Unimodal imaging foundation models and vision-language models in MRI: A systematic review of emerging methods and clinical applications.

A systematic review of 35 MRI studies finds unimodal foundation models and multimodal vision-language models are most mature for neuroimaging diagnostic classification, with segmentation and outcome prediction showing promise.

• Why it matters: MRI clinical interpretation depends on integrating multi-parametric, volumetric information across sequences with clinical context, making adaptation of foundation and vision-language models difficult.
• What they did: The review evaluates 35 studies, mapping emerging architectures and downstream uses while highlighting key barriers such as heterogeneous acquisition and limited large-scale paired image-text datasets.
• The result: Diagnostic classification is the most developed application, while automated report generation is constrained by poor clinical fidelity despite strong linguistic fluency, implying translation needs standardized multicenter data and safety-focused evaluation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.