Artificial intelligence helps diagnose oral potentially malignant disorders

Oral potentially malignant disorders (OPMDs) can lead to oral cancer, which is one of the most common cancers worldwide. Prevention is crucial in the avoidance of malignant transformations of OPMDs. Artificial intelligence (AI) provides a new and non-invasive tool for analyzing medical data, such as patient data, radiological images and clinical photographs. These AI-based tools can help in the decision-making process. However, histological examination is still the gold standard for diagnosing OPMDs. Our research aimed to investigate the diagnostic accuracy of artificial intelligence on intraoral photographs of patients with OPMDs. On November 10, 2023, a systematic search was conducted on five major databases: MEDLINE, Embase, Cochrane Library, Scopus, and Web of Science. From the available data, we performed a quantitative analysis for sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), diagnostic odds ratio (DOR), positive likelihood ratio (LR+), and negative likelihood ratio (LR-) and calculated their 95% confidence intervals (CI). Our search resulted in 6 eligible articles, in which 898 images out of a total of 4,046 were tested using AI-based architectures. In five of the six studies, at least two AI models were investigated. Among the eligible articles, two independent authors (DH, ÁF) collected data on decision making in the OPMD classification of the artificial intelligence-supported tool and filled out a customized data extraction sheet designed for this review. Our data analysis units covered sensitivity, specificity, and confusion matrix. In addition, the following information was collected: first author, publication year, study period, teaching-, validation-, testing sample size, teaching settings (epochs, batch size, and learning rate), training methods, and augmentation types. Information (number of layers and parameters) on the AI-based models used was collected from the internet if not provided in the article. Some data about the tested AI-based architectures were collected from the internet if not included in the article. The sources for this information were https://github.com and https://pytorch.org.口腔潜在恶性疾患（Oral Potentially Malignant Disorders, OPMDs）可进展为口腔癌，而口腔癌是全球最常见的恶性肿瘤之一。针对OPMDs的恶性转化进行预防，是防控该类疾病的关键环节。人工智能（Artificial Intelligence, AI）为患者数据、放射影像、临床照片等医学数据分析提供了全新的非侵入式工具，这类基于AI的工具可辅助临床决策流程的制定。但组织病理学检查仍是OPMDs诊断的金标准。本研究旨在评估人工智能基于OPMDs患者口内照片实现诊断的准确性。2023年11月10日，研究团队针对MEDLINE、Embase、Cochrane图书馆、Scopus及Web of Science五大主流数据库开展了系统检索。基于检索获得的有效数据，本研究对灵敏度、特异度、阳性预测值（Positive Predictive Value, PPV）、阴性预测值（Negative Predictive Value, NPV）、诊断比值比（Diagnostic Odds Ratio, DOR）、阳性似然比（Positive Likelihood Ratio, LR+）及阴性似然比（Negative Likelihood Ratio, LR-）进行了定量分析，并计算了各指标的95%置信区间（95% Confidence Interval, CI）。本次检索共纳入6篇符合标准的文献，其中4046张图像中有898张通过基于AI的模型架构完成了测试。6项研究中有5项至少评估了2种AI模型。在纳入的文献中，由两名独立研究者（DH、ÁF）分别提取了AI辅助工具用于OPMD分类的决策相关数据，并填写了为本综述定制的数据提取表。本研究的数据分析维度涵盖灵敏度、特异度及混淆矩阵。此外，本研究还提取了以下信息：第一作者、发表年份、研究周期、训练样本量、验证样本量、测试样本量、训练设置（训练轮次、批次大小及学习率）、训练方法及数据增强类型。若文献未提供所用AI模型的相关信息（层数与参数数量），则通过互联网进行补充收集；若文献未提及所测试的AI架构相关数据，亦通过互联网进行补充收集，信息来源为https://github.com与https://pytorch.org。