Abstract
Background/Aim: NF2-related schwannomatosis, formerly known as Neurofibromatosis type 2 (NF2) is characterized by bilateral vestibular schwannomas (VS). Managing NF2 requires balancing watchful waiting with surgical intervention, each carrying inherent risks. While these risks are acknowledged, they have not yet been subjected to systematic investigation. Accurate prognosis of tumor growth is crucial for clinical decision-making. This study investigated radiomics features from longitudinal magnetic resonance imaging (MRI) data to predict VS growth.
Patients and Methods: Radiomics features were extracted from cranial MRIs of 32 NF2 patients, each with at least two or more imaging time points. The association between these features and tumor growth was analyzed through correlation, visual inspection, and the Boruta algorithm.
Results: Correlations between growth rates and radiomics features were weak (ρ≤0.23, p<0.016). Three features exhibited a bimodal distribution, with cluster affiliation linked to tumor growth rate [cluster A: 7.9%/month, cluster B: 2.0%/month; Fisher exact odds ratio (OR)=2.55, p=0.010]. When considering only the first tumors in the MRI series, the Fisher exact OR was 2.29 (p=0.223). Boruta analysis identified wavelet.HLH_glcm_InverseVariance as a key feature, also relevant in the bimodal distribution. The Fisher exact OR of wavelet.HLH_glcm_InverseVariance for tumor growth was 2.64 (p=0.011) for all tumors and 2.21 (p=0.229) for initial tumors in the MRI series.
Conclusion: Bimodally distributed radiomics features from initial MRIs did not reliably predict rapid tumor growth (error probability: 23%) but may aid in planning MRI follow-up intervals.
Introduction
NF2 is an autosomal dominant inherited disorder predisposing individuals to neural tumors (1). The estimated incidence is approximately one in 28,000, with a prevalence of one in 50,000, which is expected to rise due to improved diagnostics and aging population (2, 3). NF2 is caused by mutations in the NF2 gene located on chromosome 22q12, which encodes the tumor suppressor protein Merlin (4). Various pathogenic variants of the NF2 gene have been identified, occurring either as germline mutations or as mosaic variants acquired during embryonic development (5, 6).
The hallmark clinical features of NF2 are bilateral vestibular schwannomas (VS), present in approximately 90% of affected individuals (7, 8). Additionally, multiple meningiomas, other schwannomas, and ependymomas are frequently observed (9). Diagnostic criteria also include cataracts, retinal hamartomas, and germline pathogenic variants in the NF2-gene (10). The most common initial symptoms of VS are tinnitus and hearing loss (7, 11).
The management of NF2 aims to balance careful conservative monitoring with surgical intervention or, in the case of inoperable VS, radiotherapy (12, 13). Both approaches carry significant risks (9, 14, 15). The off-label use of bevacizumab and brigatinib has shown potential in controlling tumor growth (16-18). The primary objective is to preserve function and maintain a high quality of life for as long as possible (11).
Prognostic factors in NF2. The age at symptom onset and the age at diagnosis are critical predictors of disease severity (19-21). These factors correlate with VS growth rate (19, 22) as well as overall mortality (23).
Radiomics and tumor growth prediction. Radiomics facilitates the quantitative analysis of medical imaging data by extracting tissue characteristics such as shape and heterogeneity (24). By integrating demographic, histological, and genetic data, radiomics provides a non-invasive and powerful tool for clinical diagnostics (25).
Radiomics feature analyses in sporadic VS. Since 2012, the prognostic value of radiomics features in sporadic VS has been investigated, yielding inconsistent results. Truong et al. (26) analyzed a cohort of 78 patients and found no correlation between MRI texture analysis and tumor growth rate. In contrast, Milchenko et al. (27) examined 229 patients and demonstrated that in automated radiomics analysis maximum and minimum intensity features provide a reliable prediction of tumor growth.
Tumor growth prediction of sporadic VS. The growth rate of sporadic VS and progressive hearing loss are critical factors in determining the optimal timing for aggressive treatment options such as microsurgical resection or radiotherapy (28-30). Similarly, early tumor growth prediction is essential for therapy planning in NF2-associated VS (31). To the best of our knowledge, the prognostic value of radiomics analysis for VS growth prediction in patients with NF2 has not yet been systematically investigated (as of March 6, 2025).
Patients and Methods
The Ethics Committee of the Hamburg Medical Council declared not having concerns on this study (Case Number: 2024-300549-WF), as the patient data and imaging used in this research were fully anonymized and could no longer be linked to any individual.
This retrospective study is based on 182 cranial MRIs from 35 consecutive patients with NF2 treated in the Neurofibromatosis Outpatient Department of the University Medical Center Hamburg-Eppendorf between 2002 and 2020. In this cohort no mosaic variants were detected by genetic analyses. NF2 was diagnosed by the criteria valid at the time of initial presentation (32). MRIs were acquired at a minimum of two time points (median: 5, range=2-15) using Siemens Avanto and Symphony 1.5T scanners (Siemens Healthineers AG, Erlangen, Germany). For an overview of the workflow, see Figure 1.
