人工智能统计建模刻画局灶性皮质发育不良的构筑扭曲Artificial intelligence‐based statistical modeling delineates architectural distortion in focal cortical dysplasia

Abstract

In epilepsy patients lacking magnetic resonance imaging‐visible lesions such as focal cortical dysplasia (FCD), surgical resections target regions with abnormal electroencephalographic (EEG) activity. However, histopathological identification of subtle cortical architectural abnormalities in these specimens remains challenging. We investigated whether artificial intelligence (AI)‐based morphometric and spatial analysis of NeuN‐stained cortical sections could detect neuronal architectural disorganization in epilepsy resections, including regions without definitive histologic dysplasia. Whole slide images were generated from 83 FCD regions and 19 neurologically normal autopsy controls. Regions of interest were annotated in QuPath as FCD, FCD‐adjacent, FCD‐distant, apparently normal (abnormal EEG without histologic dysplasia), and true normal (autopsy controls). NeuN‐positive neurons were detected using QuPath (90% sensitivity, 12% false positive rate). Spatial and morphometric features—including neuronal clustering, distribution inhomogeneity, and nuclear morphology—were extracted and used to train a multinomial, Least Absolute Shrinkage and Selection Operator‐regularized regression classifier. The classifier achieved 70.6% accuracy in subregion classification and 88.2% accuracy in overall specimen diagnosis. Notably, regions with abnormal EEG but lacking histologic dysplasia exhibited quantifiable architectural disorganization similar to those seen in areas adjacent to FCD and distinct from control tissue. AI‐driven analysis of neuronal morphology and spatial distribution reveals subtle cortical disorganization in epilepsy resections, including in histologically ambiguous regions. Further investigation is warranted to determine if this methodology can enhance the diagnostic yield of neuropathological evaluation and support more precise surgical targeting in epilepsy.

INTRODUCTION

Approximately one third of patients with epilepsy have seizures that are resistant to medical management [1] which is associated with poor socioeconomic and clinical outcomes [2]. Among patients with drug‐resistant focal epilepsy, focal cortical dysplasia (FCD) and hippocampal sclerosis are the most common underlying pathologies, both of which may be amenable to surgical resection [3]. While the majority of epileptogenic foci are detected by magnetic resonance imaging (MRI) studies, a subset of patients has MRI‐negative foci. In these MRI‐negative cases, surgical resections target foci of abnormal electroencephalographic (EEG) activity. However, surgical resection in these cases is less likely to be performed, and post‐surgical outcomes are less favorable in terms of seizure freedom [4, 5, 6].

The histopathologic classification of FCD yields considerable prognostic information, and the histologic diagnosis of FCD portends favorable surgical outcomes. The International League Against Epilepsy (ILAE) classification defines three overarching classes of FCD:

• FCD type 1 (FCD 1): Cortical architectural distortion without dysmorphic neurons or balloon cells,

• FCD type 2 (FCD 2):

  • architectural abnormalities with dysmorphic neurons (FCD 2A)

  • Architectural abnormalities with dysmorphic neurons and balloon cells (FCD 2B), and

• FCD type 3: (architectural abnormalities associated with another principal lesion) [7].

The surgically resected samples are assigned to the most inclusive class of FCD identified as the overall histopathologic diagnosis (i.e., the overall diagnosis is that of the highest class of FCD observed). Critically, surgical outcomes vary by FCD subtype; approximately 50% of patients diagnosed with FCD 1 achieve Engle class I outcomes at 1 year compared to 69% for FCD 2 [8, 9]. In MRI‐negative cases, histopathological diagnosis becomes especially critical for identifying the underlying etiology and informing prognosis.

Despite its importance, the histological diagnosis of FCD remains challenging. The ILAE criteria can be difficult to apply in routine practice, leading to substantial interobserver variability and diagnostic uncertainty [10, 11]. Moreover, histopathologic examination fails to reveal an underlying lesion in approximately 5%–10% of surgically resected focal epilepsy cases, highlighting the challenge of histologic diagnosis of FCD [8, 12]. While recent advances in molecular diagnostics have identified somatic and germline mutations associated with malformations of cortical development, these findings have primarily aided in the diagnosis of more overt FCD 2 lesions, with limited utility in detecting subtler abnormalities [13].

