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

Abstract

Artificial intelligence in medical imaging is shifting from task-specific algorithms toward adaptable foundation models (FMs) and vision-language models (VLMs). However, applying these models to MRI remains challenging because clinical interpretation requires integrating multi-parametric, volumetric imaging across multiple sequences with clinical context. This systematic review evaluates 35 studies on unimodal FMs and multimodal VLMs in MRI, summarizing emerging architectures, downstream applications, and challenges to clinical translation. Current research is concentrated in neuroimaging, with diagnostic classification representing the most mature application. Segmentation and outcome prediction also show promise, whereas automated report generation remains limited by poor clinical fidelity despite strong linguistic fluency. Broader adoption is hindered by heterogeneous MRI acquisition, limited large-scale paired image-text datasets, and inadequate safety-focused evaluation metrics. Advancing clinical translation will require standardized multicenter datasets, stronger integration with electronic health records and pathology, robust evaluation frameworks, and improved model explainability.