Improving retinal vessel segmentation with joint local loss by matting

Besides the binary segmentation, many retinal image segmentation methods also produce a score map, where a nonnegative score is assigned for each pixel to indicate the likelihood of being a vessel. This observation inspires us to propose a new approach as a post-processing step to improve existing methods by formulating segmentation as a matting problem. A trimap is obtained via a bi-level thresholding of the score map from existing methods, which is instrumental in focusing the attention to pixels of these unknown areas. A dedicated end-to-end matting algorithm is further developed to retrieve those vessel pixels in the unknown areas, and to produce the final vessel segmentation by minimizing global pixel loss and local matting loss. Our approach is shown to be particularly effective in rescuing thin and tiny vessels that may lead to disconnections of vessel fragments. Moreover, it is observed that our approach is capable of improving the overall segmentation performance across a broad range of existing methods.