Overview of the radiomics workflow from MRI preprocessing to feature extraction in NF2-associated vestibular schwannomas.
Two patients were excluded due to implausible tumor growth curves. Additionally, for two other patients, imaging of one side was excluded following neurosurgical intervention. Furthermore, MRIs from four patients were removed after undergoing treatment with Bevacizumab or Everolimus.
Study population. After exclusions, this study included 170 cranial MRTs with 232 VS (left: 112, right: 120) from 32 patients. In this selected cohort, the median age at the first MRI was 31 years (range=2-67 years), with a female-to-male ratio of 20:12.
Tumor segmentation. Despite extensive efforts to automate tumor segmentation in medical imaging datasets (33, 34), manual segmentation by an experienced clinician, with confirmation by an independent second reviewer, remains the gold standard (35-37). It has also been demonstrated for NF2-associated VS that 3D-segmented volume analysis is superior to conventional 1D methods in many cases (38).
Moreover, for smaller datasets, standardizing and organizing imaging data within the directory structure for automation would have been significantly more labor-intensive than manual tumor segmentation in our setting.
Tumor segmentation workflow. 1) Conversion of DICOM MRIs to NIfTI Format: Performed using dcm2niix in MRIcroGL (39, 40). This step includes anonymization of MRI data. 2) Brain extraction/skull stripping: Conducted using Robex (41, 42). 3) Intensity normalization: Applied by Pypi (43). 4) Coregistration of image series for each patient: Performed using FSL FLIRT (44). 5) Manual segmentation: Conducted by two experienced raters considering tumor size progression (45). A typical example is depicted in Figure 2. Segmentations from preceding time points were taken into consideration.
Segmentation of the left vestibular schwannoma in patient #17.
Radiomics feature extraction. Radiomics feature extraction was performed using PyRadiomics (46, 47). The developers of this open-source Python package aim to establish a reference standard for radiomics analysis and are actively involved in the Image Biomarker Standardization Initiative (24). However, full consistency of results across different radiomics analysis tools has not yet been achieved (48). The radiomics parameters were optimized following the recommendations in (49).
Tumor growth assessment. The primary parameter used to evaluate tumor growth was the Percentage Growth Rate per Month (PGRpM), calculated relative to the previous volume.
Utilizing PGRpM as a central metric for tumor growth enables a standardized and comparable assessment of tumor dynamics. Absolute volumetric changes alone are insufficient to accurately capture individual tumor growth patterns, particularly in cases with varying tumor sizes and differing examination intervals. PGRpM provides a normalized measure, allowing for an objective quantification of growth rates independent of initial tumor volumes and imaging time points. This approach facilitates the early identification of growth trends and supports more precise therapeutic decision-making.
Three methodological approaches were applied in the analysis: (a) Correlation analysis between tumor growth and specific features using Spearman’s correlation coefficient (50); (b) Visual inspection of tumor growth curves in relation to individual radiomics features. For visual inspection and correlation analysis, z-score normalization was applied to the radiomics features (51); (c) Importance analysis via Boruta (52, 53), a machine learning algorithm designed for large datasets and based on an extension of the random forest method (54). In Boruta analyses, features with high intercorrelation and low correlation with tumor growth were excluded, reducing the initial set from 982 to 102 features, following the approach of Escudero-Sanchez et al. (55). Given the iterative nature of the Boruta algorithm, multiple runs were performed to ensure a stable and reliable selection of relevant features (53).
Results
Manual segmentation. In the manual segmentation of contrast-enhanced T1-weighted sequences, implausible tumor size progressions of VS were repeatedly observed. These inconsistencies did not occur in T2-weighted sequences. All T2 sequences were acquired with a slice thickness of 5 mm, echo times ranging from 0.077 to 0.106 ms, and repetition times between 4.00 and 5.94 ms.
Correlations. No Spearman correlation coefficient (ρ) greater than 0.23 was observed between PGRpM and the distribution of 982 radiomics features, with p-values ranging from <0.001 to 0.016.
Visual inspection. Upon visual examination of the graphs depicting PGRpM versus the distribution of z-scored features, three radiomics features exhibited a bimodal distribution, forming distinct clusters A and B. These features (Figure 3) were: zscore_wavelet-HHH_glrlm_ GrayLevelNonUniformityNormalized; zscore_wavelet-HLH_glcm_InverseVariance; and zscore_wavelet-HHH_ glcm_InverseVariance.
Bimodal radiomics features and their relationship with percentage growth rate per month, based on 232 vestibular schwannomas.
A threshold of -1 was defined to assign features to either cluster A or B. Among the 232 tumors, 229 were clearly classified into one of the two clusters, while three tumors exhibited overlap between both clusters.
