ZonoSeg: A Zonal-aware Prostate Segmentation Framework on 3D MRI images

Accurate zonal-aware prostate segmentation in clinical 3D MRI is important for diagnosis and treatment planning. However, delineating the central-gland-peripheral-zone interface remains challenging because gradual tissue transitions, irregular internal boundaries, and pathology-induced appearance variations can lead to clinically relevant zonal boundary errors. We propose ZonoSeg, a hierarchical encoder-decoder framework that targets these ambiguous boundaries via uncertainty-aware, boundary-sensitive feature transmission in stage-wise skip connections. The ZonoSeg Block combines edge cues with learnable multi-scale Gaussian smoothing to construct a continuous probabilistic boundary field, and the ZAD head further refines zonal separation through explicit foreground-background interactions. Experiments demonstrate improved boundary localization and robust zonal segmentation on both dual-target and single-target tasks across eight multi-center prostate MRI datasets; on PROMISE MRI BMC, ZonoSeg achieves Dice scores of 83.39% for the central gland and 84.10% for the peripheral zone, with HD95 of 3.84 mm and 3.41 mm, respectively. Code: https://github.com/deepang-ai/ZonoSeg.