PET/MRI in the Presurgical Evaluation of Patients with Epilepsy: A Concordance Analysis

The aim of our prospective study was to evaluate the clinical impact of hybrid [18F]-fluorodeoxyglucose positron emission tomography/magnetic resonance imaging ([18F]-FDG PET/MRI) on the decision workflow of epileptic patients with discordant electroclinical and MRI data. A novel mathematical model was introduced for a clinical concordance calculation supporting the classification of our patients by subgroups of clinical decisions. Fifty-nine epileptic patients with discordant clinical and diagnostic results or MRI negativity were included in this study. The diagnostic value of the PET/MRI was compared to other modalities of presurgical evaluation (e.g., electroclinical data, PET, and MRI). The results of the population-level statistical analysis of the introduced data fusion technique and concordance analysis demonstrated that this model could be the basis for the development of a more accurate clinical decision support parameter in the future. Therefore, making the establishment of “invasive” (operable and implantable) and “not eligible for any further invasive procedures” groups could be much more exact. Our results confirmed the relevance of PET/MRI with the diagnostic algorithm of presurgical evaluation. The introduction of a concordance analysis could be of high importance in clinical and surgical decision-making in the management of epileptic patients. Our study corroborated previous findings regarding the advantages of hybrid PET/MRI technology over MRI and electroclinical data.

Clinical decision-making is particularly challenging in patients with discordant neuroimaging and electroclinical data, with MRI-negative results, or with the occurrence of multiple epileptic foci. Furthermore, the complexity of electroclinical and neuroimaging data challenges presurgical decision-making [2,6,14,16].
The aim of our prospective study was to evaluate the clinical impact of hybrid [ 18 F]-FDG PET/MRI on the presurgical evaluation of patients with pharmacoresistant epilepsy and to introduce a mathematical model from the multi-modality tests that may facilitate the development of artificial intelligence for the analysis of different concordance patterns.

Subjects
This prospective study was approved by the Scientific Research Ethics Committee of the Medical Research Council (008899/2016/OTIG) and carried out in accordance with the Declaration of Helsinki of the World Medical Association. Seventy patients with refractory focal epilepsy underwent a full electroclinical presurgical evaluation between June 2016 and December 2017. The inclusion criteria were: (i) pharmacoresistant focal epilepsy, (ii) MRI scans with discordant results or without noticeable morphologic epileptogenic lesion, (iii) VEEG monitoring in each patient, and (iv) age of 18-65 years. Exclusion criteria included: (i) standard contraindications for MRI examinations, (ii) acute non-epileptic neurological disorder, (iii) acute infection, and (iv) serious comorbidities. Ten of these patients were excluded from further analysis after the multidisciplinary team revealed multifocal or diffuse pathological alterations (encephalitis n = 7, vasculitis n = 2, and hydrocephalus n = 1). One more patient was removed from the current analysis because of compromised image quality. The median age of the remaining 59 patients was 33 years (range: 18-57 years), and the cohort contained 35 male and 24 female patients.

Patient Preparation
All epileptic patients were hospitalized for adaptation a day prior to the study, and a standard neurological examination, electrocardiography (ECG), and routine laboratory tests were performed. Written consent was obtained from all participants. Dual-modality [ 18 F]-FDG PET/MR imaging was performed the next day. The standardized patient preparation for the PET examination was performed according to the European guideline of 2009 [29]. Briefly, supervision of a 2 h duration and VEEG monitoring (in 10-20 EEG Placement) were performed before the intravenous tracer administration. VEEG monitoring covered the whole uptake period of the tracer to ensure the interictal state. PET/MRI acquisition started 60 min after the injection.

