Artificial intelligence-assisted imaging for oral squamous cell carcinoma diagnosis: a systematic review and diagnostic test accuracy meta-analysis

AI-assisted image analysis is increasingly studied for OSCC detection, but evidence differs by imaging modality, diagnostic endpoint, unit of analysis, and validation strategy, and prior reviews have rarely made explicit how they handle non-independent model matrices or unverifiable 2 × 2 data. Methods: We predefined and applied an explicit independence rule to prevent double-counting of models evaluated on the same test set, and restricted the primary quantitative synthesis to datasets with verifiable paired sensitivity and specificity or exact 2 × 2 data; duplicate model evaluations and non-independent matrices were excluded. The six-database search was rerun on 26 September 2026 using the same database set and search strategy; the retrieval and screening counts remained unchanged (Fig. 1). One later study identified through citation tracking is retained and transparently disclosed (see Limitations). A bivariate random-effects model was used as the primary joint synthesis, with transparent univariate random-effects estimates reported as sensitivity analyses. Results: GRADE-DTA certainty of evidence was rated low for pooled sensitivity and very low for pooled specificity, reflecting risk of bias, inconsistency, and imprecision in the underlying datasets (Table 2); the point estimates that follow should be read against this certainty rating. Ten independent datasets contributed to quantitative synthesis. The bivariate model estimated pooled sensitivity of 0.901 (95% CI 0.855–0.933) and specificity of 0.897 (95% CI 0.784–0.954). Univariate random-effects estimates were 0.898 sensitivity and 0.900 specificity. Restricting the analysis to seven datasets in which OSCC was the positive diagnostic class (though comparators varied, including non-OSCC lesions and pooled benign/OPMD categories, not only healthy mucosa) produced bivariate estimates of 0.923 sensitivity and 0.944 specificity. Heterogeneity was substantial, particularly for specificity. Conclusions: AI-assisted imaging demonstrates promising diagnostic performance, but the evidence remains heterogeneous and predominantly retrospective. High performance should not be equated with clinical readiness; prospective multicentre external validation and standardized OSCC-specific endpoints are needed.