CPNet: Cycle Prototype Network for Weakly-Supervised 3D Renal Compartments Segmentation on CT Images

Renal compartment segmentation on CT images targets extracting the 3D structure of renal compartments from abdominal CTA and is great significance to diagnosis treatment for kidney diseases. However, due unclear boundary, thin large anatomy variation images, deep-learning based a challenging task. We propose novel weakly supervised learning framework, Cycle Prototype Network, segmentation. It has three innovations: (1) A Learning (CPL) proposed learn consistency generalization. learns pseudo labels through forward process regularization reverse process. The two processes make model robust noise label-efficient. (2) Bayes Weakly Supervised Module (BWSM) cross-period prior knowledge. knowledge unlabeled data perform error correction automatically, thus generates accurate labels. (3) present Fine Decoding Feature Extractor (FDFE) fine-grained feature extraction. combines global morphology information local detail obtain maps with sharp detail, so will achieve fine structures. Our extensive experiments demonstrated our performance. achieves Dice \(79.1\%\) \(78.7\%\) only four labeled achieving significant improvement by about \(20\%\) than typical prototype PANet [16].