An efficient system for Melanoma diagnosis in dermoscopic images

Abstract

In this work, an automatic computer aided diagnosis system for Melanoma in dermoscopic images is proposed. In which, a large set of features is extracted from normalized tumor area to mimic the well-known ABCD clinical diagnosis rule. Consequently, to select the most prominent set of features, a recursive feature elimination algorithm based on random forests classifier is utilized. To alleviate classes imbalance problem which usually occurs in clinical situations, the neighborhood cleaning rule (NCL) and the borderline synthetic minority over-sampling (Borderline-SMOTE) algorithms are integrated to produce a better balanced dataset. The final diagnosis is then obtained by utilizing an extra tree classifier which builds an ensemble of classifiers using different sets of features and makes a late fusion. This proposed pipeline allows the diagnosis system to perform well in different challenging situations. Evaluation of the proposed system was performed using a recently released dataset which has a large classes imbalance. The experimental results indicate the efficiency of proposed system. It achieves the best average precision score among recent competitors who use the same dataset. Moreover, it makes a better balance between sensitivity and specificity scores.