Meningioma typing model construction using radiomics-based multi-parameter magnetic resonance imaging

doi:10.21203/rs.3.rs-1724494/v1

Purpose: To investigate the value of a cli-radiomics model based on multi-parameter magnetic resonance imaging (MRI) in differentiating fibroblastic meningiomas from non-fibroblastic meningiomas.

Methods: Clinical, imaging, and postoperative pathological data of 423 patients (128 fibroblastic meningiomas and 295 non-fibroblastic meningiomas) were randomly categorized into training (n=296) and validation (n=127) groups at a 7:3 ratio. The Selectpercentile and LASSO were used to selected the highly correlated features from 3376 radiomics features. Different classifiers were used to train and verify the model. The receiver operating characteristic (ROC) curves, ACC, SEN, and SPE were drawn to evaluate the performance. The optimal radiomics model was selected, calibration curves and decision curve analysis were used to verify the clinical utility and consistency of the nomogram constructed from the radiomics features and clinical factors.

Results: There were thirteen radiomic features selected from T1C and T2WI after dimensionality reduction. The prediction performance of RF radiomics model is slightly lower than that of the cli-radiomics model. The area under the curve (AUC), SEN, SPE, and ACC of the cli-radiomics model training set are 0.836 (95% confidence interval [CI], 0.795-0.878), 0.922, 0.583, and 0.686; the AUC, SEN, SPE, and ACC of the validation set were 0.756 (95% CI, 0.660-0.846), 0.816, 0.596, and 0.661, respectively.

Conclusion: The diagnostic efficacy of the cli-radiomics model of fibroblastic meningioma and non-fibroblastic meningioma was better than that of the radiomics prediction model alone, and can be used as a potential tool for clinical surgical planning and evaluation of patient prognosis.

Meningioma

Radiomics

Typing

Magnetic resonance imaging

Meningiomas arise from arachnoid cells and account for 39.0% of all intracranial tumors, according to the latest data from the US Brain Tumor Registry [1]. Most meningiomas are considered benign tumors with a lower histopathological grade. WHO grade I meningiomas consist of nine different subtypes [2], each of which has a different tissue composition and texture, and treatment and prognosis are not the same [3]. Surgical resection is currently the main treatment for meningioma [4], and the accurate classification of meningioma is an important factor in determining surgical planning for neurosurgery. Several studies have reported that the consistency of meningioma is one of the key factors in determining the difficulty of surgery [5-6]. In particular, meningiomas are located at the base of the skull, adjacent to important neurovascular structures [7], and soft-textured meningiomas that are easily aspirated [8]. Solid tumors are usually harder to remove and take longer to operate, and require the use of an ultrasound aspirator to extract [9-10]. Therefore, preoperative prediction of meningioma subtypes is critical for the selection of surgical options and reduction of potential complications. Surgically resected histopathology and biopsy remain the gold standard for the diagnosis of meningioma, which is invasive, limited in some specimens, and does not allow a complete and accurate assessment of tumor heterogeneity. Therefore, this study aimed to differentiate fibroblastic meningioma from non-fibroblastic meningioma based on multi-parameter MRI radiomics before surgery and to guide the choice of clinical operation, which is helpful in evaluating the prognosis of patients.

Patients

From June 2016 to May 2021, WHO grade I meningioma patients who were pathologically confirmed by our Hospital were retrospectively collected.

Inclusion criteria: 1. Patients with confirmed histopathological meningioma and definite pathological subtype; 2. Patients with meningioma resection one week after MRI examination; 3. Picture archiving and communication systems (PACS) has available pretreatment MRI images, including at least T1C and T2WI, and complete clinical data; 4. The image quality of each patient was good and there were no artifacts.

Exclusion criteria: 1. Patients who had received radiotherapy, chemotherapy, targeted therapy, or other treatments before preoperative MRI scanning; 2. Patients with different parameters of T1C and T2WI sequences in MRI images; 3. Patients with incomplete MRI sequences; 4. Patients with metallic foreign bodies or claustrophobia.

