An Interpretable Machine Learning Model for Meningioma Grade

Prediction

Ali Golbaf

, Damjan Veljanoski

, Prutha Chawda

2 c

, Swen Gaudl

C. Oliver Hanemann

and Emmanuel Ifeachor

School of Engineering, Computing and Mathematics, University of Plymouth, Plymouth, U.K.

Peninsula Schools of Medicine, University of Plymouth University, Plymouth, U.K.

Keywords: Meningiomas, Grading, Radiomics, MRI, Interpretable Techniques.

Abstract: Accurate preoperative prediction of meningioma grade is crucial for enhancing the clinical management of

these tumours. In this study, we developed a non-invasive machine learning (ML) model to predict

meningioma grade using clinical features and radiomics features from preoperative MRI scans, focusing on

interpretability to improve clinical adoption of such models. A dataset of 94 patients from The Cancer Imaging

Archive (TCIA) was analysed. Clinical features and radiomics features from T1-weighted contrast-enhanced

(T1C) and T2-weighted Fluid Attenuated Inversion Recovery (T2 FLAIR) scans were utilised. Two feature

subsets were constructed: one using radiomics features alone and the other combining clinical and radiomics

features. Feature selection was performed using a modified Least Absolute Shrinkage and Selection Operator

(LASSO) technique. Four ML models: Logistic Regression (LR), Support Vector Machine (SVM), Random

Forest (RF), and Gradient Boosting (GB), were developed. SHapley Additive exPlanations (SHAP) was

employed to address the blackbox nature of ML models by providing radiomics overall feature importance

scores and model interpretation. Results using the clinical-radiomics subset showed that the SVM

outperformed others (test AUC: 0.83), indicating its reliability for predicting meningioma grade. SHAP

highlights discriminative radiomics features and their interaction with clinical features, thereby enhancing the

clinical adoption of such models.

1 INTRODUCTION

Meningiomas, the most common primary brain

tumours, are among the most understudied tumours

within the central nervous system (Low et al. 2022).

However, a significant proportion of meningiomas

(20% - 30%), show aggressive behaviour, and high

recurrence rate (Zhang et al. 2019). These tumours

are categorised into three grades, according to 2021

World Health Organisation (WHO) guidelines (Louis

et al. 2021). High-grade meningiomas (Grades II and

III) show more aggressive behaviour than low-grade

cases (Grade I), leading to a 5-year progression free

survival probability (Wang, Nassiri, et al. 2023).

https://orcid.org/0000-0002-8104-5600

https://orcid.org/0000-0002-4951-8586

https://orcid.org/0009-0004-0345-9017

https://orcid.org/0000-0003-3116-3761

https://orcid.org/0000-0002-1951-1025

https://orcid.org/0000-0001-8362-6292

They are also challenging to be completely resected

using invasive treatment strategies and often require

adjunctive radiotherapy (Fountain, Young, and

Santarius 2020). Thus, accurate grading of these

tumours is important in enhancing the clinical

management of meningiomas.

The gold standard for grading of meningiomas

still relies on invasive methods such as

histopathology and biopsy (Herrgott et al. 2023).

However , invasive methods may not be applicable to

tumours that are surgically inaccessible and patients

with multiple diseases. Moreover, biopsies may not

accurately reflect the heterogeneity of meningiomas

due to limited sampling (Islim et al. 2020; Tagle et al.

Golbaf, A., Veljanoski, D., Chawda, P., Gaudl, S., Hanemann, C. O. and Ifeachor, E.

An Interpretable Machine Learning Model for Meningioma Grade Prediction.

DOI: 10.5220/0013114000003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 65-74

ISBN: 978-989-758-731-3; ISSN: 2184-4305

2002). Consequently, there is a growing need for the

development of non-invasive models that accurately

predict the grade of meningiomas.

Currently, MRI serves as the primary non-

invasive method in the clinical management of

meningiomas (Zhang et al. 2020). However, some

conventional MRI features of different meningioma

grades overlap, which can potentially lead to

misdiagnosis (Spille et al. 2019). In this context,

radiomic, which is a quantitative approach for

medical image analysis, has emerged as a novel way

to extract imaging features that carry valuable

biological information about tumours which are not

accessible by conventional image analysis (Lambin et

al. 2012). Machine learning has also demonstrated

potential in developing non-invasive predictive

models by capturing complex patterns within these

features (Langs et al. 2018). Such models have been

developed for the diagnosis, prognosis, and treatment

of meningiomas, and have particularly shown

promise in meningioma grading (Patel et al. 2023).

