Dual-decoder U-Net boosts trapezium bone segmentation in hand radiographs

A modified U-Net with dual decoders using contour and distance transform multi-task learning achieved a Dice score of 0.92 for trapezium segmentation on 519 hand X-rays, outperforming existing architectures in overlapping bone scenarios.

• Trained and evaluated on 519 annotated hand X-ray images; no prospective or external validation.
• Intersection over Union (IoU) reached 0.85, demonstrating improved spatial accuracy.
• Single-dataset evaluation limits generalizability; performance in other clinical settings is unknown.

RadPigeon summaries are original and for information only. They are not clinical advice.