This study included a cohort of 51 epilepsy patients, all of whom were recruited from the First Affiliated Hospital of Suzhou University (27 males and 24 females). The sample consisted of 34 patients diagnosed with temporal lobe epilepsy (TLE) and 17 with epilepsy originating from other brain regions (non-TLE). The inclusion and exclusion criteria for the epilepsy patients in this study were as follows: (1) The primary basis for the diagnosis of epilepsy was epileptiform discharges on electroencephalogram (EEG), coupled with clinical seizures and medical history. (2) Patients were excluded with absence of visible structural abnormalities or lesions on routine neuroimaging studies, that could account for the seizures. (3) Patients lacking preoperative T1-weighted images were excluded from the study. (4) Patients with MRI scans of poor quality were also excluded. The identification of the epileptogenic foci for the patients with epilepsy was determined through a combination of postoperative MRI scans and surgical records. The surgical records included details about the excised regions, such as the frontal lobe, occipital lobe, hippocampus, amygdala, middle and inferior temporal gyrus, superior temporal gyrus, and other typical areas. Demographic information and MRI lesion information of all patients is shown in Table 1.

All participants underwent the clinical standard brain MRI protocol, including T1-weighted MRI, T2-weighted MRI, FLAIR, and DWI. The brain MR images were acquired at the First Affiliated Hospital of Suzhou University using a Siemens MAGNETOM Skyra 3.0 T ultraquiet MRI system and Philips MR Achieva/Intera equipment to obtain T1-weighted MRI scans.

For the Siemens system, the scanning parameters were as follows:

Repetition Time (TR)/Echo Time (TE) = 2300.0/3.0 ms.

Slice Thickness = 1.0 mm.

Flip Angle (FA) = 8°.

For the Philips medical equipment, the scanning parameters were as follows:

Repetition Time (TR)/Echo Time (TE) = 8.1/3.7 ms.

Slice Thickness = 1.0 mm.

Flip Angle (FA) = 9°.

MRI brain images were processed using standard protocols as outlined in previous literature (21). Because the brain MR images of all patients were derived from two different devices, several preprocessing steps are recommended for consistency, including intensity normalization and spatial normalization. Intensity normalization of MR images from both devices could mitigate differences in signal strength and ensure uniform grayscale representation. Spatial normalization could transform images into a common coordinate space, facilitating anatomical comparisons across subjects. After spatial normalization, the intensity of images was normalized to 0 ~ 250. After intensity normalization, the raw MRI images have uniform dimensions (256 × 256 × 256) and a consistent spatial resolution of 1 × 1 × 1 mm³. The N3 algorithm was utilized to correct intensity inhomogeneities within the images. Subsequent to these corrections, skull stripping was performed to remove the scalp, skull, and dura mater. Following this, tissue segmentation was carried out to isolate the gray matter, white matter, and cerebrospinal fluid. The processed images were then mapped to a prelabeled automated anatomical labeling (AAL) template, enabling the segmentation of 90 regions of interest (ROIs). Last, reconstruction of the cortical surface was accomplished. Upon completion of these procedures, metrics such as the volume of gray matter, cortical thickness, and cortical surface area for each ROI were computable (Figure 1).

Following the aforementioned image preprocessing steps, morphometric features, including the volume of gray matter, cortical thickness, and cortical surface area, can be obtained for 90 brain regions in each patient’s MRI image. After consultation with clinical experts, 10 brain regions particularly relevant to temporal lobe epilepsy (TLE) were selected for analysis. These include the hippocampus, parahippocampal gyrus, amygdala, transverse temporal gyrus, polar temporal gyrus (superior temporal gyrus), polar temporal gyrus (middle temporal gyrus), superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, and fusiform gyrus (as shown in Figure 2). Moreover, by cross-referencing the descriptions of surgical removal locations in each subject’s medical records, it was found that the typical areas excised in TLE patients were included among these 10 regions. Concurrently, the literature indicates that the temporal pole, originating amygdala, and originating hippocampal head can be identified in PET images for the localization of the onset of TLE (22). Previous research based on MRI images has confirmed a high correlation between the medial temporal lobe and TLE (23). The selection of these TLE-relevant brain region features was carried out based on this existing research and following discussion with neurosurgeons.

In this study, three classifiers were chosen to serve as auxiliary diagnostic models for temporal TLE, including SVM, ELM, and cmcRVFL+. SVM is a binary classification model well suited for medium- and small-sized data samples and nonlinear and high-dimensional classification problems. In this study, we selected the radial basis function (RBF) for SVM as the kernel function. ELM is a single-hidden-layer neural network model with the number of hidden nodes set to 60. It is characterized by high efficiency, accuracy, and strong generalization performance and does not require iterative learning, thus offering fast training speed. The cmcRVFL+ was an ensemble classifier (cmcRVFL+) for small sample classification (19), which combines a series RVFL as weak classifier in order to build a more robust final classifier. It formed a cascaded model that use the predict label as privileged information, which was fed into the next RVFL learner together with morphometric features. The construction of RVFL learners prevents excessive learning from the training data resulting in a less biased model. The framework for constructing these classification models is presented in Figure 3.

