A total of 105 meningioma patients underwent surgical resections between 2019 and 2021 at Huashan Neurosurgical Center were enrolled. Histological diagnoses were reviewed according to 2016 WHO meningioma grading criteria by two experienced neuropathologists (Dr. H.C and Dr. HX.C). Clinical information including age, gender, location, treatment status, the extent of resection, surgical outcome, and neurological functions was extracted from the medical records. Patients with recurrent meningioma who underwent another opration to remove the recurrent tumor were considered as recurrent meningioma cases. Patients with multiple meningiomas were also recorded. The clinical data of 105 patients was shown in.

Table 1 30 meningioma patients from First Affiliated Hospital of Nanjing Medical University were enrolled as External verification. The clinical data of 30 patients are shown in Supplementary Material 1 . The specific research process of this study is shown in Figure 1 . This study was approved by the Human Subjects Institutional Review Board of Huashan Hospital, Fudan University.

Bidirectional sequencing was performed to detect microlesions in the NF2 gene. DNA was extracted from tumor tissue with TIANamp Genomic DNA Kit (Tiangen Biotech, Beijing, China) as instructed by the manufacturers. The whole coding sequence and the exon-intron boundaries of the gene were amplified by a standard polymerase chain reaction (PCR). Subsequently, the product was used for bidirectional sequencing, as described previously (17). The sequence data were analyzed by Sequencer 4.9 (Genecode, MI, USA) and compared with the NF2 sequence (NM_016418) from GenBank. Mutations were described according to the standard nomenclature for DNA sequence changes according to the Human Genome Variation Society (HGVS).

Each tumor sample was subjected to PCR analysis. A fraction of the PCR product (0.5 liters) was mixed with 0.1 liters of Genescan 500 size standard (PE Applied Biosystems Foster City, CA, USA) and 0.9 liters of formamide loading buffer. Combinations were eletrophoresed on a 5 percent polyacrylamide gels on an ABI 377 DNA sequencer (PE Applied Biosystems Foster city CA, USA) for 2 hours after being denaturated at 96°C for 5 minutes. Individual gel lanes were visualized using the Genotype 2.0 software. The samples were scored using strict criteria. The two highest peaks within the predicted size range were designated as alleles. A loss of heterozygosity was defined as a ratio of T1:T2/N1:N2 of less than 0.67 or more than 1.50. The majority of normal DNA amplifications yielded two PCR results, showing heterozygozity. The ratio of allelic loss to informative instances was used to calculate the LOH frequency of a locus. The average LOH frequency of the long arm of chromosome 22 was the sum of the LOH frequencies of each location.

Immunohistochemical staining was performed using monoclonal antibodies against Ki-67 [Signalway (SAB), Shanghai, China; 1:200 dilution]. The cells stained in immunohistochemistry accounting for more than 10% of all cells were considered positive(+), otherwise negative (–). Progesterone receptor (PR) level was also examined to classify the tumors into two categories: PR negative (–) or PR positive(+). H3K27me3 was examined with anti-H3K27me3 (Millipore, 07-449) on ECL Plus films (Carestream).

All patients underwent MRI scanning before operation (with or without Gadolinium enhancement). MRI scans were performed by the Trio 3.0-T scanner (Siemens, Erlangen, Germany). The imaging process included axial T1WI (TE, 15 ms; TR, 450 ms; slice thickness, 5 mm), T2WI (TE, 110 ms; TR 5800 ms; slice thickness, 5 mm), and CE scans using 0.1 mM/kg gadopentetate dimeglumine (TE, 15 ms; TR, 450 ms; slice thickness, 5 mm). Tumor location was described according to Al-Mefty’s published manuscript, such as parasagittal/falx, skull base, cerebral convexity, etc.