These challenges underscore the need for objective, reproducible tools to support the histopathological evaluation of epilepsy resections [14]. To this end, we investigated whether a machine learning classifier utilizing artificial intelligence (AI)‐quantified features of cytomorphology and neuronal distribution could assist in detecting subtle architectural abnormalities suggestive of FCD. We applied this approach to a cohort of epilepsy resections lacking definitive MRI evidence of cortical dysplasia, using AI to quantify cortical architectural and neuronal cytological features which were subsequently used to train a Least Absolute Shrinkage and Selection Operator (LASSO)‐regularized multinomial regression classifier.

MATERIALS AND METHODS

Institutional Review Board, specimens, and raw data collection

This study was approved by our Institutional Review Board and conducted in accordance with the Declaration of Helsinki. Surgical resection specimens were obtained from 29 patients with MRI‐negative, drug‐resistant focal epilepsy. Nineteen age‐matched, neurologically healthy autopsy specimens from comparable cortical regions were selected as controls. All tissue samples were obtained from our tissue archives. A blinded histopathological review of hematoxylin‐and‐eosin (H&E) stained sections from all surgical cases was performed by a board‐certified neuropathologist. Each case was classified according to the ILAE 2022 consensus classification for the most advanced degree of FCD present, as routinely performed for these cases. Additionally, because FCD lesions are spatially heterogenous with distinct tissue regions fulfilling diagnostic criteria for different classes of FCD [15, 16] each tissue subregion of interest was independently classified according to the morphologic features present only within that subregion. Representative sections (n = 83) were selected to reflect the overall lesional architecture. Autopsy control cases were confirmed to be non‐lesional by histological review, and representative sections were similarly selected.

All selected sections underwent immunohistochemical (IHC) staining for NeuN (clone A60; Zeta Corp., Sierra Madre, CA). Whole‐slide images (WSIs) of both H&E and NeuN‐stained sections were generated. Using QuPath (version 0.4.2) [17], region of interest (ROIs) were annotated as follows: (Figures 1 and Data S1):

• FCD 2A or 2B—regions with definitive FCD type 2A or FCD type 2B dysplasia,

• FCD‐adjacent—cortex adjacent to but not involved by FCD,

• FCD‐distant—cortex with normal architecture, spatially separated from dysplasia,

• Apparently normal—cortex from epilepsy resections targeting EEG ictal onset zones, without definitive histologic dysplasia, and

• True normal—histologically normal cortex from autopsy controls.

The QuPath positive cell detection algorithm was optimized for NeuN‐stained WSIs and applied to each ROI to extract positional and morphometric data for all NeuN‐positive neurons. Algorithm performance was validated against manual cell classifications performed by a board‐certified neuropathologist on a subset of 12 ROIs. Detection performance across histologic categories was compared using Kruskal–Wallis tests.

Feature generation, feature selection, and model generation

Spatial and morphologic data for all classified cells were extracted from QuPath and analyzed in R (version 4.2.1) using tidyverse and spatstat packages [18, 19]. Morphometric features of individual cells were aggregated within each ROI and summarized using statistical descriptors including mean, median, maximum, minimum, standard deviation, and range, across both NeuN‐positive and ‐negative cell populations. Spatial coordinates of NeuN‐positive cells, along with manually defined observation windows corresponding to each ROI, were used to generate planar point patterns. These patterns were used to calculate the global intensity of the neuronal distribution, and summary functions including L(r), a linear transformation of Ripley's K function, Pair correlation function, PCF(r), and the nearest‐neighbor distribution, G(r) (Figure 2 provides a brief graphical guide to the interpretation of L(r), G(r), and PCF(r)). Each summary function was estimated at 513 fixed radii across all specimens and used as input features for model development. Kernel‐smoothed estimates of local neuronal intensity were generated using a Gaussian kernel (σ = 50) with pixel resolution corresponding to 7 × 7 pixels of the original image. From these intensity maps, additional features were extracted, including global minima and maxima, number and size of intensity hotspots (defined as contiguous pixel clusters exceeding the 99th percentile of intensity), and hotspot contours (see Supporting Information S1: Data 1 for a complete feature list).

To mitigate potential confounding effects of differential cold ischemic time between surgical and autopsy specimens, morphometric features were normalized by calculating the difference and ratio between NeuN‐positive and ‐negative cell populations. These derivative features were statistically compared using Kruskal–Wallis tests, followed by post hoc pairwise comparisons using Mann–Whitney U tests. Parent features were excluded from modeling if their derivative features met the following criteria:

Exclusion Criteria A

• Significantly different Kruskal–Wallis test for at least one derivative feature,

  AND

• All pairwise comparisons between autopsy controls (true normal/non‐seizure) and surgical specimens were found to be significant,

  AND

• No additional significant pairwise comparisons were observed between surgically resected specimen groups.