However, radiomics and machine learning have

not yet been adopted in the clinical management of

meningiomas. The blackbox nature of machine

learning models make their outputs difficult to

interpret (Patel et al. 2023). The application of

interpretability techniques may mitigate the inherent

blackbox nature of machine learning models (Reyes

et al. 2020). However, only a few studies have

focused on improving the interpretability of machine

learning models in the clinical management of

meningiomas. SHAP, which is used to assess the

contribution of each radiomic feature to model

performance, has been used to interpret a machine

learning model for evaluating the post-surgical

recurrence of high-grade meningiomas (Park, Choi, et

al. 2022). Relevance-weighted Class Activation

Mapping, an explanation method for visualising class

relevance, has been employed to explain a machine

learning model for meningioma segmentation (Jun et

al. 2023). Additionally, Local Interpretable Model-

Agnostic Explanations (LIME), an estimator

technique, which approximates models locally for

interpretability, has been applied to interpret machine

learning models, for predicting glioma grades but

not for the prediction of meningioma grade (Wang et

al. 2019).

In this paper, we developed an interpretable

machine learning model for predicting meningioma

grade using both clinical and radiomics features. The

aim is to enhance the adoption of radiomics and

machine learning in the clinical management of

meningiomas by establishing links between

meningioma grade, radiomics features, and their

interactions with clinical features.

2 METHODS

2.1 Dataset

The dataset used in this study was obtained from

TCIA, a publicly available database (Clark et al.

2013). It comprises a cohort of 96 patients who were

diagnosed with meningioma between 2010 and 2019

(Vassantachart). Low-grade and high-grade

meningiomas were identified according to the 2016

WHO guidelines. Clinical features were also recorded

by two experienced neuropathologists and one

neuropathology fellow. All patients underwent pre-

operative T1C, and T2 FLAIR MRI scans. A detailed

description of the imaging protocol can be found in

(Vassantachart et al. 2022). In this study, cases with

inconsistent histopathological records and suboptimal

image qualities were excluded, yielding a final cohort

of 94 patients. The clinical features of the patients are

summarised in Table 1.

2.2 Model Development

Figure 1 shows the workflow for developing an

interpretable model. The TCIA dataset was

processed, with MRI data standardised to the Brain

Imaging Data Structure (BIDS) format for

consistency and reproducibility (Gorgolewski et al.

2016). Radiomics features were extracted, and the

most discriminative features selected. These features

trained various ML models to predict meningioma

grade. SHAP was then applied for determining

overall radiomics feature importance scores and

model interpretation. The final model can predict

meningioma grade in new cases.

2.2.1 Image Processing and Radiomics

Feature Extraction

The dataset had undergone a prior image processing

pipeline, ensuring data consistency and quality. As

detailed in (Vassantachart et al. 2022), anatomically

co-registered T1C and T2 FLAIR MRI scans were

obtained, with any misalignment corrected using the

automated rigid registration software, VelocityAI. T2

FLAIR scans were resampled into their

corresponding T1C scans, followed by isovoxel

resampling. In the present study, further image

processing techniques, including bias field correction

and normalisation, were applied based on the

HEALTHINF 2025 - 18th International Conference on Health Informatics

Table 1: Histopathological and demographic characteristics of the patients.

Features Groups

Total Female Male

94 67 (71.3%) 27 (28.7%)

Age

Min 25 25 29

Mean 55.39 54.53 57.51

Max 88 88 85

Grade

Low-grade 53 (56.4%) 46 (68.7%) 07 (26.0%)

High-grade 41 (43.6%) 21 (31.3%) 20 (74.0%)

Location

Anterior and middle cranial fossa 45 (47.9%) 33 (49.3%) 12 (44.5%)

Convexity 19 (20.2%) 09 (13.4%) 10 (37.0%)

Falx and parasagittal 16 (17.0%) 12 (17.9%) 04 (14.8%)

Posterior cranial fossa 12 (12.8%) 11 (16.4%) 01 (03.7%)

Lateral ventricle 02 (02.1%) 02 (03.0%) 00 (00.0%)

Figure 1: Study workflow.

radiomics standardisation protocol for brain MRI

scans outlined in (Carré et al. 2020), using the

SimpleITK N4BiasFieldCorrection and

NormaliseImage filters (Yaniv et al. 2018). Manually

delineated tumour lesions were also available within

T1C and T2 FLAIR MRI scans. These annotations

were created by a medical student and a radiation

oncology resident and then reviewed by a radiation

oncologist with over 5 years of experience

(Vassantachart et al. 2023). Radiomics features

including shape, first-order, and texture features were

subsequently extracted from these lesions.

2.2.2 Clinical and Radiomics Feature

Selection

Feature selection is a key step in the development of

ML models, as radiomics features often show strong

An Interpretable Machine Learning Model for Meningioma Grade Prediction

correlations, potentially resulting in redundant

information, that can detrimentally affect model

interpretability and generalisability (Reyes et al.

