深度学习诊断口腔潜在恶性疾患的两阶段方法
本研究开发并验证了一种用于诊断口腔潜在恶性疾患(OPMDs)的两阶段深度学习方法,在内部数据集(n=3305)训练,在两个外部多中心数据集(n=1756)评估。该方法在区分OPMDs与其他口腔黏膜病变及分类OPMD亚型上F1分数均显著优于不同年资临床医生,且AI辅助可使初级和中级医生诊断精确度分别提升10.8%和5.9%。
要点
- 两阶段深度学习模型诊断OPMDs的F1分数高于所有临床医生组。
- AI辅助使初级医生诊断精确度提升10.8%,达到高级医生水平。
- 该方法在外部多中心数据集上表现出稳健的泛化能力。
核心结论
- 两阶段深度学习方法在临床图像分类OPMDs上诊断性能可比或超过经验丰富的临床医生,并作为辅助工具显著提升初级和中级医生的诊断能力。
研究设计与样本
- 开发两阶段深度学习方法:CLA OPMD-OOML用于区分OPMDs与其他口腔黏膜病变,CLA OPMDs用于分类特定OPMD亚型。
- 在内部数据集(ZJUSS,n=3305)训练,在两个外部多中心数据集(WCHS和CS-SJTU,n=1756)评估泛化性。
- 采用盲法两步比较研究,九名临床医生(初级、中级、高级组)独立诊断后再在AI辅助下诊断。
主要发现
- 该方法显著优于所有临床医生组,CLA OPMD-OOML的F1分数为89.9%,临床医生为78.4-82.6%;CLA OPMDs的F1分数为92.2%,临床医生为77.7-80.0%。
- 在两个外部数据集上,CLA OPMD-OOML的F1分数分别为87.3%和86.9%,CLA OPMDs的F1分数为75.9-83.0%。
- AI辅助下,初级和中级临床医生的诊断精确度分别提升10.8%和5.9%,初级组性能提升至高级临床医生水平。
临床怎么用
- 该方法有望作为大规模OPMDs早期筛查的可靠工具,尤其在专科口腔医疗资源有限的地区。
- 可作为初级临床医生的培训和决策支持系统,帮助标准化诊断准确性。
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 develop and validate a two-stage deep learning method for diagnosing oral potentially malignant disorders (OPMDs). We also compared its diagnostic performance with that of clinicians at different levels of seniority and evaluated its utility as a clinical decision support tool. METHODS: A two-stage deep learning method was developed. CLA OPMD-OOML was designed to differentiate OPMDs from other oral mucosal lesions (OOML), while CLA OPMDs was designed to classify specific OPMD subtypes. The method was trained on an internal dataset (ZJUSS, n = 3305), and its generalizability was evaluated on two external multicenter datasets (WCHS and CS-SJTU, n = 1756). In a blinded, two-step comparative study, nine clinicians (junior, intermediate, and senior groups) performed diagnoses independently and then with AI assistance. RESULTS: The method significantly outperformed all clinician groups, achieving higher F1 scores in both CLA OPMD-OOML (89.9 % vs. 78.4-82.6 % for clinicians) and CLA OPMDs (92.2 % vs. 77.7-80.0 % for clinicians). It showed robust generalizability on the two external datasets, with F1 scores of 87.3 % and 86.9 % for CLA OPMD-OOML and 75.9-83.0 % for CLA OPMDs. With AI assistance, the diagnostic precision of junior and intermediate clinicians increased by 10.8 % and 5.9 %, respectively, raising the junior group's performance to the level of senior clinicians. CONCLUSIONS: The developed two-stage deep learning method demonstrated diagnostic performance comparable to or exceeding that of experienced clinicians in classifying OPMDs from clinical images. It functions as a powerful assistive tool that substantially enhances the diagnostic capabilities of junior and intermediate clinicians. CLINICAL SIGNIFICANCE: This AI method has the potential to serve as a reliable tool for large-scale early screening of OPMDs, particularly in regions with limited access to specialist dental care. It can also serve as a valuable training and decision support system for junior clinicians, helping to standardize diagnostic accuracy and improve patient outcomes through timely and accurate detection. A supplemental appendix to this article is available online.
引用
- Cao YQ, Zhang JY, Lu M, et al. A deep learning method for diagnosis of oral potentially malignant disorders. J Dent. 2025. PMID: 41016582. doi:10.1016/j.jdent.2025.106138
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。