Multi-classification of skin diseases for dermoscopy images using deep learning

Abstract

Skin diseases are very common in our daily life. Due to the similar appearance of skin diseases, automatic classification through lesion images is quite a challenging task. In this paper, a novel multi-classification method based on convolutional neural network (CNN) is proposed for dermoscopy images. A CNN network with nested residual structure is designed first, which can learn more information than the original residual structure. Then, the designed network are trained through transfer learning. With the trained network, 6 kinds of lesion diseases are classified, including nevus, seborrheic keratosis, psoriasis, seborrheic dermatitis, eczema and basal cell carcinoma. The experiments are conducted on six-classification and two-classification tasks, and with the accuracies of 65.8% and 90% respectively, our method greatly outperforms other 4 state-of-the-art networks and the average of 149 professional dermatologists.