Anatomy as a guide: Using learned priors for precise organ segmentation

Accurate segmentation of abdominal organs from CT is essential for clinical tasks such as radiotherapy planning but is hindered by ambiguity at organ boundaries. Standard deep learning models fail to effectively use their learned anatomical knowledge to guide boundary delineation, as their feature fusion mechanisms indiscriminately mix contextual and fine-grained information. We hypothesized that a hierarchical approach, where learned anatomical context actively guides the processing of boundary evidence, would improve segmentation accuracy. We propose Cog-Net, a 3D encoder-decoder network featuring a novel Co-axial Gated Attention (CGA) module. The CGA module redesigns the standard skip connection by using low-resolution, semantic features from the decoder to generate a dynamic spatial gate. This gate multiplicatively modulates the high-resolution feature stream from the encoder, allowing the model’s global anatomical knowledge to suppress noise and amplify subtle boundary details. We trained and evaluated our model on the Whole Abdominal Organ Dataset (WORD), a large-scale benchmark with 150 CT scans and 16 annotated organs, comparing its performance against state-of-the-art baselines using Dice Similarity Coefficient (DSC) and 95 % Hausdorff Distance (HD95). Cog-Net achieved state-of-the-art performance, with a mean DSC of 88.23 % and a mean HD95 of 6.32 mm across all organs. In terms of boundary accuracy, Cog-Net outperformed nnU-Net (mean HD95 10.82 mm) and Attention U-Net (7.86 mm). Gains were largest for challenging organs such as the spleen (HD95 2.17 mm vs. 9.76 mm for Attention U-Net). Ablation studies confirmed that the proposed CGA module was the primary driver of this improvement. By enabling learned anatomical context to hierarchically guide local boundary processing, Cog-Net significantly improves boundary delineation accuracy in abdominal organ segmentation. This architectural principle offers a path toward more precise and clinically reliable automated segmentation tools.