  OR

Exclusion criteria B

• Significantly different Kruskal–Wallis test for at least one derivative feature,

  AND

• At least five of seven significant pairwise comparisons between autopsy controls and resected specimen categories,

  AND

• At least one significant pairwise comparison of derivative features being the comparison between apparently normal and true normal/non‐seizure specimens,

  AND

• No additional significant pairwise comparisons were observed between surgically resected specimen groups.

Additionally, all features derived from NeuN‐negative cell morphometry, or from ratios or differences involving negative‐cell features, were excluded. This was because of the lack of optimization and validation of the negative‐cell detection algorithm against neuropathologist annotations (Supporting Information S2: Data 2). Spatial features were exempt from this filtering process, as they are independent of staining quality and cell preservation.

Cases and controls were randomly partitioned into training (80%) and test (20%) sets using the createDataPartition function from the caret package, ensuring representative distribution across histologic categories. Cases with multiple, spatially distinct regions of interest were not restricted to fall entirely within the test or training sets (Data S3). A LASSO‐regularized multinomial logistic regression model was trained on the extracted features using the glmnet package. The regularization parameter (λ) was optimized via five‐fold cross‐validation to minimize multinomial classification error. The final model was applied to the held‐out test set using the feature coefficients corresponding to the λ value that minimized misclassification in training. Model performance was assessed based on misclassification rates for both overall specimen diagnosis and exact histologic classification within each ROI.

Analysis and comparison of model features across histologic classifications

Neuronal morphologic features derived from QuPath, along with all inhomogeneity‐related features selected for model inclusion, were statistically analyzed using Kruskal–Wallis tests. Post hoc pairwise comparisons were performed using Mann–Whitney U tests, with Benjamini–Hochberg (BH) correction applied to control for multiple comparisons.

For features derived from secondary point pattern functions—PCF(r), L(r), and G(r)—individual estimates at specific radii were used as model features. To assess spatial relationships more comprehensively, the full distributions of these functions across all radii were compared between groups using permutation tests. Comparisons were conducted between the following regions:

• True normal and apparently normal specimens;

• True normal and each dysplasia‐associated subregion (FCD, FCD‐adjacent, FCD‐distant); and

• All dysplasia‐associated subregions against one another.

Each function was first evaluated in a group‐wise manner, followed by planned pairwise comparisons where applicable.

RESULTS

Validation and performance of the cell classifier

The performance of the optimized QuPath classifier for NeuN‐positive and ‐negative cell detection was evaluated against manual annotations performed by a board‐certified neuropathologist on 12 representative tissue regions. A nearest‐neighbor (NN) threshold of 15 μm was used to define concordant cell annotations between the algorithm and manual reference. Visual inspection demonstrated a high degree of concordance between algorithmically and manually identified NeuN‐positive neurons (Data S4). Quantitative comparison (Table 1) revealed minimal deviations in median nearest‐neighbor distances, likely attributable to minor inconsistencies in manual centroid placement. Sensitivity of the algorithm ranged from 72.5% to 96.8%, with false positive rates consistently below 20% across all regions, using manual annotation as the reference standard. Importantly, classifier performance did not significantly differ across histologic subtypes. Sensitivity and false positive rates were consistent among all specimen groups (p = 0.68 and p = 0.59, respectively; Figure 3), indicating that the algorithm performs robustly across a spectrum of cortical architectures. Collectively, these findings support that the classifier is accurate, generalizable, and lacks systematic bias.

Feature selection and model generation

From the morphologic data extracted via QuPath, 27 cell‐level features were initially derived and statistically summarized, resulting in 508 primary and derivative sample‐level features across NeuN‐positive and ‐negative cell populations. After excluding features potentially influenced by differential cold ischemic time and those derived from NeuN‐negative cells, 84 morphologic features were retained for modeling (Figure S2, Supporting Information S2: Data 2). In addition, 513 features were generated from each secondary point pattern summary function, representing function estimates at discrete radii. A further 35 features were derived to quantify inhomogeneity in the spatial distribution of cortical neurons. In total, 1662 features were included in the initial modeling dataset.

