2.1.1 Patients with temporal lobe epilepsy and negative magnetic resonance imaging (MRIn-MTLE)
We retrospectively screened MRIn-MTLE patients using the Picture Archiving and Communication System (PACS) and extracted T1W-MPRAGE images and general information of MRIn-MTLE patients who visited our hospital between January 2017 and December 2020. We also collected information on abnormal long-range video electroencephalography (VEEG) discharges during a 24-hour period. T1W-MPRAGE is a standard protocol in our center for patients with epilepsy, together with 1) T2 CUBE FLAIR(fluid attenuated inversion recovery) protocol for hippocampal signal evaluation, 2) T2 CUBE protocol for surgery guidance planning, 4) ASL(arterial spin labeling) protocol for affected brain region evaluation, 5) PET-CT, 6) VEEG. Epileptogenic foci location may or may not identified before the MRI scan, in that we need to wait for the results of PET-CT and VEEG, while the segmentation and volume evaluation will take place simultaneously to help image reading. All patients who met the inclusion and exclusion criteria were included in the MRIn-MTLE group. Prior to our study, patients were recruited through the Regional Collaborative Innovation Program of the Xinjiang Uygur Autonomous Region (Grant #2020E0275). This research was authorized by the Medical Research Ethics Board of First Affiliated Hospital of Xinjiang Medical University, and all subjects provided written informed consent.
2.1.1.1 Inclusion criteria
Patients were included if they (1) met the epilepsy diagnostic criteria of the International League Against Epilepsy (2) underwent VEEG examination and exhibited a clear origin of abnormal discharges in the left temporal lobe, (3) had no clear epileptogenic focus identified by conventional MRI and 3D T2-weighted FLAIR sequences (including SPACE and CUBE) by two imaging physicians with more than 5 years of epilepsy diagnostic experience; and (4) had less than or equal to 2 antiepileptic drugs(AEDs).
2.1.1.2 Exclusion criteria
Participants were excluded if they (1) had absolute contraindications to MRI; (2) had a mental disorder or other neurological disease; (3) exhibited cognitive impairment; and (4) reported long-term use of medicines other than antiepileptic drugs. Additionally, (5) MTLE patients whose image quality did not satisfy the criteria were excluded.
2.1.2 Healthy controls
All healthy controls received information about registering and provided informed consent prior to receiving MRI. Only the following exclusion criteria were applied:
(1) presence of brain lesions, inflammation, hemorrhage, ischemia, or a related medical history; (2) long-term smoking, alcohol consumption, or substance abuse; (3) suspected or confirmed Alzheimer's disease, Parkinson's disease, or psychiatric disease; and (4) long-term medication use.
2.1.3 Participant demography information
A total of 376 HCs and 279 MRIn-MTLE patients with VEEG-confirmed left temporal lobe discharges and a mean illness duration of 452.5 months were recruited. To account for the effects of sex and age, participants aged 18–60 years were randomly selected from the HC and MRIn-MTLE groups, resulting in a total of 190 subjects in the HC group (100 males and 90 females), and a total of 190 subjects in the MRIn-MTLE group (100 males and 90 females), as shown in Table 1.
Sex
|
HC group (\(\stackrel{-}{x}\) ± s)
|
MRIn-MTLE group (\(\stackrel{-}{x}\) ± s)
|
t
|
P
|
Table 1
Age and sex comparison of two groups (years)
Male
|
38.06 ± 12.047 (N = 100)
|
39.02 ± 11.602 (N = 100)
|
-0.574
|
0.567
|
Female
|
37.79 ± 9.579 (N = 90)
|
35.64 ± 10.747 (N = 90)
|
1.413
|
0.159
|
|
37.93 ± 10.92(N = 190)
|
37.42 ± 11.3(N = 190)
|
0.448
|
0.655
|
2.2 Image capture and computerized analysis
2.2.1 MRI equipment and parameters
MRIn-MTLE patient images were acquired using a GE Signa Architect 3.0T (General Electric, United States); the protocol parameters are outlined below. To ensure image quality and reproducibility, the sequence parameters utilized in this study adhered to the sequence quality control process of the Enhancing Neuroimaging Genetics through Meta-Analysis Consortium 2020 initiative and utilized FreeSurfer-based segmentation of brain structures and hippocampal subregions[7], with the following equipment models and parameters.
GE Signa Architect 3.0T scanner settings for T1-MPRAGE images: a 48-channel head coil, voxel size = 1.0 × 1.0 × 1.0 mm (isotropic), field of view = 256 mm (256 × 256 matrix), 156-layer axial position, slice thickness = 1 mm, phase encoding direction: anterior to posterior, readout direction: superior to inferior (3D encoding); repetition time = 6.4 ms (3.0 T), flip time = 1,000 ms (3.0 T).