Preprocessing was performed using the 3D-Slicer software (version 4.11). The MRI DICOM files of all patients were imported into 3D-slicer. T1WI, T2Flair, and DWI images were registered to the T1C sequence images; N4 bias field correction was applied to each sequence image to correct non-uniformities in intensity. Two neuroradiologists painted regions of interest (ROIs) on T1c images using the 3D-slicer software. Multiple meningiomas from the same patient were considered as a single case in ROI classification and impact feature extraction. Enhancement of the dural tail sign was included in ROI, while peritumoral edema was excluded. The neuroradiologists were not informed of the clinical and biomarker data. Pyradiomics, an open-source python package (https://github.com/Radiomics/pyradiomics), was used to extract radiomic features.

Pyradiomics, an open-source python package (https://github.com/Radiomics/pyradiomics), was used to extract radiomic features from the ROIs of each patient’s images. After that, we eliminated the time, the checked hospital, machine model and other useless information. 130 radiomics features were retained in each sequence. To avoid degradation in model performance due to overfitting and increase in feature dimension, we evaluated radiomics features in distinguishing between mut/wild type and screened all radiomic features of each sequence to generate a new feature set.

First, three pairs of samples in two groups were tested using t-test analysis. Levene test was used to test the homogeneity of variance. For the data with the homogeneity of variance greater than 0.05, a t-test was used to detect whether the characteristics of the average of the two groups of independent samples showed significant differences (P < 0.05). Only characteristics with significant differences were retained.

Second, a further feature screen was performed based on LASSO regression, which added L1 regular expression based on the least square regression. The features screened by t-test were standardized; the optimal parameter lambda was selected after ten-fold cross-validation. Thus, the corresponding coefficients of the model were trained. Features with nonzero coefficients in the LASSO regression model were selected.

Machine learning models were developed to predict the outcome of NF2 status based on different algorithm. We used five supervised machine learning algorithms to establish the prediction model. All the cases were randomly divided into training and validation cohorts in the ratio of 7:3. The algorithm was trained based on the training group, and its effectiveness was verified in the validation group. The algorithm with the highest AUC (area under the curve) in the validation cohorts was chosen as the best model. Five prediction models were generated by random forest (RF), k-nearest neighbor (KNN), support vector machine (SVM), logistic regression (LR), and extreme gradient boosting (xgboost) methods. The training and validation cohorts were divided based on the selected feature subset. Predictions were made after iterative optimization. The sensitivity, specificity, accuracy, and F1 score were evaluated. The model performance was analyzed by plotting ROC (receiver operating characteristic), P-R (analysis and precision-recall), and calibration curves. The MRI scans of 30 meningioma patients from First Affiliated Hospital of Nanjing Medical University were used as external validation to verify the accuracy of the best model.

Statistical analysis was performed using SPSS 26.0 (IBM SPSS statistics 26.0 for mac; IBM corp). For continuous variables, the Student t-test was used; for comparison of mean values of continuous variables, ANOVA was used. Categorical variables were compared using the χ2 test and the Fisher test. Continuous data were expressed as mean ± SD. P-value <0.05 was considered statistically significant.

A total of 105 patients (32 males and 73 females) with intracranial meningiomas were recruited in the training and testing cohorts, including 60 patients with NF2 mutation/loss and 45 wild type patients. The mean age of the patients was 53.17 ± 9.60 years (range 31 to 73 years). There are 90 (85.71%) grade 1, 14 grade 2 (13.33%) and 1 grade 3 patients (0.95%), respectively. The most common pathological subtype was fibrous (45, 42.86%). The median Ki-67 labeling index was 3.91 ± 2.40(range 1-12). PR was positive in 86 patients (81.9%). Loss of H3K27me3 expression was observed in 14 patients (13.3%). No diffence was observed of the compared the clinical and immunohistochemical characteristics between NF2 mutant and wild-type groups. Parasinoidal (26/60) was the most common location in NF2 mut/loss group, while NF2 wild tumours were more likely in skull base locations (25/45). Three cases of multiple meningiomas were identified in our cohort. However, only one tumor in each patient was removed, which were all belonged to the NF2 mutation group. 16 patients (4 with wildtype NF2 and 12 with NF2 mutations) were regarded as recurrent meningiomas and no difference was observed between de novo and recurrent patients regarding NF2 status (p = 0.170).

