Artificial intelligence and dental panoramic radiographs: where are we now?

Abstract

Data sources

Bielefeld Academic Search Engine (BASE), Google Scholar Association for Computing Machinery: Guide to Computing Literature (ACM) and National Library of Medicine: PubMed databases were searched for systematic reviews.

Study selection

This study addressed a structured PICO question (Population, Intervention, Comparison, Outcome). Population was panoramic radiographs in human subjects. Intervention was use of artificial intelligence (AI) diagnostics, compared to human-only diagnosis. Quantitative or qualitative AI efficiency was the outcome. Systematic reviews were considered if they stated ‘systematic review’ in their title or abstract, were published in English and were not bound by a certain time frame. No supplemental primary studies were included. Screening and removal of duplicates were performed using the Rayyan tool.

Data extraction and synthesis

Data were extracted from each systematic review by two authors, with a third author having the deciding vote in cases of inconsistency. Cohen’s Kappa co-efficient was used to measure reliability between authors, resulting in almost perfect agreement. The risk of bias was accounted for using the ROBIS method which resulted in one paper being rejected, so only 11 included in results. Data were then grouped into seven domains which were detected by AI: teeth identification and numbering, detection of periapical lesions, periodontal bone loss, osteoporosis, maxillary sinusitis, dental caries, and other tasks. The effectiveness of the AI systems was assessed by various outcome metrics - accuracy, sensitivity, specificity, and precision being the most common variables.

Results

Results of this overview show a significant increase in accuracy of AI in analysing OPTs between 1988-2023. Latest AI models are most accurate in teeth identification and numbering (93.67%) whilst caries detection and osteoporosis showed 91.5% and 89.29% accuracy, respectively. Accurate results were also observed for the detection of maxillary sinusitis and periodontal bone loss. However, given the heterogeneity of source studies used in these systematic reviews, results should be interpreted with caution.

Conclusions

With improving AI technology, its use in dental radiology can be increasingly effective in supporting dentists in the detection of different pathologies. This overview has shown that systematic reviews of AI can quickly become outdated and that results of any systematic review should be treated with caution as this field advances. As such, regular updating and ongoing research is required.