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

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Endodontic Treatment Planning

Background:

• Multimodal vision-language AI models are emerging for medical image analysis.
• Optimizing these models for specific clinical tasks like endodontic triage is crucial.
• Quantisation-aware training can improve model efficiency and performance.

Purpose of the Study:

• To evaluate the performance of quantisation-aware trained multimodal AI models in triaging endodontic treatment needs.
• To assess the models' ability to interpret endodontic radiographs with common imaging errors.
• To determine the impact of quantisation on diagnostic accuracy and computational demands.

Main Methods:

• Fine-tuning of BLIP, CLIP, Florence 2, and Paligemma multimodal models using quantisation-aware training.
• Utilisation of 3600 dental radiographs with image augmentation techniques.
• Evaluation using metrics such as BLEU, ROUGE, METEOR, and CIDEr.

Main Results:

• Quantisation-aware optimisation significantly improved evaluation metrics (BLEU-4 by ≥17.3%, METEOR by ≥11.1%, ROUGE-L by ≥9.8%, CIDEr by ≥75.5%).
• Quantisation reduced memory consumption by 87.5% while preserving diagnostic accuracy within a 0.5% error margin.
• Models correctly reproduced over 90% of practitioner-made endodontic triage assessments.

Conclusions:

• Quantisation-aware trained multimodal AI models are effective for endodontic case triage.
• These AI models demonstrate robustness against common radiographic inconsistencies and artefacts.
• The optimized AI approach offers accurate diagnostic performance with minimal computational requirements.