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

Area of Science:

• Orthopedic Surgery
• Artificial Intelligence
• Medical Imaging

Background:

• Shoulder arthroplasty implant selection relies on complex factors, including bone morphology.
• Deep learning offers potential for analyzing medical images to aid clinical decisions.

Purpose of the Study:

• To evaluate if bone morphology alone can guide shoulder arthroplasty implant selection using a deep learning framework.
• To compare the performance of an AI model against orthopedic surgeons in predicting implant type based solely on bone morphology.

Main Methods:

• A deep learning framework combining CEL-UNet for 3D bone segmentation and ArthroNet+ for pathology assessment and implant type prediction was developed.
• The model was trained and validated on preoperative computed tomography (CT) scans from 600 patients.
• A controlled study compared AI predictions to implant choices made by ten orthopedic surgeons using only bone morphology.

Main Results:

• CEL-UNet achieved high accuracy in bone segmentation (Dice scores of 0.99 for humerus, 0.98 for scapula).
• ArthroNet+ demonstrated strong performance in pathology classification (up to 95% for alignment).
• The AI model achieved 78% agreement in predicting implant choices, surpassing the surgeons' 61% accuracy and low inter-rater reliability (κ≈0.15).

Conclusions:

• Bone morphology provides a measurable but incomplete signal for shoulder arthroplasty implant selection.
• The developed AI framework can quantify the morphology-driven component of surgical decision-making.
• Further integration with multimodal clinical data is recommended for comprehensive implant selection support.