合成数据集在牙髓病学中的机遇、挑战与未来方向:范围综述
多数研究使用生成对抗网络(GAN),全景片最常见;合成数据单独保真度中等,与真实图像结合可提升多种任务准确率。
要点
- 仅4项评估临床效用,合成+真实图像可提升分类分割等任务准确率。
核心结论
- 合成数据在牙髓病学应用仍处早期但潜力较大,未来需构建多样且具临床代表性的数据集。
主要发现
- 合成影像模态以全景片最常见(n=4),口内临床图像和侧位头影测量片各n=2,根尖片、咬翼片和颞下颌关节MRI各n=1。
- 评估指标以Fréchet Inception Distance(FID)、信噪比(SNR)和视觉图灵测试最常见。
- 仅4项研究评估临床效用,合成数据单独保真度中等,与真实图像结合在根管分类、种植体分割、牙位识别和骨性模式识别等任务中持续提升准确率。
临床怎么用
- 合成数据可作为真实影像的补充用于训练AI模型,尤其在数据稀缺时可能改善诊断相关任务表现。
局限性
- 合成数据单独使用时保真度仅为中等。
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
BACKGROUND: Recent advances in Artificial Intelligence (AI) and machine learning algorithms have transformed diagnostic tasks in dentistry and endodontics. However, the development of robust AI systems remains constrained by the scarcity of annotated, high-quality datasets along with issues like class imbalance and privacy concerns. Synthetic data, replicating the statistical and visual characteristics of real data, may offer a promising solution to these limitations. OBJECTIVE: The scoping review aimed to map (i) existing synthetic datasets in dentistry/endodontics, (ii) commonly used synthetic data generation techniques, (iii) their fidelity and diagnostic utility, and (iv) current limitations and future research directions in endodontics. METHODS: A comprehensive search was conducted across PubMed, Scopus, EBSCO, and Google Scholar using keywords related to synthetic datasets in dentistry and endodontics. Data regarding synthetic image modality most used, model architecture, training dataset size, evaluation metrics, and the real-world clinical utility (use-cases) of synthetic datasets generated were extracted and analysed descriptively. RESULTS: Eleven studies met the inclusion criteria, primarily using Generative Adversarial Networks (GANs), with only one utilising a hybrid GAN-diffusion model. The most common synthetic imaging modalities were panoramic radiographs (n = 4), intraoral clinical images and lateral cephalograms (n = 2 each), followed by periapical, bitewing, and TMJ Magnetic Resonance Imaging (MRI) (n = 1 each). Training datasets ranged from 509 to 35 254 images, with stable generation achieved using as few as 1720 images in the hybrid model. Evaluation metrics varied, with Fréchet Inception Distance (FID), Signal-to-Noise Ratio (SNR), and visual Turing tests being most common. Only four studies evaluated clinical utility, using synthetic images to train AI models. While synthetic data alone showed moderate fidelity, combining it with real images consistently improved accuracy across tasks like canal classification, implant segmentation, tooth numbering, and skeletal pattern recognition. CONCLUSION: Synthetic data use in endodontics is still emerging but holds strong potential. While GANs dominate current research, diffusion and hybrid models may yield more realistic images. Future efforts should focus on generating diverse, clinically representative datasets to improve diagnostics, education, and AI-driven endodontic care.
引用
- Naved N, Umer F. A Scoping Review on Synthetic Datasets: Opportunities, Challenges and Future Directions for Endodontics. Int Endod J. 2026. PMID: 41709715. doi:10.1111/iej.70116
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。