PET/MRI Acquisition
All PET/MRI acquisitions were performed on a Biograph mMR scanner (Siemens Healthineers, Erlangen, Germany). The detailed dedicated seizure protocol of MRI acquisition is summarized in Table 1. In order to provide a complete temporally and spatially correlated PET dataset, a 20 min and 35 min list-mode 3D PET acquisition was performed simultaneously for each patient. Vendor-provided UTE sequence was used for PET attenuation correction (AC) purposes, and MR-based attenuation maps were generated automatically. Static image reconstruction was performed both for 20 min and 35 min. AC and non-AC transaxial slices were generated. For PET image reconstruction, the OP-OSEM method was applied, including PSF correction (3 iterations, 21 subsets, 4 mm full-width at half-maximum (FWHM) Gaussian filtering, and 344 × 344 × 127 imaging matrix). µMaps were checked for potential artifacts, and the completed PET raw data were archived for further evaluation. For the current assessment, a 20 min static PET image dataset was used.

Image Processing
An in-house image processing pipeline was applied to transform all individual images into the MNI152 atlas space prior to the regional analysis of the [ 18 F]-FDG PET images using Statistical Parametric Mapping (SPM). At the beginning of this procedure, we used the "recon-all" pipeline of FreeSurfer software (version 7.0) for the segmentation of T1-MPRAGE images [30][31][32]. The produced segmented T1-MPRAGE images were used for correcting the misalignment of PET/MR image pairs, global voxel intensity scaling, and calculating the transformations required by the spatial normalization. In the latter case, we applied the FSL software package (version 6.0) [33] and the Advanced Normalization Tools software (version 2.3.5) [34] for calculating the rigid body, 12-parameter affine, and non-linear transformations. After the transformations of the [ 18 F]-FDG PET images into the MNI152 space, to eliminate the inter-subject variability of the measured global-brain metabolism according to the standard PET-SPM method, we set the average of the withinbrain mask voxel-values of the PET images to 50 [35]. Finally, on the normalized [ 18 F]-FDG PET images, we applied a 10 mm and 2 mm 3D Gaussian kernel-based smoothing for the SPM and the regional analysis, respectively.
We used the spatially standardized, globally normalized, and smoothed [ 18 F]-FDG PET data and the spatially standardized T1-MPRAGE and T2-FLAIR images for calculating 15 quantitative image-processing parameters for all patients with four image-processing methods ( Table 2). The quantitative image-processing parameters were evaluated by VOI (volume of interest) analysis, asymmetry index calculations, SPM analysis, and MAP07 analysis using the spatially standardized, globally normalized, and smoothed [ 18 F]-FDG PET, and the spatially standardized T1-MPRAGE and T2-FLAIR images. During this study, two regional systems in the MNI152 space were applied: the Harvard-Oxford Cortical and Subcortical Atlas (HOVOI), containing 124 (96 cortical and 28 subcortical) regions suitable for regional analysis, and the 14 regions, combined from HOVOI's regions, used in electroclinical data evaluation (EPIREG system) [36]. All quantitative image-processing parameters were converted into these regions for the purpose of statistical and concordance analysis.
The minimum, maximum, mean, median, and standard deviation (voi.min, voi.max, voi.mean, voi.median, and voi.sd) of the regional [ 18 F]-FDG values for all HOVOI regions were estimated in the VOI analysis procedure of the [ 18 F]-FDG PET images. The VOI parameters of the overlapping HOVOI regions were used for the regional characterization of the EPIREG system by selecting the minimal value in the case of the voi.min parameter and maximum values in the other cases ( Table 2). The maximum values were applied to ensure that the highest average, median, and standard deviation HOVOI data were used to characterize the appropriate EPIREG area, thus preserving the regional variability of the [ 18 F]-FDG PET and composite z-score images.
An asymmetry index (AI) calculation of the [ 18 F]-FDG PET images was used on symmetric regions of the HOVOI system by applying the formula AI = 100 × 2 × (L − R)/ (L + R), where L and R represent the mean intensity values (ai.mean) of the corresponding left and right regions of the HOVOI system. Additionally, using a similar formula, the asymmetry of the maximum, median, and standard deviation (ai.max, ai.median, and ai.sd) were evaluated using a similar formula ( Table 2).
An HOVOI-based regional analysis of the Student-t maps was performed by the SPM12 software [37]. A Student-t map was created for each patient using the statistical comparison of their [ 18 F]-FDG PET image and the reference metabolic PET image database from our lab, which was built from a previously recorded data pool of 19 cases showing normal PET/MRI patterns. The maximum of the Student-t values and the volume of the hypometabolic region were deployed for characterizing the regional properties of the Student-t maps, sorted by an uncorrected p < 0.001 as a threshold (spm.max, spm.vol) ( Table 2).
An HOVOI-based regional analysis of the "Composite z-score" images was performed by MAP07. Morphometric analyses were applied to the T1-MPRAGE and T2-FLAIR MRI data sets of the patients using the MAP07 software [38]. The maximum, mean, median, and standard deviation estimates (map.max, map.mean, map.median, and map.sd) were used for characterizing the regional properties of the "Composite z-score" images ( Table 2).
The visual analysis of the PET images was performed and analyzed by the authors, KB and ZT, and the MRI images by the authors PB and ZV.