A LASSO‐regularized multinomial regression model was trained using 80% of the dataset, with five‐fold cross‐validation to optimize the regularization parameter (λ) for minimizing misclassification error (Data S5). Model performance was evaluated on the remaining 20% test set using the penalized coefficients corresponding to the optimal λ. Performance metrics of the final model are summarized in Table 2. The classifier achieved an overall diagnostic accuracy of 88.2% and a subregion‐level classification accuracy of 70.6%. Most misclassifications at the subregion level involved assigning a more dysplastic label than ground truth. These findings suggest that the classifier may be sensitive to subtle architectural disruptions not readily apparent on routine histology, potentially offering a complementary tool for neuropathological assessment in diagnostically ambiguous cases.

Analysis of model features

To better understand the discriminatory features used by the classifier, we examined how morphologic features varied across histologic subregions. All non‐zero model coefficients are listed in Supporting Information S3: Data 3, and a heatmap of centered and scaled median values for each morphologic feature is shown in Figure 4 with a detailed statistical summary in Supporting Information S4: Data 4. As expected, FCD 2B regions exhibited distinct cytomorphologic profiles, including elevated average minimum hematoxylin optical density (O.D.), increased maximum diaminobenzidine (DAB) standard deviation, larger median nuclear area, greater nuclear maximum caliper, and longer nuclear perimeter.

Interestingly, FCD 2B‐adjacent regions displayed cytomorphometric profiles closely resembling those of FCD 2A, suggesting that morphologic alterations may extend beyond the histologically defined lesion. Similarly, “apparently normal” regions—those lacking definitive histologic dysplasia but targeted for resection based on EEG abnormalities—showed morphologic similarity to FCD 2A‐adjacent and distant regions, particularly in nuclear staining characteristics (Figure 4, Supporting Information S4: Data 4). In contrast, true normal (autopsy‐derived) cortex appeared morphologically distinct from all surgical specimens. These findings suggest that cytomorphometric features may capture subtle, spatially graded changes associated with proximity to dysplastic foci.

We next evaluated spatial relationships between NeuN‐positive neurons using inhomogeneous L, pair correlation (PCF), and nearest‐neighbor (G) functions. These functions were plotted across radii for each histologic subregion and compared using permutation tests. Both L and PCF functions deviated markedly from complete spatial randomness (CSR) up to ~30 μm, indicating that neuronal distribution in adult cortex is inherently non‐random (Figure 5A,B). L(r) revealed strong neuronal dispersion at radii <10 μm, while PCF(r) indicated clustering between 15 and 25 μm. Finally, G(r) showed similar dispersion at small radii, converging toward CSR expectations at larger distances (Figure 5C).

Comparative analysis of spatial summary functions across histologic subregions revealed distinct differences in neuronal organization. Specifically, FCD 2A and adjacent regions exhibited greater regularization of neuronal spacing at small radii and a more abrupt transition to clustering at intermediate distances compared to true normal (non‐seizure) cortex (Figure 5A). This spatial pattern was also observed in “apparently normal” regions, as well as in cortex adjacent to and distant from FCD 2B, as reflected in the pair correlation and nearest‐neighbor distribution functions (Figure 5B,C).

In contrast, FCD 2B regions exhibited an inverse spatial pattern, characterized by a lower degree of neuronal regularization at small radii and a transition to clustering at larger radii compared to true normal cortex (Figure 5B, right). Notably, both FCD 2A and FCD 2B regions demonstrated a reduced overall magnitude of neuronal clustering compared to true normal tissue and to regions adjacent to or distant from dysplastic foci (Figure 5A,B). These findings suggest that while dysplastic regions may exhibit architectural disorganization, they are also marked by a relative loss of local clustering, potentially reflecting altered developmental or degenerative processes.

The G function further highlighted two key patterns. First, FCD 2A and FCD 2B‐adjacent regions shared similar nearest‐neighbor distance distributions. Second, FCD 2B showed a nearest‐neighbor profile closely resembling that of true normal cortex. These findings suggest a complex spatial reorganization in dysplastic cortex, where (1) dysplastic regions exhibit reduced clustering, possibly because of neuronal loss or dispersion, and (2) adjacent regions show increased clustering, potentially reflecting compensatory or transitional architecture. Critically, “apparently normal” regions mimic the spatial disorganization seen in dysplasia‐adjacent tissue, despite lacking overt histologic abnormalities.

Given that the various grades of FCD are defined by both cytologic atypia and disruption of cortical architecture—features that inherently lead to inhomogeneous neuronal distribution—we generated a set of features to quantify spatial inhomogeneity within each ROI. Of these, four features were selected by the model as discriminatory for histologic subregion classification (Figure 6, Supporting Information S1 and S5: Data 1 and 5).

