부산대학교 도서관

Current semantic segmentation methods tend to segment objects on lower resolution datasets, but are not as effective for high resolution medical images. The coronary artery stenosis detection task in this experiment, for example, needs not only to ensure the continuity of the vessel as a whole, but also to be able to identify boundary defects at the maximum level of granularity. This paper therefore proposes a Coarse-to-fine componentised model structure to achieve stenosis defect detection while ensuring vessel continuity, successfully improving the Boundary IoU to 84.53% and with only 59% of the number of parameters in ViT.