A total of 423 patients with WHO grade I meningiomas were enrolled, including 128 fibroblastic meningiomas (12 male and 116 female) and 295 non-fibroblastic meningiomas (77 male and 218 female).

MRI acquisition

Both plain and enhanced MRI images of the head were obtained using a Siemens Verio 3.0T superconducting MRI scanner (Siemens, Germany). The patient was placed in the supine position. The scanning sequence and parameters were as follows: Gradient echo (GRE): T1WI (TR=550 ms, TE=11 ms), layer thickness 5 mm, layer spacing 1.5 mm, (FOV) 260 mm×260 mm, matrix 256×256; TSE: T2WI (TR=2200 ms, TE=96 ms), echo time 10 ms, echo chain length 8, excitation twice. Enhanced scan: Gd-DTPA was injected into the elbow vein at a dose of 0.1 mmol/kg with a flow rate of 3.0 ml/s.

Image segmentation

The T1C and T2WI images of all 423 patients with meningioma were imported from a post-processing workstation in DICOM. A total of 128 fibroblastic meningiomas and 295 non-fibroblastic meningiomas were manually segmented by two radiologists (Doctors 1 and 2, with 3 and 10 years of experience, respectively) using the open-source ITK-SNAP software (www.itksnap.org ) without knowing the pathology. First, the volume of interest (VOI) of the lesion was manually segmented layer-by-layer on the axial T1C image, including tumor necrosis, cystic changes, and hemorrhage. The lesions were delineated layer-by-layer on axial T2WI, with T1C as a control. All VOI were examined by senior doctors. The process of this section is as follows (Fig. 1).

Radiomics feature extraction and filtering

After all the images were manually segmented, Z-score normalization was used to standardize the image strength normal distribution. A total of 3,376 radiomics features were extracted from T1C and T2WI using Shukun.net, including shape features extracted from the original image, first-order features, and texture features transformed by the original image filtering. Shape features, such as area, volume, diameter, and spherical degree, describe the size of the region of interest (ROI) and its spherical degree of approximation. First-order features, called histogram features, are features related to voxel intensity distribution in the ROI, such as mean, median, minimum, maximum, standard deviation, skewness, and kurtosis. Second-order features, also called texture features, are used to describe the strength of voxel spatial distribution, which mainly includes the gray level co-occurrence matrix (GLCM), gray level run long matrix (GLRLM), gray levelsize zone matrix (GLSZM), neighborhood gray-tone difference matrix (NGTDM), and some first-order features and texture features extracted by filter transformation.

In this study, all the features P<0.05 were screened by Selectpercentile, the dimension of the selected features was reduced by LASSO and 5-fold cross-validation, and the coefficient of the non-strongly correlated features was 0 by L1 regularization. Thirteen radiomics features (seven T2WI and six T1C) were screened to differentiate between non-fibroblastic and fibroblastic meningiomas. SelectKbest was used for univariate analysis of clinical features, and sex was included as the only clinical feature.

Construction and validation of cli-radiomics model

The thirteen radiomics features extracted from T2WI and T1C were fused, and radiomics models were constructed using different classifiers (Support vector machine (SVM), Random forest (RF), Decision tree (DT), Logistic regression (LR), LinearSVC, and Adaboost). Then, the diagnostic efficiencies of the different models were compared using the receiver operating characteristic (ROC) curve. The cli-radiomics model was built using the best radiomics model fused with clinical labels, and a nomogram of cli-radiomics models was constructed for predicting fibroblastic and non-fibroblastic meningiomas. The discriminant ability of the cli-radiomics models was evaluated using calibration curves of the training and validation sets. Decision curve analysis (DCA) was used to quantify the net benefit under different threshold probabilities and to assess the clinical validity of the nomogram.

