Meta-tubular-net: A robust topology-aware re-weighting network for retinal vessel segmentation

Vessel segmentation is one of the basic and vital tasks for clinical automation applications, but the expensive cost of pixel-wise labels has been hindering the development of data-driven deep networks. In recent years, the emergence of crowdsourcing platforms has provided a solution to this problem. By outsourcing the labeling works to non-specialists and medical novice groups, large human knowledge can be acquired in a cut-price way. However, crowdsourced data is bringing about new problems, i.e., the unreliability of the accuracy of outsourced data, and the huge gap between the amount of expertise data and outsourced data, which lead to the undesired bias of deep networks. Although many studies have been conducted on overcoming the bias, they are equipped with redundant computation and are not suitable for topological structure segmentation. In this paper, we propose a meta-tubular robust learning approach for vessel segmentation, leveraging a limited number of gold standard data and a large number of noisy data. The core of our method is the tubular compartmentalization of an intact vascular pattern and a novel tubular loss function which is designed for each vascular segment. Additionally, we introduce a controlled and context-appropriate method for synthesizing vascular data noise, allowing us to evaluate the robust learning performance effectively. Our experiments, conducted on two publicly available clean datasets, demonstrate that our tubular loss function yields superior performance. Moreover, on synthesized noisy datasets, our method performs at par or even surpasses other state-of-the-art robust learning approaches.