This study was approved by our Institutional Review Board (IRB no.: 10902-009). Written consent was waived because the retrospective nature of this study meant that the healthcare of the included subjects was not affected. Personal information of all included patients was anonymized and de-identified before analyses were carried out.

The inclusion criteria were patients diagnosed with WHO grade I meningiomas by means of pathological confirmation. All the included patients must have undergone pre-operative brain MRI, post-operative follow-up brain MRIs for more than 1 year, and at least one MRI performed at 3 to 6 months after surgery. Patients diagnosed with neurofibromatosis (N = 3) were excluded. From January 2007 to January 2018, a total of 128 patients (43 men and 85 women with a median age of 57.5 years) diagnosed with WHO grade I meningiomas were included according to the abovementioned criteria. No known history of pre-operative intracranial radiation was documented in any of the included subjects. The mean follow-up time was 64.2 months (ranging from 14 to 149 months). A total of 19 (19/128, 14.8%) patients were found to have P/R, and the mean time to P/R was 33.3 months (ranging from 8 to 92 months). Based on anatomic location, the tumors were classified into four subgroups: convexity, parasagittal and parafalcine (PSPF), skull base, and intraventricular meningiomas. Skull base meningiomas included tumors arising from the anterior fossa/olfactory groove, spheno-orbital region, temporal floor, sellar/cavernous sinus, and posterior fossa (23). The extent of tumor resection was determined by a review of pre-operative brain MRI and the first time post-operative MRI (3–6 months after surgery) by a neuroradiologist (C.C.K.) and a neurosurgeon (S.W.L.). Simpson grade I–III resections (considered gross total resection, GTR) were performed in 93 patients, and Simpson grade IV–V resections (considered subtotal tumor resection, STR) were performed in 35 patients. Post-operative adjuvant radiotherapy (RT) was provided for patients who underwent STR in our institution. A total of 35 patients received post-operative adjuvant RT. Post-operative adjuvant RT was carried out via stereotactic radiosurgery (SRS) (N = 28, median dose of 25 Gy, ranging from 18 to 30 Gy; median fraction of 5, ranging from 3 to 5 fractions) or fractionated stereotactic intensity-modulated radiotherapy (IMRT) (N = 7, dose ranging from 55 to 60 Gy with 30 to 33 fractions) by linear accelerators. Detailed information of post-operative RT protocols is provided in Supplementary File 1 in Supplementary Material.

P/R was evaluated by two experienced neuroradiologists (C.C.K., 7 years of work experience, and T.Y.C., 19 years of work experience) by comparing the post-operative brain MRI findings between the 3–6 months and more than 1 year follow-up. Both readers were blinded to the clinical information of the studied patients. In equivocal cases, final agreement was arrived at by consensus. P/R was defined as recurrence of tumor in Simpson grade I–III resections (GTR) or increasing residual tumor size in Simpson grade IV–V resections (STR) on contrast-enhanced T1WI. In cases of STR, the threshold of P/R was defined as a 10% increase in tumor volume in comparison with post-operative brain MRIs (10). Interobserver reliability in determining P/R with intraclass correlation coefficient (ICC) of 0.8 was obtained. For patients who received post-operative adjuvant RT, P/R was differentiated from post-irradiation effects (pseudoprogression) based on progressive tumor growth, not transient increase in tumor volume (24).

Pre-operative brain MRI images were acquired using a 1.5-T (Siemens, MAGNETOM Avanto, n = 53, or GE Healthcare, Signa HDxt, n = 58) or a 3-T (GE Healthcare, Discovery MR750) (n = 17) MR scanner, equipped with eight-channel head coils in each machine. Scanning protocols were as follows: axial and sagittal spin echo T1-weighted imaging (T1WI), fast spin-echo T2-weighted imaging (T2WI), axial fluid attenuated inversion recovery (FLAIR), axial gradient recalled echo (GRE) T2^*-weighted imaging, axial diffusion-weighted imaging (DWI), and contrast-enhanced (CE) T1WI in axial and coronal sections. Detailed MR imaging parameters can be found in Supplementary File 2 in Supplementary Material.

T2WI and CE T1WI were known to be associated with histopathology and tumor grades in meningiomas (8, 25), and the two sequences (slice thickness/spacing, 5 mm/5 mm) were consistently acquired in all subjects. Thus, they were selected for radiomics analysis in this study. Figure 1 shows the flowchart of the analysis process. For each lesion, the operator placed an initial rectangle region of interest (ROI) on axial CE T1WI exhibiting the maximal tumor diameter, locating the approximate location and also deciding the initial and final slices containing the lesion. The fuzzy c-mean (FCM) clustering-based algorithm was developed to calculate the outline of the ROI on each imaging slice (26). In cases of under- or oversegmentation, manual correction by inclusion of more tumor tissue or exclusion of unnecessary normal tissue was performed. After segmentation and correction, the ROIs gleaned from all imaging slices containing the lesion were combined to obtain the 3D information of the whole lesion. The 3D connected component labeling was applied for removing scattered voxels not connecting to the main lesion. The hole-filling algorithm was then applied to include all voxels contained within the main ROI that had been labeled as nonlesion. The final 3D tumor mask was mapped to the axial T2WI to determine the tumor ROI on corresponding imaging slices using affine transformation and linear interpolation by FMRIB's Linear Image Registration Tool (FLIRT) (27).

