ENHANCING DIAGNOSTIC ACCURACY IN DENTISTRY: MULTIMODAL AI WITH UNCERTAINTY QUANTIFICATION FOR RELIABLE DECISION- MAKING

Abstract

Objective: The objective of this research is to assess an integrated AI system together with dental X-rays and patient healthcare information, which features uncertainty quantification methods to boost diagnosis trustworthiness in dental clinics.

Methods: The research examined 400 patient cases from the 12-month period January through December 2024. Each patient case included panoramic radiograph imaging combined with essential clinical data such as age, gender, and symptoms of tooth pain. Linking possible diagnoses of dental caries and periapical lesions required the joint feature representation, which combined the analysis of clinical data with radiographs processed by a convolutional neural network (CNN) and a feed-forward neural network. The implementation of Monte Carlo dropout during inference served to provide UQ through generation of 50 stochastic predictions per case, which were followed by average prediction calculations and entropy readings for uncertainty measurement. Diagnostic metrics with 95% confidence intervals were analyzed in relation to metrics computed by an independent dentist evaluation [1,2].  The metrics included sensitivity, specificity and positive and negative predictive value with accuracy and area under the receiver operating characteristic curve [AUC]. Dental caries accuracy was examined alongside diagnosis accuracy when removing the 10% of cases with the highest uncertainty from analysis.

Results: The AI model demonstrated dental caries detection accuracy at 85.0% (95% CI: 74–93%) with values of 91.0% sensitivity and 78.3% specificity and a periapical lesion detection accuracy at 73.3% (95% CI: 60–84%), which included 96.0% sensitivity and 65.0% specificity. The tests yielded Area Under the Curve results of 0.92 and 0.85 as detected through performance evaluation. General dentists demonstrated caries detection accuracy of 78.3% through their observations at 85.0% sensitivity with 71.7% specificity but their periapical lesion detection accuracy reached only 60.0% at 90.0% sensitivity combined with 50.0% specificity [2,3]. The detection accuracy for caries increased from 85% to  92% when the most uncertain 10% of cases were excluded from analysis. Almost 89% of errors occurred in cases with high uncertainty levels.

Conclusion: The diagnostic performance of the system for general dentists increased by using multimodal data with uncertainty quantification, which gave both high sensitivity ratings and alert features to guide additional evaluations. The system demonstrates potential to facilitate clinical decisions by performing fast routine screening, which alerts clinicians to attend to difficult or less secure predictions. The system requires additional testing on extensive datasets from different backgrounds before it can be used clinically.