Semi-supervised Prostate Multi-Regional Semantic Segmentation with Patch-Based Plug-and-Play Correction Guidance.

Abstract

The high cost of medical image annotation severely restricts the clinical application of prostate precise multi-regional segmentation technologies. To address existing bottlenecks in semi-supervised learning methods, including insufficient alignment of local anatomical structures and pseudo-label noise accumulation, this paper proposes a labeled patch guidance (LPG) for patch-level interactive enhancement module that optimizes pseudo-label quality for unlabeled data by dynamically mining anatomical prior knowledge from labeled data. Specifically, 1) A labeled patch mutual correction (LPMC) mechanism first performs bidirectional exchange of annotated data across augmented states, obtaining newly labeled data with enhanced model robustness and establishing a high-confidence anatomical prior patch pool for prostate regions. 2) An unlabeled patch capture learning (UPCL) mechanism is then proposed to inject reliable anatomical information into low-confidence patches of unlabeled data through feature similarity retrieval from the high-confidence anatomical prior patch pool, thereby improving the model's ability to recognize feature distributions in unlabeled data. Comparative experiments demonstrate that our method significantly enhances the performance of mainstream semi-supervised medical image segmentation models on PROMISE12, MSD, HPH55, and ACDC datasets. In ablation studies, GradCAM-based interpretability analysis visually demonstrates that the LPG-equipped model effectively suppresses ambiguous boundaries in prostate segmentation and concentrates model attention on regions of interest. The source code associated with this work has been made publicly accessible on GitHub at https://github.com/hai-medicallab/LPG.