Within the segmented tumor on axial CE T1W images and T2W images, 107 imaging features, consisting of 32 first-order features and 75 textural features, were extracted on each modality (Figure 1). Therefore, a total of 214 descriptor features were obtained for each case. In order to evaluate the importance of these features in differentiating P/R, the sequential feature selection process was implemented via constructing multiple SVM classifiers (28). Using this method, we selected imaging features with high importance. In this process, SVM with Gaussian kernel was used as the objective function (29, 30). Ten-fold cross-validation was applied to test the model performance (31). In each iteration, the training process was repeated 1,000 times to explore the robustness of each imaging feature. After each iteration, the feature which contributed to the best performance was added into the candidate set. When the addition of features no longer improves the performance, the selection process was terminated and a final set containing the optimal features was obtained. The termination criterion for the objective function was determined at 10⁻⁶. This procedure was implemented in MATLAB 2018b. The most significant four features selected by the SVM model for the prediction of P/R were T1 gray-level co-occurrence matrix (GLCM) cluster shade, T1 gray-level size zone matrix (GLSZM) gray-level non-uniformity, T2 GLCM cluster prominence, and T2 GLCM cluster shade. The SVM score for each patient was calculated using the following equation based on the selected features.

where x is the input features and N is the length of the support vector. w_n is the parameter and b is the bias. y_n and x_n are the entries of the supporting vector. G(x_n, x) is the Gaussian kernel function that indicates the dot product in the predictor space between x and the support vectors. Herein,

For comparison with the radiomics model in the prediction of P/R in meningiomas, apparent diffusion coefficient (ADC) values (b = 1,000 s/mm²) on DWI were measured manually by two experienced neuroradiologists (C.C.K. and T.Y.C.) as in previously published works (7, 29). The circular ROI (area ranging from 35 to 78 mm²) was placed in a homogeneous area of the tumor to avoid volume averaging with calcification, necrosis, and cystic regions that might influence ADC values (Figure 2) (7, 32, 33). Due to the almost perfect reproducibility in the interobserver reliability, the subsequent statistical evaluation of ADC values was performed using the mean value calculated from both interpreters.

Statistical analyses were performed using statistical package SPSS (V.24.0, IBM, Chicago, IL, USA). For the evaluation of the clinical parameters and conventional MRI findings, chi-square (or Fisher's exact test) and Mann–Whitney U tests were performed for categorical and continuous data, respectively. The area under the receiver operating characteristic curve (ROC) curve (AUC) was calculated for SVM scores and ADC values to obtain the optimal cutoff values. Kaplan–Meier analysis was used to evaluate progression-free survival (PFS), and the log-rank test was used to assess significance. Cox hazard regression model with univariate and multivariate analyses was performed to determine independent predictors of P/R. Variables with a p < 0.05 in univariate analysis were brought forward to the multivariate analysis. For multivariate analysis and all other statistical analyses, p < 0.05 was considered statistically significant.

The clinical data and conventional MRI findings of the included 128 meningiomas are summarized in Table 1. Nineteen (19/128, 14.8%) patients were diagnosed with P/R. Statistically significant differences (p < 0.05) were observed in the extent of resection, adjacent bone invasion, ADC value, and SVM score between P/R and non-P/R groups (Table 1) (Figures 2, 3). In multivariate Cox hazards analysis (Table 2), adjacent bone invasion, low ADC value, and high SVM score were high-risk factors for P/R (p < 0.05) with hazard ratios of 7.31, 4.67, and 8.13.

The most significant four imagining features selected by the SVM model for the prediction of P/R were T1 GLCM cluster shade, T1 GLSZM gray-level non-uniformity, T2 GLCM cluster prominence, and T2 GLCM cluster shade. The reproducibility of ROI-based radiomics feature was good, and the ICCs of the four imaging features were 0.92, 0.78, 0.82, and 0.94, respectively.

For the prediction of P/R, AUCs of 0.80 and 0.73 with optimal cutoff values of 0.224 and 0.825 × 10⁻³ mm²/s were obtained in SVM score and ADC value, respectively (Figure 4). Furthermore, improved performance in predictive model was observed after combining SVM score and ADC value, with AUC of 0.88 (Figure 4). When tumor progression trends were compared, patients with adjacent bone invasion, high SVM score (more than the cutoff value of 0.224), and low ADC value (lower than the cutoff value of 0.825 × 10⁻³ mm²/s) were found to exhibit shorter PSF (p < 0.05) (Figure 5).