Developing and evaluating multimodal large language model for orthopantomography analysis to support clinical dentistry

Orthopantomography (OPG) is a primary screening tool for initial dental diagnosis, yet existing AI support typically operates in a one-way manner, lacking interactive nature like ChatGPT. To address this, we introduce ToothXpert, a dental diagnostic system with multimodal large language model. We curated a dental concept alignment dataset and a comprehensive multimodal OPG dataset, MM-OPG (comprising 131,065 question-answer [QA] pairs), covering 11 key conditions. ToothXpert facilitates simultaneous visual and language responses, fostering dynamic dentistry support. On the internal test dataset with 4,950 QA pairs, ToothXpert achieves a macro F1 score of 78.61%, outperforming LLaVA v.1.5 by 23.19%, LLaVA-Med v.1.5 by 41.39%, Qwen-VL by 56.38%, and HuatuoGPT-Vision by 26.74%. In human assessment, ToothXpert achieves an averaged performance of 3.54, surpassing compared multimodal large language models (MLLMs) with 1.46 and 1.38. Moreover, on the external dataset, ToothXpert achieves 1.96% and 2.99% higher F1 scores than two junior dentists with 3-years' clinical experience, while requiring significantly less time.