2020). LASSO, a widely used feature selection

technique for analysing high-dimensional data,

improves model performance and interpretation,

although, highly correlated features may undermine

its efficiency (Zou and Hastie 2003). To tackle this

issue, we perform a multi-level feature selection

method based on LASSO coefficient thresholds

(Wang, An, et al. 2023). Clinical features were also

analysed using the t-test for age and the chi-square

test for gender and tumour location, with p-values

below 0.05 as statistically significant.

2.2.3 ML Model to Classify Meningioma

Grade

ML models for classifying low-grade and high-grade

meningiomas were developed using LR, SVM, RF,

and GB classifiers. To ensure robustness and

generalisability, we conducted 1000 random training-

test splits (1:4 ratio), generating training-test set pairs

with 70 training and 24 test cases (An et al. 2021).

Models were trained using 10-fold cross-validation

within each training set. Model performance was

evaluated by averaging AUC, Accuracy, Precision,

Recall, and F1-score. The best-performing model was

selected for interpretability analysis.

2.2.4 Model Interpretability

SHAP is a well-established technique for enhancing

ML model interpretability. However, random

perturbation-based sampling in SHAP implies that

with different random seeds, a high ranked feature in

one iteration may be considered as a low ranked

feature in the next iteration (Xiang et al. 2023). To

mitigate this issue, we determined overall radiomics

feature importance scores by generating multiple

training-test sets. The iteration process was

terminated when the change in average importance

scores for each feature was equal or less than 0.01

between two consecutive iterations. Overall

radiomics feature importance scores were then

considered as these averages. SHAP Kernel Explainer

was utilised to determine radiomics feature

importance scores in each iteration and scores were

normalised by the sum of all feature importances.

2.2.5 Implementation

In this study, Python 3.8 was used for data conversion

to BIDS format, image processing, radiomics feature

extraction, and feature selection, as well as for ML

development, and interpretability analyses. The

HeuDiConv tool (version 0.9.0,

https://github.com/nipy/heudiconv), facilitated the

conversion of DICOM files into BIDS format

(Halchenko, Goncalves, and Castello 2020). Image

processing was implemented using SimpleITK

package (version 2.3.0). The open-source package

Pyradiomics (version 3.1.0, https://github.com/AIM-

Harvard/pyradiomics) was used for feature extraction

(Van Griethuysen et al. 2017). Clinical categorical

variables were encoded numerically. The Scikit-learn

package (version 1.3.2) was used for radiomics

feature selection, ML model development, and

evaluation. Interpretability techniques was performed

using SHAP (version 0.43.0) package.

3 RESULTS

3.1 Radiomics Feature Extraction

A total of 1576 radiomics features were extracted,

including 14 shape features describing the size and

contours of the tumours, 18 first-order features

characterising the distribution of voxel intensities

within the lesions, and 68 texture features measuring

the variation of voxel intensities across T1C and T2

FLAIR MRI scans. Texture features were extracted

using 22 Grey Level Co-occurrence Matrix (GLCM),

16 Gray-Level Run-length Matrix (GLRLM), 16

Gray-Level Size Zone Matrix (GLSZM), and 14 Gray

Level Difference Matrix (GLDM). Subsequently, 688

wavelet radiomics features were evaluated by

applying wavelet decomposition on the original

images at both high and low frequencies.

3.2 Feature Selection and Model

Performance

A subset of 18 radiomics features was identified as the

most discriminative. Significant differences in age,

gender, and tumour location between low-grade and

high-grade meningiomas were observed, with p-values

lower than 0.05. These clinical features were added to

the radiomics subset to form a clinical-radiomics

subset. The specifics of the feature subsets are outlined

in Table 2, and the performance of ML models in

training and test sets are shown in Table 3. The

classifiers exhibited high accuracy and precision in

distinguishing between tumour grades. Among the

models, the SVM using the clinical-radiomics subset

achieved the highest performance with AUC (0.90 ±

0.12 and 0.83 ± 0.07), Accuracy (0.83 ± 0.13 and 0.84

± 0.06), Precision (0.84 ± 0.18 and 0.82 ± 0.10), Recall

HEALTHINF 2025 - 18th International Conference on Health Informatics

(0.80 ± 0.21 and 0.80 ± 0.11), and F1-score (0.80 ±

0.16 and 0.80 ± 0.08) in training and test sets,

respectively.

3.3 Model Interpretability

Figure 2 depicts overall radiomics feature importance

scores. Figure 3 presents the SHAP violin summary

plot, which illustrates the distribution and variability

of SHAP values for each feature in distinguishing

between low-grade and high-grade meningiomas.

Higher SHAP values indicate greater impact on the

model output, with wider violins showing higher

density and more frequent values. Figure 4 represents

feature SHAP dependence plots for the 3 top-ranked

radiomics features. Figure 5 shows the interaction

between the top-ranked radiomics feature with

clinical features. In these figures each dot represents

a prediction related to feature values, with the x-axis

showing actual values and the y-axis showing SHAP

values.

