Discrimination between spinal meningioma and schwannoma using MRI machine learning-based radiomics texture analysis

Authors

Abstract

Aim To identify radiomics parameters that distinguish spinal meningiomas from schwannomas and evaluate machine learning models using MRI data.
Materials and Methods Patients with histopathologic diagnoses of spinal meningioma (n = 25) and schwannoma (n = 26) who underwent pre-surgical MRI were enrolled. Semantic features, including tumor location, longest diameter, foraminal extension, cystic changes, spinal cord compression, enhancement patterns, dural tail, and ginkgo leaf signs, were assessed using a single 3.0 T scanner. Radiomics features were extracted from T1-weighted, T2-weighted, STIR, and post-gadolinium T1-weighted images. Support Vector Machine (SVM) models were trained and validated using conventional MRI, radiomics features, and a combination of both.
Results A significant gender difference was found, with more females in the meningioma group (84%, p = 0.039). Other significant factors included the longest tumor dimension (p = 0.03), presence of a dural tail sign (p < 0.001), intratumoral cystic changes (p = 0.003), ginkgo leaf sign (p = 0.01), and spinal cord compression (p < 0.001). Of the 444 extracted radiomics parameters, 186 demonstrated good reproducibility (ICC ≥ 0.75). Among these, 49 were retained for model construction. Model 1 (conventional MRI) achieved an AUC of 0.902, accuracy of 0.882, sensitivity of 0.884, and specificity of 0.880. Model 2 (radiomics features) achieved an AUC of 0.909, accuracy of 0.843, sensitivity of 0.961, and specificity of 0.720. Model 3 (combined features) demonstrated the highest performance with an AUC of 0.997, accuracy of 0.960, sensitivity of 1.000, and specificity of 0.920. Pairwise DeLong’s tests confirmed that Model 3 significantly outperformed both Model 1 (p = 0.025) and Model 2 (p < 0.001).
Discussion Combining radiomics with conventional MRI improves diagnostic accuracy in differentiating spinal meningiomas from schwannomas, supporting radiomics as a valuable non-invasive tool for preoperative diagnosis. Multicenter studies are needed to validate these findings and expand clinical applications.

Keywords

Introduction

Spinal meningiomas and schwannomas are the most common intradural-extramedullary tumors in adults, comprising the majority of benign spinal neoplasms [1]. MRI with contrast is the standard imaging tool, offering excellent soft tissue contrast and anatomical detail [2]. However, conventional MRI features often overlap, limiting specificity in distinguishing these tumors [2–5]. While histopathology remains the gold standard, its invasive nature highlights the need for reliable non-invasive alternatives [3].
Accurate preoperative differentiation is clinically important, as meningiomas typically arise from the dura and often require dural excision, whereas schwannomas originate from nerve sheaths and may be treated with nerve-sparing approaches [2, 6]. Improved imaging-based diagnosis could therefore optimize surgical planning and outcomes.
Radiomics offers a promising solution by converting pixel intensity data into quantitative descriptors of tumor texture, shape, and intensity, many of which are not appreciable visually [7]. This approach has shown value in tumor characterization [7, 8], but its application to spinal tumors remains limited, and optimal parameters for distinguishing meningiomas from schwannomas are not well established [9, 10].
The aim of this study was to evaluate whether combining radiomics with conventional MRI improves differentiation between spinal meningiomas and schwannomas, and to assess the performance of machine learning models in supporting preoperative diagnosis.

