Surfaced‐based detection of focal cortical dysplasia using magnetic resonance fingerprinting and machine learning

Abstract

Objective

This study was undertaken to develop a framework for focal cortical dysplasia (FCD) detection using surface‐based morphometric (SBM) analysis and machine learning (ML) applied to three‐dimensional (3D) magnetic resonance fingerprinting (MRF).

Methods

We included 114 subjects (44 patients with medically intractable focal epilepsy and FCD, 70 healthy controls [HCs]). All subjects underwent high‐resolution 3‐T MRF scans generating T1 and T2 maps. All patients had clinical T1‐weighted (T1w) images; 35 also had 3D fluid‐attenuated inversion recovery (FLAIR). A 3D region of interest (ROI) was manually created for each lesion. All maps/images and lesion ROIs were registered to T1w images. Surface‐based features were extracted following the Multi‐center Epilepsy Lesion Detection pipeline. Features were normalized using intrasubject, interhemispheric, and intersubject z‐scoring. A two‐stage ML approach was applied: a vertexwise neural network classifier for lesional versus normal vertices using T1w/MRF/FLAIR features, followed by a clusterwise Random Undersampling Boosting classifier to suppress false positives (FPs) based on cluster size, prediction probabilities, and feature statistics. Leave‐one‐out cross‐validation was performed at both stages.

Results

Using T1w features, sensitivity was 70.4% with 11.6 FP clusters/patient and 4.1 in HCs. Adding MRF reduced FPs to 6.6 clusters/patient and 1.5 in HCs, with 68.2% sensitivity. Combining T1w, MRF, and FLAIR achieved 71.4% sensitivity, with 4.7 FPs/patient and 1.1 in HCs. Detection probabilities were significantly higher for true positive clusters than FPs (p < .001). Type II showed higher detection rates than non‐type II. Magnetic resonance imaging (MRI)‐positive patients showed higher detection rates and fewer FPs than MRI‐negative patients. Seizure‐free patients demonstrated higher detection rates than non‐seizure‐free patients. Subtyping accuracy was 80.8% for non‐type II versus type II, and 68.4% for IIa versus IIb, although limited by small sample size. The transmantle sign was present in 61.5% of IIb and 40% of IIa cases.

Significance

We developed an ML framework for FCD detection integrating SBM with clinical MRI and MRF. Advances include improved FP control and enhanced subtyping; selected model outputs may provide indicators of detection confidence and seizure outcome.

Key points.

• We developed a two‐stage ML framework for FCD detection using surface‐based processing of 3D MRF, including 44 FCD patients and 70 healthy controls.

• T1w + MRF + FLAIR yielded 71.4% sensitivity, 4.7 FPs/patient, and 1.1 FPs/control.

• The classification effectively reduced FPs in patients and controls.

• Sensitivity was higher in type II and seizure‐free patients; detection probabilities were greater for TP than FP clusters.

• Subtyping accuracy was 80.8% for non‐type II versus type II, and 68.4% for IIa versus IIb (limited by small sample size).

1. INTRODUCTION

Focal cortical dysplasia (FCD) is a frequently encountered etiology in medically intractable epilepsy, in both pediatric and adult patients undergoing surgery. ¹ , ² Although magnetic resonance imaging (MRI) features such as cortical thickness anomalies, blurred gray–white (GW) matter junction, and signal alterations on T1‐weighted (T1w), T2‐weighted (T2w), or fluid‐attenuated inversion recovery (FLAIR) sequences can be indicative, many subtle FCD lesions remain imperceptible to visual inspection. ³ , ⁴ FCD is therefore often the underlying pathology in “nonlesional” epilepsy cases. ⁵ Additionally, the identification of FCD is subject to variability among observers, necessitating a high reliance on clinicians with specialized expertise and years of experience. The consequences of failing to detect FCD are substantial, potentially leading to patients not being considered for surgery, suboptimal surgical planning, and compromised postoperative outcomes. FCD subtyping based on visual inspection of the clinical MRI also remains difficult due to overlapping radiological features across subtypes. ⁶ Addressing this challenge is important, as different subtypes have distinct histopathological and surgical outcome differences. ⁷ , ⁸ Overall, these diagnostic challenges underscore the need for advanced imaging techniques to enable noninvasive FCD identification and subtyping.

Magnetic resonance fingerprinting (MRF) is a novel quantitative MRI (qMRI) method, enabling simultaneous acquisition of multiple tissue property maps, ⁹ including T1, T2, proton density maps, as well as gray matter (GM) and white matter (WM) tissue fraction maps. ¹⁰ Compared to conventional qMRI, MRF offers superior robustness against measurement errors from hardware variations, ¹¹ , ¹² while significantly reducing scan time compared to traditional qMRI techniques. ⁹ Recent studies showcased MRF's potential in enhancing diagnostic accuracy in epilepsy, including improved detection of hippocampus sclerosis, ¹³ impaired WM in temporal lobe epilepsy, ¹⁴ and characterization of FCD. ¹⁰ , ¹⁵ Currently, whole‐brain, high‐resolution three‐dimensional (3D) MRF sequences can be achieved in 5–10 min, ¹⁰ , ¹⁶ which makes it increasingly feasible for clinical use.

