Artificial Intelligence and Human Perception in Endodontic Radiology: A Critical Narrative Review of Diagnostic Variability, Accuracy and Workflow Integration in Periapical Pathology

Abstract

Apical periodontitis is among the most prevalent chronic inflammatory conditions of the dentition, and its recognition rests almost entirely on the interpretation of radiographic images. That dependence is uncomfortable, because projection radiography detects periapical bone change only after substantial mineral loss has occurred, and because clinicians interpreting the same images frequently disagree with one another and with themselves. Convolutional neural networks and related deep learning architectures have been proposed as a corrective, and reported accuracies for automated detection of periapical radiolucent lesions are now routinely high. This critical narrative review examines whether that reported performance constitutes evidence of a genuine diagnostic advance, or whether it partly reflects the measurement conventions used to generate it. Literature was identified through programmatic bibliographic querying and structured web-based searching, with backward citation searching of retrieved reviews, and appraised for design, reference-standard construction, validation strategy and clinical relevance rather than by citation count. Four analytical themes are developed: the physical and perceptual limits that constrain any observer, human or computational; the substitution of expert consensus for histological or surgical verification and its consequences for apparent accuracy; the divergence between pooled accuracy estimates derived from different modality mixes; and the behaviour of clinician–algorithm systems, in which assistance improves specificity and reduces false-positive treatment decisions but distributes benefit unevenly across experience levels and introduces the possibility of automation bias. The available evidence indicates that current systems are competent detectors of radiographically evident lesions on the datasets that produced them, that external performance is frequently and substantially lower than internal performance, and that almost no evidence addresses whether automated detection changes patient-relevant outcomes. Priorities include reference-standard transparency, prespecified operating points matched to clinical purpose, multicentre external validation reported at the level of the individual root, and prospective evaluation of clinician–algorithm systems rather than algorithms alone.