Multiparametric Characterization of Focal Cortical Dysplasia Using 3D MR Fingerprinting

Abstract

Objectives:

To develop a multiparametric machine-learning framework utilizing high-resolution 3D MR Fingerprinting (MRF) data for quantitative characterization of focal cortical dysplasia (FCD).

Materials:

We included 119 subjects: 33 patients with focal epilepsy and histopathologically confirmed FCD, 60 age-and-gender-matched healthy controls (HCs), and 26 disease controls (DCs). Subjects underwent whole-brain 3T MRF acquisition, the reconstruction of which generated T1 and T2 relaxometry maps. A 3D ROI was manually created for each lesion, and z-score normalization using HC data was performed. We conducted 2D classification with ensemble models using MRF T1 and T2 mean and SD from GM and WM for FCD vs. controls. Subtype classification additionally incorporated entropy and uniformity of MRF metrics, as well as morphometric features from the morphometric analysis program (MAP). We translated 2D results to individual probabilities using the percentage of slices above an adaptive threshold. These probabilities and clinical variables were input into a support vector machine for individual-level classification. Five-fold cross-validation was performed and performance metrics were reported using receiver-operating-characteristic-curve analyses.

Results:

FCD vs. HC classification yielded mean sensitivity, specificity and accuracy of 0.945, 0.980, and 0.962, respectively; FCD vs. DC classification achieved 0.918, 0.965, and 0.939. In comparison, visual review detected 48% (16/33) of the lesions by official radiology report. Subgroup analysis revealed that the MRF-ML models correctly classified FCD patients from HCs and DCs in 98.3% of cross-validation trials for the MRI− and MAP− group. Type II vs. non-type-II classification exhibited mean sensitivity, specificity, and accuracy of 0.835, 0.823, and 0.83, respectively; type IIa vs. IIb classification showed 0.85, 0.9, and 0.87. In comparison, the transmantle sign was present in 58% (7/12) of the IIb cases.

Interpretation:

The MRF-ML framework presented in this study demonstrated strong efficacy in noninvasively classifying FCD from normal cortex and distinguishing FCD subtypes.

Introduction

Focal cortical dysplasia (FCD) is a common pathology in medically intractable epilepsy.¹ Accurate detection of FCD plays a pivotal role for epilepsy surgery planning. While a fair percentage of FCD lesions can be visualized on clinical magnetic resonance imaging (MRI), many patients still remain nonlesional.² On the histopathological level, FCD lesional cortex may differ only slightly from normal cortex; even when a signal difference is present, it can be quite subtle, small in size, and buried in the complex cortical convexities. Therefore, visual assessment without quantitative tools is often limited in efficacy. Noninvasive subtyping of FCD also remains a challenge. FCD encompasses a broad spectrum of histopathological abnormalities^3,4: FCD II is featured by dysmorphic neurons (IIb with balloon cells and IIa without); FCD Ia is characterized by abundant neuronal microcolumns with small neurons vertically arrayed; mild malformation of cortical development (mMCD) is defined as a mild malformation of cortical development that cannot be put into the category of type I or type II FCD, but has an excess of heterotopic neurons in the white matter.^3,4 Mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy (MOGHE) is a distinct mMCD subtype, which is characterized by an increase in oligodendroglia and heterotopic neurons in the white matter and gray-white boundary.^4,5 Many radiological features, such as abnormal cortical thickness, indistinct gray-white (GW) junction, and T1-weighted (T1w) and T2-weighted (T2w) signal abnormalities are shared by different FCD subtypes; therefore, visual determination of FCD subtypes on conventional MRI is often inadequate,⁶ the transmantle sign being the only clue to be often times associated with type IIb.⁷ Yet, noninvasive FCD subtyping is crucial for prognosis, as better post-operative seizure outcomes were reported to be associated with surgical resections involving FCD II,⁸ especially IIb;⁹ non-favorable outcomes were reported following resections of FCD Ia.¹⁰ Noninvasive FCD subtyping is also crucial for pre-surgical management. According to the recently published International League Against Epilepsy (ILAE) consensus classification of FCD⁴, for some FCD II, surgical resections with excellent results can been achieved without the need for extra-operative invasive monitoring. A recent pilot study showed D-galactose supplementation after surgery had significant response in seizure or cognitive/behavioral outcomes in patients with MOGHE, providing initial efficacy of precision medicine in this histopathology subgroup.¹¹

MR fingerprinting (MRF) is a novel quantitative MRI technique that enables efficient acquisition of multiparametric tissue property maps,¹² including T1 and T2 relaxometry maps, proton density maps, as well as gray matter (GM) and white matter (WM) tissue fraction maps.¹³ Different from conventional (weighted) MRI, the measurements from which are dependent on a mixture of proton density / relaxation times of tissue and additionally modulated by confounding factors from hardware,¹⁴ quantitative MRI techniques such as MRF provide tissue-specific measurements. Prior studies have demonstrated high repeatability and reproducibility of MRF from multiple scanners/sites/vendors.^15,16 2D MRF T1 and T2 maps showed high sensitivity and specificity in various in vivo applications in healthy individuals and patients with neurological disorders.^17–20 Recently, a whole-brain 1mm³ isotropic-resolution MRF sequence has been introduced with scan time of ~10 minutes, making it highly feasible to be employed clinically.¹³ In this study, we aim to use a machine-learning (ML) approach based on high-resolution 3D MRF to characterize FCD lesions in patients with medically intractable focal epilepsies. We analyzed the multiparametric MRF data within lesional region of interests (ROIs) (confirmed by histopathology), to identify the optimal quantitative image features for the best separation of FCD and normal tissue, as well as FCD subtypes, hoping to inform future studies to use these features for whole-brain FCD detection and delineation.

