Double-Decoder U-Net for Improved Segmentation of the Overlapping Trapezium Bone in X-ray Images

Area of Science:

• Medical imaging analysis
• Orthopedic surgery
• Artificial intelligence in healthcare

Background:

• Accurate segmentation of the trapezium bone in hand X-rays is vital for successful trapeziometacarpal (TMC) joint replacement surgery.
• Segmentation is challenging due to low contrast and bone overlap, particularly with the trapezoid bone.

Purpose of the Study:

• To develop and evaluate a multi-task segmentation approach for improved trapezium bone segmentation in hand X-rays.
• To enhance accuracy in challenging scenarios involving overlapping bones.

Main Methods:

• A modified U-Net architecture with two decoders was employed for multi-task learning.
• One decoder predicted the trapezium contour, while the other estimated a distance transform map.
• Complementary representations were fused to generate the final segmentation mask.

Main Results:

• The proposed model achieved a Dice score of 0.9203 and an IoU of 0.8498 on 519 annotated hand X-ray images.
• It outperformed several state-of-the-art segmentation architectures.
• Integrating structural priors via multi-task learning demonstrably improved segmentation accuracy.

Conclusions:

• The developed multi-task segmentation method effectively addresses the challenges of trapezium bone segmentation in hand X-rays.
• This approach shows significant promise for improving surgical planning in TMC joint replacement by providing more accurate bone segmentation.