Clinical Data
Electroclinical information and the results of the visual analysis of the PET and MRI images were extracted from patient documentation. Additional PET/MRI investigations were applied for the EPILOBE region-based statistical and concordance analysis (Table 3). According to the possible therapeutic options (resective surgery, neuromodulation, and new antiepileptic drugs), the experts of the epilepsy team (EPI team) categorized the patients by two methods using clinical decision (CD): "Grouping Method 1" (CD1): eligible for resective surgery (without icEEG investigation) and defined as "operable" (7 patients), considered for icEEG exploration and defined as "implantable" (38 patients), or not eligible for any further invasive procedures and defined as "inoperable" (14 patients). During the "Grouping Method 2" (CD2), the simplification of categorization was performed for "inoperable" (14 patients) vs. "eligible for invasive treatment" (45 patients) groups. Table 3. EPILOBE region-wide electroclinical and expert-based imaging data recorded during the study.

Semiology
Possible localization considered by semiology in the given EPILOBE region.

Statistical Comparison of Electroclinical and Image Processing Data
Non-normal distribution of the investigated variables was confirmed by the Shapiro-Wilk test. Hence, to assess the relationship between these data and the categorical clinical parameters, non-parametric Mann-Whitney or pairwise Wilcoxon tests were performed. After the statistical analysis, p-values were adjusted to control the False Discovery Rate (FDR) [39], and significant relations were selected by the corrected p < 0.05 criteria. All statistical analyses were performed by R version 3.6.3. (The R Foundation).

Concordance of the Clinical Data
Using the EPILOBE region-wide electroclinical (Semiology, interictal EEG-iiEEG, and ictal EEG-iEEG) and expert-based imaging data (MRI1, MRI2, and PET.vis), we constructed a localization observation matrix according to the 14 brain regions and the six diagnostic parameters. We excluded the most frequent iiEEG (iiEEG.mfl) and the most frequent iEEG (iEEG.mfl) localization parameters to avoid the over-representation of the ictal and interictal EEG observations. This type of data fusion is suitable for interobserver-analysis regarding different diagnostic procedures, including the independent observations and different regions of EPIREG. Gwet's AC1 statistics was chosen for the agreement analysis since it was demonstrated to be insensitive to small differences [40,41].
Gwet's AC1 parameters helped to assess the agreement between different ratings, thus enabling the definition of a new parameter for clinical data concordance (CDC). For our study, the value of the CDC was between 0 and 1, whereby 0 meant "full discordance" and 1 stood for "full concordance." The performance of the CDC parameters was assessed by means of patient categories, similar to the expert-made clinical decisions-based classification ("eligible for resective surgery," "considered for icEEG," "not eligible for any further invasive procedures"). Eight CDC values (electroclinical data (EC), EC + MRI1, EC + MRI2, EC + PET, EC + PET + MRI2, EC + MRI1 + PET + MRI2, EC + MRI1 + MRI2, and EC + MRI1 + PET) were assessed, applying two types of patient classifications (CD1 and CD2).