MRI postprocessing techniques play a vital role in enhancing the diagnostic capabilities of clinical MRI for presurgical evaluation. ⁴ Surface‐based morphometry (SBM) is a commonly used MRI postprocessing technique that has already been shown to improve FCD detection when applied to clinical MRI. ¹⁷ , ¹⁸ , ¹⁹ In this study, we aimed to establish a machine learning (ML) framework for whole‐brain FCD detection using SBM processing of clinical MRI and high‐resolution 3D MRF data. We evaluated the framework's lesion detection sensitivity, false positive rate, and subtyping accuracy. We also explored potential markers of detection confidence and seizure outcome derived from this framework, with the goal of supporting surgical planning and outcome prediction.

2. MATERIALS AND METHODS

2.1. Study design and patient selection

Beginning in 2017, we enrolled patients at the Cleveland Clinic Epilepsy Center for an MRF research scan. Inclusion criteria were as follows: (1) medically intractable focal epilepsy, (2) clinical suspicion of FCD, and (3) surgery candidate per consensus at the patient management conference (PMC). Patients with metal implants or those unable to provide informed consent/assent were excluded. The current study was a retrospective analysis that included patients from 2017 to 2024 with pathologically/radiologically confirmed FCD. Patients were excluded if they had severe imaging artifacts preventing analysis. This study was approved by the Cleveland Clinic Institutional Review Board, and all subjects provided written informed consent prior to the MRF scan. Findings from this research did not affect clinical decision‐making. We also recruited healthy controls (HCs) matched for age and gender to the patients.

2.2. MRI acquisition and reconstruction

A 3D whole‐brain MRF sequence ¹⁰ was performed on a Siemens 3‐T Prisma with a 20‐channel head coil (field of view = 300 × 300 × 144 mm³, resolution = 1 × 1 × 1 mm³, scan time = 10.4 min). A 3D whole‐brain B1 mapping sequence was performed with the same field of view and resolution as the MRF sequence to correct for B1 inhomogeneity, with a scan time of 1.8 min. T1 and T2 values were generated based on dictionary‐based reconstruction. MRF T1w images were synthesized from the T1 maps, with similar contrast to the conventional T1w magnetization‐prepared rapid gradient echo (MPRAGE) images. A clinical 3D T1w MPRAGE scan was acquired with resolution = .5 × .5 × .9 mm³, repetition time = 1930 ms, echo time = 2.57 ms, inversion time = 1100 ms, 192 slices, and scan time = 3 min 53 s. 3D FLAIR was acquired with resolution = 1 × 1 × 1 mm³, repetition time = 5000 ms, echo time = 93 ms, and scan time = 7 min 37 s.

2.3. Feature generation

We utilized the Multi‐center Epilepsy Lesion Detection (MELD) surface‐based feature extraction pipeline ¹⁷ , ²⁰ to generate SBM features. FreeSurfer version 7.2.0 was used to create cortical reconstruction ²¹ based on T1w MPRAGE images. The following surface‐based features were calculated: cortical thickness, mean curvature, intrinsic curvature, GW contrast, and sulcal depth. To calculate surface‐based signal intensity features, the MRF T1 and T2 maps as well as 3D FLAIR images were registered to the T1w images using a symmetric image normalization method in Advanced Normalization Tools. ²² Signal intensity on these sequences was sampled at different depths from 0%, 25%, 50%, and 75% of the cortex from GW junction to the GM surface, as well as .5‐ and 1‐mm projections into the WM. The asymmetry maps of these features were generated by the subtraction between left‐sided and right‐sided hemispheres. Surface‐based data were smoothed using a Gaussian kernel: 20 mm for intrinsic curvature and 10 mm for other features, following prior methods. ¹⁷ Normalization procedures with intrasubject, interhemispheric and intersubject z‐scoring using the age‐ and‐gender matched HCs were performed on all features. This workflow is shown in Figure 1A.

FIGURE 1.

Study workflow outlining (A) feature generation, (B) first‐stage classification, and (C) second‐stage classification. 3D, three‐dimensional; ANTs, Advanced Normalization Tools; CV, cross‐validation; D, dense layer; FLAIR, fluid‐attenuated inversion recovery; FP, false positive; GM, gray matter; HC, healthy control; LOO, leave‐one‐out; MRF, magnetic resonance fingerprinting; ROI, region of interest; RUS, Random Undersampling Boosting; SBM, surface‐based morphometric; T1w, T1‐weighted; TP, true positive; WM, white matter.

2.4. Lesion region of interest definition

For model training and validation, we utilized 3D regions of interest (ROIs) for lesion labels, consistent with prior research. ¹⁵ , ²³ The lesion ROIs were created by a neuroimaging expert (Z.I.W.) and were subsequently reviewed and confirmed by a board‐certified neuroradiologist (S.E.J.) with 18 years of experience in epilepsy imaging. Lesion ROI delineation was guided by the official consensus on MRI findings from the PMC, incorporating regions with signal or morphometric changes observed on preoperative MRI across multiple clinical sequences, including T1w, T2w, and FLAIR. ROI delineation was further informed by postoperative MRI and pathological confirmation when available. Notably, MRF was not used for delineating the extent of ROIs. Lesion ROIs were created in MRIcro ²⁴ on the MRF T1w images, registered to the FreeSurfer space and converted to surface‐based data.

