Combining MR Fingerprinting with Voxel-based Morphometric Analysis to Reduce False Positives for Focal Cortical Dysplasia Detection

Abstract

Objective:

We aim to improve focal cortical dysplasia (FCD) detection by combining high-resolution 3D MR fingerprinting (MRF) with voxel-based morphometric MRI analysis.

Methods:

We included 37 patients with FCD (10 IIa, 15 IIb, 10 mMCD and 2 MOGHE) and pharmacoresistant focal epilepsy and 59 healthy controls (HCs). 3D lesion labels were manually created. Whole-brain MRF scans were obtained with 1 mm³ isotropic resolution, from which quantitative T1 and T2 maps were reconstructed. Voxel-based MRI postprocessing, implemented with the morphometric analysis program (MAP18), was performed for FCD detection using clinical T1w images, outputting clusters with voxel-wise lesion probabilities. Average MRF T1 and T2 were calculated in each cluster from MAP18 output for GM and WM separately. Normalized MRF T1 and T2 were calculated by z-scores using HCs. Clusters that overlapped with the lesion labels were considered true positives (TP); clusters with no overlap were considered false positives (FP). Two-sample t-tests were performed to compare MRF measures between TP/FP clusters. A neural network model was trained using MRF values and cluster volume to distinguish TP/FP clusters. Ten-fold cross-validation was used to evaluate model performance at cluster level. Leave-one-patient-out cross-validation was used to evaluate performance at patient level.

Results:

MRF metrics were significantly higher in TP than FP clusters, including GM T1, normalized WM T1, and normalized WM T2. The neural network model with normalized MRF measures and cluster volume as input achieved mean AUC of 0.83, sensitivity of 82.1% and specificity of 71.7%. This model showed superior performance over direct thresholding of MAP18 FCD probability map at both cluster- and patient-level, eliminating ≥ 75% FP clusters in 30% of patients and ≥ 50% of FP clusters in 91% of patients.

Significance:

This pilot study suggests the efficacy of MRF to reduce false positives in FCD detection, due to its quantitative values reflecting in vivo pathological changes.

Introduction

Focal cortical dysplasia (FCD) is one of the most common pathologies underlying Magnetic Resonance Imaging (MRI)-negative epilepsy^1,2. FCD lesions are difficult to detect on MRI by visual inspection, because they can be small in size, heterogeneous in location, and often buried deep in complex sulci³. Epilepsy with FCD as the underlying pathology is often pharmacoresistant and often requires surgical intervention. Identifying the precise location of the lesion is crucial for achieving a seizure-free outcome after surgery². A typical review of a focal epilepsy case to locate the lesion involves slice-by-slice examination of hundreds of slices from multiple MRI sequences and verification of concordance with other data. This review process requires substantial time and is subject to inter-rater variability⁴. Automatic lesion detection pipelines have been developed to increase detection yield, reduce the reviewing time, and minimize rater dependency. The most widely applied pipeline thus far for individual-level voxel-based postprocessing is the Morphometric Analysis Program (MAP), which outputs gray-white (GW) junction, gray-matter (GM) extension, and cortical thickness z-score maps to highlight areas of interest for a focused re-review^5,6. The efficacy of MAP to yield detection of subtle lesions has been reported by many prior studies, retrospective and prospective^7–10. A known limitation of MAP is that many false positive (FP) findings can be present, due to artifacts and normal variants, so interpretation of the results to locate the true lesional areas still requires substantial efforts and expertise.

Quantitative MRI techniques (qMRI) have been developed in recent years to address the limitations of conventional MRI. Instead of only showing a complex mix of tissue properties as an arbitrarily scaled image, qMRI acquires maps of biophysical properties that are more specific to microstructural tissue composition. The maps are highly reproducible and thus can be used to compare between cohorts by region-of-interest or voxel-based statistical analysis^11,12. While most qMRI techniques are limited by a long scan time to acquire multiple measures, MR fingerprinting (MRF) enables fast, simultaneous acquisition of multiparametric tissue property maps, commonly including T1 and T2 relaxometry maps (T1 and T2 maps)¹³. A high-resolution isotropic-resolution (1mm³) whole-brain scan can be obtained under 10 minutes^14,15, making it highly feasible for clinical applications. MRF T1 and T2 maps demonstrated high sensitivity, specificity, and reproducibility in various in vivo clinical applications, such as differentiating brain tumor subtypes^16,17 and distinguishing subtle tissue abnormalities in the sclerotic hippocampus and ipsilateral temporal lobe^18,19. A prior study also reported visual review of the 3D MRF maps yielded additional clinically relevant information for cortical malformations such as FCD and periventricular nodular heterotopia¹⁴.