4 DISCUSSIONS

In this study, using clinical-radiomics feature

subset, SVM model was the best-performing model

with the highest average values of AUC (0.90 ± 0.12

and 0.83 ± 0.07), Accuracy (0.83 ± 0.13 and 0.84 ±

0.06), Precision (0.84 ± 0.18 and 0.82 ± 0.10),

Recall (0.80 ± 0.21 and 0.80 ± 0.11), and F1-score

(0.80 ± 0.16 and 0.80 ± 0.08) in the training and test

sets, respectively.

In the current study, the number of extracted

radiomics features from T1C, and T2 FLAIR MRI

scans was almost the same, highlighting the

importance of using multi-parametric MRI scans in

the relevant studies (Park, Shin, et al. 2022). Previous

studies used clinical features and radiomics features

from MRI scans to predict meningioma low-grade

and high-grade. Duan et al. developed a radiomics

nomogram with AUC of 0.95, using clinical and

LASSO-selected radiomics features from T1C MRI

scans (Duan, Zhou, et al. 2022). They also developed

seven ML models, with the SVM model achieving an

AUC of 0.88 (Duan, Li, et al. 2022). Similarly, Chu

et al. used LASSO-selected radiomics features from

T1, T1C, and T2 MRI scans to develop an LG model

with an AUC of 0.95 (Chu et al. 2021). While these

studies demonstrated strong predictive performance,

they lacked interpretability in their models. Although

complex machine learning models are effective in

capturing patterns in the data, they often result in

models that are difficult for clinicians to interpret,

where understanding the features influencing

predictions is crucial for clinical management.

One major hinderance to the adoption of

radiomics and ML models in the clinical management

of meningiomas is the blackbox nature of ML models.

To address this issue, SHAP was utilised to extract

the overall radiomics feature importance scores and

model interpretation. Here, GLRLM and GLSZM

radiomics features were identified as the most

discriminative features, showing high correlation

with meningioma grade prediction (Han et al. 2021).

GLSZM quantifies gray level zones within MRI

scans. A gray level zone is defined as the number of

connected voxels that share the same gray level

intensity. GLRLM features describes heterogeneity in

the distribution of run lengths, whereas GLSZM

(Traverso et al. 2020). The majority of selected

radiomics features (15 out of 18) were derived from

the wavelet-filtered MRI scans, which have been

proved to be the most discriminative features in

meningioma grade prediction (Han et al. 2021).

The violin plot depicted in Figure 3 indicates that

higher values of the first radiomics feature, wavelet-

HLH_glszm_ZoneEntropy_t1c, correspond to an

increased output probability of high-grade

meningiomas. A similar trend was observed for

the third radiomics feature, wavelet-

LHL_glszm_GrayLevelNonUniformityNormalized_

t1c. Conversely, lower values of the

second radiomics feature, wavelet-

LHH_glszm_GrayLevelNonUniformityNormalized_

t2f, were associated with an increased output

probability.

This study presented SHAP dependence plots to

illustrate the 3 top-ranked radiomics feature

interactions. The results presented in figure 4 a-c,

show that (i) higher values of the first radiomics

feature paired with lower values of the second

increase the probability of high-grade meningiomas,

while lower first feature values diminish this effect;

(ii) lower values of the first feature combined with

higher third feature values decrease probability, and

(iii) the second feature values consistently impact

probability regardless of the third feature values.

Interestingly, in this study, age emerged as a

statistically significant feature in predicting

meningioma grade based on t-test analysis. However,

as depicted in the Figure 5.a, age does not exhibit a

specific distribution that increases the output

probability, aligning with (Hu et al. 2020; Duan, Li,

et al. 2022).

An Interpretable Machine Learning Model for Meningioma Grade Prediction

Table 2: Total and statistically significant features.

Subset

Total features Statisticall

nificant features

Clinical

Radiomics

Clinical

Radiomics

Ori

inal Wavelet Ori

inal Wavelet

T1C

FLAIR

T1C

FLAIR

T1C

FLAIR

T1C

FLAIR

Radiomics - 100 100 688 688 - 2 1 8 7

Clinical-

Radiomics

3 100 100 688 688 3 2 1 8 7

Table 3: Prediction performance of ML models.

