In this retrospective study, we consecutively enrolled a total of 205 patients with LGG who underwent surgery at the Beijing Tiantan Hospital from September 2012 to December 2014. The inclusion criteria of all enrolled cases were (a) pathologically confirmed grade II gliomas according to WHO criteria 2016 (20) and (b) presurgical T2-weighted imaging. The exclusion criteria of all enrolled cases were no craniotomy or stereotactic biopsy before MRI scan. The enrolled cases were allocated to either the training or the validation cohort according to the surgery time with a 1:1 ratio, i.e., the first 102 cases enrolled formed the training cohort, and the other 103 cases formed the validation cohort. Data on routine clinical variables were collected, including age, gender, tumor pathology, and epilepsy type. The present study further utilized clinical and MRI data from all enrolled cases. The ethics committee of Beijing Tiantan Hospital approved this study, and the requirement for informed consent was waived. Our study was conducted in accordance with the Declaration of Helsinki. The study design is illustrated in Figure 1.

Patients were considered to have experienced tumor-related epilepsy when a history of at least one seizure with the presence of an enduring alteration (i.e., LGG) in the brain (21) was reported. The history and type (generalized and focal) of epilepsy were evaluated by an epileptologist based on the patient's presentation according to the classification and terminology of the International League Against Epilepsy (9, 22). The epilepsy type was determined consistently for all the enrolled patients with a history of epilepsy based on the aforementioned criteria.

All MRI examinations were performed using a Magnetom Trio 3.0 T scanner (Siemens, Erlangen, Germany) with a 12-channel receive-only head coil scan acquisition. The T2WI parameters were as follows: repetition time (TR), 5,800 ms; echo time (TE), 110 ms; flip angle, 150°; the field of view (FOV), 240 × 188 mm²; voxel size, 0.6 × 0.6 × 5.0 mm³; and matrix, 384 × 300. The MRI data were stored in DICOM format.

Regions of interest (ROIs) of the gliomas were drawn by two neuroradiologists with more than 5 years of clinical experience with ITK-snap (www.itksnap.org). The neuroradiologists were blinded to each other's results. Gliomas were segmented on each MRI slice. We defined ROIs of the LGGs as areas of the MRI images that exhibited abnormal hyperintense signals. The intraclass correlation coefficient (ICC) was used to assess whether the segmentation results of the two doctors were significantly different. No difference was defined as ICC >0.8. In the absence of a difference, each patient would obtain a segmentation result from one of the two neuroradiologists randomly.

A total of 734 radiomic features were extracted based on the research of Liu et al. (19) and Li et al. (23, 24), including 6 location features, 17 first order statistics (FOS) features, 8 shape and size features, 26 gray-level co-occurrence matrix (GLCM) features, 16 gray-level run-length matrix (GLRLM) features, 16 gray-level size zone matrix (GLSZM) features, 5 neighborhood gray-tone difference matrix (NGTDM) features, and 640 wavelet features.

The location features were extracted based on our previous research (19) using (a) polar coordinates parameters (r, θ, and Φ) based on the centroid of the tumor, and (b) City Block distance, (c) Chebyshev distance, and (d) Euclidean distance from the anterior commissure (AC) to the centroid of the tumor. The FOS features reflected the distribution of voxel values within the ROI 3D matrix and the overall information of the tumor. The shape and size features reflected the volume, surface area, and shape of the tumor. GLCM, GLRLM, GLSZM, and NGTDM were collectively referred to as texture features. The GLCM reflected the arrangement of voxels. The GLRLM reflected the arrangement of voxels with equal voxel values. The GLSZM reflected the characteristics of the homogeneous region. The NGTDM reflected the difference between each voxel in the ROI and the adjacent voxels. The wavelet features were calculated by FOS, GLCM, GLRLM, GLSZM, and NGTDM features through Coiflet 1 3D wavelet transform. The detailed information and formulas for the detection of the 734 features were published in our previous researches (19, 24).