Statistical methods

In this study, all data were analyzed using R software (version 3.4.1; http://www.Rproject.org), SPSS 25 (SPSS, Inc, Chicago, IL, USA) and Medcalc19.1(MedCalc, Mariakerke, Belgium). The chi-square test was used to compare sex, and an independent sample t-test was used to test the continuous variables, such as age, which accorded with the normal distribution. The Delong test compares the AUC values of different radiomic prediction models. The sensitivity (SEN), specificity (SPE), negative predictive value (NPV), positive predictive value (PPV), and accuracy (ACC) were calculated to distinguish fibroblastic meningiomas from non-fibroblastic meningiomas according to the confusion matrix. Calibration and decision curve analyses were drawn using R software, and the Hosmer-Lemeshow test was used to evaluate the statistical differences between the predicted and actual probabilities. P<0.05 showed significant difference.

Clinical data

A total of 423 patients with low-grade meningiomas were enrolled, including 128 fibroblastic meningiomas (12 males and 116 females), with an average age of 52.19 ± 9.03 (range 21-81) years, and 295 non-fibroblastic meningiomas (77 males and 218 females), with an average age of 51.88 ± 10.84 (range 17-81) years (Fig. 2). The training and validation sets were randomly divided according to a ratio of 7:3. The 296 training sets included 90 cases of fibroblastic meningioma and 206 cases of non-fibroblastic meningioma, 127 cases of validation set, 38 cases of fibroblastic meningioma, and 89 cases of non-fibroblastic meningioma. There were significant differences in sex and location (P<0.05) but no significant differences in age. Sex, location, age, and subtype were not significantly different between training and validation (P>0.05). The general data of all the patients are shown in Table 1. The general data for the training and validation sets are compared in Table 2.

Screening and analysis of radiomics features

A total of 3376 radiomics features were extracted from the T1C (n=1688) and T2WI (n=1688) images, including 14 shape features, 180 first-order features, 750 texture features, and 744 wavelet transform features. First, feature filtering is performed using Selectpercentile, and then the features of P>0.05 are further analyzed using LASSO. In the process of LASSO feature selection, the alpha value of the least error was selected as the optimal value of the model through 5-fold cross-validation. When -lg (alpha) = 2.0177, that is, alpha = 0.0096, six T1C features and seven T2WI features with non-zero coefficients were selected to construct the prediction model. As shown in Fig. 3 and Table 3, the thirteen features included two first-order features, two texture features, and nine wavelet transform features.

The diagnostic efficacy of different radiomics models in differentiating fibroblastic and non-fibroblastic meningiomas in the training and validation sets is shown in Table 4 and Fig. 4. The results show that the RF model has the best performance in identifying the two. The AUC, SEN, SPE, ACC, F1, PPV, NPV of the training set are 0.819 (95% CI, 0.776-0.866), 0.822, 0.665, 0.713, 0.635, 0.518, 0.895, respectively. The AUC, SEN, SPE, ACC, F1, PPV, NPV, NPV of the validation set were 0.752 (95% CI, 654-0.841), 0.711, 0.719, 0.717, 0.600, 0.519, 0.853, respectively (Fig. 5). Therefore, combining the RF radiomics model with clinical

feature (sex), a cli-radiomics model was constructed to predict the preoperative subtype of meningiomas. In the training, the AUC, SEN, SPE, ACC, F1, PPV, and NPV of fibroblastic and non-fibroblastic meningiomas were 0.836 (95% CI, 0.795-0.878), 0.922, 0.583, 0.686, 0.641, 0.491, and 0.945, respectively. The AUC, SEN, SPE, ACC, F1, PPV, NPV of the validation set were 0.756 (95% CI, 0.660-0.846), 0.816, 0.596, 0.661, 0.591, 0.463, and 0.883, respectively (Fig. 6).

Performance evaluation of cli-radiomics models

The cli-radiomics model nomogram can be used directly and visually in clinical practice (Fig.7a). The diagnostic performance in differentiating fibroblastic meningioma from non-fibroblastic meningiomas was better than that of the radiomics model alone. The calibration curve and Hosmer-Lemeshow test further verified that there was no significant difference between the actual differentiation of fibroblastic and non-fibroblastic meningiomas, which was predicted by the radiomics model, with values of 0.237 and 0.136, respectively (Fig.7b, 7c). The decision curve assessed the clinical usefulness of the cli-radiomics nomogram in differentiating between fibroblastic and non-fibroblastic meningiomas before surgery (Fig. 7d, 7e).