Classifie

Subset Set AUC Accurac

Precision Recall F1-Score

radiomics

Trainin

0.83 ± 0.15 0.75 ± 0.14 0.73 ± 0.20 0.75 ± 0.23 0.72 ± 0.18

Test 0.76 ± 0.08 0.76 ± 0.08 0.71 ± 0.11 0.77 ± 0.14 0.73 ± 0.09

Clinical-

radiomics

Training 0.88 ± 0.13 0.78 ± 0.14 0.78 ± 0.20 0.78 ± 0.22 0.75 ± 0.17

Test 0.80 ± 0.08 0.80 ± 0.08 0.76 ± 0.11 0.79 ± 0.13 0.76 ± 0.09

SVM

radiomics

Trainin

0.87 ± 0.14 0.80 ± 0.20 0.81 ± 0.20 0.75 ± 0.24 0.75 ± 0.18

Test 0.80 ± 0.07 0.80 ± 0.07 0.79 ± 0.11 0.75 ± 0.14 0.76 ± 0.09

Clinical-

radiomics

Trainin

0.90 ± 0.12 0.83 ± 0.13 0.84 ± 0.18 0.80 ± 0.21 0.80 ± 0.16

Test 0.83 ± 0.07 0.84 ± 0.06 0.82 ± 0.10 0.80 ± 0.11 0.80 ± 0.08

radiomics

Trainin

0.83 ± 0.15 0.76 ± 0.14 0.77 ± 0.22 0.72 ± 0.24 0.73 ± 0.20

Test 0.76 ± 0.08 0.77 ± 0.08 0.73 ± 0.11 0.73 ± 0.14 0.72 ± 0.10

Clinical-

radiomics

Trainin

0.86 ± 0.14 0.78 ± 0.14 0.80 ± 0.21 0.72 ± 0.24 0.73 ± 0.19

Test 0.77 ± 0.08 0.77 ± 0.08 0.75 ± 0.12 0.72 ± 0.14 0.72 ± 0.10

radiomics

Trainin

0.79 ± 0.17 0.72 ± 0.15 0.71 ± 0.23 0.69 ± 0.24 0.67 ± 0.20

Test 0.73 ± 0.09 0.73 ± 0.09 0.67 ± 0.10 0.71 ± 0.14 0.68 ± 0.10

Clinical-

radiomics

Trainin

0.79 ± 0.17 0.72 ± 0.15 0.70 ± 0.23 0.68 ± 0.25 0.67 ± 0.20

Test 0.72 ± 0.08 0.72 ± 0.08 0.67 ± 0.11 0.69 ± 0.15 0.67 ± 0.10

Figure 2: The overall radiomics feature importance scores of features extracted by SHAP. t1c: contrast-enhanced T1-

weighted; t2f: T2-weighted fluid attenuated inversion recovery.

0 0.2 0.4 0.6 0.8 1

wavelet-HLH_glszm_ZoneEntropy_t1c

wavelet-…

wavelet-HLL_glrlm_RunVariance_t2f

wavelet-…

original_firstorder_Skewness_t1c

wavelet-…

original_glrlm_HighGrayLevelRunEmphasis_t1c

wavelet-HHH_firstorder_Skewness_t1c

wavelet-HLH_firstorder_Mean_t1c

wavelet-LHH_glszm_SizeZoneNonUniformityNormalized_t2f

wavelet-LHH_glszm_SmallAreaHighGrayLevelEmphasis_t2f

wavelet-HLH_glszm_SizeZoneNonUniformityNormalized_t2f

wavelet-HHL_glszm_SmallAreaLowGrayLevelEmphasis_t2f

original_glszm_HighGrayLevelZoneEmphasis_t2f

wavelet-HHL_glszm_HighGrayLevelZoneEmphasis_t1c

SHAP Overall Importance Scores

HEALTHINF 2025 - 18th International Conference on Health Informatics

Figure 3: SHAP violin summary plot.

(a)

(b)

(

)

Figure 4: SHAP dependence plots for the three top-ranked radiomics features, illustrating the interactions between: (a) the

first and second features, (b) the first and third features, and (c) the second and third features.

An Interpretable Machine Learning Model for Meningioma Grade Prediction

(a)

(b)

(

)

Figure 5: SHAP dependence plots for clinical and the top-ranked radiomics features, illustrating the interactions between the

top-ranked radiomics features and: (a) age, (b) gender and (c) tumour location.

Conversely, concerning gender (as depicted in figure

5.b), males (red dots) show a tendency to decrease the

output probability. Moreover, it is apparent that low

values of the first radiomics feature in males, tend to

decrease the output probability. This indicates that

relying solely on statistical tests is not sufficiently

reliable for predicting the effects of features on the

model predictions. It was also shown that for females

(blue dots), it generally increases the output

probability while lower values of the first radiomics

features tend to decrease the output probability.

Figure 5.c indicates that irrelevant to the tumour

location, lower values of the first radiomics feature

decrease the output probability while the combination

of the posterior cranial fossa and middle cranial fossa

locations and higher values of the first radiomics

feature tends to increase the output probability.

