📄 Abstract
Background: Foundation models and vision-language models (VLMs) are reshaping medical artificial intelligence by enabling models to learn broad representations from heterogeneous imaging and textual data and subsequently adapt them to multiple clinical tasks. In radiology, these systems can integrate images, reports, clinical narratives and other contextual information, potentially extending artificial intelligence beyond narrowly defined classification or detection tasks. Objective: This review examines the clinical translation of foundation models and VLMs in radiology, with emphasis on multimodal medical imaging, evaluation, patient safety, workflow integration, explainability, bias, privacy, and human oversight. Methods: A narrative review was developed from contemporary peer-reviewed literature and authoritative technical and regulatory sources addressing medical foundation models, vision-language learning, multimodal radiology AI, clinical evaluation and responsible deployment. The literature was synthesized around model capabilities, clinical use cases, evaluation dimensions, safety risks and implementation requirements. Results: Foundation models can support image interpretation, report generation, visual question answering, retrieval, summarization, clinical decision support and multimodal reasoning. However, strong benchmark performance does not establish clinical effectiveness. Translation requires external validation, calibration, subgroup analysis, robustness testing, prospective workflow evaluation, monitoring for hallucinations and clinically meaningful human-AI interaction. Particular risks include dataset shift, demographic and institutional bias, automation bias, privacy leakage, unsupported recommendations and opaque failure modes. Conclusion: Radiology foundation models should be regarded as assistive clinical technologies rather than autonomous diagnostic replacements. A safe translation pathway requires clinically representative datasets, transparent evaluation, rigorous prospective testing, governance, cybersecurity, continuous monitoring and clearly defined accountability. Multimodal AI has substantial potential, but its clinical value will depend less on model scale alone and more on evidence quality, safety engineering and integration with radiologist-led care.
🏷️ Keywords
📚 How to Cite:
Sudha Y Soudi, Sacheena Kumbara, Prateek Yalawar , FOUNDATION MODELS AND VISION-LANGUAGE MODELS IN RADIOLOGY: CLINICAL TRANSLATION, SAFETY, AND EVALUATION OF MULTIMODAL MEDICAL IMAGING AI , Volume 12 , Issue 9, September 2026, EPRA International Journal of Multidisciplinary Research (IJMR) , Pages: 272 - 276 ,