Low-grade meningiomas account for more than 81% of all meningiomas [11]. A thorough understanding of the characteristics and anatomy of the tumor before surgery is a prerequisite for complete resection. The consistency of meningioma is one of the most critical factors affecting the difficulty of surgery [12], as they can be extremely soft tumors that are easily removed by aspiration or hard tumors that are difficult to resect completely [13]. Meningioma subtypes are usually diagnosed by histopathology and immunohistochemistry. Therefore, preoperative prediction of meningioma subtypes can help determine the optimal surgical strategy and predict and avoid potential complications [14]. Fibroblastic meningiomas reportedly contain a large amount of fibrous tissue and are difficult to resect. More detailed dissection is usually required, especially for tumors located at the base of the skull [15]. Therefore, accurate preoperative prediction of meningioma histological subtypes is essential for surgical planning, surgical resection, and prognostic prediction. Previous studies [16] have used conventional imaging features to predict meningioma subtypes, including irregular tumor morphology, peritumoral edema, and tumor enhancement on T1C. Jolapara. et al [17] used DTI to identify atypical, fibroblastic and other benign meningiomas. In summary, these qualitative imaging features are highly subjective, and no well-established preoperative non-invasive predictive method for meningioma typing has been established. In contrast, radiomics provides a non-invasive, comprehensive quantification of tumor phenotypes by extracting numerous microscopic features that the naked eye cannot recognize, reflecting tumor heterogeneity and pathophysiological information [18].

To knowledge, no studies have been conducted on the construction of cli-radiomics models to predict meningioma subtypes. Comprehensive information on tumor phenotypes can be obtained by radiomics features, and after feature selection and dimension reduction, a total of thirteen radiomics features, including two first-order features, which are closely related to the subtype of meningiomas, describe the distribution of voxel intensity in images using common and basic metrics, in which the coefficient of T2_auto_logarithm_firstorder_10Percentile is larger. There are two texture features, T1C_auto_logarithm_glszm_Small Area Low Gray Level Emphasis and T1C_auto_original_glszm_Small Area Low Gray Level Emphasis, which are highly correlated with meningioma classification. GLSZM represents regions with the same interconnect adjacent pixels or voxels and can be used to quantify gray level regions in an image [12]. There are significant differences between fibroblastic meningiomas and non-fibroblastic meningiomas with different texture parameters. The high-order features included five low-order wavelet features and four high-order wavelet features. By analyzing the extracted features, we explained the strength, shape, and texture of the tumor and provided quantitative parameters for the analysis of the tumor subtypes.

Meningioma consistency is the hardness of the tumor and considered a function of water and collagen content. Quantitative evaluation of the signal intensity on T2WI can predict the consistency of meningiomas more reliably. Hypointense meningiomas on T2WI are mainly fibroblastic subtypes that tend to be firm. The high signal intensity on T2WI is related to the softer texture of meningiomas [19-21]. However, conventional MRI are subjective and unstable [22]. In this study, the thirteen radiomics features were trained and validated using six classifiers (SVM, RF, DT, LR, LinearSVC, and Adaboost). The results showed that the RF radiomics model based on T2WI and T1C could distinguish fibroblastic meningiomas from non-fibroblastic meningiomas in both training and validation sets, and sex was selected as the only clinical factor, but it was not consistent with the P value of 0.010 in the chi-square test, possibly because the small proportion and corresponding coefficient of location in differentiating fibroblastic and non-fibroblastic meningiomas were not sufficient as independent predictors to identify the two. The diagnostic efficacy and sensitivity of the cli-radiomics model with clinical features (sex) were improved compared with those of the RF radiomics model. The sensitivities of the training and validation sets are 0.922 and 0.816, respectively. The calibration curve and Hosmer-Lemeshow test further confirmed the good agreement between the actual differentiation of fibroblastic and non-fibroblastic meningiomas and the prediction of meningioma subtypes using the cli-radiomics model (p=0.237 and p=0.136, respectively). Analysis of the decision curve provides an important basis for preoperative diagnosis and surgical decision-making by radiologists and clinicians. Park et al. [23] distinguished 17 cases of fibroblastic meningioma from 137 cases of non-fibroblastic meningioma using radiomics and machine learning. Niu et al. [24] used Fisher discriminant analysis to differentiate 80 cases of meningothelial meningioma, 80 cases of fibrous meningioma, and 81 cases of transitional meningioma, the accuracy of which is obviously better than that of this study. This may be related to the fact that no specific typing of non-fibroblastic meningiomas has been made in this study, which further demonstrates that radiomics has an important value in predicting meningioma subtypes before surgery.

