Anatomical prior-guided implicit segmentation of challenging thin and tubular anatomical structures in CBCT images.
Authors
Affiliations (5)
Affiliations (5)
- School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai, China. Electronic address: [email protected].
- School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai, China.
- School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai, China; Department of Computer Science, The University of Hong Kong, Hong Kong, China.
- Department of orthodontics, School and Hospital of Stomatology, Peking University, Beijing, China.
- School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai, China. Electronic address: [email protected].
Abstract
Medical image segmentation has achieved promising performance with deep learning approaches like nnU-Net. However, extremely precise segmentation of challenging topological structures, particularly thin anatomical structures, remains an unsolved problem. In digital dentistry, cortical bone in the mandible and maxillary regions and the inferior alveolar nerve (IAN) represent such thin structures and tubular topologies, where accurate segmentation from Cone-beam Computed Tomography (CBCT) scans is crucial for treatment planning. To address this challenge, we propose an anatomy-guided coarse-to-fine framework. After an initial coarse CNN-based stage, we apply an implicit refinement function to enhance segmentation, focusing on uncertain regions through point-wise selection. To guide this refinement, we leverage the Signed Distance Field (SDF) of teeth as an anatomical prior, exploiting their stable spatial relationship with the target structures. Extensive evaluations on challenging dental CBCT datasets, including both cortical bone and IAN segmentation, demonstrate our approach outperforms state-of-the-art methods across metrics. In particular, it achieves over 30% improvements in distance error metrics (ASD and HD). These results validate its robustness and generalization for precise thin structure segmentation in dental CBCT images. Our code is available at https://github.com/ShanghaiTech-IMPACT/Anatomical-Prior-guided-Implicit-Segmentation-of-Thin-Anatomical-Structures-in-CBCT-image.