Analysis of these features revealed that FCD 2A regions exhibited significantly reduced inhomogeneity compared to both apparently normal and true normal (non‐seizure) cortex. Specifically, FCD 2A regions showed lower maximum positive intensity deviation, fewer total hotspots, and hotspots per unit area, and smaller average hotspot size (Figure 6). These findings suggest that FCD 2A is characterized by relatively few, moderately sized regions of increased neuronal density that deviate less from the global intensity baseline—indicating an abnormally more uniform cortical architecture. In other words, FCD 2A appears to exhibit a reduction in the degree of spatial inhomogeneity typically seen in normal adult cortex. Conversely, apparently normal regions demonstrated a trend toward increased inhomogeneity, particularly in maximum positive intensity deviation (Figure 6A). This suggests the presence of localized areas of elevated neuronal density that exceed the variability observed in true normal cortex, potentially reflecting subtle architectural abnormalities not recognized by conventional histological evaluation.

DISCUSSION

The diagnosis of FCD relies on the identification of cytologic abnormalities and disruption of cortical neuronal architecture—features that are often difficult to detect, particularly in MRI‐negative epilepsy cases. AI offers a powerful means of quantifying both cytomorphologic and spatial features across entire tissue sections with a level of precision and scale that exceeds traditional histologic evaluation. Spatial statistical analysis, in turn, enables the detection of patterns in neuronal distribution that emerge only when considering the relative positions of thousands of cells simultaneously. In this context, the features used by our AI‐based classifier are conceptually aligned with those employed by neuropathologists—such as cortical architecture, nuclear size and shape—but differ in their precision and scope. For example, features like mean nuclear perimeter are directly analogous to visual assessments made during routine histology, while others, such as pair correlation functions or spatial inhomogeneity metrics, represent mathematically derived descriptors of neuronal organization that are not accessible through conventional microscopy. We hypothesized that a classifier integrating these AI‐derived cytologic and spatial features could detect subtle but diagnostically meaningful patterns associated with FCD.

Indeed, our LASSO‐regularized multinomial regression model achieved an overall diagnostic accuracy of 88.2% and a subregion classification accuracy of 70.6%. These results are comparable to those reported by Kubach et al., who used a convolutional neural network to distinguish FCD 2B from cortical malformations in tuberous sclerosis complex, achieving 91% accuracy in a binary classification task [20]. Notably, our model achieved similar performance despite using a simpler statistical framework and addressing a more complex multi‐class classification problem.

Importantly, the classifier was able to distinguish between true normal (non‐seizure) cortex and tissue from patients with drug‐resistant epilepsy that lacked definitive histologic abnormalities. Moreover, the model tended to classify ambiguous subregions as more closely related to dysplastic foci, suggesting that it may be sensitive to subtle architectural changes that fall below the threshold of conventional histopathologic detection. These findings may relate to the hypothesis that dysplasia‐associated changes extend beyond the boundaries of histologically defined lesions and may contribute to epileptogenicity in adjacent or apparently normal cortex.

Perhaps most compelling is the observation that spatial features in “apparently normal” cortex from epilepsy patients closely resemble those in dysplasia‐adjacent regions. Critically, these changes were not identified in the cortical regions of specimens from neurologically healthy autopsy controls. Given the consistent performance of our cell detection algorithm across surgical and autopsy specimens, these findings are less likely to be artifacts of tissue processing. Instead, they may reflect subtle cortical abnormalities that are not captured by unassisted examination of routine histology. Alternatively, they may represent secondary changes induced by chronic seizure activity. In either case, further investigation will be essential to determine the reliability and significance of these features.

If properly validated, the clinical and biological implications of these findings would be significant. From the diagnostic perspective, features derived from spatial analysis could be used to provide objective evidence of FCD lesions in cases with subtle histologic findings improving diagnostic reproducibility and confidence as well as augmenting the diagnostic yield of epilepsy resections, particularly in MRI‐negative cases. In addition to the benefits of more reliable disease classification for prognostication, reliable identification of the subtle spatial alterations consistent with FCD would facilitate more in‐depth study of the prognostic significance of residual dysplastic regions aligning with the current hypotheses that epileptogenic zones may extend beyond the core lesion. Finally, these features may enhance the ability to isolate dysplastic foci thereby facilitating simpler and more conclusive molecular characterization of subtle FCD lesions which remains a significant barrier to the etiologic understanding of FCD type 1.