In this pilot study, we incorporate quantitative MRF features with the established and widely applied MAP pipeline to enhance its performance, particularly in false positive reduction. Our hypothesis is that MRF T1 and T2 tissue property changes reflect pathologic processes underlying true epileptic lesions, e.g., demyelination, gliosis, and reduced iron content, therefore providing additional information to separate the true lesions versus false positive findings. We first investigated the differences in MRF measures between true positive (TP) clusters and FP clusters based on MAP18 neural network prediction, using the same artificial neural network employed in the validation study by David et al.²⁰, i.e. the “Validation Neural Network” (valNN). We subsequently used MRF features to train a machine-learning model that distinguishes TP from FP clusters automatically, and evaluated its performance on both cluster level and patient level.

Methods

2.1. Subjects

This study was approved by the Institutional Review Board. Patients were retrospectively included if they met the following criteria: (1) had pharmacoresistant epilepsy and underwent pre-surgical evaluation at the Cleveland Clinic Epilepsy Center; (2) had pathologically or radiologically confirmed FCD; and (3) underwent a 3D whole-brain MRF research scan. Analyses were performed retrospectively and did not influence clinical recommendations. 59 age-and-gender-matched healthy controls (HCs) were also recruited. All subjects gave written informed consent.

2.2. MRF and MRI Acquisition

A 3D whole-brain MRF sequence¹⁴ was acquired on a Siemens 3T Prisma with a 20-channel head coil (Siemens), with field of view of 300×300×144 mm³, resolution=1×1×1 mm³, axial acquisition, and scan time of 10 min 24 sec. A 3D whole-brain B1 mapping sequence was acquired with the same FOV and resolution as the MRF scan to correct for B1 inhomogeneity (scan time=1 min 50 sec)¹⁴. MRF reconstruction was performed using a dictionary that contains signal evolution patterns from a wide range of combinations of T1 and T2 generated from Bloch Equations¹⁴. T1w images with similar contrast to the conventional T1w magnetization-prepared rapid gradient echo (MPRAGE) images were synthesized from the T1 map. 3D lesion label regions of interest (ROI) were created on the synthesized MRF 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 label 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. When prospective use of MAP (without using automated detection) led to discovery of a lesional finding⁹, 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. Clinical T1w MPRAGE were acquired with the following parameters: axial acquisition, FOV=240×240 mm², resolution=0.5×0.5×0.94 mm³, interpolation ON, repetition time=1,930 ms, echo time=2.57 ms, inversion time=1,100 ms, flip angle = 10°, GRAPPA = 2, and scan time = 3 min 59 s.

2.3. Image Processing

Figure 1 shows the study workflow. Clinical T1w MPRAGE images of each patient were processed using MAP18²⁰ in MATLAB 2021b (MathWorks, Natick, Massachusetts). GM and WM segmentation maps were generated using SPM12 along with the MAP18 process. The “large average from all 1.5 and 3T scanners (61 1.5 & 3T scanners with 3,716 T1 images)” option from MAP18 was chosen as the normal database. The valNN option within MAP18 yielded voxel-wise prediction of FCD, resulting in a map of multiple clusters with >0.01 probability values in the MNI space. Clusters containing only 1 voxel were considered noise and excluded. The remaining clusters were considered valNN+ clusters.

Figure 1.

Study workflow. (a) 37 patients with FCD and pharmacoresistant focal epilepsy and 59 HCs were included. (b) For each patient, high-resolution T1w clinical MRI and MRF scans were acquired. Clinical T1w images underwent morphometric analysis by MAP18 to generate GM and WM segmentation maps as well as voxel-wise FCD probability map using the “valNN” network. MRF T1 and T2 maps were reconstructed from a dictionary of signal evolution pattern based on Bloch equations. MRF T1w images were synthesized from the T1 maps, and lesion label was created based on the synthesized T1w images. All images were registered into the MNI space. (c) Two sample T-tests were performed on cluster-wise MRF values. Cluster-wise MRF values and volume in the GM and WM were used as input features for ML model to classify TP or FP clusters, using expert-created manual lesion labels as ground truth. GM, gray matter; HCs, healthy controls; MAP18, Morphometric Analysis Program; MNI, Montreal Neurological Institute; ML, machine learning; MRF, magnetic resonance fingerprinting; nT1, normalized T1; nT2, normalized T2; T1w, T1-weighted; WM, white matter.

