Improving orbital bone segmentation with diffusion models and consensus-based refinement in facial CT images

Accurate segmentation of orbital bones in facial computed tomography (CT) images is essential for craniomaxillofacial surgery planning and the creation of bone implants. However, it has challenging issues that thin bones of orbital medial wall and floor are difficult to segment due to their ambiguous boundaries and low contrast with surrounding soft tissues. Furthermore, this issue leads to inter-observer variability in manual annotation masks. In this paper, we propose a novel segmentation framework based on a conditional diffusion model with consensus-driven correction. The framework consists of three main components: conditional diffusion model-based segmentation, consensus-driven accumulation map generation, and context-aware consensus correction. The conditional diffusion model leverages diverse annotation masks to generate multiple plausible segmentation results, addressing the inter-observer variability associated with manual annotations. These results are aggregated into a consensus-driven accumulation map, which captures the agreement among possible segmentations, offering a robust alternative to simple averaging. Finally, the segmentation is refined through context-aware consensus correction, which integrates consensus information with CT image features, considering spatial and intensity-based characteristics. Experimental results show the effectiveness of the proposed method, achieving Dice Similarity Coefficients (DSCs) of 84.38% and 90.37% and precisions of 88% and 92.28% for the medial wall and floor, respectively. Compared to CNN-based methods, the proposed framework improves precision by up to 4.74% and 4.49%, significantly reducing false positives while preserving the continuity of thin structures.