In this study, the RF radiomics prediction model was the best for differentiating fibroblastic meningioma from non-fibroblastic meningioma based on T2WI and T1C. The nomogram of the cli-radiomics model, which integrates clinical factors with sex, can improve the ability to distinguish between the two, and can help doctors to make operation plans and improve the prognosis of patients.

Our study had some limitations. First, our single-center retrospective study did not identify specific types of non-fibroblastic meningiomas, and only the differences between fibroblastic and non-fibroblastic meningiomas and meningiomas of each subtype were not analyzed and differentiated. Second, we should attempt semi-automatic and automatic segmentation technologies in future research. Finally, our study did not involve functional MRI, such as DWI, DTI, PWI, or other new techniques.

• Funding

This study was supported by grants of National Natural Science Foundation of China (No. 82071872), Lanzhou University Second Hospital Second Hospital “Cuiying Technology Innovation Plan” Applied Basic Research Project (No. CY2018-QN09), and Science and Technology Program of Gansu Province (No. 21YF5FA123).

• Competing interests:

The authors declare that they have no conflict of interest.

•Availability of supporting data:

Not applicable.

• Human and Animal Ethics：

This study was approved by the Medical Ethics Committee of the Second Hospital of Lanzhou University (approval number: 2020A-109) and informed consent was waived.

• Ethics Approval and Consent to participate

Formal consent is not required for a retrospective study.

• Consent for publication

The obligation to obtain informed consent for the publication of this study has been waived for this retrospective observational study according to China law.

• Acknowledgements

Not applicable.

• Authors' contributions

First author : Tao Han（Theoretical design, data processing, article writing, major revisions, approved of submission on behalf of all authors）

The second author : Zhendong Xu（Theoretical guidance）

The third author :Yayuan Geng（Theoretical guidance）

The fourth author : Changyou Long（Statistical analysis）

The fifth author: Bin Zhang（Statistical analysis）

The sixth author:Liangna Deng（Data processing）

The seventh author :Xiaoqiang Lin（Data processing）

The eighth author : Mengyuan Jing（Data processing）

Corresponding author : Junlin Zhou, MD, PhD.（Theoretical guidance, article revision

