Vision–Language Models in Ultrasound Imaging: A Comprehensive Review

Abstract

Ultrasound vision–language models (US-VLMs) must interpret operator-dependent images and dynamic examinations while linking visual evidence to clinical language. This review synthesizes 32 formally published studies identified through a three-database search. We apply an operational three-axis taxonomy: the interaction axis identifies the primary image–text computation, the learning axis records ultrasound-specific training objectives, and grounding granularity specifies the finest documented visual–linguistic correspondence. Study-level assignments and rationales make these distinctions reproducible. We compare task-specific results without ranking models across incompatible cohorts, endpoints, and metrics. Dataset analysis separates acquisition volume from independent clinical diversity, and a seven-domain risk-of-bias appraisal identifies limitations in reporting of patient-level partitioning and development–test overlap. We also distinguish retrospective performance, including external-cohort results, from evidence that assistance improves clinical decisions or workflows. Current systems support diverse ultrasound tasks, but clinical claims require evidence grounding, uncertainty assessment, and validation matched to the intended use.