Large language models in radiology education: a systematic review of applications and learning outcomes.

Authors

Abstract

This systematic review aimed to map how large language models (LLMs) are used in radiology education, grade their reported outcomes, and appraise evidence quality. We searched PubMed, Europe PMC, and ERIC (November 2022-October 2026) for original studies of LLMs in the radiology education of medical students or residents, including educator-rated LLM materials intended for them. Two reviewers screened titles and abstracts independently. Studies were grouped by application, outcomes were graded with a modified Kirkpatrick framework, and quality was assessed with the Medical Education Research Study Quality Instrument (MERSQI). Twenty-two studies (23 reports) were included. Twelve involved residents, five medical students, two mixed groups, and three had no learners. Applications were question generation (n = 7), feedback on learners' work (n = 6), content generation and tutoring (n = 4), simulation or personalized curricula (n = 2), and learners' own use (n = 3). LLM-generated questions matched expert questions on recall items but were weaker on clinical-reasoning items, and their apparent quality partly depended on human selection. Report feedback was moderately to highly accurate and generally welcomed. Only three studies measured learning; the rest evaluated LLM output, learner reactions, or case exposure, and none measured behavior. Mean MERSQI was 9.6 of 18 (range 6.0-12.5); 18 studies were single-center, and one was randomized. LLMs produce usable teaching materials, but evidence that they improve learning is limited and of low to moderate quality. They should be supervised until controlled multicenter studies are available.

Tags

Topics