Anatomy-driven automated subsegmentation of pelvic and proximal femoral CT into clinically relevant subregions and landmarks.

backgroundThe pelvis is among the most demanding regions in musculoskeletal imaging, and CT is its reference standard. Available automated tools, however, largely produce whole-bone or single-task masks, leaving the finer surgical subregions to slow, operator-dependent manual work. We developed a rule-based pipeline that converts pelvic and proximal femoral masks into a structured map of surgically meaningful subregions and landmarks, quantifying the manual correction burden required to achieve anatomically acceptable outputs.

methodsAnonymized pelvic CT scans were collected retrospectively; after predefined quality criteria, 757 of 1,316 screened normal adult cases were processed by a one-click pipeline in 3D Slicer. Parent structures from TotalSegmentator were divided by fixed anatomical rules into columns, acetabular regions, and proximal femoral regions. In 100 randomly selected cases, every segment was expert-reviewed and corrected where needed; correction frequency was the primary endpoint, with accuracy computed for corrected segments only. Pipeline performance was also assessed in a 60-case pathological-anatomy cohort for affected segments against manually segmented ground truth.

resultsOf 8,100 segment observations, 93.8% were accepted without correction, and 51 of 81 segments required no correction in any case. Correction clustered in a few boundary regions, mainly the ischium and femoral neck (median 4.5 segments per case). Computed only on the corrected segments, the median Dice was 0.94, indicating that errors were modest where they occurred. A secondary symmetry analysis showed median surface agreement of 90% within 2 mm and 99.9% within 5 mm. In the pathological-anatomy cohort, pooled median Dice was 0.970 across affected components. Pooled landmark error was 5.37 mm median, with 91.9% of manual clicks inside the automated landmark.

conclusionsValidated in normal adult anatomy, this rule-based pipeline converts pelvic and proximal femoral CT into a structured hierarchy of clinically relevant subregions, needing only limited, predictable correction. It substantially reduces manual workload while preserving the anatomical detail needed for morphometry and dataset generation. Performance held up in an initial pathological-anatomy sample. Extension to fracture, deformity, or clinical decision-making requires dedicated validation in abnormal anatomy.