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

Accurate segmentation of the trapezium bone in hand X-ray images is essential for surgical planning in trapeziometacarpal (TMC) joint replacement. However, this task remains challenging due to low contrast and frequent overlap with neighboring bones such as the trapezoid. In this work, we propose a multi-task segmentation approach based on a modified U-Net architecture with two decoders. The first decoder predicts the contour of the trapezium, while the second estimates a distance transform map that encodes spatial information about the bone structure. These complementary representations are fused to produce the final segmentation mask. Experiments on a dataset of 519 annotated hand X-ray images demonstrate that the proposed model outperforms several state-of-the-art segmentation architectures, achieving a Dice score of 0.9203 and an IoU of 0.8498. The results show that integrating structural priors through multi-task learning improves segmentation accuracy in challenging overlapping bone scenarios.