 88.24% and ROC area of 0.918. Comparative study carried out in this paper shows that the proposed feature extraction method outperforms three other wavelet-based approaches.