When a new case is presented, our model helps

clinicians to make better informed decisions by

providing insights into the factors influencing

predictions. Considering the interactions among

radiomics features themselves and their interaction

with clinical features may enable clinicians to

consider additional nuances in their clinical

judgments. Clinicians can also see which factors the

model considers most critical, helping them

understand the basis of the prediction. This also

enables comparison with previous cases. Clinicians

can compare the new case with similar past cases

where the model made predictions, seeing how the

new case aligns or differs, thereby validating the

model's prediction. Additionally, the model provides

detailed explanations for each prediction, breaking

down the contribution of each feature and offering a

clear rationale that clinicians can review. It also helps

identify anomalies. If a new case presents unusual

patterns or outliers in the data, interpretable models

can flag these anomalies, prompting further

investigation by clinicians to ensure the prediction is

accurate and relevant.

The current study has several limitations. TCIA,

the publicly available dataset used here was

retrospective, relatively small, and derived from a

single institution. Grade III meningiomas were also

excluded from the dataset due to their rare occurrence.

However, leveraging public datasets provides

researchers access to a diverse and extensive pool of

medical imaging data, enabling robust analysis and

enhancing the generalisability of findings across

various patient populations and clinical strategies.

Surprisingly, the utilisation of TCIA dataset accounts

for only 4% in the meningiomas-relevant studies (Patel

et al. 2023). This study only used two types of MRI

scans while other MRI scans such as ADC mapping

HEALTHINF 2025 - 18th International Conference on Health Informatics

were not considered. However, enhancing the clinical

management of meningiomas by constructing an

interpretable machine learning model that predicts

meningioma grade was the main objective of this

study.

5 CONCLUSIONS

Utilising clinical and radiomics features, the SVM

ML model, offers a reliable approach for preoperative

prediction of meningioma grade. By identifying

discriminative radiomic features and their

interactions with clinical features, SHAP supports the

potential for the enhanced clinical adoption of such

models. Future research should explore larger

datasets and diverse patients to validate and refine

these findings, further enhancing clinical adoption.

REFERENCES

An, Chansik, Yae Won Park, Sung Soo Ahn, Kyunghwa

Han, Hwiyoung Kim, and Seung-Koo Lee. 2021.

'Radiomics machine learning study with a small sample

size: Single random training-test set split may lead to

unreliable results', PloS one, 16: e0256152.

Carré, Alexandre, Guillaume Klausner, Myriam Edjlali,

Marvin Lerousseau, Jade Briend-Diop, Roger Sun,

Samy Ammari, Sylvain Reuzé, Emilie Alvarez Andres,

and Théo Estienne. 2020. 'Standardization of brain MR

images across machines and protocols: bridging the gap

for MRI-based radiomics', Scientific reports, 10:

12340.

Chu, Hairui, Xiaoqi Lin, Jian He, Peipei Pang, Bing Fan,

Pinggui Lei, Dongchuang Guo, and Chenglong Ye.

2021. 'Value of MRI radiomics based on enhanced

T1WI images in prediction of meningiomas grade',

Academic radiology, 28: 687-93.

Clark, K., B. Vendt, K. Smith, J. Freymann, J. Kirby, P.

Koppel, S. Moore, S. Phillips, D. Maffitt, M. Pringle,

L. Tarbox, and F. Prior. 2013. 'The Cancer Imaging

Archive (TCIA): maintaining and operating a public

information repository', J Digit Imaging, 26: 1045-57.

Duan, CF, N Li, Y Li, F Liu, JC Wang, XJ Liu, and WJ Xu.

2022. 'Comparison of different radiomic models based

on enhanced T1-weighted images to predict the

meningioma grade', Clinical Radiology, 77: e302-e07.

Duan, Chongfeng, Xiaoming Zhou, Jiachen Wang, Nan Li,

Fang Liu, Song Gao, Xuejun Liu, and Wenjian Xu.

2022. 'A radiomics nomogram for predicting the

meningioma grade based on enhanced T 1WI images',

The British Journal of Radiology, 95: 20220141.

Fountain, Daniel M., Adam M. H. Young, and Thomas

Santarius. 2020. 'Chapter 24 - Malignant meningiomas.'

in Michael W. McDermott (ed.), Handbook of Clinical

Neurology (Elsevier).

Gorgolewski, Krzysztof J, Tibor Auer, Vince D Calhoun, R

Cameron Craddock, Samir Das, Eugene P Duff,

Guillaume Flandin, Satrajit S Ghosh, Tristan Glatard,

and Yaroslav O Halchenko. 2016. 'The brain imaging

data structure, a format for organizing and describing

outputs of neuroimaging experiments', Scientific data,

3: 1-9.

Halchenko, Y, M Goncalves, and MVDO Castello. 2020.

'nipy/heudiconv v0. 9.0', Published online December, 23.

