Region-edge collaborative learning with fine-grained feature guidance for prostate zonal segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate segmentation of the transition zone (TZ) and peripheral zone (PZ) in transrectal ultrasound (TRUS) images is essential for early prostate cancer diagnosis and biopsy guidance. However, speckle noise and subtle acoustic impedance differences between TZ and PZ lead to blurred anatomical boundaries and low tissue contrast, making segmentation challenging. Existing deep learning methods often fail to preserve fine-grained spatial details and lack geometric constraints, resulting in boundary ambiguity and region confusion. To address these issues, we propose a Fine-Grained Region-Edge Collaborative Network (FRCNet). The Fine-Grained Boundary Localization (FGBL) module extracts high-resolution features from shallow encoder layers to enhance weak edge information, while the Region-Edge Collaborative (REC) module establishes a dual-task learning framework that jointly optimizes semantic segmentation and boundary detection through region-edge feature interaction and sigmoid-based feature modulation. Experiments on a TRUS dataset demonstrate that FRCNet achieves Dice scores of 74.28 ± 0.61% and 72.24 ± 1.08% for the inner and outer glands, respectively, with corresponding HD95 values of 3.17 ± 0.41 mm and 3.36 ± 0.52 mm, showing improved boundary localization and structural consistency in challenging clinical scenarios.