suggestions, supportive contribution）

Ostrom QT, Cioffi G, Waite K, Kruchko C, Barnholtz-Sloan JS (2021). CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2014-2018. Neuro Oncol 23(12 Suppl 2): p. iii1-iii105. https:// doi.org/10.1093/neuonc/ noab200.
Ke C, Chen HL, Lv XF, Li HJ, Zhang Y, Chen MD, Hu DK, Ruan GY, Zhang Y, Zhang YM, Liu LZ, Feng YQ (2020). Differentiation Between Benign and Nonbenign Meningiomas by Using Texture Analysis From Multiparametric MRI. J Magn Reson Imaging 51(6): p. 1810-1820. https:// doi.org/ 10.1002/jmri.26976.
Przybylowski CJ, Zhao X, Baranoski JF, Borba Moreira L, Gandhi S, Chapple KM, Almefty K K, Sanai N, Ducruet AF, Albuquerque FC, Little AS, Nakaji P (2020). Preoperative embolization versus no embolization for WHO grade I intracranial meningioma: a retrospective matched cohort study. J Neurosurg134(3): p. 693-700. https:// doi.org/10.3171/2020.1.JNS19788.
Kim L (2020). A narrative review of targeted therapies in meningioma. Chin Clin Oncol 9(6): p. 76. https:// doi.org/ 10.21037/cco-2020-mbt-01.
Al Abdulsalam HK, Aldahish AK, Albakr A, Hussain S, Alroqi A, Alromaih S, Alsaleh S, Ajlan AM (2021). Endoscopic Transnasal Resection of Midline Skull Base Meningiomas: Tumor Consistency and Surgical Outcomes. J Neurol Surg B Skull Base 82(5): p. 500-505. https:// doi.org/ 10.1055/s-0040-1714111.
Zhai Y, Song D, Yang F, Wang Y, Jia X, Wei S, Mao W, Xue Y, Wei X (2021). Preoperative Prediction of Meningioma Consistency via Machine Learning-Based Radiomics. Front Oncol11: p. 657288. https:// doi.org/ 10.3389/fonc.2021.657288.
Cepeda S, Arrese I, García-García S, Velasco-Casares M, Escudero-Caro T, Zamora T, Sarabia R (2021). Meningioma Consistency Can Be Defined by Combining the Radiomic Features of Magnetic Resonance Imaging and Ultrasound Elastography. A Pilot Study Using Machine Learning Classifiers. World Neurosurg 146: p. e1147-e1159. https:// doi.org/ 10.1016/j.wneu.2020.11.113.
Jääskeläinen J (1986). Seemingly complete removal of histologically benign intracranial meningioma: late recurrence rate and factors predicting recurrence in 657 patients. A multivariate analysis. Surg Neurol 26(5): p. 461-9. https:// doi.org/ 10.1016/0090-3019(86)90259-4.
Winter F, Furtner J, Pleyel A, Woehrer A, Callegari K, Hosmann A, Herta J, Roessler K, Dorfer C (2021). How to predict the consistency and vascularity of meningiomas by MRI: an institutional experience. Neurol Res 43(8): p. 693-699. https:// doi.org/ 10.1080/01616412.2021.1922171.
Miyoshi K, Wada T, Uwano I, Sasaki M, Saura H, Fujiwara S, Takahashi F, Tsushima E, Ogasawara K (2020). Predicting the consistency of intracranial meningiomas using apparent diffusion coefficient maps derived from preoperative diffusion-weighted imaging. J Neurosurg p. 1-8. https:// doi.org/ 10.3171/2020.6.JNS20740.
Zhang T, Yu JM, Wang YQ, Yin DD, Fang LJ (2018). WHO grade I meningioma subtypes: MRI features and pathological analysis. Life Sci 213: p. 50-56. https:// doi.org/ 10.1016/j.lfs.2018.08.061.
Kunimatsu A, Yasaka K, Akai H, Sugawara H, Kunimatsu N, Abe O (2022). Texture Analysis in Brain Tumor MR Imaging. Magn Reson Med Sci21(1):95-109. https:// doi.org/ 10.2463/mrms.rev.2020-0159.
