Ensembles of Convolutional Neural Networks for Skin Lesion Dermoscopy Images Classification

Abstract

Skin cancer is a public health problem with more than 123,000 new cases diagnosed worldwide every year. System skin cancer screening reliable automatic will provide a great help for doctors to detect skin lesions as early as possible. The efficiency of deep learning based methods has recently outperformed conventional image processing methods in terms of classification. This study applied an ensemble of CNN to classify 7 categories of skin lesions. The preprocessing stage is hair removal, image resizing, and image augmentation. Model evaluation results with 1,440 test data indicate that the ensemble model achieve the best accuracy of 91.7% with a combination of learning rate parameters of le-3 and the use of dropouts in the model architecture.