A hierarchical structure based on Stacking approach for skin lesion classification

Abstract

Malignant melanoma is the most dangerous type of skin cancer. The diagnosis of melanoma in the early stages can greatly increase the possibility of its successful treatment. In the recent years, automated systems have played an pivotal role in increasing the skin cancer diagnosis rate. The main objective of this paper was to improve the performance of a skin cancer automated diagnostic system by introducing a new approach to combining classifiers in the classification stage. Therefore, the Stacking Ensemble Method based on the Meta Learning algorithm was proposed for the skin lesion classification. To classify skin lesions as melanoma, dysplastic and benign, two new hybrid approaches of Structure Based on Stacking (SBS) and Hierarchical Structure Based on Stacking (HSBS) were introduced to combine the heterogeneous classifiers. The proposed methods for skin lesions classification were implemented and evaluated based on the dermoscopic images of two PH2 and Ganster datasets using the Five Fold Cross Validation procedure and different numbers of the selected features. The results showed that the SBS approach had a good performance in diagnosing melanoma lesions from non-melanoma lesions for both datasets. Moreover, the results indicated that the HSBS method compared to the SBS approach and other works on the same dataset offers a far better performance in classifying skin lesions as benign, dysplastic and melanoma.