Prediction of humeral torsion from proximal humerus computed tomography is enabled by three-dimensional estimation of the transepicondylar axis using deep learning.

Background: Humeral torsion is an important factor for reverse total shoulder arthroplasty planning, yet its measurement requires elbow imaging frequently absent from clinical computed tomography (CT) scans. We hypothesized that a deep learning approach could accurately predict humeral torsion from proximal humerus morphology alone. Methods: CT-based segmentations from 581 shoulders were used to train and evaluate a deep learning model within the nnU-Net framework. The transepicondylar axis was represented as a full 3-dimensional plane extending throughout the imaging volume. The model was trained on 478 cases and tested on 103, with ground-truth torsion derived from elbow segmentations. Robustness to osteophytes and reduced fields of view was assessed. Results: The model achieved a mean absolute torsion prediction error of 5.35° ± 3.96° and a strong linear correlation with ground-truth measurements (R = 0.85), outperforming prior proximal landmark-based approaches (R = 0.33-0.42). Performance was robust to osteophytes (mean error: 6.19° ± 3.95°) and was maintained with proximal field of view reductions down to 25% of humeral length, beyond which accuracy progressively declined. Conclusion: Deep learning can predict humeral torsion directly from proximal humerus morphology with clinically relevant accuracy, reducing the need for distal elbow imaging. These findings support the hypothesis that torsional alignment is largely encoded in the proximal epiphysis and offers a practical solution for reverse total shoulder arthroplasty planning when full-length CT is unavailable.