The 734 radiomic features were normalized before feature selection using the z-score method. A univariate analysis was used to screen the radiomic features. The criteria for screening features include (a) p-values of Pearson correlation coefficient <0.05 and (b) area under curve (AUC) of the radiomic features >0.6. Least absolute shrinkage and selection operator (LASSO) regression was widely used to compress the coefficients of features and select features to prevent overfitting. Logistic regression was used for data classification to build a reliable prediction model. Thereafter, LASSO and logistic regression were used to calculate the radiomic signature for epilepsy-type prediction using Glmnet package (25). Features dimension reduction and selection, i.e., univariate analyses and LASSO regression, were based on the training cohort. The optimal value of the LASSO's parameter λ was determined by leave-one-out cross-validation (LOOCV) using classification error as criterion during the training phase. We calculated the radiomic signature for the patients after determining the selected features' values using the optimal value of the LASSO's parameter λ. Radiomic signature was the linear weighting of the selected features' coefficients. Radiomic analyses of the study were implemented by MATLAB R2016a (MathWorks, Natick, MA).

Based on cohort of all patients, a multivariable logistic regression analysis was built to predict epilepsy type with clinical information, using the radiomic signature, age, gender, and tumor pathology. Akaike's information criterion was used to select the indicator with the predictive ability for building the multivariable logistic regression model (26, 27). With this radiomics-based model, we also built a novel radiomic nomogram for quantitative prediction of the epilepsy type (28).

The classification performance of the radiomic signature and radiomic nomogram was assessed by the receiver operating characteristic (ROC) curves and AUCs in each cohort. Calibration curves were plotted to assess the calibration of the radiomic signature and radiomic nomogram (29), accompanied by the Hosmer–Lemeshow test (30). Decision curve analyses (DCAs) determined the clinical usefulness of the radiomic signature and radiomic nomogram by quantifying the net benefits at different threshold probabilities in cohort of all patients (31).

Age and radiomic signature were reported as median and range. The differences between subgroups were assessed by independent samples t-test. Gender and histopathology were reported in frequencies and proportions, and differences between subgroups were assessed by Fisher's exact test. The statistical tests were two sided, and p < 0.05 were defined as significant. Nomogram building and models' validation were implemented with R software (version 3.6.1, Vienna, Austria).

The main clinical and pathological characteristics of all 205 patients are listed in Table 1. Of the 205 enrolled patients, 139 (67.8%) had generalized and 66 (32.2%) had focal seizures. Those with generalized epilepsy accounted for 72 (70.6%) and 67 (65.0%) patients, while those with focal epilepsy accounted for 30 (29.4%) and 36 (35.0%) patients in the training and validation cohorts, respectively. There were no significant differences between the two epilepsy types based on age, gender, and tumor histopathology in cohort of all patients, training cohort, and validation cohort. However, radiomic signature was significantly different between the two epilepsy types (p < 0.001) and hence a potential indicator for diagnosing the types of epilepsy.

Based on the training cohort, a logistic regression prediction model was constructed by integrating the four key radiomic features selected using the univariate analyses and LASSO regression (Table 2). The parameter λ = 0.067 was used as the optimal value. The radiomic signature for each patient in both cohorts was calculated with the train-based model. The predictive ability of the radiomic signature was interpreted from the ROC curve (Figure 2A), where it achieved a performance with classification accuracy = 80.4%, AUC = 0.859 [95% confidence interval (CI), 0.787–0.932] in the training cohort and classification accuracy = 80.6%, AUC = 0.839 (95% CI, 0.761–0.917) in the validation cohort. The radiomic signature demonstrated favorable calibration in the training and validation cohorts (Figure 2B). The p-values of the Hosmer–Lemeshow test for classification predictive ability of the radiomic signature were 0.12 and 0.10, respectively. The DCA showed that using radiomic signature to predict epilepsy type adds more benefit than either the treat-all-patients scheme or the treat-none scheme (Figure 2C).

Radiomic nomogram for epilepsy type prediction was developed based on the radiomic signature, age, and tumor pathology data (Figure 3). It showed excellent performance in predicting epilepsy type with AUC = 0.863 (95% CI, 0.810–0.916) in cohort of all patients (Figure 2D). The calibration curve and DCA of the radiomic nomogram for the epilepsy type prediction also demonstrated favorable results (Figures 2E,F). The p-value of the Hosmer–Lemeshow test was 0.11.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.