DSM: A Deep Supervised Multi-Scale Network Learning for Skin Cancer Segmentation

Abstract

The automatic segmentation of the skin lesion on dermoscopy images is an important step for diagnosing the melanoma. However, the skin lesion segmentation is still a challenging task due to the blur lesion border, low contrast between the skin cancer region and normal tissue background, and various sizes of cancer regions. In this paper, we propose a deep supervised multi-scale network (DSM-Network), which achieves satisfied skin cancer segmentation result by utilizing the side-output layers of the network to aggregate information from shallow&deep layers, and designing a multi-scale connection block to handle a variety of cancer sizes' changes. Moreover, a post-processing of the contour refinement strategy is adopted by a conditional random field (CRF) model to further improve the segmentation results. Extensive experiments on two public datasets: ISBI 2017 and PH2 have demonstrated that our designed DSM-Network has gained competitive performance compared with other state-of-the-art methods.

Keywords

• Skin cancer
• dermoscopy image
• deep supervised learning
• multi-scale feature
• conditional random field