An effective approach for the diagnosis of melanoma using the sparse auto-encoder for features detection and the SVM for classification

Abstract

Malignant melanoma is considered one of the terrible disorders causing death. The goal of the modern dermatology is the early screening of skin cancer, aiming at reducing the mortality rate with less extensive treatment. In this context, this work focuses on the problem of an automatic melanoma diagnosis. The proposed approach uses unsupervised robustness of deep learning to extract significant characteristics from pixels of the images. A preprocessing step is used to remove unwanted artifacts and to improve the contrast of the images. Then, features are extracted by a deep Sparse Auto-encoder. Finally, the classifier Support Vector Machine (SVM) is used to distinguish respectively between three populations which are Melanoma, suspicious cases, and non-melanoma. For evaluation, we test the proposed approach using images from the PH2 dataset. The results show remarkable performance in terms of specificity, sensitivity, and accuracy.