Methods

Study Design and Patient Selection

This Health Insurance Portability and Accountability Act-compliant, retrospective study was approved by the Cleveland Clinic Institutional Review Board (IRB). Patients were included if they: (1) had medically intractable focal epilepsy; (2) underwent a 3D whole-brain MRF scan (as research study) from July 2017 to Apr 2023; (3) underwent resective surgery; (4) had pathologically confirmed FCD. We excluded patients who had metal implants or could not provide informed consent/assent. Further exclusion criteria were as follows: (1) severe imaging artifacts preventing image segmentation or registration; (2) epilepsy with undetermined localization or multifocal origin based on patient management conference (PMC) consensus. MRF findings from the current study did not influence clinical recommendations. Age-and-sex-matched healthy controls (HCs) and disease controls (DCs, patients with nonlesional clinical MRI by official radiology report) were also recruited. All subjects gave written informed consent before the MRF research scan. The subject recruitment workflow is included in Figure 1.

Figure 1.

Study workflow outlining patient selection, data processing and analysis. ANTs = Advanced Normalization Tools; CSF = cerebrospinal fluid; DCs = disease controls; FAST = FMRIB’s Automated Segmentation Tool; FCD = focal cortical dysplasia; GM = gray matter; HCs = healthy controls; MAP = morphometric analysis program; MNI = Montréal Neurological Institute; MRI = magnetic resonance imaging; MRF = magnetic resonance fingerprinting; ROI = region of interest; ROC = receiver operating characteristic; SD = standard deviation; T1w = T₁-weighted; WM = white matter.

MRI Acquisition and Reconstruction

MRI scans were conducted using a Siemens 3.0 T Prisma scanner. A 3D whole-brain MRF sequence¹³ was acquired with a 20-channel head coil: axial acquisition, field of view = 300×300×144 mm³, resolution = 1×1×1 mm³, scan time = 10 min 24 sec. To correct for B₁ inhomogeneity, a 3D whole-brain B₁ mapping sequence was acquired with the same field of view and resolution as the MRF scan (scan time = 1 min 50 sec). T1 and T2 values were obtained after matching the acquired MRF data to a predefined signal evolution dictionary. T1w images with similar contrast to the conventional T1w MPRAGE images were synthesized from the T1 maps (referred to as the MRF synthesized T1w images). Conventional T1w magnetization prepared rapid gradient echo (MPRAGE) sequence with a 20-channel head coil was also acquired (resolution = 0.94 mm³, repetition time = 1,900 ms, echo time = 2.57 ms, flip angle = 10°).

MRF Data Pre-processing

The data pre-processing workflow is outlined in Figure 1. Skull stripping was performed on the MRF synthesized T1w images by the FSL Brain Extraction Tool (BET).²¹ A 3D region of interest (ROI) was manually created for each FCD lesion on the MRF synthesized T1w images by an experienced neuroimaging specialist (ZIW) and confirmed by a board-certified neuroradiologist (SEJ) with 17 years of experience in epilepsy imaging. Lesion ROI delineation was based on the official consensus of positive/negative MRI finding at the epilepsy patient management conference and included areas of signal/ morphometric changes on the pre-operative MRI, considering all available sequences from the clinical MRI protocol (including T1w, T2w and FLAIR sequences). When prospective use of the morphometric analysis program (MAP)^22,23 led to discovery of a lesional finding, MAP findings were also considered when creating the lesion label ROIs. Post-operative MRI and pathological confirmation were additionally used to inform the ROI delineation when available. MRF findings were not used to inform the ROI delineation. The MRF synthesized T1w images were registered to Montréal Neurological Institute (MNI) standard space via symmetric image normalization (SyN) in Advanced Normalization Tools (ANTs),²⁴ generating transformation matrices which were then applied to the MRF T1, T2 maps, and the corresponding ROI masks for registration to the MNI space. The FMRIB’s Automated Segmentation Tool (FAST)²⁵ was applied to generate the GM, WM, and cerebrospinal fluid (CSF) maps from the registered MRF synthesized T1w maps of each subject individually. To minimize the partial volume effect due to CSF, the CSF map was dilated by 1 mm and excluded from the ROI. Considering the location of the FCD lesion was different for each patient, and MRF was known to vary across the whole brain to represent normal anatomical/microstructural variations,^26,27 normalized MRF T1 and T2 z-scores were generated using HC MRF data. For the control subjects, the GM, WM and CSF components were individually generated before applying the lesion ROIs, reducing confounds from anatomical variations. The 60 age-and-gender-matched HCs were split into two equal-sized groups: 30 HCs were used for ROI-based z-score normalization, with the equation listed below; the remaining 30 HCs were used for model performance testing (next section). The HC data were processed with an identical workflow as the patient data.

$$zT1=\frac{{T1}_{PatientROI}–{T1mean}_{HCROI}}{{T1SD}_{HCROI}}$$

$$zT2=\frac{{T2}_{PatientROI}–{T2mean}_{HCROI}}{{T2SD}_{HCROI}}$$

Feature Generation

The feature generation workflow is shown in Figure 1. 3D ROIs were first resampled to axial 2D slices with GM and WM separated. The choice of axial resampling was consistent with the MRF acquisition plane. Small slices with fewer than 10 voxels in either GM or WM were excluded from further processing. For classifying patients and controls, mean and standard deviation (SD) of MRF zT1 and zT2 on the 2D slices were calculated. In total, there were 8 quantitative inputs to the ML models: T1 GM mean, T1 GM SD, T1 WM mean, T1 WM SD, T2 GM mean, T2 GM SD, T2 WM mean, T2 WM SD.

For FCD subtype classification, we additionally included entropy and uniformity²⁸ calculated from MRF metrics, as well as the mean and SD values of morphometric maps, specifically the junction, extension and thickness z-score maps, derived from voxel-based MRI post-processing using MAP^22,23, with clinical T1w MPRAGE as input. The morphometric maps’ z-score values were calculated from the whole ROI without separating GM and WM. In total, there were 22 quantitative inputs to ML models: T1 GM mean, T1 GM SD, T1 GM entropy, T1 GM uniformity, T1 WM mean, T1 WM SD, T1 WM entropy, T1 WM uniformity, T2 GM mean, T2 GM SD, T2 GM entropy, T2 GM uniformity, T2 WM mean, T2 WM SD, T2 WM entropy, T2 WM uniformity, MAP junction mean, MAP junction SD, MAP extension mean, MAP extension SD, MAP thickness mean, MAP thickness SD.