2.5. First‐stage ML model training

We used in‐house MATLAB (MathWorks) and Python (v3.12.2) scripts for MRF feature extraction and processing, as well as ML model development and evaluation. Computational time for each step of the process is detailed in the Supplementary Document in Data S1. A two‐stage classification workflow was employed. The first‐stage classification methodology is shown in Figure 1B. Because the lesion vertices were much fewer than normal vertices, to alleviate data imbalance, the Synthetic Minority Oversampling Technique (SMOTE) ²⁵ algorithm was applied to upsample the lesion vertices and downsample the normal vertices. To ensure no data leakage, this resampling was performed only after excluding the test patient in the leave‐one‐out (LOO) cross‐validation setup. Each vertex was treated as an independent sample during training. Neural network models with two hidden layers of 40 and 10 nodes were trained using patient and HC data. ¹⁷ To optimize model performance, various configurations were systematically tested, including different loss functions such as binary cross‐entropy, dice binary cross‐entropy, and binary focal loss. Additionally, learning rates of 1e‐1, 1e‐2, and 1e‐3 were examined. Regularization techniques, including L1 and L2 penalties, were applied with values of 1e‐2 and 1e‐3 to prevent overfitting and enhance generalization. The training was conducted using the Keras ML/deep learning library. For feature importance at this stage, we employed the permutation importance method (sklearn.inspection.permutation_importance). This approach quantifies feature importance by repeatedly shuffling the values of a single feature, evaluating the resulting change in the area under the curve (AUC) of the receiver operating characteristic (ROC) curve compared to the original model, and averaging the observed differences. ²⁶

2.6. Conversion from vertexwise to individual‐level results

The predicted vertexwise outputs were binarized with a threshold of .5. To perform surface‐based clustering, a minimum Euclidean distance of 2 mm between vertices from two different clusters was applied. ²⁷ , ²⁸ The clusters with cluster size smaller than 100 vertices were excluded. This threshold was heuristically defined as half the vertex count of the smallest lesions in our dataset. After clustering of the vertexwise prediction, LOO cross‐validation was used to assess interim model performance at this stage. True positive (TP) clusters were defined as the clusters overlapped with the lesion label (by at least one vertex), and false positive (FP) clusters were defined as the clusters outside of the lesion label, or existence of any positive clusters in HCs. We examined sensitivity, which was determined by the percentage of TP cases. To assess specificity, we quantified the mean number of FP clusters and vertices for patients and HCs.

2.7. Second‐stage ML model training

To reduce FP findings from the first‐stage results, a second‐stage clusterwise classification was performed to suppress the FP clusters and maintain the TP clusters (Figure 1C). The cluster size, mean, median, maximum, SD, 25th as well as 75th percentile of the input features, and first‐stage prediction probabilities within each cluster were used as the clusterwise input. We implemented the Random Undersampling Boosting (RUSBoost) ensemble algorithm ²⁹ for second‐stage classification, as its iterative random undersampling of the majority class effectively addresses the difference between FP and TP cluster numbers. As the final output, the prediction probabilities of each cluster were binarized by a threshold of .5. LOO cross‐validation was applied to assess performance after the second stage, with the same metrics (TP, FP, sensitivity) as the first‐stage classification (detailed in the previous section). Feature importance at this stage was calculated by aggregating each feature's contribution across all weak learners in the RUSBoost ensemble algorithm. To evaluate the effect of thresholding on classifier performance, we also varied the cluster‐level prediction probability threshold from .5 to .8, with an interval of .05. Lastly, we conducted a fivefold cross‐validation 10 times to further assess the robustness of the LOO results; training and test folds were stratified to ensure the ratio of type II and non‐type II cases were similar between the training and testing partition in each fold.

2.8. Subgroup analysis

Subgroup analyses of model performance were performed based on FCD subtypes, clinical MRI positivity, and seizure outcomes. For FCD subtypes, we combined IIa and IIb into type II. We grouped mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy (MOGHE) ³⁰ as well as mild malformation of cortical development (mMCD) ³⁰ together as non‐type II FCD. For clinical MRI positivity, we divided patients into three subgroups: MRI‐positive (lesions were visible on clinical MRI), MRI‐negative/MAP‐positive (lesions were not reported on initial clinical MRI but detected with guidance from morphometric analysis program [MAP] postprocessing and confirmed during the PMC ³¹ , ³² ), and MRI‐negative/MAP‐negative. Based on 1‐year follow‐up, patients were categorized into seizure‐free (SF; defined as International League Against Epilepsy [ILAE] 1) and non‐seizure‐free (NSF; defined as ILAE 2–6). ³³

2.9. FCD subtyping model training

For FCD subtyping, we focused on the vertices contained within the TP clusters, utilizing their surface‐based characteristics as input data. SMOTE was used to address class imbalance between different subtypes. A RUSBoost ensemble classifier was implemented to differentiate between FCD subtypes, specifically non‐type II versus type II, and type IIa versus type IIb. LOO cross‐validation was used to evaluate model performance on the individual level. Feature importance was evaluated in the same way as the first‐stage classification detailed in the previous section.

2.10. Histopathological and radiological confirmation of FCD lesions

Surgical specimens with immunohistochemical staining were microscopically reviewed by an expert pathologist (I.B.). This analysis followed the 2011 ILAE classification guidelines. ³ , ³⁰ Figure 2 presents representative clinical MRI images and MRF maps for each subtype. All MOGHE and mMCD cases underwent pathological confirmation. Although the vast majority of the FCD II cases had pathological confirmation, a few cases did not have histopathological data, primarily due to laser ablation, lesion proximity to eloquent cortex precluding surgery, or pending surgery (details in Table 1). For these cases, radiological diagnosis was used. Only cases with positive clinical MRI were included in this category. Lesions were radiologically diagnosed as IIb based on typical features of cortical thickening, FLAIR signal intensity, and transmantle sign. If no transmantle sign was seen, lesions were radiologically diagnosed as IIa.

FIGURE 2.

