Beyond the Naked Eye: Automated Detection of Digital Manipulation in Dental Radiographs Using Probabilistic Detection Model

This study evaluated how reliably humans can detect digital manipulations in dental radiographs and introduced an unsupervised, artefact-aware artificial intelligence (AI) system to identify subtle, visually imperceptible changes. A dataset of 384 anonymised periapical radiographs was used, with equal numbers of original and synthetically altered images. Manipulations included copy-move, splicing, inpainting and lesion modification using clinically relevant techniques. The AI framework combined Error Level Analysis with a probabilistic Gaussian Mixture Model and included artefact suppression layers to reduce false positives from metallic restorations and high-contrast anatomical structures. It generated pixel-level probability maps and classified images based on percentile thresholds. Two experienced endodontists independently evaluated the images, blinded to dataset composition, and agreement was measured using Cohen's Kappa. The AI system achieved 98% sensitivity, specificity and accuracy with near-perfect agreement. In contrast, human observers showed poor performance, indicating AI offers a far more reliable solution.