PanoFM: An LLM-empowered panoramic foundation model with clinical semantic integration for comprehensive dental disease diagnosis

Oral diseases impose a substantial global health burden, and panoramic radiography is widely used as a cost-effective screening modality in dental care. However, accurate diagnosis remains challenging due to complex anatomical structures, tooth overlap, and significant inter-observer variability. Moreover, the development of robust computer-aided diagnostic systems is further hindered by the variety of dental diseases and the high cost of expert annotations. To address these challenges, we introduce a foundation model, PanoFM, that integrates panoramic radiographs with large language model (LLM)-powered clinical understanding. First, to alleviate reliance on manual annotations, we propose a two-stage self-supervised pretraining framework. The initial instance-based stage integrates contrastive learning and masked image modeling to capture contextual visual representations, while the subsequent prototype-based stage employs alignment guided by a vision foundation model to refine panoramic features. Second, an Optimal Transport-based alignment module associates anatomical structures with pathological features, capturing co-occurrence patterns and spatial relationships among teeth. Third, PanoFM combines large-scale image representation learning with contextual modeling capabilities of LLMs to compensate for limitations of visual information and uncover high-dimensional correlations between diseases and individual profiles. We ensure reliability and generalizability of PanoFM by constructing a large-scale panoramic radiography dataset comprising 127,878 cases from six geographically diverse clinical centers in China, on which the model is trained and systematically evaluated. Experimental results demonstrate that PanoFM significantly outperforms baselines in diagnostic accuracy, robustness, and generalization. Specifically, with an average AUC of 0.974 across six common dental diseases, PanoFM shows strong potential as an auxiliary diagnostic system for panoramic radiography.