Yogi A, Koga T, Azama K, Higa D, Ogawa K, Watanabe T, Ishiuchi S, Murayama S (2014). Usefulness of the apparent diffusion coefficient (ADC) for predicting the consistency of intracranial meningiomas. Clin Imaging38(6): p. 802-7. https:// doi.org/ 10.1016/j.clinimag.2014.06.016.
Brabec J, Szczepankiewicz F, Lennartsson F, Englund E, Pebdani H, Bengzon J, Knutsson L, Westin CF, Sundgren PC, Nilsson M (2022). Histogram analysis of tensor-valued diffusion MRI in meningiomas: Relation to consistency, histological grade and type. Neuroimage Clin 33: p. 102912. https:// doi.org/ 10.1016/j.nicl.2021.102912.
Al Abdulsalam HK, Aldahish AK, Albakr A, Hussain S, Alroqi A, Alromaih S, Alsaleh S, Ajlan AM (2021). Endoscopic Transnasal Resection of Midline Skull Base Meningiomas: Tumor Consistency and Surgical Outcomes. J Neurol Surg B Skull Base 82(5): p. 500-505. https:// doi.org/ 10.1055/s-0040-1714111.
Tropine A, Dellani PD, Glaser M, Bohl J, Plöner T, Vucurevic G, Perneczky A, Stoeter P (2007). Differentiation of fibroblastic meningiomas from other benign subtypes using diffusion tensor imaging. J Magn Reson Imaging 25(4): p. 703-8. https:// doi.org/ 10.1002/jmri.20887.
Jolapara M, Kesavadas C, Radhakrishnan VV, Thomas B, Gupta AK, Bodhey N, Patro S, Saini J, George U, Sarma PS (2010). Role of diffusion tensor imaging in differentiating subtypes of meningiomas. J Neuroradiol 37(5): p. 277-83. https:// doi.org/ 10.1016/j.neurad.2010.03.001.
Kumar V, Gu Y, Basu S, Berglund A, Eschrich SA, Schabath MB, Forster K, Aerts HJ, Dekker A, Fenstermacher D, Goldgof DB, Hall LO, Lambin P, Balagurunathan Y, Gatenby RA, Gillies RJ (2012). Radiomics: the process and the challenges. Magn Reson Imaging30(9): p. 1234-48. https:// doi.org/ 10.1016/j.mri.2012.06.010.
Yao A, Pain M, Balchandani P, Shrivastava RK (2018). Can MRI predict meningioma consistency?: a correlation with tumor pathology and systematic review. Neurosurg Rev 41(3): p. 745-753. https:// doi.org/ 10.1007/s10143-016-0801-0.
Yamada H, Tanikawa M, Sakata T, Aihara N, Mase M (2022). Usefulness of T2 Relaxation Time for Quantitative Prediction of Meningioma Consistency. World Neurosurg 157: p. e484-e491. https:// doi.org/ 10.1016/j.wneu.2021.10.135.
Fountain DM, Young AMH, Santarius T (2020). Malignant meningiomas. Handb Clin Neurol : p. 245-250. https:// doi.org/ 10.1016/B978-0-12-822198-3.00044-6.
Haciyakupoglu E, Yilmaz DM, Walter J, Erdogan S, Haciyakupoglu S, Kuhn SA (2019). Correlation of Preoperative Radiological Evaluation of Skull Base and Non-skull Base Meningiomas with Clinical and Surgical Data. Turk Neurosurg 29(5): p. 671-676. https:// doi.org/ 10.5137/1019-5149.JTN.23078-18.2.
Park YW, Oh J, You SC, Han K, Ahn SS, Choi YS, Chang JH, Kim SH, Lee SK (2019). Radiomics and machine learning may accurately predict the grade and histological subtype in meningiomas using conventional and diffusion tensor imaging. Eur Radiol 29(8): p. 4068-4076. https:// doi.org/ 10.1007/s00330-018-5830-3.
Niu L, Zhou X, Duan C, Zhao J, Sui Q, Liu X, Zhang X. Differentiation Researches on the Meningioma Subtypes by Radiomics from Contrast-Enhanced Magnetic Resonance Imaging: A Preliminary Study. World Neurosurg (2019) 126: p. e646-e652. https:// doi.org/ 10.1016/j.wneu.2019.02.109.

