Prediction of Dermoscopy Patterns for Recognition of both Melanocytic and Non-Melanocytic Skin Lesions

Abstract

A differentiation between all types of melanocytic and non-melanocytic skin lesions (MnM–SK) is a challenging task for both computer-aided diagnosis (CAD) and dermatologists due to the complex structure of patterns. The dermatologists are widely using pattern analysis as a first step with clinical attributes to recognize all categories of pigmented skin lesions (PSLs). To increase the diagnostic accuracy of CAD systems, a new pattern classification algorithm is proposed to predict skin lesions patterns by integrating the majority voting (MV–SVM) scheme with multi-class support vector machine (SVM). The optimal color and texture features are also extracted from each region-of-interest (ROI) dermoscopy image and then these normalized features are fed into an MV–SVM classifier to recognize seven classes. The overall system is evaluated using a dataset of 350 dermoscopy images (50 ROIs per class). On average, the sensitivity of 94%, specificity of 84%, 93% of accuracy and area under the receiver operating curve (AUC) of 0.94 are achieved by the proposed MnM–SK system compared to state-of-the-art methods. The obtained result indicates that the MnM–SK system is successful for obtaining the high level of diagnostic accuracy. Thus, it can be used as an alternative pattern classification system to differentiate among all types of pigmented skin lesions (PSLs).

Keywords

• skin cancer
• pattern recognition
• computer-aided detection
• color and texture features
• support vector machine
• majority voting scheme