• Breaking Grid Constraints: Dynamic Graph Reconstruction Network for Multi-organ Segmentation

    Junhao Xiao, Yang Wei, B, Jingyu Wang, Yongchao Wang, Xiuli Bi, and Bin Xiao

    Abstract

    Morphological differences and dense spatial relations oforgans make multi-organ segmentation challenging. Currentsegmentation networks, primarily based on CNNs andTransformers, represent organs by aggregating informationwithin fixed regions. However, aggregated representationsoften fail to accurately describe the shape differences andspatial relationships of multi-organs, which leads to imprecisemorphological modeling and ambiguous boundary representation.In response, we propose a novel multi-organsegmentation network via dynamic graph reconstruction,called DGRNet. Unlike existing approaches, DGRNet employsa graph-based paradigm to reconstruct multi-organsand leverages the topological flexibility of graphs to representirregular organ morphology. Based on graph connectivity,the precise information interaction makes moreefficient multi-organ modeling. Furthermore, DGRNet introducesa category-aware guidance mechanism that utilizesorgan category-specific priors to explicitly define interorganboundaries, addressing the issue of ambiguous margindelineation in multi-organ regions. We conducted extensiveexperiments on five datasets (including CT and MRI),showing that DGRNet outperforms state-of-the-art methodsand models complex multi-organ areas better, highlightingits effectiveness and robustness. Code is available at:https://github.com/robert1818118/DGRNet.