Vision–language foundation models for multimodal medical image understanding and report generation: a narrative review with quantitative synthesis

Vision-language foundation models (VLFMs) for medical image understanding have advanced rapidly, yet evidence supporting clinical deployment remains fragmented. This narrative review synthesizes architectures, training paradigms, benchmarks, public datasets, and translation barriers. Drawing on more than 50 original studies published between 2020 and 2025, we provide a structured descriptive comparison rather than a statistical meta-analysis. In a focused audit of 13 representative studies, BLEU/ROUGE was reported by 31% (4/13), RadGraph F1 by 0% (0/13), CheXpert- or CheXbert-derived label evaluation by 46% (6/13), BERTScore by 8% (1/13), human expert evaluation by 31% (4/13), and inference latency or throughput by 0% (0/13). Because datasets, splits, candidate pools, and metrics differ, cross-task results are presented as exploratory benchmark indicators and not as a clinical-readiness scale. The synthesis identifies persistent concerns involving hallucination, limited interpretability, distribution shift, and narrow demographic and geographic coverage. We propose a minimum clinical validation and reporting standard aligned with the distinct purposes of FDA/Health Canada/MHRA and IMDRF Good Machine Learning Practice, the EU AI Act, TRIPOD+AI, and CLAIM, and identify eleven priority research directions. Medical VLFMs are promising, but progress should be judged by source-traceable, clinically grounded, and externally validated evidence.