Lesion Attributes Segmentation for Melanoma Detection with Multi-Task U-Net

Abstract

Melanoma is the most deadly form of skin cancer worldwide. Many efforts have been made for early detection of melanoma with deep learning based on dermoscopic images. It is crucial to identify the specific lesion patterns for accurate diagnosis of melanoma. However, the common lesion patterns are not consistently present and cause sparse label problems in the data. In this paper, we propose a multi-task U-Net model to automatically detect lesion attributes of melanoma. The network includes two tasks, one is the classification task to classify if the lesion attributes present, and the other is the segmentation task to segment the attributes in the images. Our multi-task U-Net model achieves a Jaccard index of 0.433 on official test data of ISIC 2018 Challenges task 2, which ranks the 5th place on the final leaderboard.