Representative clinical T1‐weighted (T1w) magnetization‐prepared rapid gradient echo (MPRAGE) images, fluid‐attenuated inversion recovery (FLAIR) images, and magnetic resonance fingerprinting (MRF) T1 maps and T2 maps for patients with focal cortical dysplasia (FCD) IIb, FCD IIa, mild malformation of cortical development (mMCD), and mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy (MOGHE). Green lines indicate the contour of lesions. First row: FCD IIb lesion in the right precentral region. Second row: FCD IIa lesion in the left occipital region. Third row: mMCD in the right orbitofrontal region. Fourth row: MOGHE in the left orbitofrontal region.

TABLE 1.

Patient demographic and clinical information.

Characteristic	Value
Patients
n	44
Age, years
Mean	27.8
Median	24.5
SD	14.2
Gender
Male	24
Female	20
FCD subtypes (pathological/radiological) a
Type IIa	10/1
Type IIb	13/7
mMCD	13/0
MOGHE	2/0
Handedness (left/right/ambidextrous)	4/39/1
Age at onset, years, mean ± SD	10.9 ± 10.7
Epilepsy duration, years, mean ± SD	17.1 ± 12.6
Seizure frequency, seizure days/year, mean ± SD	196.4 ± 138.0
Clinical MRI findings b
MRI+ and MAP+	25
MRI− and MAP+	13
MRI− and MAP−	6
Transmantle sign (pathological/radiological) a
Type IIb	8/6
Type IIa	4/0
mMCD	1/0
Seizure outcome
ILAE 1	28
ILAE 2	2
ILAE 3	2
ILAE 4	5
ILAE 5	1
Healthy controls
n	70
Age, years
Mean	25.3
Median	24
SD	5.9
Gender
Male	31
Female	39

Abbreviations: FCD, focal cortical dysplasia; ILAE, International League Against Epilepsy; MAP, morphometric analysis program; mMCD, mild malformation of cortical development; MOGHE, oligodendroglial hyperplasia and epilepsy; MRI, magnetic resonance imaging.

^a Detailed explanations of the radiologically confirmed cases are as follows. For the cases with radiologically confirmed FCD IIb, two patients did not undergo surgery due to proximity to eloquent cortex, two underwent laser ablation, and two had surgery scheduled but not yet performed; lesions were radiologically diagnosed as IIb on the positive clinical MRI, which showed the typical features of cortical thickening, fluid‐attenuated inversion recovery signal intensity, and transmantle sign. In two patients with dual lesions, one lesion was pathologically confirmed as IIa in both cases. The unresected lesion was radiologically diagnosed as IIa (MRI positive, no transmantle sign) in one patient and as IIb (MRI positive, transmantle sign) in the other. For these two patients with dual lesions, patient‐level sensitivity was defined by the detection of at least one lesion; for subtyping performance, each lesion was considered separately.

^b Clinical MRI findings (+, positive; −, negative) were defined by official radiological report. In patients who were MAP+, initial MRI reports were negative, and subtle lesions were detected with MAP (without using Validation Neural Network [valNN] automated detection) and confirmed by radiologist review, along with other multimodal data (semiology, electroencephalography, positron emission tomography, single photon emission computed tomography, magnetoencephalography) during patient management conference.

2.11. Statistical analyses

The Statistics and ML toolbox in MATLAB 2023a was used for all statistical analyses. Demographic or clinical characteristics are described using sample mean with SD or number as appropriate. To assess differences between groups, we first tested the distribution of each continuous variable using the Lilliefors test for normality. If the data followed a normal distribution, we applied two‐sample t‐tests; otherwise, we used Wilcoxon rank‐sum tests for nonparametric comparisons. Categorical variables were compared using Pearson chi‐squared tests. McNemar's test was used to evaluate whether the observed differences in patient‐level detection sensitivity were statistically significant. The level of statistical significance was set at p < .05 (two‐tailed).

3. RESULTS

3.1. Cohort overview

Forty‐four patients with histopathologically/radiologically confirmed FCD underwent presurgical evaluation and MRF during the recruitment time span; all patients had T1w MPRAGE, and 35 patients had 3D FLAIR. Seventy HC subjects were also recruited; all HCs had T1w MPRAGE, and 43 had 3D FLAIR. There were no significant differences between patients and HCs in terms of age (patients, mean = 27.8 ± 14.2 years; HCs, mean = 25.3 ± 5.9 years) or gender (patients, 24 male/20 female; HCs, 31 male/39 female). One‐year postsurgical seizure outcome was available in 38 patients, with 28 ILAE 1, two ILAE 2, two ILAE 3, five ILAE 4, and one ILAE 5. FCD subtypes included 13 pathologically confirmed and seven radiologically confirmed IIb, 10 pathologically confirmed and one radiologically confirmed IIa, and 15 pathologically confirmed non‐type II (13 mMCD, two MOGHE). Two patients had dual lesions, each with one pathologically confirmed lesion and one radiologically diagnosed lesion. Postsurgically, all patients with IIb had ILAE 1–2 outcomes; for the eight patients who had ILAE 3–5 outcomes, seven had mMCD (the remaining one had partial resection of FCD IIa due to nearby eloquent cortex). Details are provided in Table 1.

3.2. Vertexwise classification performance

In the overall cohort of 44 patients, T1w features alone yielded an average AUC of .74 ± .18 (mean ± SD). The top three ranked features (in descending order of importance) were GW contrast, cortical thickness, and intrinsic curvature (Figure S3). The combination of MRF and T1w features showed modest improvement, yielding a mean AUC of .76 ± .19. The top three ranked features were GW contrast, cortical thickness, and MRF T1 at 1 mm into the WM.