Han, Yuxuan, Tianzuo Wang, Peng Wu, Hao Zhang,

Honghai Chen, and Chao Yang. 2021. 'Meningiomas:

Preoperative predictive histopathological grading based

on radiomics of MRI', Magnetic Resonance Imaging,

77: 36-43.

Herrgott, Grayson A, James M Snyder, Ruicong She,

Tathiane M Malta, Thais S Sabedot, Ian Y Lee, Jacob

Pawloski, Guilherme G Podolsky-Gondim, Karam P

Asmaro, and Jiaqi Zhang. 2023. 'Detection of

diagnostic and prognostic methylation-based signatures

in liquid biopsy specimens from patients with

meningiomas', Nature Communications, 14: 5669.

Hu, Jianping, Yijing Zhao, Mengcheng Li, Jianyi Liu, Feng

Wang, Qiang Weng, Xingfu Wang, and Dairong Cao.

2020. 'Machine learning-based radiomics analysis in

predicting the meningioma grade using multiparametric

MRI', European journal of radiology, 131: 109251.

Islim, Abdurrahman I, Midhun Mohan, Richard DC Moon,

Nitika Rathi, Ruwanthi Kolamunnage-Dona, Anna

Crofton, Brian J Haylock, Samantha J Mills, Andrew R

Brodbelt, and Michael D Jenkinson. 2020. 'Treatment

outcomes of incidental intracranial meningiomas:

results from the IMPACT cohort',

World Neurosurgery,

138: e725-e35.

Jun, Yohan, Yae Won Park, Hyungseob Shin, Yejee Shin,

Jeong Ryong Lee, Kyunghwa Han, Sung Soo Ahn,

Soo Mee Lim, Dosik Hwang, and Seung-Koo Lee.

2023. 'Intelligent noninvasive meningioma grading

with a fully automatic segmentation using

interpretable multiparametric deep learning',

European radiology: 1-10.

Lambin, Philippe, Emmanuel Rios-Velazquez, Ralph

Leijenaar, Sara Carvalho, Ruud GPM Van Stiphout,

Patrick Granton, Catharina ML Zegers, Robert Gillies,

Ronald Boellard, and André Dekker. 2012. 'Radiomics:

extracting more information from medical images using

advanced feature analysis', European journal of cancer,

48: 441-46.

Langs, G, S Röhrich, J Hofmanninger, F Prayer, J Pan, C

Herold, and H Prosch. 2018. 'Machine learning: from

radiomics to discovery and routine', Der Radiologe, 58: 1.

Louis, David N, Arie Perry, Pieter Wesseling, Daniel J Brat,

Ian A Cree, Dominique Figarella-Branger, Cynthia

Hawkins, HK Ng, Stefan M Pfister, and Guido

Reifenberger. 2021. 'The 2021 WHO classification of

tumors of the central nervous system: a summary',

Neuro-oncology, 23: 1231-51.

Low, Justin T, Quinn T Ostrom, Gino Cioffi, Corey Neff,

Kristin A Waite, Carol Kruchko, and Jill S Barnholtz-

Sloan. 2022. 'Primary brain and other central nervous

system tumors in the United States (2014-2018): A

An Interpretable Machine Learning Model for Meningioma Grade Prediction

summary of the CBTRUS statistical report for

clinicians', Neuro-oncology practice, 9: 165-82.

Park, Chae Jung, Seo Hee Choi, Jihwan Eom, Hwa Kyung

Byun, Sung Soo Ahn, Jong Hee Chang, Se Hoon Kim,

Seung-Koo Lee, Yae Won Park, and Hong In Yoon.

2022. 'An interpretable radiomics model to select

patients for radiotherapy after surgery for WHO grade

2 meningiomas', Radiation Oncology, 17: 147.

Park, Yae Won, Seo Jeong Shin, Jihwan Eom, Heirim Lee,

Seng Chan You, Sung Soo Ahn, Soo Mee Lim, Rae

Woong Park, and Seung-Koo Lee. 2022. 'Cycle-

consistent adversarial networks improves

generalizability of radiomics model in grading

meningiomas on external validation', Scientific reports,

12: 7042.

Patel, Ruchit V, Shun Yao, Raymond Y Huang, and Wenya

Linda Bi. 2023. 'Application of radiomics to

meningiomas: a systematic review', Neuro-oncology,

25: 1166-76.

Reyes, Mauricio, Raphael Meier, Sérgio Pereira, Carlos A

Silva, Fried-Michael Dahlweid, Hendrik von Tengg-

Kobligk, Ronald M Summers, and Roland Wiest. 2020.

'On the interpretability of artificial intelligence in

radiology: challenges and opportunities', Radiology:

artificial intelligence, 2: e190043.

