Reducing contrast use in brain MRI begins with a specific imaging question: can non-contrast sequences identify tumour regions that would otherwise enhance? A retrospective multi-cohort study tested AI models on that task. Its findings provide a basis for further evaluation, but do not establish which patients can safely forgo a contrast-enhanced examination.
The study brought together 11,089 MRI studies from 10 international datasets, using non-contrast T1-weighted, T2-weighted and FLAIR images. In a held-out set of 1,109 studies, the strongest model achieved 83% balanced accuracy, 91.5% sensitivity and 74.4% specificity. These are measures of enhancement detection under the study conditions, not general brain tumour diagnostic accuracy or direct measures of patient benefit.
The clinical question also depends on who is being examined and what follows an error. UCL’s account describes variation across tumour types and populations. Its paediatric subgroup contained only 20 cases. The researchers suggest under-representation in training data may have affected performance; the small subgroup cannot establish dependable performance in children.
The versions of the available evidence should also remain clear. The preprint was first submitted on 19 August 2025 and updated to version 3 on 23 June 2026. UCL announced publication in The Lancet Medical Imaging & Theranostics on 16 September. This article checks the public preprint abstract and institutional account, rather than claiming review of the final published full text.
For a clinical team considering comparable technology, relevant questions include similarity to local patients, scanners and sequences, and how missed enhancement would be addressed through review. The researchers state that the tool is not yet accurate enough to replace contrast-enhanced MRI. Prospective evaluation within a real workflow is needed to connect an imaging prediction with a safe decision about contrast use.
References