Among all 105 patients, 52 patients (49.52%) had NF2 mutations; 8 patients (7.62%) showed loss of NF2 gene due to partial deletion of chromosome 22q. The remaining 45 patients had wild-type NF2. Allelic deletion of NF2 and mutations were classified as NF2 mutation group, amounting to 60 cases. Of all patients with NF2 mutations, 23 were nonsense mutations; 16 were frameshift mutation; 8 were splice site mutation; 4 were missense mutation. Exon 1 and Exon 6 was detected the highest mutation frequency, accounting for 6 (11.54%) and 7 (13.46%) of all mutants, respectively. However, we did not find any obvious hot spot mutations. The most common copy number deletion occurred in 22q11.21- q13.33. Details of the NF2 mutation status were shown in Supplementary Table 1 .

A total of 130 radiomic features were extracted from each sequence. 390 radiomic features were included in the screening process. 147 radiomic features showed statistically significant differences between the NF2 mut/loss and wild-type groups. Only 9 features had nonzero coefficients in the LASSO regression model. The screening process of Lamda is shown in Figure 2 . In Figure 2A , The red line represents the standard deviation of the mean square error (MSE) of λ. The blue bar indicates the range of the mean square error. The lambda with the lowest standard deviation is the most suitable for classification and the model is the simplest. Therefore, We chose the position with the lowest lambda standard deviation (red line) as the most appropriate λ Value. The details and p-values for these 9 features are shown in Table 2 . 4 features were from the CE-T1Flair, 3 from T1WI, and 2 from T2WI sequences. 7 features could describe the texture of tumors and 2 described the wavelet of tumors. The 9 radiomics features were selected for the model building. We tried to cluster nine radiomics features through unsupervised hierarchical cluster analysis. PCA (Principal Component Analysis) is used to reduce the dimension and describe the distribution of data ( Figure 3 ).

Finally, we decided to incorporate the 9 features into the model. Gray Level Dependence Matrix (GLDM) quantifies gray level dependencies in an image. included descriptors of the three-dimensional size and shape of the ROI. First-order statistics describe the distribution of voxel intensities within the image region defined by the mask through commonly used and basic metrics.

Gray Level Co-occurrence Matrix (GLCM) describes the second-order joint probability function of an image region constrained by the mask.

We randomly divided all patients into training and validation groups in a ratio of 7:3. 73 patients were included in the training group. 9 radiomic features of 73 patients were used to train 5 supervised machine learning algorithms (RF, SVM, LR, KNN, and xgboost). The data of a total of 32 patients in the validation cohort was used to evaluate the effectiveness of the algorithms. The AUCs of each algorithm in training and validation cohorts were calculated and compared as shown in Table 3 and Figure 4 . SVM (support vector machine) had the highest AUC of 0.89 in the training cohort and 0.85 in the validation cohort. F1 score is the harmonic average of accuracy and recall, while the value was 0.80 in the SVM model. LR (logistic regression) had an AUC of 0.84 in the training cohort and 0.82 in the validation cohort. Most of the algorithms had AUCs above 0.7.

ROC and P-R curve analysis of SVM in training and validation cohorts were shown in Figure 5 . AUC of SVM were 0.89 and 0.85 in the training cohort and validation cohort respectively. Figure 6 shows the predicted value and their actual mutation of each validation group sample. Samples with a predicted value greater than 0 were predicted to be NF2 mut-type by the SVM model. The actual mutation of the sample were showen by color, Green represents mut-type and blue represents wild-type. 30 meningioma patients from First Affiliated Hospital of Nanjing Medical University were enrolled as external validation to verify the accuracy of the SVM model. the AUC of SVM model is 0.82 in the external validation cohorts. The calibration curve analysis and Hosmer-Lemeshow test for SVM model demonstrated the observations and predictions in validation cohorts were in good accordance ( Figure 7 ).