Cascade learning in multi-task encoder-decoder networks for concurrent bone segmentation and glenohumeral joint...

Area of Science:

• Orthopedic Surgery
• Radiology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Osteoarthritis (OA) is a degenerative joint disease impacting bone and cartilage, leading to structural changes like osteophyte formation, bone density loss, and joint space narrowing.
• Glenohumeral (GH) joint osteoarthritis can significantly impair shoulder functionality, necessitating effective diagnostic and treatment strategies.
• Current treatment options for shoulder OA range from conservative management to surgical interventions, with preoperative planning being crucial for optimal outcomes.

Purpose of the Study:

• To introduce an innovative deep learning framework for processing shoulder CT scans to aid in the diagnosis and surgical planning of glenohumeral osteoarthritis.
• To enable semantic segmentation of the proximal humerus and scapula, 3D bone surface reconstruction, and identification of the GH joint region.
• To accurately stage common osteoarthritic conditions: osteophyte formation (OS), GH space reduction (JS), and humeroscapular alignment (HSA).

Main Methods:

• Development of a deep learning pipeline utilizing two cascaded Convolutional Neural Network (CNN) architectures: 3D CEL-UNet for segmentation and 3D Arthro-Net for classification.
• Retrospective analysis of 571 shoulder CT scans from patients with varying degrees of GH osteoarthritic pathologies.
• Evaluation of 3D reconstruction accuracy using Root Mean Squared Error (RMSE) and Hausdorff distance, and classification performance for OS, JS, and HSA.

Main Results:

• The framework achieved high accuracy in 3D bone reconstruction with RMSE and Hausdorff distance median values of approximately 0.22-0.24 mm and 1.48 mm, respectively, outperforming existing methods.
• Classification accuracy for osteophyte formation, joint space reduction, and humeroscapular alignment consistently reached around 90% across all severity stages.
• The entire inference pipeline completed in under 15 seconds, demonstrating significant efficiency for clinical application in orthopedic radiology.

Conclusions:

• The developed deep learning framework offers accurate 3D bone reconstruction and robust classification of key shoulder osteoarthritis features.
• The high accuracy and rapid processing time make this AI tool a promising advancement for streamlining preoperative planning in shoulder arthroplasty.
• This technology has the potential to support orthopedic surgeons by providing detailed joint condition analysis and guiding the selection of patient-specific surgical approaches.