'One wrong word in a novel': How AI can spot brain abnormalities in epilepsy patients

Searching for abnormalities on brain scans can be like looking for one very slightly wrong word in a 200-page novel. 

But now scientists in London are using AI to detect previously 'invisible' brain abnormalities – too subtle for standard detection methods – that cause epilepsy.

The tool, called MELD Graph, spots around two-thirds of the tiny abnormalities that human radiologists often miss when studying a patient's MRI scan.

"An MRI scan is maybe 10 million pixels and they're looking for something that's maybe 100 pixels," explains Dr Konrad Wagstyl, a Senior Lecturer at the School of Biomedical Engineering & Imaging Sciences, King's College London, before giving an illuminating comparison for this needle-in-a-haystack search.

"An example might be a novel – a 200 page novel has 100,000 words in it and they're looking for one word that's slightly the wrong size or slightly the wrong font," Wagstyl explained.

A decade's development

The team has spent 10 years developing MELD Graph, which was trained on MRI data from more than 700 people with focal cortical dysplasia (FCDs), a major cause of epilepsy. FCDs can be subtle and difficult to see with the human eye and up to half of these lesions are missed by radiologists.

The system highlights what it thinks might be an abnormality and also explains why. The analysis can then be checked by a radiologist, speeding up a diagnosis and access to surgery.

"If you can find the brain abnormality, you can offer surgery which might cure the seizures," said Wagstyl. "And what our AI tool can do is find two-thirds of those abnormalities that doctors might miss – and that's really important because it means that more children can get the best possible form of treatment."

Epilepsy affects around one in 100 people globally, with one in five epilepsy cases being caused by a structural abnormality in the brain.

While the tool is not yet clinically available, the team has made the software open-source and is training clinicians and researchers worldwide on its use.