Publication

Abstract : Background: The dermatologist widely uses digital dermoscopy for the detection of melanoma. The accurate detection of melanoma by clinicians is subjective and further depends on their experience. Fully automated computer-aided diagnosis systems are necessary to eliminate the inter-operator variability inherent in the personal analysis of dermoscopy images. Objective: Automated skin lesion classification is challenging because of the fine-grained difference in the appearance of these lesions on the skin surface. Deep convolutional neural network (CNN) have shown great separability across many fine-grained object classes. Methods: This article presents two novel hybrid CNN models with an SVM classifier at the output layer for classifying dermoscopy images into either benign or melanoma lesions. The features extracted by the first CNN and second CNN models are concatenated and fed to the SVM classifier for classification. The labels obtained from an expert dermatologist are used as a reference to evaluate the performance of the proposed model. Results: The proposed models displayed better results over the state-of-the-art CNN models on the publicly available ISBI 2016 dataset. The proposed models achieved 88.02% and 87.43% accuracy, which remain higher than the traditional CNN models. Conclusions: The proposed framework could attain considerable improvement in accuracy to classify dermoscopy images.