A new AI model could help radiologists identify brain abnormalities in MRI scans for neurological conditions such as stroke, multiple sclerosis and brain tumours, in a new study from researchers at UCL and King’s College London.
The study, published in Radiology AI , shows the potential of AI to help to clear waiting times in neurology departments due to a shortage of radiologists, as well as increasing demand for MRI scans that are vital for diagnosing and monitoring a range of brain conditions including tumours, strokes and aneurysms.
The AI model could help to ease the pressure on radiology departments by triaging cases according to their likely severity and decreasing the time it takes to return results. To test its capabilities, the model was first asked to distinguish between ’normal’ and ’abnormal’ scans, which it did accurately (getting it right approximately 19 out of 20 times) when compared to assessments made by expert radiologists.
It was then tested on specific conditions (using new MRI scans which weren’t included in the training data) such as stroke, multiple sclerosis and brain tumours, and was able to recognise these conditions accurately (getting it right approximately nine out of 10 times).
Most AI models are currently built with large datasets that have been manually labelled by expert radiologists, which are expensive and time-consuming to produce.
To overcome this, the team built an AI model that trained itself - without the need for expert radiologists - on over 60,000 existing brain MRI scans using their corresponding radiology reports simultaneously.
The researchers also designed the model so that when given a scan or textual query like ’glioma’ (a type of brain tumour), the system could search and retrieve similar cases, potentially supporting diagnostic review or teaching.
Professor James Cole, an author of the study from the UCL Hawkes Institute and UCL Computer Science, said: "The combination of language models and vision models has made a massive impact in many domains, and it’s exciting to have this demonstrated for brain-MRI scans using real-world clinical data. The study highlights the potential that automated tools have for optimising neuroradiological practice. With further validation, I can see a clear path to these types of models benefitting patients in the near future."
The study indicates that the model could be used at the time of scanning to flag abnormalities and support clinical decision-making by suggesting findings to radiologists, detecting potential errors in reports, or retrieving similar cases from past examinations. The researchers hope that this would speed up diagnoses and reduce reporting delays, helping to improve patient outcomes.
Dr Thomas Booth, senior author of the study from King’s College London and Consultant Neuroradiologist at King’s College Hospital, said: "By training the system on scans and the language radiologists use to describe them, we can teach it to understand what abnormalities look like."
"The next step is to run a randomised multicentre trial across the UK to see how abnormality detection improves workflows in practice. We are pleased to say that this trial will start in hospitals in 2026."
Chris Lane
/ +44 (0) 7717 728648
E: chris.lane [at] ucl.ac.uk