MRF synthetic T1w images were registered to MNI space using the symmetric image normalization method (SyN) in advanced normalization tools (ANTs)²¹. SyN was chosen because it was shown to provide consistently high registration accuracy across subjects²². The warping transformation from synthetic T1w images was applied to T1 and T2 maps as well as the lesion label ROIs. Cerebrospinal fluid (CSF) voxels were excluded by thresholding at WM + GM fraction of 0.95. Voxels with MRF values beyond 2.5 standard deviations plus the maximal value, or below 2.5 SD minus the minimal value of whole-brain normal MRF atlases²³ were considered artifacts and therefore removed.

2.4. Calculation of MRF Metrics

The average MRF T1 and T2 were calculated in each valNN+ cluster for GM and WM separately. The GM and WM assignment of each voxel were determined by a threshold of 0.5 probability on the SPM segmentation maps, to include as many pathological voxels at the GM/WM junction as possible. Given that the location of the FCD lesion varies for each patient, and each region of the brain has its own characteristic mean MRF value^23,24, we also generated anatomically normalized MRF T1 and T2 values, using a z-score map based on HC MRF data. The HC data were processed with an identical workflow as the patient data. For each lesion ROI, the homologous regions of HCs were used for calculating the MRF mean and standard deviation (SD) within the regions. The mean MRF value of all voxels in the homologous regions from all HCs was considered the baseline HC mean (mean_{HC ROI}), and standard deviation of HC MRF values (SD_{HC ROI}) was computed for the same set of voxels. The normalized T1 and T2 were calculated as follows:

$$nT1=\frac{T1\mathit{\text{\hspace{0.17em}mea}}{n}_{\mathit{\text{Patient\hspace{0.17em}ROI}}}-T1\mathit{\text{\hspace{0.17em}mea}}{n}_{\mathit{\text{HC\hspace{0.17em}ROI}}}}{T1S{D}_{\mathit{\text{HC\hspace{0.17em}ROI}}}}$$

$$nT2=\frac{T2\mathit{\text{\hspace{0.17em}mea}}{n}_{\mathit{\text{Patient\hspace{0.17em}ROI}}}–T2\mathit{\text{\hspace{0.17em}mea}}{n}_{\mathit{\text{HC\hspace{0.17em}ROI}}}}{T2S{D}_{\mathit{\text{HC\hspace{0.17em}ROI}}}}$$

2.5. Statistics and Machine Learning

The Statistics and Machine Learning toolbox in MATLAB 2021b was used for all statistical analyses and ML model implementation. Clusters identified by MAP18 valNN were assessed for overlap with the lesion label ROIs. Clusters that overlapped with the lesion label ROI by at least one voxel were considered TP²⁰, while those that did not overlap with the lesion ROI were considered FP. Two-sample T-tests assuming unequal variance were performed to compare the mean MRF values between FP and TP clusters, with the significance level set at p<0.05.

An MRF-based neural network (MRF-ML model) was trained in Python Scikit-learn and Imbalanced-learn to separate TP and FP clusters, using MRF metrics (T1 and T2, or nT1 and nT2) and cluster volume in GM and WM separately. The network contains two hidden layers of 10 nodes with tanh activation functions. The optimization method was adaptive moment estimation (ADAM) and max iteration was 2,000. Ten-fold cross-validation was used to evaluate the model performance at cluster level. Random shuffling was performed before each 10-fold cross-validation and the whole process was repeated 5 times to ensure repeatability and reproducibility. Because FP clusters outnumber TP clusters by about twelve-fold, TP clusters in the training folds were oversampled with the Adaptive Synthetic (ADASYN) oversampling technique to generate synthetic samples based on 5 nearest neighbors. Training and test folds were stratified to ensure that the ratio of TP and FP clusters were similar between the training and testing partition in each fold. Sensitivity and specificity analyses were performed using Receiver Operating Characteristic (ROC) curve. The average area under the curve (AUC) was calculated for each training/testing set and averaged across the 5 trials to produce an overall evaluation of the classifier.

