文献速递口腔正畸

超越几何形变:ControlNet引导生成式AI实现高保真正畸侧貌合成

Beyond geometric deformation: High-fidelity orthodontic profile synthesis via ControlNet-guided generative AI.
PubMed 文献 · J Dent · 2026-07-15

本研究开发并验证了一种基于扩散模型的正畸侧貌可视化框架,在头影测量轮廓引导下合成逼真的下面部治疗结果。平均标志点误差为1.38±0.13 mm,95.0%的偏差在2.0 mm临床阈值内;图像质量指标显示高保真度(LPIPS: 0.089±0.024; SSIM: 0.951±0.009; PSNR: 28.69±2.47 dB)。

要点

  • ControlNet引导的扩散框架可合成与治疗后头影轮廓吻合的侧貌图像。
  • 平均标志点误差1.38±0.13 mm,95.0%偏差在2.0 mm临床阈值内。
  • 外行人视觉图灵测试准确率51.3%±13.0%,无法区分合成与真实图像。

核心结论

  • 所提出的ControlNet引导扩散框架能有效合成正畸可视化图像,对治疗后头影测量轮廓有良好依从性且感知逼真度高。

研究设计与样本

  • 使用Stable Diffusion修复模型,以ControlNet为条件合成下面部三分之一。

主要发现

  • 平均标志点定位误差为1.38±0.13 mm,95.0%的偏差在2.0 mm临床阈值内。
  • 图像质量指标:LPIPS为0.089±0.024,SSIM为0.951±0.009,PSNR为28.69±2.47 dB。
  • 视觉图灵测试中,外行人准确率为51.3%±13.0%(P > 0.05),正畸医生检测准确率为71.0%±9.9%(P < 0.001)。
  • 生成图像与真实图像的真实感评分经Wilcoxon符号秩检验无差异(P = 0.579)。

临床怎么用

  • 该方法可作为患者沟通和期望管理的辅助工具,将用户定义的轮廓转换为逼真的侧貌可视化图像。

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: To develop and validate a clinically oriented, diffusion-based framework for orthodontic profile visualization that synthesizes photorealistic lower-facial outcomes under explicit cephalometric contour guidance. METHODS: Retrospective paired post-treatment cephalograms and profile photographs of ten adult Asian female patients were collected. Soft-tissue profiles were traced and rigidly registered to photographs. The lower facial third was synthesized using a Stable Diffusion inpainting model conditioned by ControlNet. Performance was evaluated quantitatively via landmark localization error and image quality metrics, including Learned Perceptual Image Patch Similarity (LPIPS), Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). Subjective evaluation included a Visual Turing Test (VTT) and realism ratings by orthodontists and laypeople. RESULTS: The mean landmark error was 1.38 ± 0.13 mm, with 95.0% of deviations within the 2.0 mm clinical threshold. Image quality metrics demonstrated high fidelity (LPIPS: 0.089 ± 0.024; SSIM: 0.951 ± 0.009, PSNR: 28.69 ± 2.47 dB). In the VTT, laypeople performed at chance level (accuracy: 51.3% ± 13.0%; P > 0.05), whereas orthodontists showed higher detection accuracy (71.0% ± 9.9%; P < 0.001). Realism ratings did not differ between generated and real images using the Wilcoxon signed-rank test (P = 0.579). CONCLUSION: The proposed ControlNet-guided diffusion framework effectively synthesizes orthodontic visualizations that show favorable adherence to post-treatment cephalometric contours and high perceptual realism. CLINICAL SIGNIFICANCE: This profile synthesis approach may serve as an adjunct for patient communication and expectation management by converting user-defined contours into realistic profile visualizations.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Wang F, Lyu Y, Yang Y, et al. Beyond geometric deformation: High-fidelity orthodontic profile synthesis via ControlNet-guided generative AI. J Dent. 2026. PMID: 42457116. doi:10.1016/j.jdent.2026.106916

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。