Sensitivity Evaluation of Deep Learning-Based Models for Dental Caries Detection in Bitewing Radiographs: A Systematic Review and Meta-analysis

Abstract

Aims: To analyse the sensitivity of deep learning models in detecting carious lesions in bitewing radiographs through a systematic review and meta-analysis. Study Design: Systematic review and meta-analysis. Place and Duration of Study: The review was conducted using the Cochrane Library, LILACS, PubMed, Scielo, and the Virtual Health Library databases, covering articles published until March 2024. Methodology: The research question was developed using the PICOT framework. Paired and independent searches with no filters applied were performed using the registered Medical Subject Headings "(machine learning) OR (deep learning) AND (dental caries) AND (radiography, bitewing)" and its related entry terms, covering articles relevant to the study's theme with no language or time restrictions. A total of 2,841 articles were initially identified, with 8 selected after title and abstract screening and removal of duplicates. After full-text review, 6 articles were included for analysis. The risk of bias was assessed using the QUADAS-2 tool, and the data were analysed with Review Manager 5.4 and R software. Results: All included studies were classified as having a low risk of bias. The meta-analysis revealed that deep learning models demonstrated moderate sensitivity (0.77) in detecting carious lesions, with an overall performance AUC of 0.779. In comparison, dentists achieved higher performance with an AUC of 0.886. Significant heterogeneity was observed, largely due to variations in model architecture, training datasets, and possibly image preprocessing techniques. These factors impacted the models' diagnostic accuracy, emphasising the need for standardisation. Conclusion: Deep learning models currently exhibit moderate sensitivity, and their variable performance underscores the importance of not relying on them in isolation for caries detection. Instead, they should serve only as supplementary diagnostic tools to assist clinicians. Further research is necessary to enhance model reliability, address heterogeneity, and establish standardised evaluation frameworks.