文献速递牙体牙髓病口腔修复与种植口腔颌面外科

可控制的全景放射片生成模型PRGen

Controllable Panoramic Radiograph Synthesis Using a Generative Model.
PubMed 文献 · J Dent Res · 2026-07-18

本研究开发了疾病引导、解剖可控的生成模型PRGen,使用50,127对文本-图像样本训练,可从文本描述合成逼真的全景放射片及配对掩码。将合成图像加入训练使分割Dice提升47.59%、AUC提升11.53%,超过82%的合成图像经10名放射科医生评估忠实反映所描述疾病。外部多中心验证显示分割Dice平均提升25.58%。

要点

  • PRGen可从文本疾病描述生成带掩码注释的逼真全景放射片,减少人工标注依赖。
  • 合成图像加入训练使分割Dice提升47.59%、AUC提升11.53%。
  • 外部多中心验证分割Dice平均提升25.58%,泛化性良好。

核心结论

  • PRGen能够生成高质量、带掩码注释的全景放射片,降低对人工标注的依赖,支持更可靠的自动化分析。

研究设计与样本

  • 开发生成模型PRGen,使用50,127对文本-图像样本进行训练。
  • 在留出的内部测试集、公开数据集及10名放射科医生的评估中综合评价生成图像的质量与临床效用。
  • 进行外部多中心验证以评估泛化性。

主要发现

  • 将PRGen合成图像纳入训练使分割Dice分数提升47.59%,AUC提升11.53%。
  • 专家评估中超过82%的合成全景放射片被判定忠实反映所描述疾病,并保持临床逼真的解剖结构和放射学表现。
  • 外部多中心验证显示分割Dice平均提升25.58%,诊断任务中亦有持续增益。

临床怎么用

  • 可将PRGen合成图像用于扩充训练数据,缓解数据稀缺、隐私限制和标注不平衡问题。
  • 合成图像带配对掩码,可用于分割和诊断模型的开发与验证。

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

Panoramic radiography (PR) is the one of the most widely prescribed diagnostic imaging modalities in dentistry. Achieving clinical-level automated interpretation of PR is critical for improving diagnostic efficiency, reducing radiologist workload, and ensuring interpretive consistency. Artificial intelligence (AI) has demonstrated significant utility and promise in PR interpretation, improving diagnostic efficiency, reducing radiologists' workload, and enhancing interpretive consistency. However, developing AI-based PR interpretation methods remains hindered by persistent challenges of data scarcity, privacy constraints, and annotation imbalance. Facing these limitations, the aim of this study is to develop a disease-guided, anatomy-controllable generative model, named PRGen (for "panoramic radiograph generation"), designed to mitigate challenges related to data scarcity, privacy constraints, and annotation imbalance. PRGen is developed using 50,127 paired text-image samples and overcomes the uncontrollability of existing generative models, synthesizing both realistic PRs and paired masks from textual disease descriptions, enabling precise control of dental anatomy through simple or detailed sketches, and effectively mitigating barriers to large-scale, well-annotated dataset construction. We comprehensively evaluated the quality and clinical utility of PRGen-generated PRs across held-out internal test sets, publicly available datasets, and assessments by 10 radiologists. Incorporating PRGen-synthesized images into training led to a 47.59% improvement in Dice score for segmentation and a 11.53% increase in the area under the curve. Upon expert review, more than 82% of synthesized PRs were judged to faithfully reflect the described diseases while maintaining clinically realistic anatomical structures and radiographic appearances. External multicenter validation confirmed robust generalizability, with an average Dice improvement of 25.58% for segmentation and consistent gains in diagnostic tasks. Collectively, these results demonstrate that PRGen enables the generation of high-quality, mask-annotated PRs, thereby reducing the reliance on manual annotations and supporting more reliable and automated analysis of PRs, ultimately facilitating both model development and clinical translation.

来源:J Dent Res,查看原文。版权归原作者所有。

引用

  1. Lin YY, Fu WT, Guo XL, et al. Controllable Panoramic Radiograph Synthesis Using a Generative Model. J Dent Res. 2026. PMID: 42470285. doi:10.1177/00220345261450592

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。