Multi-dimensional consistency learning between 2D Swin U-Net and 3D U-Net for intestine segmentation from CT volume

• DOI: 10.1007/s11548-024-03252-6
• Publication Date: 2025-02-22T11:30:37Z
• Authors: Qin An; Hirohisa Oda; Yuichiro Hayashi; Takayuki Kitasaka; Hiroo Uchida; Akinari Hinoki; Kojiro Suzuki; Aitaro Takimoto; Masahiro Oda; Kensaku Mori
• Keywords: Labeled data; F1 score; Large intestine; Net (polyhedron)

Abstract Purpose The paper introduces a novel two-step network based on semi-supervised learning for intestine segmentation from CT volumes. folds in the abdomen with complex spatial structures and contact neighboring organs that bring difficulty accurate labeling at pixel level. We propose multi-dimensional consistency method to reduce insufficient results caused by limited labeled dataset. Methods designed two-stage model segment intestine. In stage 1, 2D Swin U-Net is trained using data generate pseudo-labels unlabeled data. 2, 3D create final model. comprises two networks different dimensions, capturing more comprehensive representations of potentially enhancing model’s performance segmentation. Results used 59 volumes validate effectiveness our method. experiment was repeated three times getting average as result. Compared baseline method, improved 3.25% Dice score 6.84% recall rate. Conclusion proposed involves training both U-Net. mitigates impact maintains consistncy outputs improve accuracy. previous methods, demonstrates superior performance.