文献速递口腔修复与种植口腔正畸

AI驱动牙冠生成:下颌第一磨牙点云补全模型对比分析

AI-driven crown generation: A comparative analysis of point cloud completion models for mandibular first molar restoration.
PubMed 文献 · J Dent · 2026-02-16

本研究将PF-Net、PCN和PoinTr三种点云补全模型适配于下颌第一磨牙牙冠生成,使用120例口内扫描数据,以36缺失的部分牙列模型为输入、天然牙冠为真值进行训练与评估。结果显示PoinTr在几何精度和临床适用性上表现最佳,近远中径偏差最低(0.14±0.33 mm),表面误差小,且兼顾效率(5.52秒/牙冠)。

要点

  • PoinTr生成牙冠的近远中径偏差最低(0.14±0.33 mm),几何精度优于PF-Net和PCN。
  • PF-Net生成速度最快(3.22秒/牙冠)但形态偏差大于6 mm,临床不可接受。
  • PoinTr兼顾精度与效率(5.52秒/牙冠),可辅助种植规划中的修复体设计。

核心结论

  • PoinTr生成的牙冠对天然牙形态保真度最佳且尺寸控制精确,符合修复引导种植原则。
  • 将其整合到数字化工作流程可提升手术规划精度与效率,并减少对技师的依赖。

研究设计与样本

  • 采用120例非病理性牙列的口内扫描数据,处理为3D点云。
  • 以36缺失的部分牙列模型为输入,对应天然牙冠为真值。
  • 数据集划分为训练集90例、验证集20例、测试集10例。
  • 三种AI模型(PF-Net、PCN、PoinTr)使用Chamfer Distance损失进行相同训练。
  • 评估几何保真度、临床适用性(颊舌径/近远中径)和计算效率,统计方法为ANOVA/Tukey HSD或Kruskal-Wallis/Dunn检验(α=0.05)。

主要发现

  • PoinTr的近远中径偏差最低(0.14±0.33 mm,p=0.008 vs. PCN)。
  • PoinTr表面误差最小(Chamfer Distance:0.10 mm;Hausdorff Distance:0.22 mm)。
  • PF-Net生成时间最快(3.22秒/牙冠),但形态偏差大于6 mm,临床不可接受。
  • PoinTr在所有几何指标上显著优于PF-Net和PCN(p<0.05),生成时间为5.52秒/牙冠。

关键数值与推荐

  • PoinTr近远中径偏差0.14±0.33 mm(p=0.008 vs. PCN)。
  • PoinTr Chamfer Distance为0.10 mm,Hausdorff Distance为0.22 mm。
  • PF-Net生成时间3.22秒/牙冠,PoinTr生成时间5.52秒/牙冠。
  • 统计显著性水平α=0.05。

临床怎么用

  • PoinTr可用于修复引导种植手术规划,生成高形态保真度和尺寸精度的虚拟牙冠。
  • 生成的牙冠可直接指导最佳种植体定位和手术导板设计。
  • 该方法也可扩展至传统固定修复中牙支持式牙冠的自动化设计,提升数字化修复的准确性和一致性。

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

OBJECTIVES: This study aimed to adapt and evaluate three established artificial intelligence (AI)-driven point cloud completion models-PF-Net, PCN, and PoinTr-for generating anatomically accurate dental crowns for mandibular first molars, thereby advancing prosthetically guided implant surgery planning. METHODS: A dataset of 120 intraoral scans from non-pathological dentitions was processed into 3D point clouds. Partial dentition models (with tooth 36 missing) served as inputs, while the corresponding natural crowns served as the ground truth. The dataset was partitioned into training (90), validation (20), and test (10) sets. All models were trained identically using Chamfer Distance loss and evaluated for geometric fidelity (Chamfer Distance, Hausdorff Distance, Root Mean Square Error, Mean Negative Deviation), clinical applicability (buccolingual/mesiodistal diameters), and computational efficiency. Statistical analyses included ANOVA/Tukey's HSD or Kruskal-Wallis/Dunn's tests (α = 0.05). RESULTS: PoinTr demonstrated superior clinical efficacy and geometric accuracy, exhibiting the lowest mesiodistal diameter deviation (0.14 ± 0.33 mm, p = 0.008 vs. PCN) and minimal surface errors (Chamfer Distance: 0.10 mm; Hausdorff Distance: 0.22 mm). PF-Net exhibited the fastest generation time (3.22 s/crown) but clinically unacceptable morphological deviations (>6 mm). PoinTr balanced accuracy (significantly outperforming PF-Net/PCN in all geometric metrics, p < 0.05) and efficiency (5.52 s/crown). CONCLUSION: PoinTr generates crowns with optimal morphological fidelity to natural teeth and precise dimensional control, enabling virtual prostheses that align with prosthetically guided implant principles. Its integration into digital workflows can enhance surgical planning precision and efficiency while reducing technician dependency. CLINICAL SIGNIFICANCE: The proposed PoinTr-based approach provides a clinically significant tool for prosthetically guided implant surgery by generating crowns with high anatomical fidelity and dimensional precision, directly informing optimal implant positioning and surgical guide design. Furthermore, the methodology extends beyond implant planning, offering an efficient and automated solution for designing tooth-supported crowns in conventional fixed prosthodontics, thereby enhancing accuracy, consistency, and workflow efficiency in digital restorative dentistry.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Shu Z, Tian M, Wei H, et al. AI-driven crown generation: A comparative analysis of point cloud completion models for mandibular first molar restoration. J Dent. 2026. PMID: 41707826. doi:10.1016/j.jdent.2026.106581

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。