Boosting Performance of Machine Learning based Melanoma Detection using GLCM Features Extracted with Luminance Chromaticity Color Spaces

Abstract

One of the deadliest types of cancers is Melanoma skin cancer, one of the best solutions to fight this disease is to detect it in the early stages. Manual detection can sometimes be time-consuming and error-prone. Automated detection may also sometimes fail to distinguish different types of skin abnormalities. But by working towards increasing the accuracy of automated detection one can reduce the abnormalities. The proper selection of features for these abnormalities has a big role in getting more accurate prediction results. Therefore in this work, the “Gray Level Co-occurrence Matrix (GLCM)” features to extract the textural features of melanoma skin images are proposed with  different color spaces and using these features the skin image of melanoma is classified into benign or malignant. The paper explores melanoma detection methods as existing based on GLCM applied on RGB’, proposed ‘GLCM applied on YCbCr color space’ and ‘proposed GLCM applied on CIE L*u*v color space’. These extracted features are used further to train different models of machine learning and ensemble of models of machine learning. The experimentation is done using benchmark “International Skin Imaging Collaboration (ISIC)” dataset of melanoma images. The ensemble of Random Forest-Multi Layer Perceptron (MLP) with majority voting logic gave the best results among the variations of models of machine learning. Among the experimented color spaces the YCbCr color space has shown promising performance closely followed by CIE L*u*v color space proving the worth of the proposed modification.