All patients were randomly split into training and testing datasets. During the training phase, all training samples underwent image preprocessing and feature extraction before being fed into the classifiers for training. To address the issue of a small sample size, 5-fold cross validation was conducted for each split and the training and testing experiments were repeated five times. The average results from these experiments were used as the final outcome. The classification performance of the models was evaluated by accuracy (ACC), specificity (SPE), sensitivity (SEN), and area under the curve (AUC) values.

The study analyzed 32 epilepsy patients, including 13 males and 19 females, ranging in age from 7 to 59 years. Fifteen of the patients had epileptic foci located in the left hemisphere, and 17 had epileptic foci in the right hemisphere. Using t-tests, we analyzed the age, sex, and location of the epileptic foci in these 32 patients. The results indicate that there were no significant differences between the temporal lobe epilepsy (TLE) group and the group with epileptic foci in other locations in terms of sex, age, or the lateralization of the epileptic foci. The statistically significant differences were assessed also for gray matter volume (transverse temporal gyrus with p = 0.005960 and superior temporal gyrus with p = 0.040157), cortical thickness (middle temporal gyrus with p = 0.032215 and fusiform gyrus with p = 0.022513), and cortical surface area (amygdala with p = 0.032215 and fusiform gyrus with p = 0.019514). Detailed results of the morphometric features were listed in Tables 2–4.

To validate the efficacy of using the ratio features of 20 selected brain regions for distinguishing between TLE and epilepsies of other localizations, three classification models were employed, including SVM, ELM, and cmcRVFL+. The classification performance was evaluated using metrics such as accuracy, sensitivity, specificity, area under the curve (AUC), and the receiver operating characteristic (ROC) curve. The performance of the three classifiers is shown in Table 5, The accuracy of the ELM classifier with gray matter volume feature reached the highest accuracy of 92.79%. The classification accuracy was higher when using gray matter volume features than when using cortical thickness or surface area features.

The ROC curves were illustrated in Figure 4. We compare the AUC value of the gray matter volume feature using SVM (AUC = 0.8235), ELM (AUC = 0.8019), and cmcRVFL+ (AUC = 0.7667). For cortical thickness, the ROC curves of SVM (AUC = 0.7276), ELM (AUC = 0.7028), and cmcRVFL+ (AUC = 0.9064) were compared. For cortical surface area, the ROC curves of SVM (AUC = 0.7245), ELM (AUC = 0.7988), and cmcRVFL+ (AUC = 0.7405) were compared. The highest AUC was obtained using cortical thickness with cmcRVFL+ classifier.

The visualization results of the statistically significant analysis of gray matter volume, cortical thickness, cortical surface area in the temporal related ROIs between TLE and non-TLE are shown in Figure 5. The color of each brain region represents the p-value of the statistical T-test for morphological features between two groups, where red indicates significant differences and yellow indicates less significant differences. The statistically significant brain regions of gray matter volume locate at transverse temporal gyrus (p = 0.005960) and superior temporal gyrus (p = 0.040157). The statistically significant brain regions of cortical thickness locate at middle temporal gyrus (p = 0.032215) and fusiform gyrus (p = 0.015641). The statistically significant brain regions of cortical surface area locate at amygdala (p = 0.022513) and fusiform gyrus (p = 0.019514). Details of the statistical analysis results of gray matter volume, cortical thickness, cortical surface area are shown in Tables 2–4.

Some patients with temporal lobe epilepsy (TLE) exhibit no discernible lesions on MRI, termed MRI-negative TLE. The diagnosis of TLE relies on imaging studies, requiring the exclusion of hippocampal sclerosis and other structural anomalies. Previous research typically utilized 1.5 T MRI scans for case selection, which risks overlooking subtle lesions and consequently misclassifying them as MRI-negative patients. The widespread use of 3.0 T high-resolution cranial MRI in recent years has increased the detection rate of epileptogenic foci in TLE patients by approximately 20–48%. While some MRI-negative cases can be accurately diagnosed with a structural abnormality using 3.0 T MRI, 20–30% of TLE patients still do not show any obvious lesions on MRI and are considered to be strictly MRI-negative TLE. The debate continues as to whether TLE is a distinct spectrum of the disease; some scholars suggest the possible presence of minor hippocampal sclerosis, visible only on a histopathological level. Studies have found that certain brain regions in TLE patients show a reduced volume of white matter compared to the control group, while other studies indicate atrophy in the ipsilateral entorhinal cortex in MRI-negative TLE patients. Therefore, the clinical and morphometric characteristics of TLE remain ambiguous.

