| Reliable segmentation of hip implants in radiographic images is essential for assessing prosthesis condition and detecting potential complications. In this study, we investigate uncertainty-aware optimisation for implant segmentation using Monte Carlo Dropout and entropy-based regularisation. We first compare multiple segmentation architectures under identical training conditions and select U-Net as the baseline model based on validation performance (Dice = 0.937). Predictive entropy derived from Monte Carlo Dropout is incorporated both as a pixel-wise uncertainty measure and as a regularisation term within the training objective. By varying the entropy regularisation weight, we observe a clear trade-off between segmentation accuracy and predictive confidence. Moderate regularisation (weight = 0.05) preserves peak Dice performance while improving convergence stability, whereas stronger penalisation degrades segmentation quality. Qualitative analysis shows that predictive entropy localises near anatomically ambiguous regions, particularly along prosthesis–bone interfaces. Moreover, removing the 10 percent most uncertain pixels increases Dice on the retained regions, indicating that entropy effectively identifies pixels prone to misclassification. These results suggest that mild entropy-based regularisation can enhance segmentation reliability without sacrificing overlap accuracy in orthopaedic radiography. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.