All features were generated on 2D slice level. Overall, the sample size included 550 slices in the 33 FCD patients (202 in FCD IIb, 123 in FCD IIa, 208 in mMCD, and 17 in MOGHE⁵), 15,870 slices in the 30 test HCs and 13,728 slices in the 26 DCs. Since patients’ ROIs had different locations, all ROI regions were applied to the MRF maps of each control subject independently (therefore, controls had many more ROIs than patients). For each control subject, all ROIs needed to be correctly classified as a control, for the individual to be considered as a control.

2D-level and Individual-level Classification

We first performed 2D-level classification utilizing an ensemble model, which integrates multiple models (weak learners) to perform the same task. This ensemble approach helps reduce overfitting and enhance predictive robustness. For classifying patients from HCs or DCs, features generated from the previous step served as input to a Random Undersampling (RUS) Boosting ensemble classifier²⁹ with a learning rate of 0.1 to address class imbalance. For classifying lesion ROI from the contralateral homotopic ROI in the same brain, or classifying FCD subtypes, given the data between categories was relatively balanced, a Robust Boosting ensemble classifier³⁰ was used. The RUS boosting directly address class imbalance through random under-sampling the majority class during each boosting iteration, while Robust boosting is designed to handle noisy but relatively balanced training data. An adaptive 2D-level threshold determined by the Youden index calculated from the training dataset was applied to the testing dataset for classification.

To convert 2D-level classification to individual level, the percentage of slices which surpass the 2D-level threshold was calculated to indicate the overall probability of the 3D ROI to belong to a certain class. A second-stage individual-level classification was performed using support vector machine (SVM) with the input of 3D-ROI-level probability and clinical variables, such as age, epilepsy duration (ED) and onset age. Specifically, age was used as input for classifying patients and controls; age, ED and onset age were used as input when performing FCD subtyping.

Five-fold cross-validation was performed; the proportion of each class in each fold was kept roughly equal to its proportion in the entire data set. Slices from the same subject were kept together and not used for both training and testing purposes. The whole procedure was repeated 10 times to ensure repeatability and reproducibility. Individual-level performance metrics were reported using mean and SD of the receiver operating characteristic (ROC) curves and the corresponding area under curve (AUC) for the cross-validation trials. The entire workflow is outlined in Figure 1, and Figure S1 provides detailed steps of the ML model building and performance evaluation.

Subgroup Analysis

We divided patients into three subgroups: MRI+ and MAP+ (lesions were already visible on clinical MRI and also detected by MAP), MRI− and MAP+ (lesions were missed on initial clinical MRI but were detected by MAP), and MRI− and MAP− (negative clinical MRI or MAP). Classification accuracy in the subgroups was assessed by comparing the percentage of successful classifications in the cross-validation trials. We also retrained the models separately using MRI+ and MRI− groups and re-evaluated performance metrics. To maximize the sample size for model training, we combined the MRI− and MAP+ subgroup with the MRI− and MAP− subgroup into a single MRI− group.

Histopathological Confirmation

Surgical specimens with immunohistochemical staining were microscopically reviewed by an expert pathologist (IB) for FCD subtypes, using the 2011 ILAE classification guidelines³, with considerations of the critical updates in 2022⁴. For subgroup analysis, we included FCD type IIa and type IIb for type II. MOGHE⁵ and mMCD were grouped together as non-type-II FCD.

Statistical Analyses

The demographic or clinical variables were described using sample mean with SD or number as appropriate. Continuous variables were compared using two-sample t-test or Wilcoxon rank sum test, whereas categorical variables were compared using Pearson’s chi-square test. The level of statistical significance was set at P < .05 (two-tailed). The Statistics and Machine Learning toolbox in MATLAB 2021b was used for all statistical analyses and ML model implementation.

Results

Cohort Overview

Ninety-three patients underwent pre-surgical evaluation and MRF research scan during the recruitment time span, of whom 71 patients underwent surgery. In the surgical patients, 26 patients had non-FCD pathology and 8 underwent laser ablation with no pathology. Four patients had to be excluded due to significant imaging artifacts. A total of 33 patients with histopathologically confirmed FCD were included in the study. There was no evidence of differences between HCs for testing (n=30) and patients, with respect to age (HC mean, 25.8 ± 6.7 years; patients mean, 28.6 ± 13.7 years, p = 0.810) or gender (HC, 14 males/16 females; patients, 21 males/12 females, p = 0.176). Besides, no significant difference between DCs (n=26) and patients was shown, with respect to age (DC mean, 25.2 ± 7.8 years, p = 0.674) or gender (DC, 13 males/13 females, p = 0.293). The HCs for normalization and testing were equal in size (both n=30), as well as age-and-gender-balanced (HC for normalization, mean age 25.2 ± 4.8 years, gender 14 males/16 females). Postsurgical seizure outcome at one year was available in 30 patients, with 23 ILAE-1, 2 ILAE-2, 1 ILAE-3, 3 ILAE-4 and 1 ILAE-5 outcomes. FCD subtypes included 12 IIb, 8 IIa, and 13 non-type II (11 mMCD and 2 MOGHE). Post-surgically, all patients with FCD IIb had ILAE 1–2 outcomes; for the 5 pts who had ILAE 3–5 outcomes, 4 had mMCD (the remaining one had partial resection of FCD IIa due to nearby eloquent cortex). The lesion sizes among all patients averaged 4,279 mm³ (SD = 3,435mm³).

Multiparametric MRF Values

Figure 2 shows representative clinical images and MRF maps in each FCD subtype. Table 1 shows the individual-level MRF characteristics in patient and control groups, as well as FCD subtype groups. Patients showed significant T1 increase than HCs in GM (p < 0.001) and WM (p = 0.025), as well as T2 increase in GM (p = 0.006). Additionally, patients showed significant T1 increase than DCs in the GM (p < 0.001). When comparing subtypes, FCD non-type-II (mMCD and MOGHE) showed significantly higher T1 than FCD type II in both GM and WM (p = 0.014 and 0.049, respectively), as well as higher T2 in GM (p = 0.031). When comparing FCD type IIa and IIb, there was no significant difference in T1 or T2, in the GM or WM.

