Oral and lip cancers contribute substantially to preventable mortality in low- and middle-income countries (LMICs), where screening coverage is limited and diagnostic delays are common. Foundation models such as large vision–language and promptable segmentation architectures trained on broad corpora can support zero-shot or minimally supervised recognition of lesions using textual prompts, potentially reducing dependence on large, locally annotated datasets. We evaluate the state of foundation models in oral cancer screening, focusing on the intersection of diagnostic performance and real-world viability. By analyzing adaptation strategies and governance requirements, this review provides a roadmap for the equitable deployment of these tools in resource-limited settings. We conducted a narrative review of peer-reviewed studies, preprints, and policy guidance published between January 2020 and June 2025. We searched PubMed, Scopus, and Google Scholar using terms spanning oral oncology, vision–language models (e.g., CLIP), promptable segmentation (e.g., SAM/MedSAM), adaptation (prompt engineering, test-time adaptation, few-shot learning, federated learning), and equity/governance (bias, fairness, LMIC, data sovereignty). Evidence was summarized qualitatively due to heterogeneity of datasets, imaging conditions, and metrics. While general-purpose AI models show promise as a baseline for clinical, histopathologic, and cytologic imaging, their performance remains inconsistent. Zero-shot foundation models often miss small, faint lesions, particularly early or subtle epithelial changes and perform worse than task-specific supervised models across segmentation and anomaly-detection benchmarks, with degradation most pronounced under domain shift. Lightweight adaptation strategies, including prompt refinement, training-time text augmentation, and test-time wrappers, yield only modest performance gains and do not consistently achieve clinically acceptable sensitivity for early lesions. Liu et al (Find Assoc Comput Linguist EMNLP 2024:9978–92, 2024), Ma (Nat Commun 15:654, 2024), Marzullo (arXiv preprint. arXiv:2411.09310, 2024). While foundation models offer a scalable path for oral cancer triage in LMICs, they are currently best suited as decision-support tools within supervised workflows. Moving safely toward real-world translation will require more than just better code; it demands representative datasets and rigorous evaluation across diverse phenotypes and capture conditions. Ultimately, minimal local calibration and adaptive oversight will be essential to ensure these tools close rather than widen, existing diagnostic gaps.
Phillips et al. (Mon,) studied this question.