2.2.2 Automated segmentation of brain regions
The preprocessed images were automatically segmented on a computer running Ubuntu 24.0 using FreeSurfer, an open-source software package developed by the Computational Neuroimaging Laboratory of the Athinoula A. Martinos Center for Biomedical Imaging. The program is publicly accessible at https://surfer.nmr.mgh.harvard.edu and is used for extensive analysis and display of structural and functional imaging data. The version utilized in this study was 7.1.1. (Released in May 2020). Freesurfer recommends T1W-MPRAGE as preferred imaging protocol. The "recon-all" script allows FreeSurfer's segmentation procedure to be executed with a completely automated workflow, it contains more than 29 steps, the major steps are: 1)2 steps of intensity normalization, to correct for fluctuations in intensity that would otherwise make intensity-based segmentation much more difficult, 2)computing the affine transform from the original volume to the MNI305 atlas using Avi Snyders 4dfp suite of image registration tools, 3)skull stripping, 4)B1 bias field correction, 5) gray-white matter segmentation, segmentation is performed using image intensity and probability profiles and local spatial interactions between subcortical structures[8], 6)labeling of regions on the cortical and subcortical surface, in this part we used Desikan-Killiany Atlas for labeling. Further description of recon-all script was enumerated elsewhere. The processing of all T1W-MPRAGE brain volumes yielded a comprehensive morphometric description, and all volumes were utilized in training and testing of the classification model. FreeSurfer (version 7.1.1) image analysis software was used to perform cortical reconstruction and volume segmentation in the normalized space for each participant.
2.3 Volume-based diagnostic classification model
All HC and MRIn-MTLE patients’ age, and volume of Freesurfer volume segmentation results were used as classification variables for model training. Since the patients could not tolerate too long scans, the patients' scans were mainly T1-MPRAGE, and also the volume segmentation results were interpretable and helped the interpretation of the results of the classification model outputs, only the segmentation results of T1W-MPRAGE were included as classification variables.
For the classification model based on all participants, the training dataset consisted of 393 cases (167 positive/226 negative), an additional 262 cases were chosen as an independent validation dataset (112 positive/150 negative). For the classification model based on male participants, the training dataset consisted of 120 cases (60 positive/60 negative), an additional 80 cases were chosen as an independent validation dataset (40 positive/40 negative). For the classification model based on female participants, the training dataset consisted of 108 cases (54 positive/54 negative), an additional 72 cases were chosen as an independent validation dataset (36 positive/36 negative). Each case had 70 features including age and brain subregion volumes of both hemispheres.
The model based on different classifier was trained and tested with follow steps.
2.3.1 Data balancing
We repeatedly entered positive and negative samples out of datasets into randomly selected cases, to alleviate data imbalance.
2.3.2 Matric normalization
As the mean and standard deviation for each feature vector were determined before training, each feature vector was subtracted from the mean and then divided by the standard deviation. Each vector was centered at zero after normalization, and the standard deviation is expressed in units. Due to the high dimensionality of the feature space, only each pair of features was compared. If the classification accuracy (Pearson correlation coefficient value, PCC)of a pair of features was above the threshold, one of them was eliminated. Following this procedure, the dimensionality of the feature space was reduced, resulting in a set of distinct features.
2.3.3 Feature selection
Recursive feature elimination (RFE) was used to select features prior to model construction. The RFE approach selects features based on the classifier by iteratively evaluating a reduced set of features.
2.3.4 Cross-validation and Classifier evaluation
In the training dataset, five-fold cross-validations were carried out to identify the model's hyperparameters (e.g., the number of features). We divided the training set into 5 subsets randomly and equally and used each subset as the validation set and the other 4 subsets as the training set, respectively, and run the modeling process 5 times. The hyperparameters were determined based on the model's performance on the validation dataset. We employed support vector machine, logistic regression, random forest for model training and testing, the evaluation of the performance of models on the classification task based on each classifier was according to the receiver operating characteristic (ROC) curve. For quantification, the area under the ROC curve (the AUC) was computed. At the cutoff value that optimized, the Youden index, the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were computed. A total of 1,000 bootstrap samples were used to calculate 95% confidence intervals. All of the aforementioned procedures were implemented using Python (3.7.6) and the Feature Explorer Pro (FAEPro, version 0.4.3) program[9].
2.Statistical analysis
Since the differences in brain structure volumes between MRIn-MTLE patients and healthy controls have been little reported in previous studies, we conducted an exploratory comparative analysis of structural brain volumes in patients and healthy controls by gender. In order to understand the differences in structural brain volumes between genders in the healthy population, we also performed a comparative analysis of structural brain volumes in healthy controls by gender. In addition, our pretest suggested that the changes in brain structural volume were different between male and female MRIn-MTLE patients, so we also performed a comparative analysis of brain structural volume in patients of different genders. All comparison above used independent samples t-tests on SPSS 19.0.