Image Resolution's Impact on Artificial Intelligence & Human Accuracy in Full-Mouth Radiographic Analysis.
Authors
Affiliations (7)
Affiliations (7)
- The Ruth and Bruce Rappaport Faculty of Medicine, Technion - Israel Institute of Technology, Haifa, Israel; Department of Periodontology, Rambam Health Care Campus, Haifa, Israel. Electronic address: [email protected].
- The Ruth and Bruce Rappaport Faculty of Medicine, Technion - Israel Institute of Technology, Haifa, Israel; Department of Periodontology, Rambam Health Care Campus, Haifa, Israel.
- Department of Endodontics, Rambam Health Care Campus, Haifa, Israel.
- Department of Prosthodontics, Rambam Health Care Campus, Haifa, Israel.
- Department of Periodontology, Rambam Health Care Campus, Haifa, Israel.
- Department of Public Dental Health, Semmelweis University, Budapest, Hungary.
- The Ruth and Bruce Rappaport Faculty of Medicine, Technion - Israel Institute of Technology, Haifa, Israel; Department of Prosthodontics, Rambam Health Care Campus, Haifa, Israel.
Abstract
Artificial intelligence (AI) is increasingly used for dental radiographic interpretation, yet the effect of image resolution on diagnostic accuracy compared with human evaluators remains unclear. This study evaluated how medium- versus high-resolution full-mouth radiographs influence diagnostic performance of an AI system and experienced clinicians. In this retrospective comparative study, 200 full-mouth series radiographs were divided equally into medium-resolution (96-300 dpi) and high-resolution (≥720 dpi) groups. Three independent human examiners and an AI system (Diagnocat, San Francisco, CA, USA) assessed six pathological conditions. The reference standard was defined as agreement between at least two human examiners. Diagnostic metrics (sensitivity, specificity, predictive values, accuracy, F1-score) were calculated with 95% confidence intervals. Inter- and intra-examiner reliability and AI-human agreement were assessed using Cohen's kappa. Differences in diagnostic accuracy between resolutions were tested using chi-square tests for independent proportions (or Fisher's exact test where appropriate), with Bonferroni correction for multiple comparisons. Human inter-examiner agreement ranged from moderate to substantial across pathologies (κ = 0.27-0.87), with the highest agreement for missing teeth and the lowest for root resorption. AI-gold standard agreement varied from essentially none to substantial (κ = 0.00-0.90) and increased with higher resolution for most conditions. AI accuracy ranged from 75.3% to 99.1%, with consistently high specificity (73.5%-99.8%) and variable sensitivity (0.0%-93.6%). High-resolution imaging significantly improved AI diagnostic accuracy for caries (+4.3%), furcation involvement (+1.4%), dental calculus (+9.8%), and missing teeth (+3.9%) (all P < .001, after Bonferroni correction), while changes for periapical lesions and root resorption were not significant. High-resolution radiographs enhance diagnostic accuracy for AI and human evaluators. AI achieved clinically acceptable performance, though sensitivity differed across pathologies. High resolution full mouth radiographs improve diagnostic accuracy for both artificial intelligence and human evaluators. This finding underscores the importance of optimal image quality in clinical practice to enhance diagnostic confidence and support AI assisted decision making in dentistry.