Figure 2.

Representative clinical T1w MPRAGE images, MRF synthetic T1w images, T1 maps, T2 maps and zoom-in views for patients with FCD IIb, FCD IIa, mMCD and MOGHE. Arrows indicate location of lesions. First row (coronal plane): FCD IIb lesion in the left posterior basal temporal region (P5 in Table S1). Second row (coronal plane): FCD IIa lesion in the left mesial parietal region (P15 in Table S1). Third row (axial plane): mMCD in the right orbitofrontal region (P24 in Table S1). Fourth row (axial plane): MOGHE in the left orbitofrontal region (P32 in Table S1). Subtle increased signals can be visualized on MRF T1 map for all lesion subtypes, and on T2 map for FCD II. P15 (second row) is also an example of an FCD lesion that was missed on first clinical visual assessment and MAP (MRI− and MAP−), which was successfully classified (FCD vs. HC) by the MRF ML framework.

T1w = T1-weighted; MRF = MR fingerprinting; FCD = focal cortical dysplasia; mMCD = mild malformation of cortical development; MOGHE = mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy.

Table 1.

Comparison of MRF characteristics within ROIs between different groups on individual level.

	T1_GM (ms)	T1_WM (ms)	T2_GM (ms)	T2_WM (ms)
PT (all subtypes)	1342.83 ± 101.45	975.73 ± 76.38	56.04 ± 7.14	45.1 ± 5.77
HC	1212.19 ± 36.08	936.04 ± 38.4	52.93 ± 2.04	45.24 ± 1.69
DC	1218.46 ± 52.54	953.69 ± 48.6	53.61 ± 2.99	45.6 ± 2.12
P value a - PT vs. HC	< 0.001*	0.025*	0.006*	0.831
P value a - PT vs. DC	< 0.001*	0.255	0.052	0.577
FCD IIa	1313.94 ± 122.2	959.76 ± 112.81	55.03 ± 12.21	44.98 ± 10
FCD IIb	1303.88 ± 95.11	955.82 ± 59.85	53.83 ± 4.74	43.54 ± 3.17
mMCD	1384.9 ± 71.29	993.13 ± 55.34	57.78 ± 3.47	45.64 ± 3.59
MOGHE	1460.77 ± 48.13	1063.29 ± 46	63.75 ± 1.62	51.9 ± 0.11
FCD II (IIa + IIb)	1307.9 ± 103.75	957.4 ± 82.26	54.31 ± 8.26	44.12 ± 6.57
Non-II (mMCD + MOGHE)	1396.57 ± 72.38	1003.92 ± 58.51	58.7 ± 3.9	46.6 ± 4.04
P value a – II vs. Non-II	0.014*	0.049*	0.031*	0.101
P value a - IIa vs. IIb	1	0.847	0.616	0.563

^aWilcoxon rank sum test with statistical threshold of p < 0.05.

^*indicating significance

FCD = focal cortical dysplasia; PT = patients; HC = healthy control; DC = disease control; ROI = region of interest; GM = gray matter; WM = white matter; mMCD = mild malformation of cortical development; MOGHE = mild malformation of cortical development with oligodendroglial hyperplasia in epilepsy.

Classification Performance: FCD vs. HCs

Using the MRF-ML models to classify FCD patients and HCs, the mean ROC curve is shown in Figure 3A, with AUC of 0.976 ± 0.009. Sensitivity, specificity, and accuracy were 0.945 ± 0.026, 0.98 ± 0.031 and 0.962 ± 0.016, respectively. The top 50% ranked features (in descending order of importance) were T1 GM mean, T1 GM SD, T1 WM SD, and T1 WM mean. The addition of MAP-related features along with MRF features did not further improve the performance (Table 2). In comparison to the ML model classification performance, initial MRI visual review only detected 48% (16/33) of the lesions per official radiology report; assisted by MAP post-processing, 82% (27/33) of the lesions were detected per official PMC report (Table S1).

Figure 3.

Individual-level performance for classifying PTs vs. HCs and PTs vs. DCs with MRF features. ROC curves are shown, with the solid lines referring to the average and the gray shaded area indicating the standard error among 10 trials.

AUC = area under curve; FCD = focal cortical dysplasia; MRF = MR fingerprinting; HCs = healthy controls; PTs = patients; ROC = receiver operating characteristic.

Table 2.

Individual-level performance summary using different input data.

PTs vs HCs
Input Data	Sensitivity	Specificity	Accuracy	AUC
MRF	0.945 ± 0.026	0.98 ± 0.031	0.962 ± 0.016	0.976 ± 0.009
MRF + MAP	0.952 ± 0.031	0.968 ± 0.019	0.959 ± 0.015	0.978 ± 0.011
T1w + MAP	0.485 ± 0.084	0.771 ± 0.124	0.616 ± 0.042	0.576 ± 0.051
Type II vs Non-type-II
MRF	0.84 ± 0.099	0.777 ± 0.1	0.815 ± 0.052	0.833 ± 0.045
MRF + MAP	0.835 ± 0.1	0.823 ± 0.103	0.83 ± 0.036	0.855 ± 0.035
T1w + MAP	0.755 ± 0.076	0.808 ± 0.086	0.776 ± 0.041	0.787 ± 0.041
Type IIa vs IIb
MRF	0.675 ± 0.146	0.825 ± 0.115	0.735 ± 0.09	0.734 ± 0.114
MRF + MAP	0.85 ± 0.133	0.9 ± 0.094	0.87 ± 0.06	0.886 ± 0.055
T1w + MAP	0.742 ± 0.164	0.8 ± 0.16	0.765 ± 0.1	0.758 ± 0.128

AUC = area under curve; HCs = healthy controls; MAP = morphometric analysis program; MRF = magnetic resonance fingerprinting; MRI = magnetic resonance imaging; PTs = patients.