In the 35 patients with 3D FLAIR, a substantial enhancement was observed when FLAIR features were combined with T1w features (mean AUC = .82 ± .17). The top three ranked features were cortical thickness, GW contrast, and intrinsic curvature. The combination of all three inputs of T1w, FLAIR, and MRF demonstrated similar performance (mean AUC = .82 ± .15). The top three ranked features were cortical thickness, MRF T2 at 1 mm into WM, and intrinsic curvature. Interim performance evaluation showed a very high number of FP findings in all feature sets at this stage, precluding clinical utility (Table S1).

3.3. Patient‐level classification performance: FCD detection

Following the second‐stage classifier, in the overall cohort of 44 patients, utilizing only T1w features resulted in 70.4% sensitivity, with an average of 11.6 FP clusters/patient (Table S1). In HCs, the average FP clusters were 4.1/patient. The top three ranked features were the maximum of cortical thickness, cluster size, and SD of cortical thickness (Figure S4). With MRF features implemented, the sensitivity was comparable at 68.2%, with a much reduced 6.6 FP clusters/patient on average. In HCs, the average FP clusters dropped to 1.5/HC. The top three ranked features were cluster size, maximum of cortical thickness, and maximum of cortical thickness asymmetry.

In the 35 patients with 3D FLAIR, with FLAIR features added to T1w features, the sensitivity was 71.4%, with 3.6 FP clusters/patient and .6 FP clusters/HC on average (Table S1). The top three ranked features were the maximum of cortical thickness, cluster size, and mean of the first‐stage prediction probability. When T1w, MRF, and FLAIR were all used as input, the sensitivity was 71.4%, with 4.7 FP clusters/patient and 1.1 FP clusters/HC on average. The top three ranked features were the maximum of cortical thickness, cluster size, and the SD of the first‐stage prediction probability. McNemar's test revealed no statistically significant differences in sensitivity between any pair of feature combinations. Figure S1 shows the ROC curve of the second‐stage clusterwise performance. Figure 3 shows subjectwise lesion detection for eight example patients whose prediction output overlapped with lesion labels (two for each feature combination).

FIGURE 3.

Individual‐level lesion detection in eight example patients using different feature combinations. Each panel shows results from a different input setting: T1‐weighted (T1w) only (upper left), T1w + magnetic resonance fingerprinting (MRF; upper right), T1w + fluid‐attenuated inversion recovery (FLAIR; lower left), and T1w + MRF + FLAIR (lower right). For each patient (P1–P8), the lesion label is shown on the left and the corresponding prediction on the right. Multiple colors in the prediction output represent different clusters that were detected. Patients include a range of pathologies such as mild malformation of cortical development (mMCD), focal cortical dysplasia (FCD) IIa/IIb, and mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy (MOGHE).

Fivefold cross‐validation performance (details in Table S2) was consistent with the LOO cross‐validation approach, with the highest average patientwise sensitivity (73.4%) achieved using T1w + MRF + FLAIR features and an increased number of FP clusters across all feature sets.

3.4. Confidence index and varying threshold analyses

To assess whether the cluster‐level probabilities can serve as a confidence index for the detection results, we compared the probabilistic distribution of TP and FP clusters across various feature combinations (Figure 4A). The probabilities within TP clusters were significantly higher compared to FP clusters (p < .001) in all feature sets. Figure 4B shows model performance metrics across a range of cluster‐level probability thresholds using the model with all inputs (T1w + FLAIR + MRF). Increasing the threshold from .5 to .6 substantially reduced the mean number of FPs/patient (from 4.7 to 1.6), with a corresponding decrease in sensitivity (from 71.4% to 54.3%).

FIGURE 4.

Evaluation of second‐stage classification cluster probabilities and performance evaluation across threshold levels. (A) Violin plots represent the distribution of cluster probabilities for true positive (TP) and false positive (FP) clusters across different feature sets, including T1‐weighted (T1w) images, T1w + magnetic resonance fingerprinting (MRF), T1w + fluid‐attenuated inversion recovery (FLAIR), and T1w + MRF + FLAIR, with white areas indicating TP clusters and gray areas indicating FP clusters. Dashed lines denote the 25th, 50th, and 75th percentiles. Asterisks mark statistically significant differences (p < .001) between TP and FP cluster probabilities. (B) Performance metrics include patientwise detection sensitivity and FP clusters (mean ± SD) across a range of cluster‐level probability thresholds (.50–.80, in .05 intervals) using all features.

3.5. Patient‐level detection performance: Subgroup analysis

FCD type II lesions were detected with higher sensitivity than non‐type II lesions, particularly with the T1w + MRF + FLAIR combination (83.3% vs. 45.5%). When examining model performance based on clinical MRI positivity, sensitivity followed the following general trend: MRI‐negative/MAP‐negative < MRI‐negative/MAP‐positive < MRI‐positive. FPs by cluster count showed the opposite trend: MRI‐negative/MAP‐negative > MRI‐negative/MAP‐positive > MRI‐positive. The SF subgroup exhibited higher sensitivity than NSF cases across all feature sets; for instance, with T1w + MRF + FLAIR, sensitivity was 72.7% in SF cases compared to 50.0% in NSF cases. Details are provided in Table 2.

TABLE 2.

Patient‐level detection performance based on subgroups.