Materials and Methods

Study Design
This retrospective, single-center study was conducted following the principles outlined in the Declaration of Helsinki (2013) [11]. Due to its retrospective nature, the IRB waived the requirement for informed consent. The study adhered to the CheckList for Evaluation of Radiomics (CLEAR) reporting guidelines to ensure methodological rigor, reproducibility, and transparency of the findings [12].
Study Population
The study population comprised patients who underwent spinal surgery at Ankara City Hospital for spinal canal masses between February 2019 and June 2022. Eligible cases were identified from the hospital’s electronic medical record system, which documented cases histopathologically confirmed as spinal meningioma or schwannoma. Initial screening identified 101 cases; however, patients were excluded if they had undergone prior spinal surgery (n = 2) or if they were imaged at external facilities (n = 34). Of the remaining 65 patients, those with incomplete imaging sequences (n = 3), image artifacts that obscured interpretation (n = 4), or MRI scans performed with a 1.5 Tesla MRI device (n = 7) were also excluded. This left a final cohort of 51 patients (15 males, 36 females) who met the study’s inclusion criteria. The study flowchart is summarized in Figure 1.
MRI Acquisition
All scans were performed on a 3T system (SIGNA™ Pioneer, GE Healthcare, USA) using routine spinal protocols. Gadolinium- based contrast (0.1 mmol/kg) was administered intravenously. Sequences included T1-weighted, T2-weighted, STIR, and post-contrast T1-weighted imaging. Acquisition parameters (TR, TE, FOV, matrix, slice thickness) are summarized in Table 1.
Conventional MRI Findings
Evaluated features included tumor location (cervical, thoracic, lumbar), craniocaudal diameter, cystic changes, intraspinal location (anterior, posterior, lateral), spinal cord compression, neural foramen extension, dural tail sign, and cord signal changes. Additional findings were:
• Ginkgo leaf sign: Characteristic cord indentation caused by stretching of the dentate ligament, described by Yamaguchi et al. as a specific marker for spinal meningiomas [13].
• Tumor enhancement pattern: Homogeneous, rim, or heterogeneous after gadolinium.
Two radiologists (Y.C.G., 5 years; S.D., 24 years) reviewed all features independently; disagreements were resolved by consensus or by the senior radiologist.
Texture Analysis
Texture analysis was conducted using Olea Sphere 3.1 SP- 28 software (Olea Medicals, La Ciotat, France, https://www. olea-medical.com). For each lesion, regions of interest (ROIs) were manually delineated on each imaging slice containing the tumor. The voxel dimensions of the ROIs were interpolated to a resolution of 1×1×1 mm³ to standardize spatial resolution across all sequences and minimize partial volume effects. The radiomics workflow is summarized in Figure 2.
Tumor Segmentation and ROI Delineation
Two radiologists (Y.C.G. and S.D.) manually segmented tumors using a freehand ROI approach on sagittal images. Each slice was carefully outlined to exclude adjacent structures (vessels, fat, muscle, bone), ensuring accurate delineation (Figure 3). To maintain consistency across sequences, VOIs generated from contrast-enhanced sagittal T1-weighted images were propagated to corresponding T1-, T2-, and STIR sequences, thereby standardizing tumor boundaries and minimizing variability in texture analysis.
Feature Extraction
Radiomics features were computed across shape, first-order, and texture categories. Shape features (n = 17) quantified tumor geometry (e.g., volume, sphericity, major axis length). First-order statistics (n = 19) described intensity distribution (mean, variance, skewness, kurtosis). Texture features included GLCM (24), GLRLM (16), GLSZM (16), NGTDM (5), and GLDM (14). In total, 111 features were extracted per sequence.
Feature Reduction and Selection
A multi-step pipeline ensured reproducibility and minimized redundancy. Two radiologists independently extracted features, retaining those with high interobserver agreement (ICC ≥ 0.75), yielding 186 features. Pearson correlation (r > 0.85) reduced this to 120 variables. Mann-Whitney U testing with Benjamini- Hochberg correction identified 49 discriminative features (p < 0.05) for model construction.
Machine Learning and Model Configuration
Three linear-kernel SVM models were developed for classification.
• Model 1: Included only conventional MRI features with significant group differences (p < 0.05).
• Model 2: Included radiomics features retained after feature reduction, requiring high interobserver agreement (ICC ≥ 0.75) and statistical significance (p < 0.05).
• Model 3: Combined conventional MRI features (Model 1) with selected radiomics features (Model 2) to maximize diagnostic accuracy.
All models were trained with min–max normalized features. Hyperparameters were optimized via cross-validation, with C = 1.0 and tolerance = 0.1.
Model Validation
The leave-one-out cross-validation (LOOCV) method was applied to validate the models. In LOOCV, each case was systematically left out as a test case while the model was trained on the remaining cases. This approach is particularly suitable for small datasets, ensuring robust and unbiased model performance evaluation.
Performance Metrics and Statistical Analysis
Model performance was assessed using AUC, accuracy, sensitivity, specificity, PPV, and NPV. AUCs were compared with DeLong’s test. Continuous variables were analyzed with the Mann-Whitney U test or t-test, depending on distribution, and categorical variables with Chi-Square or Fisher’s exact test. Interobserver agreement was measured using ICC, with ≥0.75 indicating good reproducibility. Optimal cut-off values for radiomics features were derived from ROC analysis using the Youden Index.
Software Utilized
All statistical analyses were conducted using R software (https://www.R-project.org) and IBM SPSS version 23 (IBM Corp., Released 2015), with a significance level set at 0.05 to ensure statistical robustness and reproducibility. Additionally, feature selection and machine learning model development were performed in R and Python’s Scikit-Learn library.
Ethical Approval
This study was approved by the Ethics Committee of Ankara Bilkent City Hospital (Date: 2022-09-07, No: E2-22-2323).

