An artificial intelligence (AI) tool is able to identify focal cortical dysplasias (FCDs) — the microscopic lesions responsible for refractory epilepsy in children — with up to 94% accuracy.

Dubbed the “AI Epilepsy Detective” the system is able to identify these subtle abnormalities, which are frequently missed on conventional imaging.

The tool’s developers believe it has the potential to reduce the need for children to undergo invasive diagnostic testing and allow for more accurate and rapid diagnosis.

“Identifying the cause early lets us tailor treatment options and helps neurosurgeons plan and navigate surgery,” said lead author Emma Macdonald-Laurs, PhD, MBChB and research officer at the Murdoch Children’s Research Institute in Parkville, Australia, in a press release.

“With more accurate imaging, neurosurgeons can develop a safer surgical roadmap to avoid important blood vessels and brain regions that control speech, thinking, and movement and removing healthy brain tissue,” she added.

The study was published online on September 30 in Wiley Epilepsia.

Lesions Missed on MRI

Bottom-of-sulcus dysplasia (BOSD) is a subtype of FCD that often underlies drug-resistant epilepsy in children and can be difficult to detect. Investigators noted that approximately 60% of BOSD cases go undetected on initial MRI.

BOSD can be treated with surgery if identified, underscoring the need for more accurate diagnosis. To improve detection, the researchers trained their model on MRI and FDG PET scans from 54 pediatric patients (54% boys) with confirmed BOSD, most of which were missed on initial MRI.

They extracted 12 imaging features — including cortical thickness, gray-white matter contrast, and regional metabolism — and used machine learning to distinguish dysplastic tissue from normal cortex.

To strengthen the analysis, the model was validated in two independent cohorts — 17 newly diagnosed pediatric patients and 12 previously published adult cases.

Because there were no healthy controls, the researchers generated “pseudocontrols” for each patient (median number per patient, 26) from the training set.

The detector used a single hidden layer neural network classifier, trained on these pseudocontrol-normalized MRI and FDG PET features from each patient to predict the BOSD location, based on cluster probability maps.

The detector identified a top cluster as the most likely lesion site and also ranked the five most probable regions. An imaging expert then reviewed these clusters to confirm the presence of BOSD.

“Having the option of the top five cluster output offers up more areas to check but ultimately several of these areas will be an incorrect solution, unless that patient has multiple lesions,” Macdonald-Laurs told Medscape Medical News in an email.

Advancing Diagnostic Accuracy

Among the training and test cohorts, 81% of patients initially had normal MRIs, and most BOSDs measured < 1.5 cm³ in size. Combining MRI and PET scans achieved the highest detection performance in the test group, with the detector identifying BOSD overlap in one of the top five clusters in 94% of patients (top cluster, 88%).

In contrast, MRI alone was substantially less effective, with only 35% of BOSDs overlapping one of the top five clusters (top cluster, 24%). Of the 17 children in the test group, 12 had surgery and 11 are now seizure-free.

The investigators emphasized that the tool is not intended to replace expert radiologist review but to assist in highlighting lesions, potentially improving BOSD detection rates.

“The detector is not yet ready for clinical use. Our next plan is to perform a prospective study of the detector to see how it performs on new patients across several Australian hospitals. We also will survey what the children, parents, and clinicians think about using AI in epilepsy care so that we can better understand barriers or challenges to clinical implementation,” Macdonald-Laurs said.

The study’s limitations include a lack of healthy control data and its small cohort size.

The authors reported no disclosures. The full list of funding can be found in the original article.