Position-Encoded Pixel-to-Prototype Contrastive Learning for Aortic Vessel Tree Segmentation

Hyeongyu Kim, Yejee Shin, Dosik Hwang

Abstract: The SegA challenge concentrates on segmenting the aorta in CTA scans. Our approach is divided into two main phases: a broad, preliminary segmentation to identify the aorta’s general vicinity, and a subsequent detailed segmentation for precision. The distinguishing feature of our method is the use of prototype-based learning during the detailed segmentation. By studying specific examples from the foreground, the algorithm captures the nuanced differences in aortic structures, leading to more precise segmentation results. Our technique demonstrated its robustness by achieving the 5th rank in the phase 2 of the challenge. Such advancements in segmentation techniques not only prove effective in competitions but also have the potential to revolutionize medical image analysis, paving the way for improved diagnostic and treatment planning in the clinical realm.

External IDs:dblp:conf/miccai/KimSH23a