An accelerated system for melanoma diagnosis based on subset feature selection

Abstract

In this paper we present an accelerated system for diagnosing skin lesions based on digitized dermatoscopic color images. This system is composed mainly of three levels: lesion detection, lesion description (features selection) and decision. The lesion detection level consists in the preprocessing of the lesion image in order to remove the undesired objects from the original image. Then, the extraction of the lesion is done by separating it from the healthy surrounding skin. The lesion description level is based on the extraction of a set of features modeling clinical signs of malignancy. The decision level is based on the produced vector of features scores, which is used as input to a multi-layer perceptron classifier in order to assign the lesion to the class of benign lesions or to the one of malignant melanomas. We focus particularly in this paper on the critical step of the features selection allowing to select a reasonable reduced number of useful features while removing redundant information and approximating the properties of melanoma recognition. This permits to reduce the dimension of the lesion's vector, and consequently the computing time, without a significant loss of information. In fact, a large set of features was investigated by the application of relevant features selection techniques. Then, the number of features for classification was optimised and only five well-selected features were used to cover the discriminatory information about lesions malignancy. With this approach, for reasonably balanced training/test sets, we record a good classification rate of 77.7% in a very promising CPU time.