Accurate Humeral Version Angle Prediction From Limited Anatomical Data Using Statistical Shape Models

Shoulder replacement is challenged by the need to approximate the patient's pre-morbid humeral version angle (HVA) with limited information. Statistical Shape Models (SSMs) can reconstruct bone shapes and predict HVA by identifying the best global shape match. However, SSMs require visible humeral anatomy, and optimal global reconstruction does not necessarily ensure accurate HVA. Multisegmented SSMs (MSSMs) combine adjacent bones to leverage inter-bone shape correlations, but their potential remains largely unexplored. This project uses scale-free humeral SSM and three-shoulder-bone MSSM to estimate HVA from clinically available surfaces and examine how HVA accuracy relates to global and local bone reconstruction. Sixty-one registered bones were segmented (29 males and 32 females, 36.1 ± 14.3 years) from medical images. A leave-one-out method tested the models' ability to reconstruct the humerus and HVA under surgical scenarios: the SSM for cases with visible humeral parts (distal- or proximal-limited), and the MSSM for cases with no humeral information (solely on adjoining segments). SSM-based cases produced higher HVA accuracy than MSSM scenarios (max error: 3.3° ± 2.3° vs. 6.6° ± 5.4°). In both models, errors were driven primarily by marker-level discrepancies rather than global shape error. These results highlight that, when predicting anatomical measurements, SSM predictions should prioritize anatomical accuracy over achieving the best-fitting global shape.