Cortical Segmentation Flaws Explain Major Detection Errors in Epilepsy AI Pipeline

A detailed retrospective investigation into the automated neuroimaging tool MELD Graph has revealed that a substantial portion of its diagnostic mistakes stems from upstream cortical surface reconstruction errors rather than the decision-making network itself. Published in the journal Brain Informatics, the independent failure-mode study examined the computational pipeline across two multicenter magnetic resonance imaging cohorts comprising two hundred and forty-one total participants. Focal cortical dysplasia represents one of the most critical structural abnormalities associated with drug-resistant epilepsy, yet its subtle radiological features often evade human visual inspection during standard clinical reviews. While deep learning algorithms such as graph neural networks hold considerable promise for flagging subtle lesions that human experts overlook, their performance in real-world settings is tightly constrained by preliminary image processing routines. The investigative team determined that focal segmentation breakdowns produced by the underlying FreeSurfer software were directly associated with twenty-one percent of false-negative patient assessments and up to twenty-five percent of false-positive lesion clusters. Furthermore, repairing these defective cortical segmentations allowed the algorithm to correctly identify sixty-seven percent of previously overlooked dysplastic lesions, demonstrating that automated lesion discovery depends fundamentally on the geometrical integrity of the reconstructed cerebral boundaries.

Epilepsy persists as a major contributor to neurological morbidity across both pediatric and adult populations, actively affecting approximately zero point six percent of the global population. When seizures arise from focal cortical dysplasia, surgical resection of the dysplastic cortical tissue offers a potentially curative outcome, yielding seizure freedom categorized as International League Against Epilepsy Class One in up to seventy percent of operated individuals. Achieving these curative outcomes, however, depends entirely on precise radiological localization and prompt referral to specialized epilepsy centers for comprehensive surgical planning. Unfortunately, focal cortical dysplasia frequently presents with exceptionally subtle, ambiguous structural alterations on magnetic resonance imaging scans, causing expert human readers to miss the lesion in more than half of all clinical evaluations. In a benchmark study assessing specialized human detection, expert radiologists achieved an overall sensitivity of only forty-nine percent when reviewing magnetic resonance scans. Because an undetected lesion frequently leads to prolonged medical therapy and missed surgical windows, approximately sixty percent of patients suffering from drug-resistant epilepsy never receive appropriate surgical intervention, compounding the risk of progressive neurocognitive decline in memory, processing speed, and academic functioning.

To overcome the limitations of human visual evaluation, computational neuroscientists have developed automated artificial intelligence systems designed to scan magnetic resonance volumes and flag abnormal cortical patches. Among these emerging tools, MELD Graph has achieved prominence as an open-source, interpretable pipeline that evaluates surface-based morphological metrics and cortical intensity properties using graph neural networks. Trained across heterogeneous imaging archives collected from twenty-three international epilepsy centers, the model was engineered to identify all pathological grades of focal cortical dysplasia. In this newly published failure-mode analysis, researchers evaluated the latest publicly released model implementation, designated version 2.2.2, without performing site-specific model retraining, hyperparameter adjustments, or scan harmonization, thereby mimicking genuine clinical deployment. The evaluation leveraged two distinct public repositories: the Bonn Open Presurgery MRI Dataset, which contributed eighty-five patients with focal cortical dysplasia type two and eighty-five healthy controls, and the Imaging Database for Epilepsy and Surgery, known as IDEAS, which supplied fifty-one confirmed surgical patients and one hundred control examinations. After removing one Bonn participant due to complete surface reconstruction failure and excluding fifteen IDEAS patients whose lesions involved non-neocortical structures like the hippocampus, the primary study cohort comprised one hundred and twenty patients and one hundred and twenty-one matched control subjects.

