AI helped reveal epilepsy-causing lesions hidden in apparently normal brain scans

An AI system found signs of epilepsy-causing brain abnormalities in 51 of 80 patients whose MRI scans had previously been reported negative. The 63.7% detection result came from a retrospective test of MELD Graph, published in JAMA Neurology in February 2025. These were patients with focal cortical dysplasia, rather than an unselected population receiving a first epilepsy scan.[1]

Focal cortical dysplasia (FCD) is an area of brain tissue that developed abnormally. It can drive seizures that medication does not control, yet remain hard to distinguish on an MRI. Locating it can help an epilepsy team assess whether surgery is possible. For a family living through repeated seizures and investigations, finding a plausible physical target is a consequential step.[2]

Looking at the shape of the cortex

The model treats the folded outer brain as a connected surface. Measurements at one location are considered alongside surrounding tissue, allowing the network to learn spatial relationships across a hemisphere. This addresses a weakness of earlier methods that examined isolated points or small patches and could produce numerous unrelated areas for a radiologist to investigate.[3]

The architecture also learns how far points lie from a lesion and whether a hemisphere contains one. These additional training tasks help organise its predictions when the hand-drawn outlines used for training are imperfect. The goal is to identify a coherent suspicious region within the wider anatomy.[3]

The report can explain what caught the model’s attention: unusual cortical thickness, abnormal folding or a blurred boundary between grey and white matter. It shows the suspected region on the patient’s scan, identifies its location and size, and displays the features that contributed most to the prediction. A clinician receives a place to inspect and an explanation to examine.[4]

A report that supports another look

The project’s own examples include a lesion missed by five expert raters that the system highlighted with only 7% confidence. The finding illustrates why a probability score is one part of the evidence: an uncertain candidate can still direct attention to a subtle abnormality. The original images and clinical findings determine whether that suggestion is useful.[7]

MELD’s interpretation guidance asks users to overlay predictions on the original scan and inspect the reconstructed brain surfaces. A cluster can arise because the reconstruction followed the wrong tissue boundary or because the scan contains an artefact. Showing the prediction in context makes that checking possible, instead of presenting a diagnosis without a visible basis.[4]

The 2025 study did not establish that introducing MELD Graph increased seizure-free surgery. Its surgical follow-up information helped identify and characterise existing cases; the AI evaluation examined whether the lesions could be located.[1]

Taking the method to new scans

A September 2026 external study tested the approach in 70 children at one centre. Within the MRI-positive group, MELD Graph’s lesion-level precision was 0.86 and recall was 0.63: most highlighted lesions matched expert review, while some known lesions remained undetected. A competing 3D-nnUNet system made fewer false-positive findings in MRI-negative scans, measured against the specialists’ image-review assessment. That comparison reinforces the value of a specialist interpreting the suggestions.[6]

The current documentation identifies MELD Graph as a research tool for suspected focal cortical dysplasia. Scanner differences matter: the project recommends adjusting measurements to the scanner’s characteristics. It now advises using the T1-weighted scan as the primary input, with an additional FLAIR sequence interpreted cautiously because it can introduce more false-positive clusters.[5]

The code and training materials are publicly accessible. In December 2025, the project reported research use in more than 100 hospitals. That makes the method available for further independent evaluation across different patients, scanners and clinical workflows.[7]

The useful advance is the ability to revisit information already present in a scan. A difficult visual search becomes a smaller set of explicit, inspectable questions about particular regions of the brain. Connecting those findings to the rest of an epilepsy assessment is how detection can support a treatment decision.

Report an error or suggest a correction