人工智能在影像学图像上检测牙周炎的诊断准确性:系统综述与诊断试验准确性meta分析
患者水平汇总敏感度0.93、特异度0.88、SROC AUC 0.96;牙水平汇总敏感度0.90、特异度0.94、SROC AUC 0.97。AI模型诊断准确性优异,但需前瞻性研究、外部验证和透明报告后方可常规整合。
要点
- AI模型在患者水平检测牙周炎的汇总敏感度0.93、特异度0.88、SROC AUC 0.96。
- 牙水平汇总敏感度0.90、特异度0.94、SROC AUC 0.97。
- 算法架构、成像方式和外部验证均非异质性来源,但需前瞻性研究验证。
核心结论
- AI模型对牙周炎检测表现出优异的诊断准确性,但常规临床应用前需设计良好的前瞻性研究、稳健的临床参考标准、外部验证和透明报告。
- 当前证据基础对中低收入国家代表性不足,而这些地区牙周炎负担最重且AI辅助筛查的公共卫生效益可能最大。
研究设计与样本
- 系统综述与诊断试验准确性meta分析,检索PubMed、Web of Science和Cochrane Library至2026年4月。
- 采用双变量随机效应模型合并敏感度和特异度,用meta回归探索异质性。
主要发现
- 患者水平汇总敏感度0.93(95%CI 0.86-0.97),汇总特异度0.88(95%CI 0.77-0.94),SROC曲线下面积AUC为0.96(95%CI 0.94-0.97)。
- 牙水平汇总敏感度0.90(95%CI 0.83-0.94),汇总特异度0.94(95%CI 0.91-0.96),SROC曲线下面积AUC为0.97(95%CI 0.96-0.98)。
- meta回归分析表明算法架构、成像方式或外部验证均非异质性来源。
临床怎么用
- AI辅助影像学筛查可能尤其惠及全科牙医和基层或资源有限环境中的非专科临床医生,这些场景牙周专科医生可及性差且诊断变异性最大。
- 此类工具可在椅旁数秒内标记疑似中重度牙周炎,加速转诊和分诊决策,使专科医生能更早进行确定性检查和制定治疗计划,并减少常规牙科就诊中晚期病变患者的漏诊。
局限性
- 需设计良好的前瞻性研究、稳健的临床参考标准、外部验证和透明报告后方可负责任地将这些工具整合到常规牙科诊疗中。
- 当前证据基础对中低收入国家代表性不足,而这些地区牙周炎负担最重且AI辅助筛查的公共卫生效益可能最大。
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: To determine the diagnostic accuracy of artificial intelligence (AI) models for the periodontitis on radiographic images, and to investigate the factors influencing diagnostic performance. METHODS: PubMed, Web of Science, and Cochrane Library were searched until April 2026. A bivariate random-effects model was used to synthesise sensitivity and specificity, and meta-regression explored heterogeneity. RESULTS: Fourteen studies comprising 50,080 radiographic images were included. Seven studies performed diagnosis at the patient level and seven at the tooth or image level. At the patient level, the pooled sensitivity was 0.93 (95%CI, 0.86-0.97), pooled specificity was 0.88 (0.77-0.94), and the area under each curve (AUC) of summary receiver operating characteristic curves (SROC) was 0.96 (0.94-0.97). At the tooth level, pooled sensitivity was 0.90 (0.83-0.94), pooled specificity was 0.94 (95%CI, 0.91-0.96), and the AUC of SROC was 0.97 (0.96-0.98). The meta-regression analysis indicates that none of the algorithm architecture, imaging modality, or external validation are sources of heterogeneity. CONCLUSIONS: AI models demonstrate excellent diagnostic accuracy for periodontitis detection. Well-designed prospective studies with robust clinical reference standards, external validation, and transparent reporting are urgently required before these tools can be responsibly integrated into routine dental care. The current evidence base underrepresents low- and middle-income countries, where the burden of periodontitis is greatest and AI-assisted screening could offer the greatest public health benefit. CLINICAL SIGNIFICANCE: AI-assisted radiographic screening may particularly benefit general dental practitioners and non-specialist clinicians in primary care or resource-limited settings, where access to periodontists is scarce and diagnostic variability is greatest. By flagging suspected moderate-to-severe periodontitis at the chairside within seconds, such tools could accelerate referral and triage decisions, enabling earlier definitive examination and treatment planning by specialists, while reducing the number of patients with advanced disease who are overlooked during routine dental visits.
引用
- Zhang X, Liu Z, Zhang Y, et al. Artificial intelligence for detection of periodontitis on radiographic images: a systematic review and diagnostic test accuracy meta-analysis. J Dent. 2026. PMID: 42753854. doi:10.1016/j.jdent.2026.107054
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。