Classification Performance: Broader signal changes

To test if the MRF-ML models detected lesional changes or broader signal changes due to epilepsy, we examined the model performance to classify patients vs. DCs, using ROIs at the exact same anatomical locations. Classification performance remained high: the AUC of the ROC curve was 0.971 ± 0.011, with sensitivity of 0.918 ± 0.027, specificity of 0.965 ± 0.021 and accuracy of 0.939 ± 0.011 (Figure 3B). The top 50% ranked features were T1 GM mean, T1 WM SD, T1 GM SD, and T1 WM mean.

Within the patient cohort, we examined the homotopic ROI from the contralateral side in the same brain. Lesion ROIs and their contralateral homotopic ROIs can be classified with extraordinary performance, with sensitivity of 0.994 ± 0.018, specificity of 0.997 ± 0.009, accuracy of 0.995 ± 0.01, and AUC of 0.995 ± 0.015. When we used data from the contralateral homotopic ROI as input to the MRF-ML models to classify patients vs. HCs. The performance was much reduced, showing sensitivity of 0.742 ± 0.249, specificity of 0.518 ± 0.158, accuracy of 0.639 ± 0.079, and AUC of 0.6 ± 0.101.

Classification Performance: FCD subtypes

Figure 4 shows the ROC analysis of classification performance for subtyping (detailed MRF input comparisons are shown in Table S6). Classification of FCD II and non-type-II exhibited AUC of 0.855 ± 0.035, with sensitivity of 0.835 ± 0.1, specificity of 0.823 ± 0.103 and accuracy of 0.83 ± 0.036 (Figure 4B). The top 25% ranked features (in descending order of importance) were T1 WM SD, T2 GM mean, T1 WM mean, MAP extension mean, and MAP thickness mean, T1 GM uniformity. Classification of FCD IIa and IIb showed an AUC of 0.886 ± 0.055, with optimal sensitivity of 0.85 ± 0.133, specificity of 0.9 ± 0.094 and accuracy of 0.87 ± 0.06 (Figure 4D). The top 25% ranked features (in descending order of importance) were T2 WM mean, MAP extension mean, T2 GM SD, MAP extension SD, and T1 WM mean, and T2 GM mean. The addition of MAP-related measures to the ML model boosted performance, as evidenced by the decreased performance when these measures were omitted (Table 2, Figure 4A, Figure 4C). In comparison to the ML model performance, clinical evaluation of the subtype was restricted to the identification of the transmantle sign, which was visible only in 58% (7/12) of the IIb cases, while being present in 38% (3/8) of the IIa cases.

Figure 4.

Individual-level performance for subtyping using different image data inputs. (A) and (B) show FCD type II vs. non-type-II classification performances using MRF alone, and MRF + MAP features respectively. (C) and (D) shows FCD type IIa vs. IIb classification performances using MRF alone, and MRF + MAP features respectively. ROC curves are shown, with the solid lines referring to the average and the gray shaded area indicating the standard error among 10 trials.

AUC = area under curve; FCD = focal cortical dysplasia; MRF = MR fingerprinting; MAP: morphometric analysis program; ROC = receiver operating characteristic.

Classification Performance: MRF vs. clinical T1w

To determine whether the classification performance came from quantitative MRF features or the use of ML models, we compared model performances using conventional clinical T1w scans (with MAP features) as inputs. Performance metrics markedly declined as compared to those using MRF and MAP input, across all classification tasks (Table 2).

Subgroup Analysis

In the three subgroups, MRI+ and MAP+ (N=16), MRI− and MAP+ (N=11), and MRI− and MAP− (N=6, example shown in Figure 2, second row), the FCD vs. control classification accuracy was not impacted by the different grouping. When comparing to HCs, FCD patients were correctly classified in 86.3%, 98.2% and 98.3% of the cross-validation trials, respectively for these three groups. When comparing to DCs, FCD patients were correctly classified in 91.9%, 92.7% and 98.3% of the cross-validation trials, respectively for these three groups. As shown in Table 3, the models separately trained using MRI+ and MRI− groups showed similar performances for the FCD vs. HC, and FCD vs. DC classifications. However, for FCD subtyping, the performances were markedly lower in the MRI− group.

Table 3.

Individual-level performance summary for subgroup analyses.

MRI-positive Subgroup
	Sensitivity	Specificity	Accuracy	AUC
PTs vs. HCs	0.9 ± 0.041	0.937 ± 0.071	0.924 ± 0.044	0.952 ± 0.018
PTs vs. DCs	0.875 ± 0.068	0.915 ± 0.054	0.9 ± 0.03	0.938 ± 0.029
Type II vs. Non-type-II	0.983 ± 0.033	1.0 ± 0.0	0.988 ± 0.025	0.99 ± 0.021
Type IIa vs. IIb	0.712 ± 0.168	0.75 ± 0.194	0.725 ± 0.149	0.714 ± 0.148
MRI-negative Subgroup
PTs vs. HCs	0.919 ± 0.049	0.937 ± 0.053	0.93 ± 0.027	0.956 ± 0.019
PTs vs. DCs	0.9 ± 0.053	0.938 ± 0.058	0.923 ± 0.026	0.951 ± 0.012
Type II vs. Non-type-II	0.838 ± 0.098	0.7 ± 0.141	0.765 ± 0.059	0.731 ± 0.07
Type IIa vs. IIb	0.575 ± 0.225	0.675 ± 0.16	0.625 ± 0.125	0.6 ± 0.166

AUC = area under curve; DCs = disease controls; HCs = healthy controls; MRF = magnetic resonance fingerprinting; MRI = magnetic resonance imaging; PTs = patients.