Model performance at the patient level was evaluated by the number of patients with TP lesion cluster successfully predicted by the MRF-ML model, as well as the number of FP clusters remaining. Prediction was considered successful for the patient if at least one of the TP clusters was detected by the MRF-ML model. The same model architecture as cluster-level classification was used for patient-level classification but trained/tested differently using leave-one-patient-out cross-validation, i.e., in each iteration, all clusters from one patient was left out for validation while the clusters in the rest of the patients formed the training set. ADASYN was similarly applied in the training set to address class imbalance. The result was compared to direct thresholding of MAP18 valNN probability map using various thresholds, with the 0.5 threshold being the benchmark as established previously²⁰. All analyses were first performed in the overall cohort and then in patients with pathologically confirmed FCD as subgroup analysis.

2.6. Histopathological Confirmation

Surgical specimens with immunohistochemical staining were microscopically reviewed by an expert pathologist (IB) for FCD subtypes using the 2011 ILAE classification guidelines²⁵, with considerations of the critical updates in 2022²⁶. For subgroup analysis, we included FCD type IIa and type IIb for type II. Mild malformation of cortical development (mMCD) and mild malformation of cortical development with oligodendroglial hyperplasia in epilepsy (MOGHE)²⁷ were grouped together as non-type II subgroup.

Results

3.1. Clinical Characteristics

A total of 37 patients were included in the study (demographic and clinical information summarized in Table 1, details in Table S1). FCD patients and HCs were similar for age (HC=25±5 years; patients=28±14 years, mean±std) and gender (HC=30 male/ 29 female; patients=20 male / 17 female). Eighteen patients had positive MRI; 6 patients had negative MRI; 13 patients were MAP+, i.e., their initial MRI reports were negative, but subtle lesions were detected with visual analysis of MAP (without using automated detection) and confirmed by radiologist review during PMC, similar to prior studies^8,9. Pathology subtypes included 10 type IIa (pathologically confirmed), 15 type IIb (11 pathologically confirmed, 4 radiologically confirmed [MRI-positive with transmantle sign, laser ablation or pending surgery]), 2 MOGHE (pathologically confirmed), and 10 mMCD (pathologically confirmed). Lesion locations were 20 frontal, 7 temporal, 2 temporo-occipital, 2 occipital, 2 cingulate, 1 parietal, 1 parieto-occipital, 1 central and 1 paracentral lobule. One-year postoperative seizure outcome data was available for 30 patients (Table S1): 25 patients (25/30, 83%) were seizure-free (ILAE-1 or −2), with 11 IIb, 7 IIa, 5 mMCD and 2 MOGHE as pathology subtypes; 5 patients (5/30, 17%) were not seizure-free (ILAE-4 or −5), with 2 IIa and 3 mMCD as the underlying pathology.

Table 1.

Demographic and clinical information for patients included in the study.

Number of Patients		37
Age (years)
	Mean	28.2
	Median	24
	SD	14.6
Gender
	Male	20
	Female	17
FCD Subtypes
	Type IIa	10
	Type IIb*	15
	mMCD	10
	MOGHE	2
Handedness (Left/right/ambidextrous)		4/31/2
Age at onset, years, mean ± SD (range)		11.7 ± 10.9 (Range: 1 ~ 49)
Epilepsy duration, years, mean ± SD (range)		16.6 ± 13.2 (Range: 2 ~ 58)
Clinical MRI a	Positive	18
	Negative	6
	MAP+	13
Post-operative seizure outcomes b	ILAE 1–2	25
	ILAE 4–5	5

^*For FCD IIb, 11 were pathologically confirmed, 4 were radiologically confirmed [MRI-positive with transmantle sign, with laser ablation or pending surgery]). All other cases were pathologically confirmed.

^aMRI findings (positive/negative) were defined by official radiological report; in patients who were noted as MAP+, initial MRI reports were negative, and subtle lesions were detected with MAP (without using automated detection) and confirmed by radiologist review, along with other multimodal data (semiology, EEG, PET, SPECT, MEG) during patient management conference.

^bPostoperative seizure outcome was classified using the ILAE classifications in patients with > 1 year postoperative follow-up.

