Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation

Research Overview

This scoping review provides a systematic assessment of the current landscape of vision foundation models (VFMs) developed exclusively for radiological imaging. As foundation models—large-scale, task-agnostic neural networks—gain traction in medical AI, the authors sought to categorize how these models are built, trained, and evaluated. By analyzing 67 peer-reviewed studies published between 2017 and 2026, the review maps the field across three pillars: data scale and heterogeneity, architectural and pretraining scalability, and downstream transferability.

Methodology and Pillars

The authors utilized the PRISMA-ScR framework to identify studies that move beyond single-task models toward generalizable representation learning. The three-pillar conceptual framework organizes the findings as follows:

• Data Scale and Heterogeneity: The review highlights a wide spectrum of dataset sizes, ranging from fewer than 100,000 samples to multi-million-image cohorts. While brain MRI, thoracoabdominal CT, and chest X-rays are the most common modalities, the diversity of data sources remains a critical factor for model generalization.
• Scalable Pretraining and Architecture: Transformer-based architectures have become the standard, with self-supervised learning (SSL) serving as the primary pretraining strategy. Masked image modeling and contrastive learning are the most prevalent techniques for extracting features from unlabeled data.
• Downstream Transferability: The review evaluates how these models adapt to clinical tasks like segmentation and classification. While many models demonstrate strong performance, the authors note that cross-center and cross-scanner validation—essential for real-world clinical deployment—is inconsistently reported.

Clinical Translation and Future Directions

A significant portion of the review assesses alignment with the FUTURE-AI principles (fairness, universality, traceability, usability, robustness, and explainability). The authors found that adherence to these principles is uneven across the literature. The study concludes that while radiology-specific VFMs offer a powerful new paradigm for medical imaging, the field must prioritize standardized benchmarking, better reporting of external validation, and a stronger focus on deployment-oriented evaluation to bridge the gap between research and clinical practice.