CRANIOSEG：基于深度学习的锥形束计算机断层扫描颅颌面解剖结构自动分割

文献信息

中文摘要

目的： 开发并评估一种基于nnU-Net v2的深度学习模型，用于在锥形束计算机断层扫描（CBCT）上对27个颅颌面解剖结构进行全自动多类别分割。
方法： 这项回顾性研究纳入了106名成年患者的CBCT扫描。采用分层标注工作流程，在3D Slicer中手动分割27个解剖结构，以建立共识参考标准。数据集在患者层面被划分为训练集（n = 96）和留出的内部测试集（n = 10）。训练并评估了一个三维全分辨率nnU-Net v2模型，评估指标包括Dice相似系数（DSC）、Jaccard指数（IoU）、精确率、召回率、第95百分位Hausdorff距离（HD95）以及平均对称表面距离（ASSD）。
结果： 在所有27个解剖结构中，该模型达到了平均DSC为0.850 ± 0.132、平均IoU为0.773 ± 0.159、平均精确率为0.863、平均召回率为0.868、平均HD95为2.34 mm。类别层面表现最佳的是鼻旁窦（DSC 0.961）和颅面骨性框架（DSC 0.961），而颅底裂隙和窝的表现最低（DSC 0.694），主要原因是翼腭窝分割困难（DSC 0.377）。大型、边界清晰的结构，包括下颌骨（DSC 0.976）和上颌窦（DSC 0.977），分割精度极佳，而较小、低对比度的结构如下颌管（DSC 0.654）仍更具挑战性。
结论： 本研究证明，在选定的内部数据集上，使用单一nnU-Net v2模型从CBCT图像中同时分割27个颅颌面解剖结构在技术上是可行的。总体而言，该模型取得了较强的分割性能，对于下颌骨、上颌窦、蝶窦和上颅骨等大型、边界清晰的结构，重叠值尤其高。不同结构之间的性能存在差异，较低的值主要见于若干较小、低对比度且形态复杂的结构。

英文摘要

OBJECTIVE: To develop and evaluate an nnU-Net v2-based deep-learning model for fully automated multiclass segmentation of 27 craniomaxillofacial anatomical structures on cone-beam computed tomography (CBCT).
METHODS: This retrospective study included CBCT scans from 106 adult patients. Twenty-seven anatomical structures were manually segmented in 3D Slicer to establish a consensus reference standard using a hierarchical annotation workflow. The dataset was divided at the patient level into training (n = 96) and held-out internal test (n = 10) sets. A three-dimensional full-resolution nnU-Net v2 model was trained and evaluated using the Dice similarity coefficient (DSC), Jaccard index (IoU), precision, recall, the 95th percentile Hausdorff distance (HD95), and the average symmetric surface distance (ASSD).
RESULTS: Across all 27 anatomical structures, the model achieved a mean DSC of 0.850 ± 0.132, a mean IoU of 0.773 ± 0.159, a mean precision of 0.863, a mean recall of 0.868, and a mean HD95 of 2.34 mm. The highest category-level performance was observed for the paranasal sinuses (DSC 0.961) and the craniofacial osseous framework (DSC 0.961), whereas the skull-base fissures and fossae showed the lowest performance (DSC 0.694), primarily because of the difficulty in segmenting the pterygopalatine fossa (DSC 0.377). Large, well-defined structures, including the mandible (DSC 0.976) and maxillary sinus (DSC 0.977), were segmented with excellent accuracy, whereas smaller, low-contrast structures such as the mandibular canal (DSC 0.654) remained more challenging.
CONCLUSIONS: This study demonstrates the technical feasibility of simultaneously segmenting 27 craniomaxillofacial anatomical structures from CBCT images using a single nnU-Net v2 model on a selected internal dataset. Overall, the model achieved strong segmentation performance, with particularly high overlap values for large, well-defined structures such as the mandible, maxillary sinus, sphenoid sinus, and upper skull. Performance varied among structures, with lower values observed mainly for several small, low-contrast, and morphologically complex structures.