Quantitative PET and MRI Analysis
Examples of the results of the presurgical evaluation tests with pathologic findings and the corresponding circular plots of the presurgical data demonstrating different patterns of concordance are shown in Figure 1A-D. The statistical analysis resulted in 28 significant (FDR-corrected p < 0.05) regional associations between the image processing data and clinical data (Table 4). Visual PET investigations (PET.vis) of the regional data correlated with the metabolic-PET asymmetry parameters and the maximal Student-t value of the SPM analysis. The visually localized lesions in the MRI component of the PET/MRI (MRI2) measurements correlated with the PET asymmetry indexes; however, they did not correlate with the MAP07 data. The interictal EEG (iiEEG) localization significantly correlated with the VOI analysis data and the MAP07 regional maximum values, while the iiEEG.mfl localization presented a The statistical analysis resulted in 28 significant (FDR-corrected p < 0.05) regional associations between the image processing data and clinical data (Table 4). Visual PET investigations (PET.vis) of the regional data correlated with the metabolic-PET asymmetry parameters and the maximal Student-t value of the SPM analysis. The visually localized lesions in the MRI component of the PET/MRI (MRI2) measurements correlated with the PET asymmetry indexes; however, they did not correlate with the MAP07 data. The interictal EEG (iiEEG) localization significantly correlated with the VOI analysis data and the MAP07 regional maximum values, while the iiEEG.mfl localization presented a statistically significant association with the SPM-detected relative volume of hypometabolism. Semiology-or iEEG-based localization did not show any significant association with the image processing data. The iiEEG and iiEEG.mfl activity localization significantly correlated with the [ 18 F]-FDG regional maximum value asymmetry, the [ 18 F]-FDG regional mean, median, and standard deviation, the MAP07 generated "composite z-score" maximum, and the SPMbased estimation of the hypometabolic region of the temporal and frontolateral lobes.
We found that the asymmetry score of the regions was highly correlated with the visually identified lesions, mostly in the temporal and the frontal lobes. Despite the low amount of the cardinality of the normative [ 18 F]-FDG PET database (n = 19), we could demonstrate that the results of SPM analysis, in the cases of temporal lobe hypometabolism, correlated with the visual findings.

Concordance Analysis
The eight concordance parameters in the CD1-type classification statistical analysis by FDR-corrected p-values revealed that neither CDC variant could significantly separate the group pairs ( Figure 2). However, a tendency was present; in the case of PET-related CDCs, the "inoperable" group showed a borderline significant difference compared to the "operable" or "implantable" groups. In contrast, when the "operable" and "implantable" groups were integrated into the "invasive" group (CD2 classification), only CDC variants containing PET were able to statistically differentiate between the "invasive" and "inoperable" categories ( Figure 3A). Figure 3B illustrates the clinical decision differentiation capabilities of the introduced eight CDC parameters by the p-values of the Mann-Whitney applied on the CDC-CD2 analysis tests (controlled for FDR).

Concordance Analysis
The eight concordance parameters in the CD1-type classification statistical analysis by FDR-corrected p-values revealed that neither CDC variant could significantly separate the group pairs ( Figure 2). However, a tendency was present; in the case of PET-related CDCs, the "inoperable" group showed a borderline significant difference compared to the "operable" or "implantable" groups. In contrast, when the "operable" and "implantable" groups were integrated into the "invasive" group (CD2 classification), only CDC variants containing PET were able to statistically differentiate between the "invasive" and "inoperable" categories ( Figure 3A). Figure 3B illustrates the clinical decision differentiation capabilities of the introduced eight CDC parameters by the p-values of the Mann-Whitney applied on the CDC-CD2 analysis tests (controlled for FDR).

