From Accuracy to Reliability: What Prevents Artificial Intelligence from Being Clinically Reliable for Periapical Lesion Detection? A Review

Artificial intelligence (AI), particularly deep learning, has shown potential in detecting periapical radiolucencies on dental radiographs. However, accuracy on a retrospective research dataset does not guarantee clinical performance. This narrative review explores the discrepancy between algorithmic accuracy and real-world clinical reliability, including dataset representativeness, spectrum and selection bias, cropped-image evaluation, lesion size and anatomical complexity, reference-standard uncertainty, domain shift in imaging systems, external validation, explainability of AI, and the limited integration of radiographic AI with clinical information. Recent systematic reviews report high pooled diagnostic performance, but consistently cite substantial heterogeneity, risk of bias, limited prospective evidence, and insufficient external validation. The argument for this review is that the next phase of research should change the question from 'Can AI detect a periapical lesion?' to 'Can AI reliably detect an unexpected lesion in an unselected clinical image and improve the clinician's decision?' Clinical reliability should therefore be viewed as a multidimensional concept incorporating technical performance, generalizability, calibration, interpretability, workflow compatibility, and impact on clinician and patient outcomes. A staged framework of internal validation, external validation, prospective clinical validation, human-AI evaluation and post-deployment monitoring is proposed. AI should be regarded as an adjunctive second reader and decision-support tool rather than an autonomous diagnostic replacement.