Assessing the full unharmonized cohort of two hundred and forty-one individuals, the MELD Graph pipeline demonstrated an overall patient-level sensitivity of sixty-eight percent, an overall specificity of forty-three percent, and a positive predictive value of seventy-five percent. The investigational workflow incorporated the automated quality control routine established in the original MELD framework, which searches for severe structural deviations or extreme feature values indicative of widespread imaging artifacts or catastrophic surface failures. This automated quality filter flagged and excluded only three participants—one patient and two controls—leaving a final post-quality-control cohort of two hundred and thirty-eight subjects. Following this exclusion, the system recorded a modest metric shift, displaying a sensitivity of sixty-nine percent, a specificity of forty-four percent, and an identical positive predictive value of seventy-five percent. These operational metrics closely mirror the non-harmonized test results documented in the foundational multicenter trial of the algorithm, confirming that the current architecture maintains stable generalization across external cohorts but continues to generate a substantial burden of false-positive detections and overlooked lesions.

To understand why the algorithm stumbled on specific scans, two experienced neuroradiologists conducted a systematic manual audit within the Freeview interactive visualization environment, overlaying algorithmic predictions and reconstructed boundaries onto three-dimensional anatomical volumes. The reviewers established rigorous, objective failure criteria to distinguish true algorithmic misclassifications from anatomical reconstruction defects generated by FreeSurfer version 7.2.0. A focal FreeSurfer segmentation failure was formally logged whenever more than five millimeters of brain tissue was erroneously omitted from between the anatomical pial surface and the reconstructed boundary, or whenever fifty percent or more of the dysplastic lesion volume was excluded from the reconstructed cortical ribbon. The five-millimeter threshold was deliberately chosen to avoid penalizing minor surface variations that routinely occur during automated boundary extraction. Discrepant visual assessments between the two primary readers were resolved through consensus consultations involving a third independent neuroradiologist. The team observed that when dysplastic lesions were incorrectly segmented, the white matter mask frequently overexpanded outward toward the pial surface, causing abnormal cortical tissue to be incorrectly categorized as subcortical white matter and corrupting the quantitative geometric features supplied to the neural network.

The quantitative audit established that FreeSurfer segmentation breakdowns were responsible for a considerable fraction of all incorrect pipeline outputs. In the pre-quality-control cohort, focal FreeSurfer failures accounted for four out of nineteen patient false negatives, representing twenty-one percent of all missed diagnoses. Similarly, segmentation failures coincided with seven of twenty-eight false-positive predicted clusters in patients, or twenty-five percent, as well as fourteen of eighty-eight false-positive clusters in control participants, representing sixteen percent of control errors. When examining the cohort following automated quality control, focal segmentation failures remained associated with four of nineteen false-negative patients, corresponding to twenty-one percent, six of twenty-seven false-positive clusters in patients, or twenty-two percent, and ten of eighty-two false-positive clusters in controls, representing twelve percent. Crucially, the standard automated quality control mechanism failed to identify these focal errors because it was configured to detect only gross, widespread morphological abnormalities across whole hemispheres. Consequently, automated screening created an unwarranted impression of segmentation adequacy while permitting localized boundary errors to quietly compromise downstream neural network predictions.

Statistical evaluations revealed striking relationships between human-salient radiological features and the likelihood of successful algorithmic detection. Evaluating continuous conspicuity scores on a standardized zero-to-ten scale, the investigators discovered that lesion visibility on T1-weighted images exhibited a statistically significant positive correlation with successful MELD Graph detection, yielding a Pearson correlation coefficient of zero point thirty-five across one hundred and nineteen evaluated cases. Conversely, lesion visibility on T2-weighted fluid-attenuated inversion recovery sequences, commonly known as T2-FLAIR, demonstrated a weaker positive association that failed to attain conventional statistical significance, recording a correlation coefficient of zero point twenty-one across eighty-three patients. More remarkably, calculating the relative conspicuity gradient between T2-FLAIR and T1-weighted sequences yielded a significant negative correlation of negative zero point twenty-four across seventy-six subjects. This negative association confirmed that dysplastic lesions that stood out sharply on T2-FLAIR scans but remained faint or indistinct on T1-weighted acquisitions were significantly more prone to being overlooked by the artificial intelligence model, despite their obvious visibility to human clinical specialists.