Figure 2.
Clinical data concordance (CDC) of "Grouping Method 1." Boxplots of the eight CDCs grouped by three-way clinical decisions. PET-related measurements showed a slight, but not significant, difference between the "inoperable" versus the "operable" or the "implantable" groups, while PET-independent methods showed relatively less accuracy. Analyzed by Mann-Whitney tests with FDR-correction, p-values are shown on the intervals in the charts.  and "implantable" groups were integrated into the "invasive" group, only PET-related CDC variants were able to significantly differentiate between "invasive" and "inoperable" categories.

Discussion
In our study, the electroclinical data of patients with drug-resistant epilepsy presented a widely discordant pattern. The aim of our prospective study was to test the performance of dual-modality [ 18 F]-FDG PET/MRI in patients with pharmacoresistant epilepsy. Using an objective statistical method, we demonstrated that metabolic hybrid PET/MRI technology may significantly contribute to the clinical decision-making in patients with discordant electroclinical and imaging data. The decisions of the EPI team were based on professional knowledge and skills. However, the decision-making was subjective and carried the potential for diagnostic uncertainty among patients with discordant data, which could be even more challenging in the case of MRI-negative patients [2,14]. For this purpose, we introduced a novel concordance analysis method, which demonstrated that PET matrices are of high importance and well-suited to support clinical decisions, especially the matrices including both PET and 3T MRI.
Numerous previous publications suggested the advantages of simultaneous PET/MRI technology over the diagnostic algorithm, with only MRI and electroclinical data [2,[11][12][13]15,17,[26][27][28]42,43]; however, recent studies have applied a mathematical model to confirm its reliability. Both statistical and concordance analysis highlighted the role of PET imaging for the non-invasive localization of epileptogenic foci, especially in patients with discordant electroclinical data and MRI scans without a definitive epileptogenic lesion or patients with multiple abnormalities.
A concordance analysis demonstrated that PET/MRI examination is able to differentiate between the "invasive" (eligible for invasive treatment) and "inoperable" groups. PET was particularly important in the selection of inoperable patients and confirmed MRIpositive lesions. MRI-assisted PET post-processing techniques (such as the brain atlas-based asymmetry index calculation and SPM analysis) also held additional supportive value for defining clinical decisions. Comparing the visual PET assessment and quantitative PET data, an association between the asymmetry index parameters and visual PET localization proved to be significant, especially for both temporal lobes.
Moreover, besides its good feasibility and proper applicability, the hybrid PET/MRI was justified by reductions in radiation exposure, time savings, anesthesia, simplification of study-related organizational and design factors, a range of personalized diagnostic tests, and a range of comorbidity and medication data that may arise [18,23,43,44].
Another non-invasive alternative for localizing epileptogenic foci is simultaneous fMRI and EEG recording. In our study, the positive predictive value of interictal epileptiform discharges, associated with BOLD changes within 2 cm of the epileptogenic zone, was 78%, and the negative predictive value was 81% [45]. Additionally, EEG-fMRI can distinguish between ictal onset-, spread-, and preictal-related BOLD changes [46][47][48]. Besides the singlepass EEG/PET/fMRI [20], the recently reported sub-second analysis method [49] and the topography-related EEG-fMRI [50] may also improve the detection rate of epileptic foci.
In summary, our model confirmed the relevance of simultaneous PET/MRI for epileptic treatment planning. Additionally, the proposed clinical concordance calculation could support the development of a novel artificial intelligence-based decision system in the near future.

Conclusions
The fully integrated hybrid [ 18 F]-FDG PET/MRI has demonstrated a significant impact on the presurgical evaluation workflow of patients with pharmacoresistant epilepsy. The diagnostic algorithm of presurgical evaluation should not miss the comprehensive compliance of PET/MRI, mainly for precarious, subtle lesions or uncertain metabolic patterns. The introduction of a concordance analysis may help the EPI team in clinical decisionmaking in the future.