文献速递口腔颌面外科口腔修复与种植

深度学习用于颞下颌关节自动分割

Deep learning for automated segmentation of the temporomandibular joint.
PubMed 文献 · J Dent · 2023-03-02

本研究开发并验证基于3D U-net的深度学习算法,在154例CBCT数据集上训练,用于自动分割下颌髁突和关节窝。测试集8例CBCT显示,AI分割髁突和关节窝的IoU分别为0.955和0.935,耗时3.6秒,显著优于两位观察者的人工分割精度和速度。结论认为该工具具有高精度、高速度和高一致性,但存在泛化性风险。

要点

  • AI分割髁突IoU 0.955、关节窝IoU 0.935,高于两位观察者手动分割的0.895和0.928(p<0.05)。
  • AI分割平均耗时3.6秒,显著快于观察者的378.9秒和571.6秒(p<0.001)。
  • 算法仅在正颌手术患者、单一CBCT扫描仪的数据上训练,泛化性风险不能排除。

核心结论

  • 基于AI的自动分割工具分割下颌髁突和关节窝具有高精度、高速度和高一致性。

研究设计与样本

  • 开发基于3D U-net的三步深度学习算法,使用154例手动分割的CBCT图像进行训练和验证,测试集为8例CBCT,由两位独立观察者和AI算法分别分割。

主要发现

  • AI分割髁突的IoU为0.955,关节窝的IoU为0.935。
  • 两位观察者手动分割髁突的IoU分别为0.895和0.928,均低于AI(p<0.05)。
  • AI分割平均耗时3.6秒(SD 0.9),观察者1耗时378.9秒(SD 204.9),观察者2耗时571.6秒(SD 257.4),AI显著快于观察者(p<0.001)。

临床怎么用

  • 将AI分割工具整合到诊断软件中,可促进临床环境下颞下颌关节的3D定性和定量分析,尤其用于颞下颌关节紊乱病的诊断和纵向随访。

局限性

  • 算法仅在正颌手术患者的扫描数据上训练,且仅来自一种CBCT扫描仪,潜在有限的鲁棒性和泛化性风险不能排除。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVE: Quantitative analysis of the volume and shape of the temporomandibular joint (TMJ) using cone-beam computed tomography (CBCT) requires accurate segmentation of the mandibular condyles and the glenoid fossae. This study aimed to develop and validate an automated segmentation tool based on a deep learning algorithm for accurate 3D reconstruction of the TMJ. MATERIALS AND METHODS: A three-step deep-learning approach based on a 3D U-net was developed to segment the condyles and glenoid fossae on CBCT datasets. Three 3D U-Nets were utilized for region of interest (ROI) determination, bone segmentation, and TMJ classification. The AI-based algorithm was trained and validated on 154 manually segmented CBCT images. Two independent observers and the AI algorithm segmented the TMJs of a test set of 8 CBCTs. The time required for the segmentation and accuracy metrics (intersection of union, DICE, etc.) was calculated to quantify the degree of similarity between the manual segmentations (ground truth) and the performances of the AI models. RESULTS: The AI segmentation achieved an intersection over union (IoU) of 0.955 and 0.935 for the condyles and glenoid fossa, respectively. The IoU of the two independent observers for manual condyle segmentation were 0.895 and 0.928, respectively (p<0.05). The mean time required for the AI segmentation was 3.6 s (SD 0.9), whereas the two observers needed 378.9 s (SD 204.9) and 571.6 s (SD 257.4), respectively (p<0.001). CONCLUSION: The AI-based automated segmentation tool segmented the mandibular condyles and glenoid fossae with high accuracy, speed, and consistency. Potential limited robustness and generalizability are risks that cannot be ruled out, as the algorithms were trained on scans from orthognathic surgery patients derived from just one type of CBCT scanner. CLINICAL SIGNIFICANCE: The incorporation of the AI-based segmentation tool into diagnostic software could facilitate 3D qualitative and quantitative analysis of TMJs in a clinical setting, particularly for the diagnosis of TMJ disorders and longitudinal follow-up.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Vinayahalingam S, Berends B, Baan F, et al. Deep learning for automated segmentation of the temporomandibular joint. J Dent. 2023. PMID: 36870441. doi:10.1016/j.jdent.2023.104475

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。