A new color assessment methodology using cluster-based features for skin lesion analysis

Melanoma is considered the most dangerous form of skin cancer, however if detected in early stages there are high success rates of recovery, making prevention essential. The risk assessment of skin lesions usually follows the ABCD rule (asymmetry, border, color and dermoscopic structures). This paper presents a methodology to assess the number of ABCD rule colors of skin lesion images. It starts by extracting 660 color features and several feature selection and machine learning classification methods are tested in order to minimize classification errors. The developed methodology has the advantage of being adaptable to the dataset used. Two publicly available dermoscopic image datasets and one mobile acquired dataset were used to test the methodology, achieving accuracy rates of 77.75%, 81.38% and 93.55%, respectively, for ABCD rule color feature assessment.