Edge-competing Pathological Liver Vessel Segmentation with Limited Labels

The microvascular invasion (MVI) is a major prognostic factor in hepatocellular carcinoma, which one of the malignant tumors with highest mortality rate. diagnosis MVI needs discovering vessels that contain carcinoma cells and counting their number each vessel, depends heavily on experiences doctor, largely subjective time-consuming. However, there no algorithm as yet tailored for detection from pathological images. This paper collects first liver image dataset containing $522$ whole slide images labels vessels, MVI, grades. essential step automatic accurate segmentation vessels. unique characteristics images, such super-large size, multi-scale blurred vessel edges, make challenging. Based collected dataset, we propose an Edge-competing Vessel Segmentation Network (EVS-Net), contains network two edge discriminators. network, combined edge-aware self-supervision mechanism, devised to conduct limited labeled patches. Meanwhile, discriminators are introduced distinguish whether segmented background residual features adversarial manner. In training stage, compete predicted position edges. Exhaustive experiments demonstrate that, only patches, EVS-Net achieves close performance fully supervised methods, provides convenient tool segmentation. Code publicly available at https://github.com/wang97zh/EVS-Net.