Spille, Dorothee Caecilia, Peter B Sporns, Katharina Hess,

Walter Stummer, and Benjamin Brokinkel. 2019.

'Prediction of high-grade histology and recurrence in

meningiomas using routine preoperative magnetic

resonance imaging: a systematic review', World

Neurosurgery, 128: 174-81.

Tagle, Patricio, Pablo Villanueva, Gonzalo Torrealba, and

Isidro Huete. 2002. 'Intracranial metastasis or

meningioma?: an uncommon clinical diagnostic

dilemma', Surgical neurology, 58: 241-45.

Traverso, Alberto, Michal Kazmierski, Ivan Zhovannik,

Mattea Welch, Leonard Wee, David Jaffray, Andre

Dekker, and Andrew Hope. 2020. 'Machine learning

helps identifying volume-confounding effects in

radiomics', Physica Medica, 71: 24-30.

Van Griethuysen, Joost JM, Andriy Fedorov, Chintan

Parmar, Ahmed Hosny, Nicole Aucoin, Vivek Narayan,

Regina GH Beets-Tan, Jean-Christophe Fillion-Robin,

Steve Pieper, and Hugo JWL Aerts. 2017.

'Computational radiomics system to decode the

radiographic phenotype', Cancer research, 77: e104-e07.

Vassantachart, A., Cao, Y., Shen, Z., Cheng, K., Gribble,

M., Ye, J. C., Zada, G., Hurth, K., Mathew.

"Segmentation and Classification of Grade I and II

Meningiomas from Magnetic Resonance Imaging: An

Open Annotated Dataset (Meningioma-SEG-CLASS)

(Version 1)." In, edited by The Cancer Imaging

Archive.

Vassantachart, April, Yufeng Cao, Michael Gribble,

Samuel Guzman, Jason C Ye, Kyle Hurth, Anna

Mathew, Gabriel Zada, Zhaoyang Fan, and Eric L

Chang. 2022. 'Automatic differentiation of Grade I and

II meningiomas on magnetic resonance image using an

asymmetric convolutional neural network', Scientific

reports, 12: 3806.

Vassantachart, April, Yufeng Cao, Zhilei Shen, Karen

Cheng, Michael Gribble, Jason C Ye, Gabriel Zada,

Kyle Hurth, Anna Mathew, and Samuel Guzman. 2023.

'A repository of grade 1 and 2 meningioma MRIs in a

public dataset for radiomics reproducibility tests',

Medical Physics.

Wang, Justin Z, Farshad Nassiri, Alexander P Landry,

Vikas Patil, Jeff Liu, Kenneth Aldape, Andrew Gao,

and Gelareh Zadeh. 2023. 'The multiomic landscape of

meningiomas: a review and update', Journal of Neuro-

oncology, 161: 405-14.

Wang, Ke, Ying An, Jiancun Zhou, Yuehong Long, and

Xianlai Chen. 2023. 'A novel Multi-Level feature

selection method for radiomics', Alexandria

Engineering Journal, 66: 993-99.

Wang, Xiuying, Dingqian Wang, Zhigang Yao, Bowen

Xin, Bao Wang, Chuanjin Lan, Yejun Qin, Shangchen

Xu, Dazhong He, and Yingchao Liu. 2019. 'Machine

learning models for multiparametric glioma grading

with quantitative result interpretations', Frontiers in

neuroscience, 12: 1046.

Xiang, Xu, Hong Yu, Ye Wang, and Guoyin Wang. 2023.

'Stable local interpretable model-agnostic explanations

based on a variational autoencoder', Applied

Intelligence: 1-15.

Yaniv, Ziv, Bradley C Lowekamp, Hans J Johnson, and

Richard Beare. 2018. 'SimpleITK image-analysis

notebooks: a collaborative environment for education

and reproducible research', Journal of digital imaging,

31: 290-303.

Zhang, Guobin, Yunsheng Zhang, Guijun Zhang, Da Li,

Zhen Wu, Yonggang Wang, and Junting Zhang. 2019.

'Outcome and prognostic factors for atypical

meningiomas after first recurrence', Journal of Clinical

Neuroscience, 63: 100-05.

Zhang, Jing, Kuan Yao, Panpan Liu, Zhenyu Liu, Tao Han,

Zhiyong Zhao, Yuntai Cao, Guojin Zhang, Junting

Zhang, and Jie Tian. 2020. 'A radiomics model for

preoperative prediction of brain invasion in

meningioma non-invasively based on MRI: A

multicentre study', EBioMedicine, 58: 102933.

Zou, Hui, and Trevor Hastie. 2003. 'Regression shrinkage

and selection via the elastic net, with applications to

microarrays', JR Stat Soc Ser B, 67: 301-20.

HEALTHINF 2025 - 18th International Conference on Health Informatics