Diagnostic Accuracy of Artificial Intelligence Software in Intraoral, Panoramic, and CBCT Imaging: A Pilot Study

Background/Objectives: Artificial intelligence (AI) is increasingly used in dental radiology, but its performance may depend on the imaging modality and the analytical model. This pilot (feasibility) study evaluated the diagnostic accuracy of AI-based software (Diagnocat) in intraoral, panoramic, and cone-beam computed tomography (CBCT) images against a single-expert reference standard, accounting both for indications of pathology and for indications of its absence. Methods: In this retrospective analysis, 45 anonymized examinations (15 intraoral radiographs, 15 panoramic radiographs, and 15 CBCT) were assessed. The reference standard was established by an expert in dental and maxillofacial radiology with over 30 years of experience, who took no part in selecting the examinations and was blinded to the software output. The primary analysis was performed per diagnostic indication, defined as the combination of one examination, one diagnostic category, and one anatomical location. A secondary binary "pathology versus no pathology" analysis was performed with one observation per tooth or anatomical site. Because observations were clustered within examinations and patients, the analysis was descriptive; no hypothesis tests were performed, and no confidence intervals were reported. Results: In the primary per-indication analysis, 999 diagnostic indications were recorded in the reference standard and 1017 by the software; 791 were concordant, comprising 45 of 71 reference indications for intraoral radiographs, 358 of 452 for panoramic radiographs, and 388 of 476 for CBCT. In the secondary binary analysis, intraoral radiographs yielded sensitivity 0.79, specificity 0.59, precision 0.71, and F1-score (the harmonic mean of sensitivity and precision) 0.75; panoramic radiographs 0.73, 0.88, 0.79, and 0.75; and CBCT 0.80, 0.86, 0.78, and 0.79. Sensitivity and F1-score were numerically highest for CBCT. The three modality groups comprised different patients, anatomical regions, and distributions of diagnostic categories, so these differences cannot be attributed to the imaging modality alone. Conclusions: In this pilot study, the performance of the software varied across imaging modalities and diagnostic categories. Because the set of negative units was defined by the locations charted by the expert rather than by a complete anatomical inventory, specificity, negative predictive value (NPV), and accuracy are reported as secondary descriptive outcomes, are conditional on that operational definition, and are not compared between modalities. The findings are exploratory and hypothesis-generating. The software may support the interpretation of dental radiographs, but should not replace expert assessment.