Table 1 Clinical Characteristics

	Fibroblastic meningioma	Non-fibroblastic meningioma	x²	t	p
Age（y）	51.88 ±10.84	52.19 ±9.03	—	-0.283	0.777
Sex			15.033	—	< 0.001
Male	77（26.1%）	12（9.4%）
Female	218（73.9%）	116（90.6%）
Location			11.322	—	0.010
Convexity of brain	65（22.0%）	17（13.3%）
Cerebral falx	115（39.0%）	62（48.4%）
Skull base	25（8.5%）	3（2.4%）
Other	90（30.5%）	46（35.9%）

A Student’s t-test was used to compare the difference in age, while the chi-square test was used to compare the difference in sex and location.

Table 2 Comparison of clinical characteristics between training set and validation set

Parameter	Training	Validation	x²	t	p
subtype			0.010	—	0.921
Fibroblastic meningioma	90（30.4%）	38（29.9%）
Non-fibroblastic meningioma	206（69.6%）	89（70.1%）
Sex			0.728	—	0.393
Male	59（19.9%）	30（23.6%）
Female	237（80.1%）	97（76.4%）
Location			2.931	—	0.402
Convexity of brain	52（17.6%）	30（23.6%）
Cerebral falx	127（42.9%）	50（39.4%）
Skull base	22（7.4%）	6（4.7%）
Other	95（32.1%）	41（32.3%）
Age	51.33±10.21	53.46±10.46	—	-1.956	0.051

A Student’s t-test was used to compare the difference in age, while the chi-square test was used to compare the difference in sex, location and subtype.

Table 3 Radiomics features extracted from T1C and T2WI

T1C	T2WI
logarithm_glszm_SmallAreaLowGrayLevelEmphasis	lbp-2D_firstorder_10Percentile
original_glszm_SmallAreaLowGrayLevelEmphasis	logarithm_firstorder_10Percentile
wavelet-HHH_glszm_SmallAreaEmphasis	wavelet-HHL_firstorder_RobustMean Absolute Deviation
wavelet-HLL_firstorder_90Percentile	wavelet-LHH_glcm_Correlation
wavelet-HLL_glszm_GrayLevelVariance	wavelet-LLH_glcm_ClusterShade
wavelet-LLL_gldm_LargeDependenceLowGray LevelEmphasis	wavelet-LLH_glszm_SmallAreaEmphasis
	wavelet-LLL_glszm_GrayLevelVariance

T1C: Contrast enhanced T1 weighted imaging; T2WI: T2 weighted imaging

Table 4 The diagnostic efficacy of different radiomcs models in differentiating fibroblastic meningioma from non-fibroblastic meningioma in training set and validation set

	Model	AUC（95%CI）	SEN	SPE	ACC	F1	PPV	NPV
Training	LR	0.755（0.701,0.812）	0.600	0.811	0.747	0.590	0.581	0.823
	SVM	0.739（0.678,0.796）	0.689	0.733	0.720	0.599	0.530	0.844
	RF	0.819（0.776,0.866）	0.822	0.665	0.713	0.635	0.518	0.895
	LinearSVC	0.746（0.686,0.801）	0.667	0.714	0.699	0.574	0.504	0.831
	Adaboost	0.990（0.982,0.996）	0.978	0.942	0.953	0.926	0.880	0.990
	DT	0.607（0.567,0.649）	0.922	0.291	0.483	0.520	0.362	0.896
Validation	LR	0.698（0.594,0.788）	0.421	0.798	0.685	0.444	0.471	0.763
	SVM	0.636（0.523,0.733）	0.421	0.708	0.622	0.400	0.381	0.741
	RF	0.752（0.654,0.841）	0.711	0.719	0.717	0.600	0.519	0.853
	LinearSVC	0.670（0.665,0.764）	0.447	0.730	0.646	0.430	0.415	0.756
	Adaboost	0.591（0.470,0.702）	0.526	0.708	0.654	0.476	0.435	0.778
	DT	0.609（0.553,0.669）	0.947	0.270	0.472	0.518	0.356	0.923

LR:Logistic regression; SVM:Support vector machine; RF:Random forest; DT:Decision tree; AUC:Area under curve; SEN:Sensitivity; SPE:Specificity; ACC:Accuracy; PPV:Positive predictive value;NPV:Negative predictive value

No competing interests reported.

Meningioma typing model construction using radiomics-based multi-parameter magnetic resonance imaging

Status:

Version 1

Abstract

Figures

Introduction

Materials And Methods

Results

Discussion

Declarations

References

Tables

Additional Declarations

Status:

Version 1