Can MRI machine learning detect focal cortical dysplasia?
A proof-of-concept study distinguished focal-cortical-dysplasia epilepsy from healthy controls with a highest internally cross-validated AUC of 0.791. There was no external clinical validation, and the model found only three of six MRI-negative cases.
The short version
- What happened
- The final analysis included 167 people from one public dataset: patients with focal cortical dysplasia type II-associated epilepsy and healthy controls scanned at one centre.
- Who it affects
- People with difficult-to-detect epilepsy lesions may eventually benefit from a carefully validated second reader, but this study does not justify relying on an automated negative result.
- What remains unproven
- Only 167 people were analysed, with hundreds of imaging features and no independent external cohort.
Keep in mind
Only 167 people were analysed, with hundreds of imaging features and no independent external cohort.
The task compared focal-dysplasia epilepsy with healthy controls rather than a real differential-diagnosis population.
Nearly two-thirds of lesions were frontal, limiting evidence for rarer locations.
Only six MRI-negative, histopathology-positive patients were available; the model detected three.
No lesion-localisation accuracy, reader impact, treatment decision or patient outcome was tested.
Source analysis · 1 linked source. Primary analysed source date: 14 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: a research signal, not a clinical detector
A machine-learning pipeline separated people with focal cortical dysplasia type II-associated epilepsy from healthy controls with a highest internally cross-validated area under the receiver operating characteristic curve of 0.791. That is a meaningful proof-of-concept signal from structural MRI, but it is not evidence that the system can diagnose a new patient in routine care.
The result came from one public dataset and was evaluated through nested cross-validation rather than an independent hospital. The model's numerical lead over a second approach was not statistically significant. More importantly, the task was patient-versus-healthy-control classification, not localisation of a lesion among people referred with difficult-to-interpret epilepsy scans.
The hardest subgroup illustrates the gap. Six patients had initially been described as MRI-negative but were later confirmed to have focal cortical dysplasia on histopathology. The model identified three of them, an exploratory sensitivity of 50%. Six people are far too few for a reliable performance estimate, but missing half does not support replacing specialist review.[1]
The study analysed 167 people scanned at one centre
The original dataset contained 85 people with focal-cortical-dysplasia type II-associated epilepsy and 85 healthy controls. Severe motion artefact and failed cortical reconstruction excluded one patient and two controls, leaving 167 people. The patient group's mean age at scanning was 28.9 years, with a range from 12 to 65, and 61 of the 85 original patients had drug-resistant epilepsy.
All scans were acquired at the Life & Brain Center in Bonn on a 3-tesla Siemens system, although two head-coil and acquisition settings were used. FreeSurfer software reconstructed cortical surfaces and extracted thickness, area and volume measurements. Trained operators visually inspected reconstructions and manually corrected inaccuracies.
The lesion distribution was uneven: 62.4% were in the frontal lobe, while temporal, occipital and insular lesions were uncommon. A model can exploit patterns that are frequent in its dataset and perform less well on rare locations. The authors acknowledge that larger international datasets are needed to represent the condition's broader anatomical range.[1]
Nested cross-validation reduced leakage but cannot test transport
The researchers compared 11 supervised classifiers and two feature-reduction approaches. Processing, feature selection, scaling and hyperparameter tuning were fitted inside a nested design with 10 outer folds and five inner folds. This is a careful way to reduce leakage: each outer test participant is evaluated by a pipeline whose fitting did not use that person.
Random forest combined with Elastic Net feature selection produced the highest observed AUC, 0.791 with a 95% confidence interval from 0.720 to 0.855. Random forest with principal-component analysis reached 0.732. Their AUC difference was 0.059, but the confidence interval ran from -0.016 to 0.130 and the reported p value was 0.110, so the data did not establish that the Elastic Net pipeline was superior.
LightGBM with Elastic Net reached AUC 0.765 and was also not significantly different from the leading model. That instability in rankings is common in modest high-dimensional datasets. Nested cross-validation estimates performance within the observed sample; it cannot show that scanner differences, referral patterns, demographics or clinical mimics at another hospital will leave performance intact.[1]
The comparator is easier than the diagnostic decision clinicians face
Healthy controls provide a clear comparison, but a real epilepsy service must distinguish focal cortical dysplasia from other lesions, normal variation, motion artefact and other causes of seizures. It may also need to locate a subtle abnormality precisely enough to inform multidisciplinary review and surgery. A whole-person classification does not demonstrate either differential diagnosis or lesion localisation.
Age distributions differed significantly between the focal-dysplasia and control groups. The researchers tried to remove age-related variation from MRI features inside the cross-validation pipeline and ran a sex sensitivity analysis. They note that linear adjustment may leave nonlinear age effects. Residual confounding can be especially important when a model works with hundreds of correlated anatomical measurements.
The reported neuroanatomical contributors are exploratory associations. Some features were selected consistently across folds, but they should not be read as a map of the causal biology or as evidence that the algorithm has found the lesion itself. A distributed anatomical pattern could help classification while remaining clinically nonspecific.[1]
What this means for clinicians and patients
For neuroradiologists and epilepsy teams, the study supports continued research on quantitative MRI features as a second-reader aid. It does not support autonomous screening, a negative rule-out or surgical planning. Any future tool should display the region and evidence behind an alert, preserve access to the source images and allow a specialist to reject the output.
For a patient with drug-resistant epilepsy, a missed subtle lesion can delay specialist assessment and possible surgery, while a false positive can lead to anxiety and further invasive investigation. Evaluation must therefore report sensitivity for genuinely subtle and MRI-negative lesions, false positives per scan, localisation accuracy and effects on multidisciplinary decisions — not only an overall AUC against healthy controls.
Prospective testing should place the software in the intended clinical workflow and compare decisions with and without assistance. Researchers should record whether the model finds lesions that specialists missed, whether it distracts readers with false alerts, and whether any change improves treatment or patient outcomes. Performance should be stratified by scanner, site, age and lesion location.[1]
Evidence limits and what would change the assessment
The study is peer reviewed and describes a reproducible pipeline using an open dataset, public code and an archived software release. The authors declare no competing interests. Its nested cross-validation and fold-specific preprocessing are stronger than evaluating a fitted model on the same observations used for selection.
The main limitations are the 167-person sample, single-centre acquisition, uneven lesion locations, healthy-control comparator and absence of independent external validation. Manual correction of FreeSurfer outputs also means the input pipeline was not fully automatic. The MRI-negative analysis involved six people and was post hoc. No reader study, diagnosis, lesion-localisation decision, surgical outcome or patient benefit was tested.
The assessment would change with locked-model validation across multiple clinical centres and scanners, using consecutive epilepsy referrals and appropriate disease controls. The decisive evidence would show sensitivity and false-positive burden for subtle lesions, reliable localisation, benefit to specialist readers and improved patient-important outcomes. Until then, this is promising research infrastructure rather than a clinical detection system.[1]