Subgroup performance	T1w	T1w + MRF	T1w + FLAIR	T1w + MRF + FLAIR
Type II
Sensitivity (%)	21/29 (72.4%)	20/29 (69.0%)	19/24 (79.2%)	20/24 (83.3%)
FP cluster, n	11.9	6.5	3.5	4.9
Non‐type II
Sensitivity (%)	10/15 (66.7%)	10/15 (66.7%)	6/11 (54.5%)	5/11 (45.5%)
FP cluster, n	11.0	6.9	3.6	4.2
MRI+
Sensitivity (%)	17/25 (68.0%)	20/25 (80.0%)	14/19 (73.7%)	15/19 (78.9%)
FP cluster, n	12.0	5.5	3.3	4.3
MRI−, MAP+
Sensitivity (%)	10/13 (76.9%)	7/13 (53.8%)	8/12 (66.7%)	8/12 (66.7%)
FP cluster, n	10.9	7.7	3.5	4.5
MRI−, MAP−
Sensitivity (%)	4/6 (66.7%)	3/6 (50.0%)	3/4 (75.0%)	2/4 (50.0%)
FP cluster, n	11.5	9.2	5.3	6.8
SF
Sensitivity (%)	20/28 (71.4%)	19/28 (67.9%)	16/22 (72.7%)	16/22 (72.7%)
FP cluster, n	12.7	6.8	3.6	4.7
NSF
Sensitivity (%)	5/10 (50.0%)	6/10 (60.0%)	4/8 (50.0%)	4/8 (50.0%)
FP cluster, n	9.8	6.6	3.6	4.6

Abbreviations: FLAIR, fluid‐attenuated inversion recovery; FP, false positive; MRF, magnetic resonance fingerprinting; MRI, magnetic resonance imaging; NSF, non‐seizure‐free; SF, seizure‐free; T1w, T1‐weighted.

3.6. Patient‐level FCD subtyping performance

For distinguishing non‐type II from type II lesions, the best performance was observed for T1w + MRF + FLAIR, with sensitivity of 80.0%, specificity of 83.3%, accuracy of 80.8%, and AUC of 86.7% (Table S3). For differentiating type IIa from IIb (noting the small sample size), the optimal performance was achieved with T1w + FLAIR, resulting in a sensitivity of 63.6%, specificity of 75.0%, accuracy of 68.4%, and AUC of 78.4% (Table S3). The ROC analysis results are shown in Figure S2. The important features for subtyping were morphometric measures such as mean curvature, intrinsic curvature, sulcal depth, and thickness, followed by signal measures such as MRF T1/T2/FLAIR (Figure S5). As a comparison with clinical benchmark, the transmantle sign was visible in only 61.5% of the 13 pathologically confirmed IIb cases (8/13) and was also present in 40% of the 10 pathologically confirmed IIa cases (4/10) as well as 7.7% of the 13 mMCD cases (1/13).

3.7. Comparison with existing surface‐based pipeline

We compared our MRF‐SBM model performance with the publicly available MELD FCD classifier published in 2022. ¹⁷ When applied to the current cohort using input features without MRF (i.e., T1w and T1w + FLAIR), the MELD classifier yielded lower sensitivity (52.3% and 48.6%, respectively) despite achieving lower FPs/patient (1.2 and 1.9 clusters, respectively).

4. DISCUSSION

In this study, we developed a surface‐based MRF pipeline incorporating a two‐stage ML scheme for whole‐brain FCD detection. The main advances of this approach include more robust control of FPs and improved subtype prediction. Additionally, the models' prediction probabilities may be used as an index of detection confidence. Patient‐level sensitivity showed potential relevance to postoperative seizure outcome. These results demonstrate the framework's initial efficacy and show competitive performance as compared to clinical MRI report, performance from prior surface‐based pipeline, ¹⁷ and average human performance. ³⁴ The model's performance was driven by various combinations of parameters, highlighting the value of an integrated multiparametric framework.

4.1. Contribution to literature

Human performance in FCD detection varies significantly across expertise levels, with expert readers achieving a 68% detection rate compared to 45% for nonexperts reported by a prior study, ³⁴ highlighting the challenges in visual assessment. Advanced ML approaches offer advantages over human readers through consistent performance metrics and the ability to detect subtle imaging features, while also providing quantitative confidence measures that complement clinical expertise. ⁴ Our study is the first to incorporate SBM with MRF‐generated multiple quantitative tissue property maps, which were obtained through a single MRF sequence. The inherent coregistration of these quantitative tissue property maps may enhance detection accuracy for subtle abnormalities. ¹⁰ , ¹³ , ¹⁴ , ¹⁵ The investigation of whole‐brain FCD detection, FCD subtyping, and subgroup analysis of seizure outcomes also distinguished this study from previous ROI‐based MRF studies. ¹⁰ , ¹⁵ , ²³ Previous research has demonstrated surface‐based FCD identification using ML methods applied to T1w and FLAIR images. ¹⁷ , ¹⁸ , ¹⁹ , ³⁵ , ³⁶ In the 2022 MELD study, a large cohort of 260 patients was analyzed, demonstrating an overall lesion detection sensitivity of 59%, with an average of two clusters per patient. ¹⁷ When we applied the pretrained MELD pipeline to the current cohort, with T1w and FLAIR input features (but without MRF), there was a lower number of FPs/patient, although with reduced patient‐level sensitivity. ¹⁷ The balance between sensitivity and specificity is always a key component in the design of lesion detection algorithms. We additionally varied the classification probability threshold to assess this balance and demonstrate the potential flexible application of our MRF‐based approach (more details are provided in section 4.7).