3.2. MAP18 FCD Detection Performance

Using MAP18 valNN unthresholded output, FCD lesions were detected in 23 patients (valNN+); the pathology subtype included 8 IIa (8/10, 80%), 9 IIb (9/15, 60%), 5 mMCD (5/10, 50%), and 1 MOGHE (1/2, 50%). In 14 patients, the valNN+ regions were present but did not overlap with the lesion label; the pathology subtype included 2 IIa (2/10, 20%), 6 IIb (6/15, 40%), 5 mMCD (5/10, 50%), and 1 MOGHE (1/2, 50%). At the 0.5 benchmark threshold for valNN, 19 of the 23 patients remained valNN+, which includes 6 IIa (6/10, 60%), 7 IIb (7/15, 47%), 5 mMCD (5/10, 50%) and 1 MOGHE (1/2, 50%). An average of 2.9 FP clusters were seen per patient, at the 0.5 threshold. The percentage overlap between all TP clusters and lesion label averaged to be 13% at the 0.5 threshold. In the valNN+ patients, there was no significant association between seizure outcomes and complete inclusion of TP clusters in resection; nor was there significant association between seizure outcomes and overlap of TP cluster with lesion label; this could be related to the small sample size with highly unbalanced outcome data. The number of FP clusters were also not significantly different in the seizure-free and non-seizure-free groups. Figure 2 shows an example of a patient with a right frontal lobe FCD which was detected by one of the clusters generated by the MAP18 valNN; the TP cluster, although detected, was not the one with the highest mean probability.

Figure 2.

Example patient with a right frontal lobe FCD lesion, with images shown on two different slices: (a) Clinical T1w MPRAGE (input to MAP18) overlapped with the FCD probability map using “valNN”. (b) GM segmentation map. (c) WM segmentation map. (d) MRF T1 map. (e) MRF T2 map. (f) Multiple clusters were present on the MAP18 output FCD probability map, including true-positive (TP) and false-positive (FP) clusters. The cluster with highest probability may not be the TP cluster. As shown in this case, the TP cluster (circled) had lower probability than some of the FP clusters on the more ventral slice.

3.3. Comparison of MRF Measures between TP and FP Clusters

As shown in Figure 3a–b, unnormalized GM T1 values were significantly higher in the TP clusters (mean±SD=1,313±105 ms, n=27) than the FP clusters (mean±SD=1,256±147 ms, n=355, p=0.014). Other measures such as WM T1, GM T2 and WM T2 showed similar trend but did not reach significance. For normalized MRF values (Figure 3c–d), WM nT1 was significantly higher for TP (mean±SD=0.79±0.7, n=27) than FP clusters (mean±SD=0.24±0.96, n=353, p<0.001). The normalized WM nT2 was also significantly higher for TP (mean±SD=0.43±0.86) than FP clusters (mean±SD=0.02±0.89, p=0.026). Other normalized measures such as GM nT1 and GM nT2 showed similar trend but did not reach significance. More details are in Table S2.

Figure 3.

(a-b) MRF T1 and T2 comparison in TP and FP clusters. (c-d) Normalized T1 and T2 comparison in TP and FP clusters. Error bars show standard errors. *p<0.05, ns: non-significant.

3.3. MRF-ML Model Performance

On cluster level, the MRF-ML model trained on unnormalized MRF metrics and cluster volume achieved mean±SD AUC of 0.73±0.17 for separating TP and FP clusters, with optimal sensitivity of 82.1% and specificity of 39.9% (Figure 4a). The MRF-ML model trained on normalized MRF metrics and cluster volume showed mean±SD AUC of 0.83±0.13 for separating TP and FP clusters, with sensitivity of 82.1% and a markedly higher specificity of 71.7% (Figure 4b). The relative feature importance for the latter model, ranked by coefficients generated by sci-kit learn were: WM nT2 (1), GM volume (0.91), WM nT1 (0.76), GM nT1 (0.48) and GM nT2 (0). Direct thresholding of the valNN probability map achieved lower cluster-level sensitivity of 41.9%−59.8% across all thresholds, although specificity increased up to 86.6% as threshold became stricter (Table S3).

Figure 4.

(a) Receiver Operating Characteristic (ROC) curve of the 10 -fold cross-validation of the NN model trained on unnormalized MRF metrics (GM T1, WM T1, GM T2, WM T2) with cluster volume in GM and WM; the confusion matrix in (a) corresponds to the threshold marked as red dot. (b) ROC curve of the 10-fold cross-validation of the NN model using normalized MRF metrics (GM nT1, WM nT1, GM nT2, WM nT2) with cluster volume in GM and WM; and the confusion matrix in (b) corresponds to the threshold marked as red dot. FNR, False negative rate; FPR, False positive rate; TNR, True negative rate (specificity); TPR, True positive rate (sensitivity).