Discussion

Our study proposed a novel multiparametric machine-learning framework based on high-resolution 3D MR fingerprinting for FCD characterization. This approach showed initial efficacy in automatically classifying FCD from normal cortex and amongst FCD subtypes. The top-ranked features that contributed to the model performances were brought by different combinations of MRF measures and morphometric measures, which speaks to the complementary nature of the multiparametric framework. Although the current study uniquely has the largest cohort with pathologically validated FCD with MRF data, the sample size is still relatively small, making an independent validation dataset not possible. While future larger-cohort studies are needed to replicate our findings, the clinical implications from these initial data are multi-faceted. On an everyday-practice level, the noninvasive nature of MRF makes it feasible to be used prior to surgery for providing subtyping information, which can be an additional indicator to inform post-surgical seizure outcomes. This is echoed in the outcome data of our cohort, where mMCD was seen with most of the patients with ILAE 3–5 outcomes. Our findings are also pertinent and timely for the recently updated ILAE consensus classification of FCD,⁴ which proposes a multilayered approach with histopathological, genetic and imaging; the MRF-based quantitative MRI analysis framework can provide objective imaging characterization with quantifiable results, ready to be plugged into the neuroimaging layer.⁴

Contribution to literature

Noninvasive MRI subtyping of FCD has been previously reported using advanced whole-brain post-processing algorithms based on 3D T1w images,³¹ or multimodal data including T1w, FLAIR, diffusion MRI and resting-state fMRI,^32,33 which were sequentially acquired as different sequences. Our study differs from the prior studies by using a 3D MRF-based framework, which allows for the acquisition and reconstruction of multiple tissue property maps with one single sequence. The spatially-registered maps ensure perfect co-registration across the high-resolution multiparametric data,³⁴ thereby further increasing the sensitivity and specificity for detection of subtle abnormalities.^13,19,20 While we employed T1 and T2 maps under the current framework, there is ample flexibility to include other maps relevant to the epileptic process, including magnetic transfer, myelin and diffusion.³⁵ Furthermore, the high reproducibility of MRF also allows for multi-site study without confounding factors commonly found in conventional clinical MRI (e.g., from scanner and other measurement tools), thereby could facilitate future larger studies on quantitative imaging and outcome predictions.

Data augmentation using 2D Reslicing

Due to the relatively small sample size especially for the patient group (N=33), we used resliced 2D ROI instead of 3D ROI as input for the ML models, which offers the advantage of augmenting the available data samples to enhance model performance. We converted 2D-level classification to individual level, by calculating the percentage of slices which surpass the 2D-level threshold, to come up with an overall probabilistic classification of the overall 3D ROI. A similar approach was adopted by other studies in the literature, e.g., in one prior study, ML techniques were used on 2D brain MRI slices for tumor classification;³⁶ in another study, extracted 2D MRI slices from axial, coronal, and sagittal orientations were used to train convolutional neural network models for early diagnosis of Alzheimer’s disease.³⁷ As future work, more advanced techniques, such as deformable augmentation and deep learning-based approaches,³⁸ could be potentially implemented in our data set to boost model performance.

Important Features: FCD vs. Normal Cortex

Our models exhibited remarkable performance for separating FCD lesions with normal cortex in healthy control and disease control cohorts. Notably, we employed stringent criteria for this classification task. Since patients’ FCD ROIs differ in anatomical locations, for HCs or DCs, all the corresponding ROIs would need to be correctly classified as control, for the individual to be considered a control. Considering this stringent criteria, the individual-level classification accuracies, being 96.2% for FCD vs. HCs and 93.9% for FCD vs. DCs, were remarkable. In comparison, only 48% of the lesions were noted by visual review of the clinical MRI, and MRI visual review assisted by MAP only detected 82% of the lesions, i.e., MRI signal or morphometric abnormalities were present in 82% of the cases at the best, even upon focused re-review during PMC guided by other modalities. The high accuracies brought by the MRF-based ML classification speak to the high sensitivity and specificity of the MRF measures. Notably, compare to using merely MRF features, no significant improvement was shown after adding the morphologic features. The top 4 ranked features were all MRF features, not MAP-related features. This suggest that MRF features by themselves were highly effective to produce differentiative results to separate FCD and normal cortex.

Were the detected MRI signal changes related to specific lesions, or due to broader, disease-related brain abnormalities? Our results showed that the latter has a lesser impact. The MRF metrics were highly effective to separate FCD patients and DC, with a comparable yet slightly decreased accuracy (93.9%) than HC (96.2%). This slight decrease was not surprising, as quantitative MRI changes in epileptic individuals have been previously reported, even away from the presumed epileptogenic zone.^39,40 A prior study using MRF, for example, detected group-level abnormalities located on the ipsilateral side of the epilepsy in a conventional-MRI-negative cohort.⁴¹ However, these broader changes only minimally decreased the model performance and a high accuracy was maintained. In the same patient, the lesion ROI can be almost perfectly distinguished from its corresponding region on the contralateral side. This separation mimics the lesion detection task performed by neuroradiologists and demonstrates the ability of MRF to differentiate lesional from non-lesional tissue in the same brain. When the contralateral ROIs served as input to the same ML framework, the performance was much reduced (63.9% accuracy, slightly better than chance). Taken together, these data suggest that the MRF signals within the lesion ROIs were predominantly capturing lesion-related changes, with less influence from seizure-related changes.

T1 and T2 values were reported in the literature to be highly relevant to epileptic lesions.⁴² Increased T1 was reported in hippocampal sclerosis as well as cortical malformations in patients with epilepsy.^13,19,43 Increased T2 was also reported at the site of cortical malformations and in patients with negative conventional MRI, with the indicated minor structural changes concordant with the electro-clinical profile.⁴⁴ T2 voxel-based relaxometry studies reported increased signals in the ipsilateral temporal lobe and mesial structures in patients with temporal lobe epilepsy.^45,46 Our data showed a significant increase of T1 in FCD as compared to normal cortex in the GM and WM, as well as a significant increase of T2 in the GM, which is consistent with the general notion of T1/T2 increase in pathological brain tissue. The top 50% ranked features identified for FCD vs. HC or DC classifications were almost identical and included all MRF T1 features, i.e., T1 GM mean and SD, as well as T1 WM mean and SD. This finding is consistent with prior studies, as MRF T1 was shown to have robust measurements with low variability,²⁷ and consistently showed higher sensitivity than MRF T2 to indicate subtle pathologies in the epileptic brain.^13,41,47 The top 50% ranked features included both GM and WM metrics, which is consistent with the common location of FCD lesions being at the GW junction, with varying distributions on either side. The SD measures of T1 GM and WM being ranked high as important features is also consistent with the disruption of cortical lamination observed across the FCD spectrum.³