The researchers also uncovered an unexpected paradox regarding how lesion appearance affects automated surface reconstruction: the very radiological features that render a lesion conspicuous to human neuroradiologists frequently trigger cortical segmentation breakdowns. Higher human conspicuity scores on both T1-weighted and T2-FLAIR sequences correlated positively with the occurrence of focal FreeSurfer segmentation errors, registering correlation coefficients of zero point thirty-three and zero point thirty-one, respectively. Because focal cortical dysplasia characteristically blurs the histological and radiological boundary between gray and white matter, the intensity-based edge-detection algorithms embedded within FreeSurfer often misplace the cortical ribbon at the lesion epicenter. Morphological analysis further showed that non-bottom-of-sulcus dysplasias carried a dramatically elevated vulnerability to segmentation failure compared to bottom-of-sulcus lesions, exhibiting a phi coefficient of zero point thirty-one and a relative risk of seven point eighty-two. In addition, lower overall T1-weighted image quality correlated significantly with higher rates of segmentation failure both before and after automated quality control, emphasizing that high-fidelity scanner acquisition protocols are essential for stable automated boundary extraction.

To establish whether focal segmentation failures directly caused incorrect algorithmic classifications, the investigators performed a targeted interventional experiment on affected subjects. The neuroradiologists manually edited the white matter segmentations slice by slice across all three anatomical planes using Freeview, adding and removing voxels according to standard editing protocols to restore the true gray-white matter boundary. After recalculating the cortical surface meshes and reapplying MELD Graph under identical operational settings, four out of six previously missed dysplastic lesions—amounting to sixty-seven percent of the segmentation-associated false negatives—were successfully detected by the algorithm. Concurrently, manual surface correction eliminated or reduced segmentation-associated false-positive clusters in six out of eight control subjects, achieving a seventy-five percent correction rate. These restorative results demonstrate that FreeSurfer errors play an active, causal role in misleading the neural network, and confirm that supplying clean, geometrically accurate anatomical surfaces substantially recovers the underlying diagnostic capacity of the machine learning classifier.

The authors emphasized that while these findings highlight immediate pathways for technical refinement, several methodological limitations must be considered. Because manual corrections were conducted with prior knowledge of true lesion coordinates, the findings reflect proof-of-principle efficacy rather than the automated performance gains of an untended pipeline. The authors suggested that replacing legacy surface reconstruction pipelines with modern deep learning segmentation frameworks—such as nnU-Net, SegResNet, or SwinUNETR—could potentially eliminate many localized surface defects without requiring laborious human intervention. Given the observed specificity of forty-four percent, the researchers concluded that automated lesion detection tools should currently serve as sensitivity-enhancing secondary readers rather than autonomous diagnostic decision-makers in clinical practice. Prioritizing sensitivity remains clinically justified when searching for surgically treatable causes of disabling epilepsy, provided that expert neuroradiologists review flagged cortical clusters alongside raw multi-sequence examinations. Integrating T2-FLAIR features directly into the core architectural representations of future graph models will likely prove essential for closing the gap between artificial intelligence performance and expert human perception.

Subject of Research: Failure-mode evaluation of automated artificial intelligence detection of focal cortical dysplasia in magnetic resonance imaging

Article Title: Pipeline evaluation of a state-of-the-art AI algorithm for detection of focal cortical dysplasia: insights into potential failure sources

Article References: Esmeraldo, M. A., Chambers, S., Kravutske, Y., Reis, E. P., Kasprian, G., Geraldo, A. F., Gatidis, S., & Soares, B. P. (2026). Pipeline evaluation of a state-of-the-art AI algorithm for detection of focal cortical dysplasia: insights into potential failure sources. Brain Informatics, 13(1), Article 13. https://doi.org/10.1186/s40708-026-00299-w

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00299-w

Keywords: artificial intelligence, focal cortical dysplasia, epilepsy, neuroimaging, magnetic resonance imaging, FreeSurfer, graph neural networks, medical image segmentation, Pipeline, evaluation, state-of-the-art, algorithm