Error-Tolerant Multimodal Vision-Language Models for Endodontic Triaging: A Cross-Sectional Study

To assess the performance of multimodal vision-language artificial intelligence models, optimised using quantisation-aware training, in triaging endodontic treatment needs. The focus is on the ability to interpret endodontic radiographs while tolerating common image capture errors, including cone cutting, elongation, foreshortening, horizontal misalignment, over- and under-exposure and artefacts.

In total 3600 dental were obtained across a 1-year period. Image augmentation techniques were applied to enhance model generalisability. Bootstrapped Language-Image Pretraining (BLIP), CLIP, Florence 2 and Paligemma multimodal models were fine-tuned using quantisation-aware training and evaluated using Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), Metric for Evaluation of Translation with Explicit ORdering (METEOR), Consensus-based Image Description Evaluation (CIDEr) and Loss Trends and Convergence.

Quantisation-aware optimisation improved BLEU-4 by at least 17.3%, METEOR by at least 11.1%, ROUGE-L by at least 9.8% and CIDEr by at least 75.5% across all models. Quantisation reduced memory consumption by 87.5% while preserving diagnostic accuracy within a 0.5% error margin while correctly reproducing over 90% endodontic triage assessments made by practitioners.

Multimodal AI demonstrates tolerance to imaging inconsistencies and is capable of accurately triaging endodontic cases with minimal computational demands, without compromising diagnostic performance.