基于锥形束CT作为参考标准评估商用人工智能平台在根尖片上判断牙髓治疗结局的诊断准确性
这项回顾性诊断准确性研究分析了4项前瞻性临床试验中的376颗牙(860个根),以锥形束CT为参考标准,比较人工智能平台(Diagnocat)与2名校准牙髓专科医生在根尖片上评估牙髓治疗结局的表现。人工智能平台在牙和根水平的敏感性均更高但特异性均更低,牙水平总体准确性相当而根水平略低,作者认为其可作为辅助工具但需临床监督以防过度诊断。
要点
- 人工智能平台牙水平敏感性67.3%高于临床医生的49.3%,特异性82.3%低于临床医生的92.5%。
- 根水平敏感性人工智能平台54.3%高于临床医生的43.8%,特异性86.7%低于临床医生的94.5%。
- 总体准确性牙水平人工智能平台76.3%与临床医生75.3%相当,根水平人工智能平台78.5%略低于临床医生的81.6%。
核心结论
- 人工智能平台有潜力作为评估牙髓治疗结局的辅助工具,尤其在检测可能被人为评估遗漏的病变方面,但其较低的特异性提示需要临床监督以防止过度诊断。
研究设计与样本
- 回顾性诊断准确性研究,分析4项前瞻性临床试验中的376颗牙(860个根)。
- 由2名校准牙髓专科医生和人工智能平台独立评估根尖片上的治疗结局,以锥形束CT扫描为参考标准。
- 计算敏感性、特异性、准确性和受试者工作特征曲线下面积。
主要发现
- 牙水平:人工智能平台敏感性67.3% vs 临床医生49.3%(P < .001),特异性82.3% vs 92.5%(P < .001)。
- 根水平:人工智能平台敏感性54.3% vs 临床医生43.8%(P = .003),特异性86.7% vs 94.5%(P < .001)。
- 总体准确性牙水平:人工智能平台76.3% vs 临床医生75.3%(P = .716);根水平:人工智能平台78.5% vs 临床医生81.6%(P = .021)。
- 受试者工作特征曲线下面积:牙水平人工智能平台0.75 vs 临床医生0.71,根水平人工智能平台0.71 vs 临床医生0.69。
临床怎么用
- 作者认为人工智能平台有潜力作为评估牙髓治疗结局的辅助工具,尤其在检测可能被人为评估遗漏的病变方面。
- 作者指出其较低的特异性提示需要临床监督以防止过度诊断。
局限性
- 原文未报告局限性。
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
INTRODUCTION: Artificial intelligence (AI) has shown promise in dental diagnostics; however, its accuracy in assessing endodontic treatment outcomes compared to experienced clinicians remains unclear. This study evaluated the performance of an AI-driven platform (Diagnocat) against experienced clinicians in assessing endodontic treatment outcomes on periapical radiographs, using cone-beam computed tomography as the reference standard. METHODS: This retrospective diagnostic accuracy study analyzed 376 teeth (860 roots) from 4 prospective clinical trials. Treatment outcomes were assessed using periapical radiographs, independently evaluated by 2 calibrated endodontists and the AI-driven platform. Cone-beam computed tomography scans served as the reference standard. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve were calculated. RESULTS: The AI-driven platform demonstrated higher sensitivity but lower specificity than clinicians at both tooth (sensitivity: 67.3% vs 49.3%, P < .001; specificity: 82.3% vs 92.5%, P < .001) and root levels (sensitivity: 54.3% vs 43.8%, P = .003; specificity: 86.7% vs 94.5%, P < .001). Overall accuracy was comparable at the tooth level (AI: 76.3%, clinicians: 75.3%, P = .716) but slightly lower for the AI-driven platform at the root level (78.5% vs 81.6%, P = .021). Receiver operating characteristic curve analysis showed comparable area under the curve values between the AI-driven platform and clinicians at both tooth (0.75 vs 0.71) and root levels (0.71 vs 0.69). CONCLUSIONS: While the AI-driven platform demonstrated potential as an adjunctive tool for assessing endodontic treatment outcomes, particularly in detecting lesions that might be missed by human assessment, its lower specificity highlights the need for clinical oversight to prevent overdiagnosis.
引用
- Allihaibi M, Koller G, Mannocci F, et al. Diagnostic Accuracy of a Commercial AI-based Platform in Evaluating Endodontic Treatment Outcomes on Periapical Radiographs Using CBCT as the Reference Standard. J Endod. 2025. PMID: 40180318. doi:10.1016/j.joen.2025.03.007
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。