Classification of skin lesions using multi-task deep neural networks

Abstract

Skin cancer is the most prevalent type of cancer disease. The most of skin cancer deaths are caused by melanoma, despite being the least common skin cancer. Early and accurate detection and treatment is the best healing, however detection of this type of malignancy in the early stages is not obvious. Data-driven solutions for malignant melanomas detection can make treatment more effective. Convolutional neural networks have been successfully applied in different areas of computer vision, also in the classification of cancer types and stages. But in most cases, images are not enough to reach robust and accurate classification. Such metadata as sex, age, nationality, etc. could also be applied inside the models. In this paper, we propose an end-to-end method for the classification of melanoma stage using convolutional neural networks from an RGB photo and persons' metadata. Also, we provide a method of semi-supervised segmentation of the region of melanoma appearance. From the experimental results, the proposed method demonstrates stable results and learns good general features. The main advantage of this method is that it increases generalization and reduces variance by using an ensemble of the networks, pretrained on a large dataset, and fine-tuned on the target dataset. This method reaches ROC-AUC of 0.93 on 10982 unique unseen images.