4.2. Patient‐level detection performance

The inclusion of MRF features alongside T1w features showed significant benefits in reducing FP while maintaining sensitivity, which is consistent with prior studies. ¹⁵ , ²³ FLAIR also proved beneficial in FCD detection (which is not surprising), with T1w + FLAIR showing comparable sensitivity to T1w + MRF. This underscores 3D FLAIR's known value in highlighting FCD. The important role of 3D FLAIR is endorsed by the ILAE recommendation of its inclusion in the HARNESS (Harmonized Neuroimaging of Epilepsy Structural Sequences) protocol. ³⁷ When all three modalities were combined (T1w + MRF + FLAIR), the method achieved sensitivity of approximately 71%. These results support integrating all modalities for optimized performance and open an interesting future direction of utilizing MRF‐synthesized FLAIR images in the established pipeline. Although the pairwise differences in sensitivity across feature sets did not achieve statistical significance, likely due to small sample size, the patient‐level detection differences still remain clinically meaningful, as correct identification in an individual case can directly impact that patient's management. As a comparison to our model performance, clinical MRI only detected 57% lesions based on official radiology report. The unique contribution of MRF (as compared to FLAIR) was also demonstrated in cases where FCDs were not detected by T1w + FLAIR but identified by T1w + MRF (Figure S6). Taken together, these data highlight the potential of MRF to reveal subtle FCDs and support its integration into clinical protocols for more comprehensive lesion detection.

4.3. Performance comparison using LOO and fivefold cross‐validation

Our main results used LOO cross‐validation. With fivefold cross‐validation, sensitivity was comparable or slightly higher across several feature combinations, suggesting the model's robustness across different training/testing partitions. However, fivefold cross‐validation also showed increased FPs, possibly reflecting the effect of a smaller amount of training data in each fold. Future research with larger cohorts is necessary for additional validation.

4.4. Subgroup analysis

Consistently across all feature combinations, FCD type II lesions were detected with higher sensitivity compared to non‐type II lesions, aligning with findings in existing literature indicating that type II FCDs are more easily detectable. ³ , ³⁸ Subgroup analysis by clinical MRI positivity showed the best performance in the MRI‐positive group, followed by the MAP‐positive and then MAP‐negative groups. This pattern is expected; MRI‐positive lesions were visible on visual inspection, MRI‐negative/MAP‐positive lesions became apparent after MAP's guidance, and MRI‐negative/MAP‐negative cases likely represented truly unidentifiable lesions. As subgrouping reduces sample size in each category, these findings should be interpreted with caution. In the SF subgroup, our approach demonstrated notably higher sensitivity compared to the NSF group across all feature sets. This marked difference suggests our method's potential in indicating postoperative seizure outcomes, which could provide additional information for informing prognosis.

4.5. Important features

In the first‐stage vertexwise classification, GW contrast, cortical thickness, and intrinsic curvature consistently ranked as important features in separating normal and lesional vertices across various feature sets. These findings align with established FCD characteristics; GW contrast abnormalities often reflect disrupted cortical lamination, ³⁹ increased cortical thickness is a hallmark of many FCD subtypes, ³ and intrinsic curvature captures abnormal local gyrification patterns associated with FCDs. ¹⁷ , ¹⁸ When MRF data was incorporated, T1 and T2 values at 1 mm into the WM also showed high importance, suggesting that quantitative tissue properties in the juxtacortical WM play a key role in distinguishing FCD lesions. These findings align with the multiparametric feature characterization observed in our earlier study with an ROI‐based approach. ¹⁵

In the second‐stage patient‐level classification, we examined cluster‐level features for separating TP and FP clusters. The maximum of cortical thickness consistently ranked high, likely capturing the focal nature of cortical thickening in FCDs. Cluster size was also an important feature, potentially helping to differentiate true lesions from spurious findings due to artifacts, normal variants, and registration errors. The inclusion of statistical measures of vertex features within clusters (e.g., SD of cortical thickness, mean of first‐level prediction probability) suggests the importance of capturing heterogeneity within abnormalities. Taken together, feature importance at both vertex and cluster levels highlights the complementary value of local and regional analyses, demonstrating the strength of our two‐stage approach in capturing fine and broad‐scale image features for accurate FCD detection.

4.6. FP considerations

FP findings present a major challenge for FCD detection. ⁴ , ⁴⁰ In a prior study, a cascaded classification system was proposed, where the first model focuses on maximizing lesional voxel detection, whereas the second model aims to minimize FPs without compromising sensitivity. ⁴¹ In our method, the initial vertexwise neural network model yielded high sensitivity accompanied by numerous FPs. A second‐stage clusterwise classification was then developed based on cluster characteristics to effectively suppress FPs while maintaining sensitivity. Despite these reductions, residual FP clusters still remained. Visual inspection revealed that in ~60% there was a lack of clear anatomical or imaging correlates (Figure S7A–D). Potential causes could be identified in the remaining ~40% of cases, including anatomical boundaries (e.g., cerebellum, interhemispheric fissure), sulcal irregularities, hemispheric asymmetries, and motion artifacts (Figure S7E–H). Interpreting these findings, FPs likely arise from: (1) genuine artifacts due to image quality; (2) nonepileptogenic morphological deviations from normal anatomy; and (3) some identified clusters, particularly those near the lesion labels, possibly carrying clinical significance related to lesion extent or gradient, with their clinical significance requiring long‐term outcome data to clarify.