In terms of patient-level performance, the MRF-ML model achieved higher sensitivity in the 23 valNN+ patients than thresholding at 0.5 (Figure 5a), leaving an average of 3 FP clusters per patient. This performance was consistent in the overall cohort and different subtypes. To approach similar sensitivity of the MRF-ML model, the direct thresholding method would have to use no threshold, and therefore drastically sacrifice specificity, leaving 10.1 FP clusters per patient (Table S3). As shown in Figure 5b, in the 23 valNN+ patients, the MRF-ML model eliminated ≥ 75% FP clusters in 30.4% of patients, and ≥ 50% of FP clusters in 91.3% of patients. The performance was similar for type II and non-type-II subgroups (Figure 5b). The percentages of FP cluster elimination across different lobes showed no significant preference for a specific lobar region (Table S4).

Figure 5.

(a) Comparison of patient-level detection rate between MRF-ML model and direct thresholding of the MAP18 valNN output for different subtypes. Only the 23 valNN+ patients were included. (b) Patient-level FP elimination percentage chart in valNN+ patients, for all subtypes, type II, and non-type II. Colors indicate the percentages of FP clusters eliminated for a given patient.

3.4. Subgroup Analysis

When considering the subgroup of patients with pathologically confirmed FCD (N=33), 21 patients were valNN+. Similar to the overall cohort, GM T1 and GM nT1 were significantly higher in TP clusters than FP clusters. WM nT2 in WM were trending towards but not significantly different between TP and FP clusters (p=0.07, Table S5). Nevertheless, the model performance was similar to the overall cohort, at both cluster level and patient level. On cluster level, the model trained on normalized MRF metrics and cluster volume still showed superior performance as compared to the unnormalized MRF metrics, with mean±SD AUC of 0.83±0.14, sensitivity of 84% and specificity of 64.7% (Figure S1). On patient level (Figure S2), in the 21 valNN+ patients, the MRF-ML model eliminated ≥ 75% FP clusters in 15.3% of patients, and ≥ 50% of FP clusters in 80.9% of patients, while maintaining sensitivity.

Discussion

Our study incorporated MR fingerprinting with the most widely adopted voxel-based morphometric MRI analysis program (MAP) to improve focal cortical dysplasia detection. We first showed the significant differences in MRF metrics in clusters representing true pathological and false positive findings (supposedly representing normal variants or imaging artifacts). Using quantitative MRF measures as input to a machine learning model resulted in a remarkable separation between TP and FP clusters generated by post-processing of conventional MRI. The MRF-ML model outperformed direct thresholding of the valNN FCD probability map (a commonly employed approach) on both cluster- and patient-level. The performance was consistent in the overall cohort and the subgroup of patients with pathologically confirmed FCD. Patient-level model performances were also consistent amongst type II and none-type-II FCDs, demonstrating the robustness of our findings.

MRF T1 and T2 values

The sensitivity of quantitative MRI to subtle variations in tissue properties has made it an increasingly utilized tool to characterize pathological tissue in various neurological diseases^19,28,29. Prior quantitative MRI studies showed T1 relaxation time is mainly affected by myelin and water content^30–32, while T2 is affected by water mobility, iron deposition, myelin and glial cell count³³. On histopathology, WM myelin reduction was clearly seen in FCD type IIb and to a lesser extent in FCD IIa^34,35; association between the severity of hypomyelination and disease duration has been reported³⁶. The role of demyelination in seizure generation and epileptogenesis has been recently proposed - demyelination of both WM and GM axons was thought to be associated with increase in hyperexcitability³⁷. In the GM, studies reported disorganized cortical myeloarchitecture in FCD IIb³⁴ and thinner myelin sheaths of layer V in FCD IIa³⁸. The reduction of myelin content in both WM and GM could therefore result in increased T1. T2 is closely related to motion restricted water protons in the myelin, therefore could also increase when myelin content is reduced. Additionally, abnormal neuronal content could also play a role in generating different qMRI signals. Balloon cells, the hallmark of type IIb FCD, are giant in size, with opalescent, glassy eosinophilic cytoplasm²⁵, which are drastically different than normal neurons. Dysmorphic neurons, seen in both IIa and IIb, are significantly larger than any typical pyramidal cells in controls; glial cells are also dysmorphic and often enlarged²⁶. These abnormal cellular compositions would represent higher water content²⁵, which may also lead to increased T1. Since iron co-localizes with myelin, reduced iron content caused by demyelination will cause T2 increase³⁹. Last but not least, gliosis was histologically confirmed to increase both T1 and T2⁴⁰. Overall, the increased MRF signals in both GM and WM could be explained by the myriad of pathophysiological changes underlying the FCD lesions.