Results

Study Population and Conventional MRI Findings
The study comprised 51 patients, including 25 (49%) with spinal meningioma and 26 (51%) with schwannoma. The mean age was higher in the meningioma group (53.16 ± 19.38 years) than in the schwannoma group (45.08 ± 19.94 years), though not significantly (p = 0.134). Meningiomas were most frequently located in the thoracic (52%) and cervical (28%) regions, whereas schwannomas were most often found in the cervical (34.6%) and lumbar (34.6%) regions, with a broader distribution. Within the canal, meningiomas showed a balanced distribution across anterior, lateral, and posterior compartments, while schwannomas favored posterolateral (38.4%) and lateral (23.1%) sites. Female predominance was marked in meningiomas (84%, p = 0.039). Significant group differences were observed in maximum tumor size (p = 0.03), dural tail sign (p < 0.001), cystic changes (p = 0.003), ginkgo leaf sign (p = 0.01), and spinal cord compression (p < 0.001). Contrast enhancement patterns and spinal cord signal changes did not differ significantly. Full demographic and conventional MRI data are summarized in Table 2.
Interobserver Reproducibility
Of the 444 radiomics features assessed, 186 achieved ICC ≥ 0.75, indicating good interobserver agreement: 45 from STIR, 45 from post-contrast T1WI, 62 from T2WI, and 34 from non- contrast T1WI. Among these, 49 features showed both statistical significance and high reproducibility and were retained for model construction. Reproducibility was strongest in T2-derived features.
Diagnostic Performance of Individual Parameters
The 49 retained features demonstrated AUCs ranging from 0.640 to 0.822. GLSZM Gray Level Non Uniformity from T1WI achieved the highest AUC (0.822; sensitivity 77%, specificity 66%), while Total Energy from STIR also performed strongly (AUC 0.785–0.814) (Table 3). Cut-off values were determined using the Youden Index. GLSZM Gray Level Non-uniformity from T1-weighted images demonstrated the highest individual diagnostic performance (AUC = 0.822), capturing intratumoral signal heterogeneity related to microtextural complexity. Total Energy (TE) from STIR sequences also showed strong discriminative capability (AUC = 0.814), reflecting the overall voxel intensity distribution that corresponds to tumor cellularity and cystic composition. In addition, Variance (VAR) from T2-weighted images (AUC = 0.778) effectively represented intensity dispersion within the lesion, which tends to be greater in schwannomas due to their heterogeneous histologic architecture and cystic degeneration.
Together, these parameters capture complementary structural and intensity-based characteristics of spinal meningiomas and schwannomas, supporting the high diagnostic accuracy achieved by the combined radiomics model.
Performance of SVM Models
Model 1 (conventional MRI) achieved an AUC of 0.902, accuracy of 0.882, sensitivity of 0.884, and specificity of 0.880. Model 2 (radiomics only) showed an AUC of 0.909, accuracy of 0.843, high sensitivity (0.961), but lower specificity (0.720). Model 3 (combined features) outperformed both, with an AUC of 0.997, accuracy of 0.960, sensitivity of 1.000, and specificity of 0.920. Pairwise DeLong’s tests revealed no significant difference between Models 1 and 2 (p = 0.083), but Model 3 significantly outperformed both Model 1 (p = 0.025) and Model 2 (p < 0.001), confirming its superior diagnostic power.

