Deep Learning Models for Detection of Periapical Radiolucent Lesions on Panoramic Radiographs: A Systematic Review and Meta‐Analysis
Ibrahim Ali Ahmad, Raidan Ba‐Hattab, Sadeq Ali Al‐Maweri, Faleh TamimiABSTRACT
Background
Panoramic radiographs are used routinely to screen dental conditions and treatment patterns. Recently, numerous studies have suggested that deep learning (DL) models can be utilized for analysing panoramic radiographs.
Objective
This review aimed to evaluate the accuracy of DL models in detecting periapical radiolucent lesions (PRLs) on panoramic radiographs, as compared to a reference standard.
Methods
An electronic search was conducted across six databases (PubMed, Embase, Scopus, Web of Science, IEEE, and ArXiv) and grey literature (through Google Scholar) from January 1st, 2012, to March 18th, 2026. Studies were selected according to the pre‐defined eligibility criteria. Data extraction from included studies was conducted using self‐constructed tables. The risk of bias and applicability concerns were assessed using the Quality Assessment and Diagnostic Accuracy (QUADAS‐2) tool. Quantitative analysis of studies that reported complete confusion matrices was conducted using the split component synthesis method. The certainty of evidence for studies included in the meta‐analysis was evaluated using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) tool.
Results
A total of 30 studies, mostly from Europe and the Middle East, were included in the systematic review. Of these, six studies were included in the meta‐analysis. Sixteen studies (53.3%) had risk of bias, while applicability concerns were found in 12 studies (40%). Compared to the reference standard, DL models had pooled sensitivity of 0.80 (95% confidence interval [CI]; 0.49–0.94), specificity of 0.98 (95% CI 0.87–1.00), diagnostic odds ratio (DOR) of 176.37 (95% CI 20.48–1515.70), and area under receiving operating curve (AUROC) of 0.93 (95% CI 0.82–0.98). There was high heterogeneity among studies included in the meta‐analysis, while publication bias was not detected. The certainty of evidence was rated as “moderate”.
Conclusion
Overall, the DL models demonstrated high accuracy in detecting PRLs on panoramic radiographs. However, this finding should be interpreted with caution due to high heterogeneity in the conduct and reporting across studies. Future studies using large and diverse datasets and validated reference standards are warranted to confirm the reliability and generalizability of DL models. Standardized reporting is also critical for fair comparability between these models.
Trial Registration
PROSPERO database (CRD42024608038)