A strong association was observed between cluster membership and PGRpM. Tumors in cluster A had a mean PGRpM of 7.9% per month, whereas those in cluster B exhibited a mean PGRpM of 2.0% per month. The odds ratio (OR) was 2.54 [95% confidence interval (CI)=1.1-5.4], indicating that tumors in the cluster A had a 2.54-fold higher risk of faster growth (Fisher’s exact test: OR=2.55, p=0.010).
When considering only the first imaging set from each patient’s MRI series for these three features, the OR was 2.25 (95%CI=0.5-8.9; Fisher’s exact OR=2.286, p=0.223). When considering the wavelet.HLH_glcm_InverseVariance alone in the initial examination the Fisher OR was 2.214, p=0.229.
Boruta analysis. The 102 selected features were analyzed using 1,000 iterations of the Boruta algorithm, with a significance threshold of p<0.05.
Features classified as “confirmed” in more than 600 iterations included: wavelet.LLH_firstorder_Skewness; wavelet.HLH_glcm_InverseVariance; wavelet.LLH_glcm_ ClusterShade. These features exhibited mean importance values ranging from 4.30 to 5.27 (Table I). The feature wavelet.HLH_glcm_InverseVariance was identified as relevant in both the Boruta analysis and the bimodal distribution.
Radiomics features present in >600 runs of Boruta.
Higher wavelet.HLH_glcm_InverseVariance values suggest smooth, uniform regions with minimal intensity fluctuations, commonly found in homogeneous surfaces or soft-textured tissues in medical imaging. Lower values indicate high-contrast, textured, or rough regions, which may correspond to structural details, edges, or areas with fine-grained variations (56, 57).
Beyond the analysis of NF2 VS, glcm-Inverse-variance emerged as the feature most strongly correlated with local recurrence of lung cancer (58). In another study of histological findings in lung cancer, previous research has reported a significant correlation between glcm-Inverse-variance and the Ki-67 proliferation index (p<0.001) (59).
Discussion
The bimodal distribution of wavelet.HLH_glcm_Inverse Variance emerges as a significant predictor of VS growth across all tumors (Fisher’s exact OR=2.64, p=0.011), and in the initial examination (Fisher’s exact OR=2.214, p=0.229). The high importance of this feature is further emphasized by the Boruta analysis, where it was confirmed in 691 out of 1,000 iterations, with a mean importance score of 4.61.
Limitations. This study has several significant limitations. As a retrospective analysis, it is inherently subject to limited generalizability (60, 61). Furthermore, the relatively small patient cohort of only 32 individuals necessitates a critical interpretation of the findings. The slice thickness of 5 mm in the segmented T2-weighted sequences represents a major potential source of error due to partial volume effects (62). To mitigate this issue, the study accounted for the segmentation from the preceding time point in each case.
The pronounced differences between cluster A and B were largely driven by two tumors with exceptionally high growth rates in cluster A. After excluding these outliers, the OR for wavelet.HLH_glcm_InverseVariance between clusters decreased to 1.72 (95%CI=0.80-3.52; Fisher’s exact test OR=1.73, p=0.113). However, these extreme growth rates were thoroughly reviewed and deemed valid.
Interpretation. To the best of our knowledge, this study is the first to investigate the association between radiomics features and NF2-associated VS. In previous studies on sporadic acoustic neuromas, only a limited number of features were analyzed (27). Furthermore, the relatively new Boruta algorithm, which enhances feature selection, has not yet been utilized in this context (26, 63, 64).
Conclusion
This study aimed to predict the growth behavior of NF2-associated VS based on radiomics features. While certain features, such as wavelet.HLH_glcm_InverseVariance, appear to have prognostic relevance. Given an error probability of 23%, the predictive value for rapid tumor growth based on the first MRI alone is insufficient for clinical decision-making. However, this feature may still be useful in determining the optimal frequency of MRI follow-up examinations.
Acknowledgements
The MRI series were kindly provided by Prof. Victor F. Mautner from the Department of Neurology, University Medical Center Hamburg-Eppendorf (UKE). This study was funded by the German lay organisation Bundesverband Neurofibromatose e.V. (Nina Boe and Johannes A. Koeppen).
Footnotes
Authors’ Contributions
Study conception and design: Nina Boe, Johannes A. Koeppen, Victor F. Mautner; Data collection: Victor F. Mautner; Data analysis: Nina Boe, Johannes A. Koeppen; Results interpretation: Reinhard E. Friedrich, Victor F. Mautner, Nina Boe, Johannes A. Koeppen; Drafting the manuscript: Nina Boe, Johannes A. Koeppen; Manuscript revision and approval: All Authors.
Conflicts of Interest
SF has received speaker honoraria from Alexion and compensation for advice or lecturing from SpringWorks and Alexion not related to this study. All other Authors declare that they have no conflicts of interest.
- Received March 11, 2025.
- Revision received March 21, 2025.
- Accepted March 31, 2025.
- Copyright © 2025 International Institute of Anticancer Research (Dr. George J. Delinasios), All rights reserved.
This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY-NC-ND) 4.0 international license (https://creativecommons.org/licenses/by-nc-nd/4.0).
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.