Discussion

This study is among the first to apply radiomics for differentiating spinal meningiomas from schwannomas, demonstrating that combining radiomics with conventional MRI yields superior diagnostic performance (AUC 0.997, sensitivity 100%, specificity 92%). Radiomics extracts quantitative descriptors of tumor heterogeneity, shape, and intensity patterns that extend beyond visual inspection. Among individual features, GLSZM Gray Level Non Uniformity from T1- weighted images achieved the highest diagnostic performance (AUC 0.822), reflecting signal heterogeneity associated with histopathological differences. Shape-based metrics, particularly Major Axis Length, also consistently differentiated tumor types across all sequences (AUC 0.717), underscoring its robustness as a structural parameter. Sequence-specific analyses further highlighted the discriminative power of radiomics. In STIR and T1WI, Total Energy and Energy were top-performing features (AUC 0.814–0.785), reflecting global voxel intensity distribution. On post-contrast T1WI, Variance and Mean Absolute Deviation (AUC 0.777) were most effective, consistent with the broader intensity spread seen in schwannomas. While conventional MRI features alone provided good accuracy (AUC 0.902), the addition of radiomics markedly enhanced diagnostic precision, in keeping with contemporary oncologic imaging strategies [14, 15]. Importantly, the radiomics model by itself also demonstrated strong independent value (AUC 0.909). Previous radiomics studies in intracranial meningiomas and schwannomas have shown that texture- and shape-based features can effectively differentiate tumor types [16–18]. However, spinal canal tumors differ anatomically and radiologically, and radiomics research in this region has largely focused on bone lesions [19–21]. Xu et al. reported a combined MRI-clinical nomogram with high diagnostic accuracy (AUCs = 0.994, 0.962, 0.949) [10]. Our combined model achieved comparable performance (AUC = 0.997; accuracy 96%; sensitivity 100%; specificity 92%), confirming its robustness. Whereas Vychopen et al. linked sphericity to postoperative recovery in meningiomas [9], our inclusion of higher-order texture metrics such as GLN and GLDM further demonstrates the added value of radiomics for non-invasive spinal tumor characterization. Deep learning approaches, such as convolutional neural networks (CNNs), have also shown promise in tumor differentiation [22, 23]. Maki et al. reported CNN-based classification of cranial meningiomas and schwannomas with AUCs of 0.870 and 0.876 using sagittal T1- and T2-weighted sequences [22]. Yet, deep learning typically demands large datasets for reliable training, which may not be feasible in rare clinical populations. Radiomics combined with machine learning algorithms such as SVM can provide comparable performance with smaller datasets, as demonstrated in our study. In summary, radiomics can be integrated into routine MRI workflows for spinal canal tumors as a decision-support tool. Beyond diagnostic accuracy, it can aid surgical planning—for example, guiding dural excision in meningiomas versus nerve-sparing strategies in schwannomas. Moreover, features such as GLN and Total Energy may serve as imaging biomarkers of tumor heterogeneity and potentially correlate with genetic subtypes, supporting personalized treatment and advancing non-invasive tumor subtyping.

Limitations

This study has several limitations. Its retrospective, single- center design with a relatively small sample size may introduce selection bias and limit generalizability. The exclusive use of a 3T MRI scanner constrains applicability to other field strengths and manufacturers; validation across multicenter cohorts with 1.5T and 3T systems is needed. Manual ROI segmentation, though reproducible, remains subjective and time-intensive; automated or semi-automated methods could improve efficiency and consistency. Advances in AI, particularly convolutional neural networks, may further enhance segmentation accuracy. Future research should focus on expanding radiomics-based spinal tumor analysis through semi-automated or fully automated segmentation workflows, which would enhance reproducibility and reduce operator dependency. Additionally, deep learning– based feature extraction and classification methods hold promise for capturing more complex spatial and textural relationships within spinal lesions. Integrating these approaches in multicenter, prospective studies could further improve generalizability, reduce manual workload, and facilitate clinical translation of radiomics models into preoperative decision- support systems.

Conclusion

This study shows that combining radiomics with conventional MRI improves accuracy in differentiating spinal meningiomas from schwannomas, offering a valuable non-invasive approach for preoperative diagnosis. The enhanced performance of the combined model supports radiomics as an adjunct to conventional imaging. Further multicenter studies are recommended to confirm these findings and advance clinical applications.