However, in their current state, the findings of this study are preliminary and as such, their significance is somewhat speculative. Validation of these findings in a large cohort of epilepsy resections including patients with diverse pathologies, autopsy controls, and if possible non‐seizure brain tissue from surgical resections for other conditions will be needed to characterize the extent to which the observations in the present study are specific to, and sensitive for FCD. As a corollary, such studies may identify changes that are associated with chronic seizure activity as opposed to a specific underlying etiology. If validated, the clinical significance of these observations must be explored through detailed correlation of spatial parameters with post‐operative patient outcomes.

Our findings also highlight the potential of spatial analysis as a tool for understanding cortical organization in epilepsy. While the spatial distribution of neurons in FCD has not been extensively studied, prior work has demonstrated altered spatial patterns of glial cells and vasculature in FCD 2 lesions [21, 22, 23]. For example, studies have shown reduced oligodendroglial progenitor cell density and increased microvascular density [24] in dysplastic cortex compared to perilesional tissue. These observations, together with our findings, suggest that spatial relationships among neurons, glia, and vasculature may provide a rich source of diagnostic and mechanistic insight.

While the classifier developed in this study performed well on a relatively small training and test set, several avenues remain for improving model performance and generalizability. In the current approach, neuronal point patterns were modeled as stationary, inhomogeneous, isotropic point processes. This assumes isotropy—that spatial relationships are directionally uniform. Future models could incorporate anisotropic spatial features to capture potential directionality in neuronal interactions [25]. Similarly, our classifier utilized discrete estimates from secondary point pattern summary functions (e.g., L(r), PCF(r)) as features. In follow‐up analyses, we observed that the rate of change of these functions varied across histologic subregions, suggesting that derivatives of these functions may offer further discriminatory power. Expanding the feature space to include such dynamic spatial metrics could enhance model sensitivity to subtle architectural differences.

From a modeling perspective, we used LASSO‐regularized multinomial regression because of its interpretability and suitability for high‐dimension data with limited sample size. However, more sophisticated classifiers—such as random forests, gradient‐boosted trees, or convolutional neural networks—may yield improved performance, particularly in larger datasets. As our understanding of this feature space matures, these approaches could be explored to develop more robust models.

Another limitation lies in the manual annotation of ROIs, which is time‐intensive and potentially introduces variability and dilutes dysplasia‐associated features. Automating ROI selection through algorithmic segmentation or weakly supervised learning could improve reproducibility and scalability for clinical application.

The use of autopsy‐derived tissue as “true normal” controls also warrants consideration. While autopsy specimens were selected to ensure the absence of neuropathology and to match cortical regions sampled in surgical cases, differences in tissue processing and postmortem interval may influence cytomorphometric features. Although we attempted to mitigate this by excluding features sensitive to cold ischemic time, further validation—potentially in animal models—would help clarify the impact of postmortem changes on the model. Similarly, the use of additional control material such as non‐lesional tissue obtained from resections for causes other than epilepsy, in addition to autopsy controls, will likely help to identify the significant confounders that would be present in each of these types of material.

Notably, FCD type I was not explicitly represented in this dataset because of the inherent difficulty in confidently diagnosing these regions in routine surgical specimens. To address this, we included “apparently normal” and “dysplasia‐adjacent” regions to capture a spectrum of cortical abnormalities that may fall below the threshold for definitive diagnosis of FCD. A larger, consensus‐classified cohort incorporating unsupervised clustering may be necessary to define features specific to FCD 1, microdysgenesis, and related entities. Finally, the current study focused on a relatively narrow set of pathologies. Future work should expand the classifier to include additional epilepsy‐associated lesions such as hippocampal sclerosis, mesial temporal sclerosis, FCD type III, and tuberous sclerosis complex. This would allow for broader diagnostic applicability and further test the generalizability of the feature set.

In summary, we developed a classifier of FCD histomorphology based on AI‐quantified neuronal cytomorphometric and spatial distribution features. The model demonstrated promising performance in training and test sets, suggesting the potential for clinical utility of this methodology. Moreover, analysis of the most discriminatory features revealed trends across histologic subregions, which may prove to relate to the underlying biology of FCD upon further investigation. With continued refinement and validation in larger, more diverse cohorts, this approach may ultimately enhance the diagnostic precision and prognostic value of neuropathologic evaluation in epilepsy surgery.