To address these challenges, this study proposes a TLE identification method based on morphological features of temporal lobe regions and machine learning. Three machine learning models, namely, SVM, ELM, and cmcRVFL+ were employed for classification experiments. SVM is a supervised learning model commonly used for pattern recognition, classification, and regression analysis. It is well suited for binary classification tasks involving medium-to-small size datasets, nonlinearity, and high dimensionality. Compared with SVM, the ELM consists of a simple three-layer architecture comprising input, hidden, and output layers, forming a single-hidden-layer feedforward neural network. It only requires setting the number of hidden layer nodes, obviating the need to adjust the input weights and thresholds of the hidden layer, eventually producing an optimal solution. ELM thus demonstrates better adaptability using gray matter volume (ACC = 92.79%, AUC = 0.8019). The cmcRVFL+ is an ensemble classifier for small sample classification, which combines a series RVFL as weak classifier in order to build a more robust final classifier. It formed a cascaded model that use the predict label as privileged information, which was fed into the next RVFL learner together with morphometric features. The construction of RVFL learners prevents excessive learning from the training data resulting in a less biased model. The classification performance of cmcRVFL+ were improved (ACC = 90.07%, AUC = 0.9064) using cortical thickness. All these classification models were used to evaluated the effectiveness of morphometric features for MRI-negative TLE detection, the classifier based on ELM outperformed in distinguishing TLE from epilepsies using temporal lobe gray matter volume features.

The t test method was utilized to analyze the statistical significance of volumetric features of gray matter and cortical surface area in the bilateral brain regions of three clinically TLE-relevant areas: the hippocampus, the transverse temporal gyrus, and the middle temporal gyrus. As indicated in Table 4, the gray matter volume features in the bilateral regions of the hippocampus showed a significant difference with a p value of 0.0086. Additionally, the gray matter volume features in the transverse temporal gyrus displayed a significant difference between the TLE group and the group with epilepsies originating in other brain regions, with a p-value of 0.045. Ultimately, we selected the ratio of three features in the bilateral brain regions of 10 areas within the surgical zone as features for machine learning classification, and this yielded the most effective results.

Recent research indicates that subtle structural changes in subregions of the hippocampus could lead to neurological and psychiatric disorders, including epilepsy, Alzheimer’s disease, major depressive disorder, posttraumatic stress disorder, and schizophrenia (24). Temporal lobe epilepsy with hippocampal sclerosis (TLE-HS) is a common subtype of temporal lobe epilepsy and a classical focal epilepsy syndrome. Epileptic seizures in these patients generally originate from the medial temporal lobe or simultaneously involve limbic structures and are accompanied by hippocampal sclerosis (25). For patients with temporal lobe epilepsy where the hippocampus is the frequent epileptogenic zone, 75% ultimately progress to drug-resistant epilepsy. The seizure types in TLE-HS patients are diverse, with causes ranging from perinatal hypoxia-ischemia-induced brain damage, brain malformations, cerebral vascular malformations, trauma, and infection-induced scar tissue, leading to hippocampal degenerative sclerosis and forming a new epileptogenic focus. Reports indicate that hippocampal sclerosis is also a significant cause of epilepsy, ultimately resulting in a “dual-source” epileptogenic focus (26).

Studies have found that MRI-negative TLE patients exhibit atrophy in the transverse temporal gyrus area. The possible mechanism involves the reduction of output neurons in the epileptogenic zone, resulting in diminished afferent input and consequently reduced volume in brain regions that are synaptically connected to the epileptogenic zone (27). Epileptic discharges are not confined to the epileptogenic focus but propagate widely to other brain areas. Local neurotoxins can cause damage to neurons not only in the epileptogenic zone but also in remote areas, leading to neuronal loss. Furthermore, there may be subtle cortical developmental anomalies and other pathological changes in brain tissue. The collective impact of these multiple factors is likely the pathophysiological mechanism underlying the reduction in gray matter volume seen in epilepsy patients (28).

Despite achieving promising research results, this study has several limitations: (1) The small sample size is a main concern for overfitting. We have taken actions to deal with this issue that an ensemble classifier was used. The cmcRVFL+ combines multiple RVFL classifiers to build a more robust classifier. It formed a cascaded model that use the predict label as privileged information, which was fed into the next RVFL learner together with features. The construction of RVFL learners prevents excessive learning from the training data resulting in a less biased model. The overfitting issue was further ensured by the use of 5-fold cross validation and repeated experiment. Future studies will include additional samples for validation, addressing the overfitting limitations. (2) We only used imaging data from a single modality. If data from multiple modalities were available, it would be possible to extract features from more dimensions for more accurate classification. Future experiments could be conducted using data from multiple modalities, and other types of information will be used to improve the model’s classification accuracy, such as neuropsychological assessments.