Related Experiment Video
Updated: May 7, 2026
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
Deep learning pipeline for trapezium segmentation in thumb radiographs
Victor Maigné1, Youssef Frikel2, Félix Barbier3
1Orthopaedic Surgery Department, Hôpital Avicenne, AP-HP, Université Paris Sorbonne Nord, Bobigny, France. victor.maigne@aphp.fr.
Objective:
Accurate identification of the trapezium is crucial for trapeziometacarpal (TMC) arthroplasty but remains challenging on standard radiographs due to overlapping anatomy. Artificial intelligence has shown promise in musculoskeletal imaging, yet its application to small joints is limited.
Materials And Methods:
We retrospectively analyzed 624 thumb radiographs, of which 519 met the inclusion criteria. Radiographs of insufficient quality-blurred images or non-centered TMC joints-were excluded by consensus of two hand surgeons. Manual trapezium annotations performed by an expert surgeon were reviewed by two additional surgeons. Inter-observer agreement was assessed on 10% of cases using Cohen κ. We developed a two-stage deep learning pipeline combining You Only Look Once (YOLO)v8 for trapezium detection with U-Net for segmentation. Its performance was compared with the standalone U-Net, segment anything model (SAM), and Mobile-SAM. Detection accuracy was measured using mean average precision (mAP), while segmentation was evaluated with Dice similarity coefficient (DSC) and intersection over union (IoU).
Results:
YOLOv8 achieved a detection mAP of 99.5%. The combined YOLOv8 + U-Net model yielded a DSC of 94.2% and an IoU of 89.1%, outperforming U-Net (DSC 89.5%, IoU 81.2%), SAM (Dice 88.8%, IoU 80.3%), and Mobile-SAM (Dice 88.9%, IoU 80.5%). Inter-observer agreement was excellent (κ = 0.89, DSC = 93.8%).
Conclusion:
The proposed two-stage pipeline provides accurate, reproducible trapezium segmentation on radiographs, outperforming widely used models. This approach may enhance preoperative planning and intraoperative guidance in TMC arthroplasty.
Relevance Statement:
This two-stage AI pipeline enables precise trapezium segmentation on thumb radiographs, supporting improved surgical planning and intraoperative guidance in TMC arthroplasty, with potential to enhance implant placement accuracy and patient outcomes.
Key Points:
A two-stage AI pipeline (YOLOv8 + U-Net) accurately detects and segments the trapezium on thumb radiographs. The method outperforms popular segmentation models and achieves expert-level reproducibility. This tool may enhance surgical planning and intraoperative guidance for TMC arthroplasty.