Important Features: FCD Subtyping

Comparing FCD type II and non-type-II lesions, T1 GM, T1 WM and T2 GM were significantly higher in the non-type-II group. This finding is intriguing, since no prior studies specifically examined the T1 and T2 relaxometry values for the mMCD and MOGHE pathologies, it is difficult to have a comparison benchmark. The higher T1 and T2 values in the non-type-II subgroup could be driven by the MOGHE subtype, due to the marked increase of oligodendroglia and heterotopic neurons in the WM and GW boundary, a key feature of this pathology.^4,5 The WM T1 and T2 increase could be explained by the underlying pathology, but the GM T1 and T2 increase is especially interesting. We speculate that the higher GM values could be due to the increased cellular density at the GW boundary. Since the lesions were segmented to GM and WM compartments, those located at the blurred GW boundary could contribute to either side, explaining the higher GM values. When comparing the IIa and IIb cases, although a clear difference of neuronal composition is present on a pathological level (IIb with balloon cells and IIa without), no significance was seen in the T1/T2 GM/WM measures on individual-level, highlighting the challenges of measuring these subtle differences with noninvasive in vivo imaging.

Although the T1 and T2 metrics in the GM and WM brought high accuracy in separating FCD lesions and normal cortex, merely using these metrics was inadequate for subtyping classification performance, which is consistent with the general inability of predicting subtypes solely dependent on MRI signal changes, especially on individual-patient level. We therefore expanded the features to include additional calculation of entropy and uniformity based on MRF metrics, as well as morphological features based on MAP. These additions, particularly the morphometric features (many ranked top 25%), provide information that cannot be obtained solely from the signal-based MRF metrics. With the additional features, accuracy for FCD non-type-II and type II classification was achieved at 83%, and for FCD IIa and IIb classification at 87%. Notably, the results outperformed the transmantle sign, which was only visible only in 58% of the IIb cases in the same cohort (low sensitivity) and was also present in a non-negligible percentage of IIa cases (low specificity). Interestingly, when examining the top 25% ranked features, many optimal features were based on variation/homogeneity metrics (e.g., T1 WM SD, T1 GM uniformity, T2 GM SD, MAP extension SD) providing measurements of disorganization of tissue properties, or morphometric features; these features are relatively difficult to visually extract and can only be brought out with a quantitative approach. Many optimal features were WM features (e.g., T1 WM SD, T1 WM mean, T2 WM mean, T1 WM mean), consistent with the differing involvement of WM in the FCD spectrum, e.g., in the WM of FCD type II, hypomyelinated neurons and balloon cells were reported; in mMCD, excessive heterotopic neurons can be seen.^3,4,48 The WM contribution to FCD subtypes was also echoed in a number of previous studies,^49,50 with d-MRI studies identifying alterations in tissue microstructure.³²

Additional Yields

When using conventional clinical T1w images as inputs to the ML framework, even with MAP features, the performance metrics markedly declined comparing with MRF input. This decline was consistently seen in all the classification tasks. Since the ML process remained the same, these results suggest that the efficacy of the MRF-ML framework was largely driven by the quantitative MRF features rather than the ML framework itself. Our subgroup analysis based on MRI and MAP positivity also illustrated there was added value of the MRF-ML methodology in the most difficult cases. In the MRI− and MAP− group, FCD patients were correctly classified from HCs and DCs in 98.3% of the cross-validation trials. Retraining the model using MRI+ and MRI− groups did not affect FCD and HC classification performance. However, FCD subtyping showed lower performance in the MRI− group, possibly due to the smaller sample sizes when dividing into subtypes, or the subtle lesion appearances that made subtyping inherently more challenging.

Clinical Utilization: Promises and Caveats

The MRF-ML approach outlined in this study offers various clinical applications. For instance, when structural MRI or other modalities like PET and magnetoencephalography suggest focal abnormalities, a manually masked 3D ROI on the MRF data can be created and input into the MRF-ML models. The classification results can inform whether the 3D ROI represents FCD or normal cortex, leveraging the high classification accuracy demonstrated by the current data. Similarly, the hypothesis ROI could serve as input to the subtype classification ML process, to make predictions of the FCD subtypes. This data could be used to inform seizure outcome prognosis and surgical management strategies. Caveats of the current study should also be noted. As the current study focuses on FCD characterization, all analyses were lesion-ROI based; within the scope of the study, we do not directly address whole-brain lesion detection for individual patients. The ROI approach also does not directly address the lesion extent question, as there could be epileptogenic tissue outside of the manually delineated ROIs, especially in the seizure-recurring cases. Manual delineation of the ROI could also introduce subjectivity regarding the lesion border, especially the subtle lesions such as MOGHE and mMCD. However, the quantitative image features discovered in the current study could provide valuable information for future studies on whole-brain FCD detection and delineation based on MRF.

Limitations

First, our study was limited by the small sample size. The detailed pathological subtype can only by obtained by rigorous pathology re-review, especially for the non-type-II subtypes; therefore, we included only patients who underwent surgery and had surgical specimens available for pathology review. The sample size limits the subtype classification tasks that could be performed. Therefore, we were only able to perform binary classification of type II vs. non-type-II, as well as type IIa vs IIb. With a larger sample size, the MRF-based ML models may be better trained, especially for the MOGHE subgroup, which will eventually allow for multi-subtype classification. The small sample size also prevented us from further splitting the cohort to a training/cross-validation and an independent cohort for validation. Future larger-cohort, preferably multi-center studies are needed to replicate findings from the current study. Second, the retrospective nature inevitably brings selection bias. Many of the included FCD lesions were subtle in nature and small in size, and only half were visible on the conventional MRI on first pass; a third of our cohort were non-type-II (mMCD and MOGHE) cases. Using MAP with valANN automated lesion detection²³ in this same cohort, the sensitivity of FCD detection was only 42.4%, which was similar to the 48.5% sensitivity of using T1w and MAP input to separate patients and HCs with our ML framework (Table 2). This cohort, therefore, may not be directly comparable to cohorts consisting of obvious FCD II lesions which are reported to yield ~80% sensitivity of lesion detection²³. Thirdly, in both the healthy and disease control cohorts, the absence of accessible pathology data poses a limitation to the study. FCD cannot be excluded in 100% of the control subjects, so some positive detections in these controls might have been correct.

