Artificial Intelligence in Temporomandibular Joint Imaging: Diagnostic Performance Across Input Modality-Target Pairings-A Systematic Review and Diagnostic Test Accuracy Meta-Analysis.

Authors

Abstract

To estimate standalone artificial intelligence (AI) diagnostic performance in temporomandibular joint (TMJ) imaging across input modality-target pairings. Four databases were searched through 18 June 2026. Eligible studies assessed standalone imaging AI and supplied an exact test-set 2 × 2 table in the report or after author contact. Four pairings were analysed by bivariate meta-regression. Two reviewers applied QUADAS-3 at estimate level. Sixteen studies contributed 17 estimates. For disc displacement, magnetic resonance imaging input yielded pooled sensitivity 0.924 (95% CI, 0.859-0.960) and specificity 0.812 (0.709-0.885), compared with 0.765 (0.618-0.867) and 0.741 (0.592-0.850) for panoramic or cone-beam computed tomography input. The sensitivity difference was 0.159 and remained 0.152-0.198 in leave-one-out and replicate analyses; the omnibus pairing test was not significant (χ²(6) = 10.38, P = .110). Reference-standard relation and pairing partitioned the disc-displacement evidence identically. No estimate had low overall risk of bias, and several otherwise eligible estimates lacked exact published 2 × 2 data. Sensitivity was higher when disc displacement was depicted directly on magnetic resonance imaging than when inferred from imaging that does not depict the disc. However, the cross-domain group contained only three estimates, the omnibus pairing test was not significant, and design factors were entangled. Performance should therefore be interpreted by pairing, while direct depiction versus proxy inference requires prospective testing. TMJ imaging AI performance should be interpreted in the context of the input modality-target pairing rather than as a property of the algorithm alone; exact test-set confusion matrices are essential for verification.

Tags

Topics