HGSF-Net: Automatic Pelvic Fracture CT Segmentation with 2.5D Neighborhood Context and Adaptive Detail Guidance

Accurate assessment of pelvic fractures is crucial for preoperative planning and postoperative follow-up. While CT is widely used due to its high contrast for osseous structures, manual delineation is time-consuming and subject to inter-observer variability, underscoring the need for automation. Addressing challenges in pelvic CT-namely sparse small targets, thin boundaries, and high similarity to surrounding soft tissue-we propose a Hierarchically Guided Semantic Fusion Network (HGSF-Net). Built on a convolutional encoder-decoder backbone, HGSF-Net introduces two key advances: (i) 2.5D adjacent-slice context encoding with a slice-consistency constraint, where a five-slice stack (center ±2) is fed to explicitly model inter-slice dependencies at shallow stages; the primary loss is back-propagated only to the central slice to enforce cross-slice coherence for thin lamellar and fissure-like structures. (ii) Adaptive detail guidance along two shallow/mid-level guided paths: content-aware detail hint maps are generated as auxiliary supervision to emphasize true edges and textures, while the high-level semantic path is equipped with a lightweight CB (GAP+GMP) attention refinement to improve boundary separability. At inference, the guided side branches are removed, yielding a computational cost comparable to standard 2D CNNs. Evaluated on CTPelvic1K with mIoU and Accuracy, HGSF-Net achieves consistent improvements over peer methods in overall accuracy, boundary quality, and the small-object subset, demonstrating strong practicality and robustness without increased deployment cost.