4.7. Usage of clusterwise classification probabilities

We observed significantly higher probabilities for TP clusters compared to FP clusters across all feature sets. This difference in probability profiles indicates its potential as a confidence index for distinguishing TP from FP findings, and the higher probabilities may serve as a proxy for more reliable detections. The varying‐threshold analysis supports a flexible way of utilizing our developed framework; by adjusting the threshold, users can balance sensitivity and specificity depending on clinical priorities, favoring sensitivity for subtle lesion identification, or specificity for reducing FP burden.

4.8. FCD subtyping performance and important features

The optimal accuracy of our model was 80.8% for separating non‐type II from type II and 68.4% for separating IIa and IIb. Although improvement is still needed, these results exceed the current clinical benchmark of the transmantle sign for FCD IIb. Feature importance analysis revealed that subtype classification was driven by morphometric measures such as intrinsic curvature, cortical thickness, asymmetry of GW contrast for II versus non‐II, and mean curvature, sulcal depth, asymmetry of cortical thickness for IIa versus IIb. Morphometric features such as sulcal depth and curvature are sensitive to local cortical folding alterations, which could differ for subtypes. ¹⁷ , ¹⁸ , ³⁶ Cortical thickness is known to increase in type II lesions, ³ and GW contrast asymmetry captures laminar disruption that may be more severe in type II lesions. ¹⁸ Of slightly lesser importance were signal‐based measures such as MRF T1 (II vs. non‐II) and FLAIR (IIa vs. IIb), which likely can be attributed to different cellular composition between subtypes. ¹⁵ The higher FLAIR signal in WM being a discriminative feature for type IIb lesions is consistent with prior literature. ³ Together, these results underscore the value of integrating morphometric, intensity‐based, and asymmetry‐derived multimodal features for maximized subtype classification performance. Future work could explore novel MRF sequence designs to generate additional maps beyond T1 and T2, such as magnetization transfer, myelin, and diffusion, potentially offering richer multimodal information for both FCD detection and subtyping. Additionally, enhancing input MRI contrasts (e.g., MRF) together with advancing ML frameworks to incorporate broader contextual information and extract more complex features may provide substantial benefits. ⁴² , ⁴³

4.9. Limitations

First, the relatively small sample size limits the generalizability and stability of the findings. This constraint also restricted our subtype classification to binary comparisons, precluding multisubtype classification. A larger cohort, ideally including multicenter data, would improve model training and validation. The high reproducibility of MRF makes it inherently well‐suited for multisite studies by minimizing confounding factors commonly encountered in conventional clinical MRI (e.g., scanner‐related variability). Although data use agreements and institutional approvals remain necessary, MRF offers a technical advantage for enabling larger‐scale studies focused on quantitative imaging and outcome prediction in epilepsy. Second, the retrospective nature of our study introduces potential selection bias. Our cohort predominantly comprised subtle, small FCD lesions, with fewer than 60% being visible on clinical MRI at initial assessment. This composition may limit direct comparability with studies focusing on more conspicuous FCD II lesions. Third, manual lesion labels have inherent limitations in precisely delineating lesion borders. Inter‐ and intrarater reliability of manual segmentation was not evaluated as a main scope of this study but may introduce biases. Fourth, not all subjects had 3D FLAIR data, which could contribute to performance variations across feature combinations. To address this, future studies might consider employing synthesized MRF FLAIR images, which could help mitigate this limitation. Finally, the absence of pathological data in unsampled brain regions is a limitation, as FCD cannot be conclusively ruled out in these areas, and some FPs might be correct. Future studies incorporating long‐term seizure outcomes could address this limitation.

5. CONCLUSIONS

We established a novel ML framework for whole‐brain FCD detection using surface‐based processing of clinical MRI and high‐resolution 3D MRF data. The main advances include more robust control of FPs and improved subtype prediction. Additionally, selected model outputs may serve as indicators of detection confidence and seizure outcome. Future research with larger, multicenter cohorts is warranted to further validate this approach for its reproducibility, interpretability, and suitability for prospective clinical application.

AUTHOR CONTRIBUTIONS

Conception and design of the study: Ting‐Yu Su, Xiaofeng Wang, Dan Ma, and Zhong Irene Wang. Acquisition and analysis of data: Ting‐Yu Su, Siyuan Hu, Xiaofeng Wang, Sophie Adler, Konrad Wagstyl, Zheng Ding, Joon Yul Choi, Ken Sakaie, Ingmar Blümcke, Hiroatsu Murakami, Stephen E. Jones, Andreas V. Alexopoulos and Imad Najm. Drafting/editing a significant portion of the manuscript or figures: Ting‐Yu Su, Sophie Adler, Konrad Wagstyl, Joon Yul Choi, Ken Sakaie, and Zhong Irene Wang.

FUNDING INFORMATION

This study is supported by NIH R01 NS109439, 2R01NS109439. Sophie Adler was supported by the Rosetrees Trust (A2665) and Epilepsy Research Institute UK (P2208). Konrad Wagstyl is supported by the Wellcome Trust (301991/Z/23/Z).

CONFLICT OF INTEREST STATEMENT

I.N. is on the speakers' bureau of SK Life. S.E.J. received travel and speaker fees from Siemens Healthineers. D.M. has MRF patents licensed by Siemens. None of the other authors has any conflict of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Data S1

Su T‐Y, Hu S, Wang X, Adler S, Wagstyl K, Ding Z, et al. Surfaced‐based detection of focal cortical dysplasia using magnetic resonance fingerprinting and machine learning. Epilepsia. 2026;67:257–271. 10.1111/epi.18667

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.