Quantifying the Contribution of Bone Morphology to Implant Selection in Shoulder Arthroplasty Using CT-Based Deep Learning

We investigated whether bone morphology alone can inform implant selection in shoulder arthroplasty using a hypothesis-driven deep learning framework applied to preoperative computed tomography (CT) scans. The proposed approach extends a previously validated segmentation and pathology-staging pipeline by introducing implant-type prediction and a controlled human–AI comparison. The workflow combines CEL-UNet for 3D bone segmentation with ArthroNet+, a multi-task network assessing osteophytes, joint-space narrowing, humeroscapular alignment, and implant type. Trained on a multicenter cohort of 600 patients, CEL-UNet achieved Dice scores of 0.99 for the humerus and 0.98 for the scapula. ArthroNet+ achieved high performance in pathology classification (up to 95% for alignment tasks). Under morphology-only conditions, ten orthopedic surgeons achieved 61% accuracy with low inter-rater agreement (Fleiss’ κ≈0.15), while the model reached 78% agreement with the implant choices observed in the dataset, reflecting the ability to reproduce clinical decision patterns rather than to identify an optimal implant selection. This performance is characterized by a class-dependent asymmetry, with higher recall for reverse implants than for anatomical ones. These findings indicate that bone morphology provides a measurable but incomplete signal for implant selection, and should therefore not be interpreted as reflecting clinical decision-making performance. The framework quantifies the morphology-driven component of surgical decision making under controlled conditions, supporting future integration with multimodal clinical data.