Clinical image-based artificial intelligence models for early detection and classification of oral cancer: A systematic review and meta-analysis

Abstract

Oral squamous cell carcinoma (OSCC) and oral potentially malignant disorders (OPMDs) are major challenges in global health. Convolutional neural network (CNN)-based deep learning and artificial intelligence (AI) models have demonstrated superior results when used automatically. However, no prior meta-analysis examines the pooled diagnostic performance of AI models. An electronic search was performed in MEDLINE via PubMed, Scopus, Web of Science, Springer Link, and Elsevier up to 30 June 2025, without a period restriction. The articles that investigated OSCC and OPMDs using AI and CNN-based deep learning on clinicopathological images were included for review. Quality assessment of articles was performed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). A bivariate random-effects model was used to estimate the pooled sensitivity and specificity, I² values, receiver-operating characteristic (ROC) curves, and diagnostic odds ratios (DOR) were calculated using RevMan 5.4 and Open Meta-Analyst for Sierra software. Publication bias was evaluated using Deeks’ funnel plot. Only ten studies, totalling more than 47,124 clinical photos, satisfied the inclusion requirements out of the 175 original records that were found. Transfer learning is a common training method, and the majority of research employed CNN-based models for analysis, such as ResNet, EfficientNet, VGG, InceptionV3, and DenseNet. The results showed that the pooled diagnostic odds ratio was 87.43 (95% confidence interval [CI]: 42.18–181.26) with a pooled sensitivity of 0.86 (95% CI: 0.84–0.89), and pooled specificity of 0.92 (95% CI: 0.89–0.95). There is no statistically significant evidence of publication bias according to the Deeks funnel plot asymmetry test ( P = 0.31).