Causes for False Positive Findings

False positive findings on MAP can be due to a variety of reasons. In a previous study, we illustrated some commonly encountered examples⁴¹. A main culprit is imaging artifact, e.g., field inhomogeneity, motion, or pulsation from vasculature. MRF provides an innate advantage to address many of these issues commonly seen on conventional MRI. Prior studies reported the MRF images are less prone to field inhomogeneity and motion^13,42–44, providing a valuable avenue for extracting quantitative and tissue-specific information in the brain region of interest. False positives on conventional MRI can also be due to partial volume effect and/or registration errors in brain regions with small and complex gyri, especially when comparing to a normal control database. High-resolution 3D MRF could help provide additional information for the post-processing, due to its low across-subject variability on normal controls²³, and the perfectly co-registered T1 and T2 maps from a single scan enables multiparametric quantitative analysis. These factors may explain why incorporating MRF data can help separate true positive and false positive findings generated by conventional MRI post-processing in the current study.

Important Features for MRF-ML Model Performance

The feature importance coefficients of the MRF-ML model indicated that normalized WM T1, WM T2 and GM volume were the largest contributors to the prediction. It is reasonable that a cluster needs sufficient GM volume for its classification as a TP cluster, as FCD lesions primarily reside in cortical GM and GW junction²⁵. Although WM volume itself was not as important as GM volume for the model prediction, normalized WM T1 and T2 signals were both shown to be the largest contributors, likely reflecting the degree of microstructural changes in the subcortical WM – e.g., in the WM of FCD type II, hypomyelinated neurons and balloon cells were reported; in mMCD, excessive heterotopic neurons can be seen^25,26,45. Curiously, GM MRF features were not as important for the model prediction. This may be explained by the larger anatomical intersubject variability of the cortical GM, which makes this measure more prone to imperfect spatial normalization with the HCs²⁴. Interestingly, our data showed that the MRF-ML model achieved similar performance in identifying TP clusters for all included FCD subtypes, demonstrating generalizability of the approach.

Effect of Normalization

Our study showed better classification performance when normalized MRF metrics were used as input to the MRF-ML classifier. Previous studies investigating normal brain T1 and T2 relaxometry atlases reported that T1 and T2 values varied across cortical regions, and their spatial inhomogeneity was more pronounced than the inter-subject variation^23,24. In our prior MRF normal atlas study, significantly lower T1 and T2 were identified in the precentral, postcentral, paracentral lobule, transverse temporal, lateral occipital, and cingulate areas²³. Region-specific baseline is therefore especially needed for pathology detection in these regions to alleviate the spatial confounding factors. Consistent with this rationale, our data showed that the ROI-based anatomical normalization led to better classification performance for separating TP and FP clusters.

Other FP Reduction Approaches for FCD Detection

Many prior studies investigated automated FCD prediction using machine/deep learning with voxel-wise or vertex-wise MRI-based features as input^20,46–48. FP findings were commonly reported⁴⁹. The desired balance between sensitivity and specificity is dependent on the cohort being evaluated and the user’s experience. At multimodal patient management conferences, semiology, video-EEG monitoring or PET can provide lobar localization or lateralization, and FP findings on MRI can be quickly ruled out based on these other findings. Therefore, weighing sensitivity more than specificity may be helpful in evaluating patients with an apparently normal MRI. However, too many FPs will inevitably make MRI review inefficient and case discussions unfocused. In our experience, up to 3–5 regions to review per case seems to be practical. In the current study, with optimized performance, the MRF-ML model still left an average of 3 FP clusters per patient. To achieve similar sensitivity to the MRF-ML model, directly thresholding the valNN output would have to drastically sacrifice specificity, leaving 10 FP clusters per patient, which may be cumbersome to review in a clinical setting. In a recent study⁴⁷, multicenter clinical MRI data from 618 patients with focal FCD-related epilepsy were used to construct a neural network based on surface-based features. A sensitivity of 85% was achieved in the sub-cohort of seizure-free FCD IIB patients, while in the entire withheld test cohort the sensitivity was 59% and specificity was lower at 54%⁴⁷. In another study⁴⁶, data from 148 patients with histologically verified FCD were trained to develop a deep convolutional neural network (CNN) with 3D patches of T1w and FLAIR images as input⁴⁶. An overall sensitivity of 83% was achieved in the independent testing cohort and specificity of 89% in healthy and disease controls. To handle false positive findings, a bayesian uncertainty estimation was proposed as a measure of confidence, and the true FCD was indeed among the clusters with the highest confidence in 73%⁴⁶. In the current study, we essentially used MRF metrics to build a second-stage classifier on voxel-wise prediction based on the clinical MRI, and drastically decreased FP clusters per patient while maintaining sensitivity.

