Answers are grounded in this paper and cite their sources.
The rapid evolution of medical vision-language models (VLMs) has produced a vast technical literature, yet a critical gap remains between achieving high engineering benchmarks and demonstrating true clinical utility. To map this landscape, we conducted a systematic review of medical VLM research published between 2020 and July 2026 across six databases. We identified 2,822 eligible works, classifying them into 8 distinct model families. From this broad map, we performed an in-depth, full-text analysis on a stratified subset of 153 works, extracting comprehensive profiles of their architectures, clinical tasks, data provenance, evaluation protocols, and open-source releases. We organize this literature along architectural frameworks and clinical tasks to synthesize the datasets, metrics, and scoring conventions that define current performance. Our analysis reveals that report generation dominates the field, heavily anchored by chest radiography datasets. Crucially, while technical transparency is growing, with 148 of the 153 works explicitly stating their evaluation protocols and 99 releasing code, clinical translation lags severely. Only 41 works report any form of clinical study, and a mere 1 includes a prospective trial. Ultimately, while medical VLM research is flourishing at the level of algorithmic development, robust evidence supporting real-world clinical generalization remains limited.
No takes yet. Share an insight, caveat, or question.
Shiwakoti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: