Deep Level Set with Confidence Map and Boundary Loss for Medical Image Segmentation

Level set method is widely used for image segmentation. Recent work combined traditional level set method with deep learning architecture for image segmentation. However, it is limited when dealing with medical images because of the blurred edges and complex intensity distribution, which leads to the loss of spatial details. To address this problem, we propose a deep level set method to refine object boundary details and improve the segmentation accuracy. We integrate augmented prior knowledge into inputs of CNN, which can make the level set evolution result has more accurate shape. In addition, to consider the spatial correlation of the object, we combine a boundary loss with deep level set model for preventing the reduction of details. We evaluate the proposed method on two medical image data sets, which are prostate magnetic resonance images and retinal fundus images. The experimental results show that the proposed method achieves state-of-the-art performance.