Foundation Models Applied to Infectious Disease Imaging: A Scoping Review of Architectures, Applications, and Gaps

Foundation models — large-scale pretrained architectures including vision transformers, vision-language models, and segment anything models — have reshaped medical imaging artificial intelligence since 2020. No review has examined their application specifically to infectious disease imaging, a domain with distinct challenges including emerging pathogens, resource-limited deployment settings, and diverse imaging modalities.

This scoping review followed JBI methodology and PRISMA-ScR reporting guidelines. Six databases (PubMed, Embase, Scopus, IEEE Xplore, arXiv, medRxiv/bioRxiv) were searched up to March 2026. Title/abstract screening used AI-assisted methods with 20% manual validation (Cohen's kappa = 0.858). Data were charted on disease, imaging modality, model architecture, clinical task, and geographic origin.

From 5,480 identified records, 634 studies were included after deduplication and two-stage screening. COVID-19 dominated (61.2% of studies), followed by pneumonia (47.3%) and tuberculosis (16.1%). Chest radiography (62.6%) and computed tomography (30.1%) accounted for 93% of imaging modalities, while microscopy — the primary diagnostic tool for malaria and parasitic infections — represented only 5.5%. Vision transformers were the most common architecture (53.6%), with self-supervised learning (16.2%) and Swin transformers (14.7%) following. Segment anything models (0.6%) and CLIP-based approaches (1.3%) remained rare. China (22.9%) and India (17.2%) contributed the most studies, while Sub-Saharan Africa — bearing the highest infectious disease burden globally — contributed fewer than 2.5%. No prospective validation or deployed clinical studies were identified.

Foundation model research in infectious disease imaging is growing rapidly but remains concentrated on COVID-19, chest imaging, and vision transformer architectures. Critical gaps exist for neglected tropical diseases, non-radiographic modalities (microscopy, dermoscopy), and newer foundation model paradigms (SAM, CLIP, diffusion models). The geographic mismatch between research activity and disease burden raises equity concerns.