References

1. De Verdelhan O, Haegelen C, Carsin-Nicol B, et al. MR imaging features of spinal schwannomas and meningiomas. J Neuroradiol. 2005;32(1):42-9.
   Article PubMed Google Scholar
2. Lee JH, Kim HS, Yoon YC, Cha MJ, Lee SH, Kim ES. Differentiating between spinal schwannomas and meningiomas using MRI: A focus on cystic change. PLoS One. 2020;15(5):e0233623.
   Article PubMed Google Scholar
3. Liu WC, Choi G, Lee SH, et al. Radiological findings of spinal schwannomas and meningiomas: Focus on discrimination of two disease entities. Eur Radiol. 2009;19(11):2707-15.
   Article PubMed Google Scholar
4. Takashima H, Takebayashi T, Yoshimoto M, et al. Differentiating spinal intradural-extramedullary schwannoma from meningioma using MRI T2 weighted images. Br J Radiol. 2018;91(1092):20180262.
   Article PubMed Google Scholar
5. Zhai X, Zhou M, Chen H, et al. Differentiation between intraspinal schwannoma and meningioma by MR characteristics and clinic features. Radiol Med. 2019;124(6):510-21.
   Article PubMed Google Scholar
6. Setzer M, Vatter H, Marquardt G, Seifert V, Vrionis FD. Management of spinal meningiomas: Surgical results and a review of the literature. Neurosurg Focus. 2007;23(4):E14.
   Article PubMed Google Scholar
7. Koçak B, Durmaz EŞ, Ateş E, Kılıçkesmez Ö. Radiomics with artificial intelligence: a practical guide for beginners. Diagn Interv Radiol. 2019;25(6):485- 95.
   Article PubMed Google Scholar
8. Cay N, Mendi BAR, Batur H, Erdogan F. Discrimination of lipoma from atypical lipomatous tumor/well-differentiated liposarcoma using magnetic resonance imaging radiomics combined with machine learning. Jpn J Radiol. 2022;40(9):951- 60.
   Article PubMed Google Scholar
9. Vychopen M, Arlt F, Wilhelmy F, et al. Association of quantitative radiomic shape features with functional outcome after surgery for primary sporadic dorsal spinal meningiomas. Front Surg. 2023;10:1303128.
   Article PubMed Google Scholar
10. Xu Z, Wang YH, Cheng ZY, et al. Combined radiomics nomogram of different machine learning models for preoperative distinguishing intraspinal schwannomas and meningiomas: A multicenter and comparative study. Clin Radiol. 2024;79(9):e1108-e1116.
    Article PubMed Google Scholar
11. World Medical Association. World Medical Association Declaration of Helsinki: Ethical principles for medical research involving human subjects. JAMA. 2013;310(20):2191-4.
    Article PubMed Google Scholar
12. Kocak B, Baessler B, Bakas S, et al. CheckList for EvaluAtion of Radiomics research (CLEAR): A step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights Imaging. 2023;14(1):75.
    Article PubMed Google Scholar
13. Yamaguchi S, Takeda M, Takahashi T, et al. Ginkgo leaf sign: a highly predictive imaging feature of spinal meningioma. J Neurosurg Spine. 2015;23(5):642-6.
    Article PubMed Google Scholar
14. Shur JD, Doran SJ, Kumar S, et al. Radiomics in oncology: A practical guide. Radiographics. 2021;41(6):1717-32.
    Article PubMed Google Scholar
15. Lohmann P, Galldiks N, Kocher M, et al. Radiomics in neuro-oncology: Basics, workflow, and applications. Methods. 2021;188:112-21.
    Article PubMed Google Scholar
16. Brunasso L, Ferini G, Bonosi L, et al. A spotlight on the role of radiomics and machine-learning applications in the management of intracranial meningiomas: A new perspective in neuro-oncology: A review. Life (Basel). 2022;12(4):586.
    Article PubMed Google Scholar
17. Langenhuizen PPJH, Zinger S, Leenstra S, et al. Radiomics-based prediction of long-term treatment response of vestibular schwannomas following stereotactic radiosurgery. Otol Neurotol. 2020;41(10):e1321-e1327.
    Article PubMed Google Scholar
18. Gu H, Zhang X, di Russo P, Zhao X, Xu T. The current state of radiomics for meningiomas: promises and challenges. Front Oncol. 2020;10:567736.
    Article PubMed Google Scholar
19. Zhang Z, Li N, Qian Y, Cheng H. Establishment of an MRI-based radiomics model for distinguishing between intramedullary spinal cord tumor and tumefactive demyelinating lesion. BMC Med Imaging. 2024;24(1):317.
    Article PubMed Google Scholar
20. Gitto S, Bologna M, Corino VDA, et al. Diffusion-weighted MRI radiomics of spine bone tumors: Feature stability and machine learning-based classification performance. Radiol Med. 2022;127(5):518-25.
    Article PubMed Google Scholar
21. Lang N, Zhang Y, Zhang E, et al. Differentiation of spinal metastases originated from lung and other cancers using radiomics and deep learning based on DCE-MRI. Magn Reson Imaging. 2019;64:4-12.
    Article PubMed Google Scholar
22. Maki S, Furuya T, Horikoshi T, et al. A deep convolutional neural network with performance comparable to radiologists for differentiating between spinal schwannoma and meningioma. Spine (Phila Pa 1976). 2020;45(10):694-700.
    Article PubMed Google Scholar
23. Cao J, Wang X, Qiao Y, et al. Differentiation of benign and malignant spinal schwannoma using guided attention inference networks on multi-source MRI: Comparison with radiomics method and radiologist-based clinical assessment. Acta Radiol. 2023;64(3):1184-93.
    Article PubMed Google Scholar

Declarations

Additional Information

Publisher’s Note
Bayrakol MP remains neutral with regard to jurisdictional and institutional claims.

Rights and Permissions

Creative Commons License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). To view a copy of the license, visit https://creativecommons.org/licenses/by-nc/4.0/

View full license details →

About This Article