Diagnostic accuracy of deep learning for detecting periapical lesions on radiographs and cone-beam computed tomography: a systematic review

Introduction: Deep learning models for detecting periapical lesions report near-perfect accuracy under internal validation but much lower sensitivity in clinical conditions, and the magnitude of this gap is unresolved. Methods: A systematic review was conducted following PRISMA 2020. Scopus, PubMed and Crossref were searched up to 2026 for diagnostic accuracy studies of deep learning for periapical lesion detection on periapical radiographs, panoramic radiographs and cone-beam computed tomography against an expert reference standard. Screening was assisted by artificial intelligence and verified by the authors. Results: Twenty-eight primary studies were included. Reported accuracy ranged from 0.84 to 0.96 under internal validation but fell to sensitivities between 33.3% and 93.9% in external or clinical evaluations, with the lowest figures on panoramic radiographs using a commercial tool. Model performance on cone-beam computed tomography was generally higher than on two-dimensional radiographs. Conclusions: Deep learning detects periapical lesions with high accuracy under controlled conditions, but its reported performance is inflated relative to real-world use, particularly on panoramic radiographs. External validation on clinically representative datasets is required before these tools can be recommended for routine diagnosis.