Large AI Models in Dental Healthcare: From General-Purpose Systems to Domain-Specific Foundation Models

Abstract

Background: While oral pathologies impact approximately 3.5 billion individuals globally, the comparative clinical utility of expansive AI architectures within dentistry is not well characterized. Three primary model categories have arisen: language-generative systems, discriminative vision foundation models, and specialized dental foundation models. Currently, there is no comprehensive review that elucidates the interrelationships among these categories or their collective limitations.

Methods: Adhering to PRISMA-ScR guidelines, we conducted a systematic search across four databases—PubMed, Google Scholar, Scopus, and arXiv. Two reviewers independently screened the results. Following the application of specific inclusion and exclusion criteria, 97 studies published between 2020 and 2026 were selected for analysis. We developed a two-dimensional classification framework to organize these models based on their architectural paradigm and the degree of dental specialization.

Results: Language-generative models demonstrate superior capabilities in text-centric applications, such as clinical reasoning, licensing examination preparation, and patient communication; however, their performance in image-based diagnostics remains variable. Conversely, modified versions of SAM and CLIP have shown robust efficacy in detecting lesions and segmenting teeth. Specialized models, including DentVFM, DentVLM, and OralGPT, exhibit the highest performance levels for complex multimodal dental tasks. Integrated pipelines consistently surpass single-model strategies in effectiveness. A notable data asymmetry exists: pretraining for dental-specific models is predominantly focused on the visual domain, a trend driven by the scarcity of large-scale dental textual corpora.

Conclusions: General-purpose and dental-specific models serve complementary functions, with the most efficient systems integrating both within structured workflows. For safe autonomous deployment, three critical barriers must be addressed: the hallucination issues inherent in generative models, the shortage of annotated dental datasets, and the lack of standardized benchmarks for clinical evaluation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC