Medical Vision Foundation Models: A comprehensive review from technical evolution to clinical translation.
Authors
Affiliations (4)
Affiliations (4)
- School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jianshe North Road, Chenghua District, Chengdu, 610054, China.
- Network and Data Security Key Laboratory of Sichuan Province, School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jianshe North Road, Chenghua District, Chengdu, 610054, China; Yangtze Delta Region Institute (Quzhou), University of Electronic Science and Technology of China, No. 1 Chengdian Road, Kecheng District, Quzhou, 324003, China. Electronic address: [email protected].
- Network and Data Security Key Laboratory of Sichuan Province, School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jianshe North Road, Chenghua District, Chengdu, 610054, China. Electronic address: [email protected].
- Network and Data Security Key Laboratory of Sichuan Province, School of Information and Software Engineering, University of Electronic Science and Technology of China, No. 4, Section 2, Jianshe North Road, Chenghua District, Chengdu, 610054, China; Ningbo WebKing Technology Joint Stock Co., Ltd, Ningbo, 315000, Zhejiang, China.
Abstract
Medical Vision Foundation Models (MVFMs) mark a shift from task-specific artificial intelligence (AI) toward general-purpose visual intelligence in medical imaging. However, clinical deployment remains constrained by tensions among data availability, technological development, and integration into clinical workflows. To address these issues, this review provides a systematic analysis of the field's evolution from 2019 to 2025. We introduce the Data-Technology-Translation Triad as a unifying conceptual framework to structure the progress, key factors, and challenges of MVFMs. Our analysis traces the field's development and identifies a convergence toward Transformer-based and hybrid architectures, alongside hybrid pre-training paradigms. We present a unified technical framework that covers pre-training strategies, including generative, contrastive, and hybrid learning, as well as adaptation mechanisms such as full fine-tuning, parameter-efficient fine-tuning, and prompt learning. Furthermore, we examine how data scarcity, heterogeneity, and privacy constraints in medical imaging drive the adoption of self-supervised learning and highlight domain generalization as a critical research direction. Although MVFMs demonstrate strong performance in data-efficient and multi-task settings, their clinical translation is hindered by limited generalization, insufficient interpretability, misalignment with clinical workflows, and regulatory uncertainty. In response, we propose practical directions for developing reliable general-purpose medical vision systems, with emphasis on causal reasoning, dynamic data learning, interactive human-AI collaboration, and coordinated efforts among researchers, clinicians, and regulators. This review provides a structured reference to guide MVFMs from technical validation to clinically effective, reliable, and ethically sound deployment. An updated collection of MVFM papers and related resources is available at our website https://github.com/ice-TuoShui/MVFM-Review.