Deep Residual Network-based Melanocytic Lesion Classification with Transfer Learning

Abstract

Melanoma is fatal cancer that can develop from melanocytes, which is also called malignant melanoma. Melanomas usually occur due to ultra-violate (UV) radiation in the skin. A benign tumor can also develop from melanocytes. But it is not deadly like malignant melanoma. Deep learning has been used successfully in case of dermatological diagnosis. Here, we present a deep learning-based scheme to classify melanocytic lesion from dermoscopic images. Utilizing deep neural networks requires huge data. Here, a limited dataset problem was solved with transfer learning. For classifying malignant melanoma, a deep residual network architecture was used for the purpose of feature extraction. Using those features, supervised learning methods, like support vector machine (SVM) and decision tree was used for classification. Test accuracy of 95% was found with the best model. It is expected that the findings of this study will be helpful for cancer diagnosis.