Deep learning pipeline for trapezium segmentation in thumb radiographs.

Area of Science:

• Musculoskeletal imaging
• Artificial intelligence in radiology
• Orthopedic surgery

Background:

• Accurate trapezium identification is vital for successful trapeziometacarpal (TMC) arthroplasty.
• Overlapping anatomy on standard radiographs complicates trapezium visualization.
• Current AI applications in small joint imaging are limited.

Purpose of the Study:

• To develop and evaluate a deep learning pipeline for precise trapezium segmentation on thumb radiographs.
• To compare the performance of the proposed AI model against existing segmentation techniques.
• To assess the potential of AI in improving preoperative planning for TMC arthroplasty.

Main Methods:

• Retrospective analysis of 519 thumb radiographs meeting quality criteria.
• Development of a two-stage deep learning pipeline: YOLOv8 for detection and U-Net for segmentation.
• Comparison with standalone U-Net, Segment Anything Model (SAM), and Mobile-SAM.

Main Results:

• The YOLOv8 detector achieved 99.5% mean average precision (mAP).
• The combined YOLOv8 + U-Net model demonstrated superior segmentation performance (DSC 94.2%, IoU 89.1%) compared to other models.
• Expert-level inter-observer agreement (κ=0.89) was achieved.

Conclusions:

• The proposed two-stage AI pipeline offers accurate and reproducible trapezium segmentation on radiographs.
• This AI approach surpasses the performance of widely used segmentation models.
• The tool shows promise for enhancing preoperative planning and intraoperative guidance in TMC arthroplasty.