Malignant melanoma detection using multi-scale image decomposition and a new ensemble-learning scheme

Malignant melanoma is one of the most serious and deadly types of skin cancer, fortunately it is treatable if detected at an early stage. Many Computer-Aided Diagnosis (CAD) systems are proposed in the literature to assist in detecting this type of cancer. The vast majority of them are performed in four main steps, which are preprocessing, image segmentation, features extraction, and classification. In this paper, we propose a new ensemble-learning scheme that uses three classifiers each one based, separately, on a different type of features: color, texture, and shape. Moreover, in order to be able to efficiently identify the lesion and safe skin region and thus extract the features, we proposed using, as a preprocessing step, a multi-scale image decomposition based on Partial Differential Equations (PDE) that will allow us to isolate the object and the texture from the image. Color and texture features will be extracted from both, the safe skin and the extracted lesion, while shape features will be extracted from only the detected lesion. We proposed using the SVM algorithm for the design of each of the three classifiers and the final diagnostic decision is obtained by applying the majority voting approach. The proposed model was validated and tested on the public PH2 database, achieving a classification accuracy of 98%, rappel of 95%, and precision of 95%; we can conclude that the proposed approach is very promising.