Conclusion

We demonstrated initial strong efficacy of a multiparametric machine-learning framework based on high-resolution 3D MRF, to classify FCD from normal cortex, FCD type II from non-type-II, and FCD IIa from IIb. The high accuracies of the classification performance suggest the usefulness of this quantitative imaging approach to augment FCD characterization, which may contribute to the noninvasive pre-surgical evaluation for individuals with epilepsy, as well as an integrated clinical-pathological-imaging understanding of the FCD spectrum.

Supplementary Material

Fig S1

Figure S1. Detailed steps of ML model building and cross validation. A: Illustration of 5-fold cross-validation and the composition of training and testing data sets. The proportion of each class in each fold was kept roughly equal to its proportion in the entire data set. Slices from the same subject were kept together and not used for both training and testing purposes. The whole procedure was repeated 10 times to ensure repeatability and reproducibility. B: 2D-level classification and optimal 2D-level threshold determination. For the training set, all features were generated on 2D slice level and 2D-level classification was first performed. For classifying patients from HCs or DCs, a Random Undersampling (RUS) boosting ensemble classifier with a learning rate of 0.1 was used to address class imbalance. For classifying FCD subtypes, given the data between subtypes was relatively balanced, a Robust Boosting Ensemble classifier was used. An optimal 2D-level threshold was calculated from the training dataset, determined by the Youden index. An example 2D-level threshold of 0.5 is shown in the dashed box. C: Individual-level classification & threshold determination for classifying patients and controls (including healthy controls and disease controls). For the testing set, similar to the training set, all features were generated on 2D slice level and 2D-level classification was first performed. To convert 2D-level classification to individual level, the percentage of slices which surpass the optimal 2D-level threshold (determined by the training set) was calculated to indicate the overall probability of the 3D ROI to belong to a certain class. Since patients’ ROIs had different locations, all ROI regions were applied to the homotopic region of each control subject independently (therefore, controls had many more ROIs than patients). For each control subject, the probability of any homotopic ROIs to be lesional was calculated, and the maximal probability was taken. The probability of 3D ROI and clinical variable (age) served as input to a second-stage SVM classifier, which generates a final individual-level prediction. D. Individual-level classification & threshold determination for classifying subtypes (including IIa vs. IIb, and type II vs. non-type-II). This process is similar to the classification task described in C, the only difference being there was only one ROI per subtype. The probability of 3D ROI and clinical variable (age, epilepsy duration and onset age) served as input to a second-stage SVM classifier, which generates a final individual-level prediction. In both C and D, individual-level performance metrics were reported using mean and SD of the area under curve (AUC) for the receiver operating characteristic (ROC) curves for the 10 trials that the five-fold cross-validation was performed.

Fig S2

Figure S2. MRF T1 template from 30 HC subjects in the current study. The Z values correspond to the axial coordinate in MNI space.

Fig S3

Figure S3. MRF T2 template from 30 HC subjects in the current study. The Z values correspond to the axial coordinate in MNI space.

2. What is the current knowledge on the topic?

Focal cortical dysplasia (FCD) is a common pathology in medically intractable epilepsy, and its detection and characterization are crucial for surgical planning and prognosis. Many FCD lesions are not detected on conventional MRI, and noninvasive subtyping is challenging due to subtle histopathological differences.

3. What question did this study address?

The study aimed to determine if a multiparametric machine-learning framework using 3D MR Fingerprinting (MRF) can provide quantitative characterization of FCD lesions.

4. What does this study add to our knowledge?

The study shows that machine learning models, using a 10-minute whole-brain MRF scan yielding high-resolution quantitative T1 and T2 tissue properties, can accurately classify FCD from normal cortex and distinguish between different FCD subtypes. The added value of the MRF-ML methodology is demonstrated by the highly accurate classification performance in challenging cases where conventional MRI and post-processing techniques were negative.

5. How might this potentially impact on the practice of neurology?

The MRF-ML approach offers noninvasive, highly accurate classification of FCD and its subtypes, which can be used pre-surgically to inform seizure prognosis and surgical strategies. This methodology aligns with the updated International League Against Epilepsy consensus classification of FCD, providing objective imaging data to support a multilayered diagnostic approach.

Draft Tweet

New study shows an efficient, whole-brain MR Fingerprinting scan with ML accurately classifies FCD & subtypes, aiding pre-surgical planning and prognosis. Effective even in challenging cases. #epilepsysurgery #MRI

Glossary

AUC

area under curve

ANTs

Advanced Normalization Tools

BET

brain extraction tool

CSF

cerebrospinal fluid

DCs

disease controls

ED

Epilepsy duration

FAST

FMRIB’s Automated Segmentation Tool

FCD

focal cortical dysplasia

GM

gray matter

HCs

healthy controls

ILAE

International League Against Epilepsy

IRB

Institutional Review Board

MAP

morphometric analysis program

ML

machine learning

MNI

Montréal Neurological Institute

MRI

magnetic resonance imaging

MRF

magnetic resonance fingerprinting

mMCD

mild malformation of cortical development

MOGHE

mild malformation of cortical development with oligodendroglial hyperplasia and epilepsy

MPRAGE

magnetization prepared rapid gradient echo

PMC

patient management conference

RUS

random undersampling

ROI

region of interest

ROC

receiver operating characteristic

SD

standard deviation

SyN

symmetric image normalization

SVM

support vector machine

T1w

T1-weighted

T2w

T1-weighted

WM

white matter

Data Availability

Anonymized data are available on request to the corresponding author from qualified investigators.