Sensitivity Considerations

Our study did not address the sensitivity of FCD prediction. Only 62% (23/37) of the included patients were valNN+, which was lower than the sensitivity reported previously (81% on independent validation data)²⁰. Some of the included cases were not valNN+, but successfully detected by MAP visual analysis with a lowered z-score threshold, guided by focal finding from other multimodal data at patient management conference. This lower sensitivity could be because the MAP18 valNN was trained on FCD II cases that were clearly MRI visible, but a third of our cohort were non-type II (mMCD and MOGHE) cases. In our current study, we chose to use unthresholded valNN outputs for the next level of analysis, in order to maximize sensitivity of the overall MRF-ML patient-level detection. Improving the sensitivity of FCD detection, especially for the cases with visually negative MRI, is an extremely important topic beyond the scope of the current study. MRF features, in combination with morphometric features, may lead to a more sensitive, quantitative-MRI-based framework for FCD detection (a parallel study is currently ongoing). More MRF data of FCD cases (than what was available from the current study) are likely needed for training these networks; as in the case of valNN, 113 patient data sets were used for training.

Limitations

First, our pilot study consists of a relatively small cohort. With a larger sample size, the MRF-ML model may be better trained, especially for the non-type-II subgroups. The smaller sample size also prevented us from further splitting the cohort to a training/cross-validation and an independent cohort for validation. Secondly, brain tissue sampling can affect the final pathology diagnosis, therefore impacting the subtype groups. Lastly, it cannot be ruled out that some of the FP clusters located ipsilateral to the lesion label ROIs may indeed correspond to dysplastic tissue that escaped the labeling of the ROIs, or in principle resides below the detection threshold for human reading of MRI. Therefore, some of the false-positive findings may not be necessarily incorrect.

Conclusion

By incorporating MRF features into voxel-based morphometric post-processing, we successfully reduced most of the FP clusters in a high percentage of patients while maintaining patient-level sensitivity. Significant increase in normalized MRF T1 and T2 were observed in TP clusters compared to FP clusters, especially in the WM, reflecting tissue property changes of the underlying pathology. Our work demonstrates the potential of using MRF and machine learning techniques to aid epileptic lesion detection, which is a crucial part of noninvasive presurgical evaluation for individuals with pharmacoresistant focal epilepsy.

Key Points.

• We used three-dimensional (3D) magnetic resonance fingerprinting (MRF) to reduce false positive (FP) clusters from automated FCD detection using MAP18.

• MRF metrics were significantly higher in true-positive (TP) clusters than FP clusters, including GM T1, normalized WM T1, and normalized WM T2.

• A machine-learning (ML) model using MRF metrics and cluster volume was developed to distinguish TP and FP clusters, with mean AUC of 0.83, sensitivity of 82.1% and specificity of 71.7%.

• The MRF-ML model showed superior performance over direct thresholding of the MAP18 FCD probability map at both cluster- and patient-levels, eliminating ≥ 75% FP clusters in 30% of patients and ≥ 50% of FP clusters in 91% of patients.

• Integrating MRF into presurgical evaluation may facilitate accurate localization of FCD lesions.

Abbreviations:

FCD

focal cortical dysplasia

FP

false positive

GM

gray matter

HC

healthy control

ILAE

International League Against Epilepsy

MAP

morphometric analysis program

ML

machine learning

MNI

Montréal Neurological Institute

MRI

magnetic resonance imaging

MRF

magnetic resonance fingerprinting

TP

true positive

WM

white matter

valNN

validation neural network

Data Availability

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