A single MRI scan contains sufficient imaging information for accurate prediction of meningioma growth risk

Highlights

• •Single contrast-enhanced MRI enables meningioma growth risk prediction.
• •Deep network features and SVM obtain F1-scores up to 0.97 on 192 single-tumor cases.
• •High accuracy for small tumors; medium-sized tumors remain challenging to classify.
• •Performance varies by vendor, field strength, and location; 3 T performs better.
• •Non-invasive AI-driven single-scan assessment complements MRI monitoring pathway.

Abstract

Neurosurgical strategies for monitoring meningiomas and evaluating their growth risk largely rely on serial imaging or invasive sampling, practices that place considerable burdens on both patients and clinical resources. In this study, we present a novel framework for predicting meningioma growth risk using only a single contrast-enhanced MRI scan. Our approach compares a custom-trained fully convolutional neural network encoder and PyRadiomics features at both tumor- and whole-image scale, capturing tumor-specific and peritumoral image features, evaluated using conventional machine learning classifiers. The study cohort includes 192 patients with single meningiomas, categorized as growing, stable, or shrinking, based on volumetric assessments by expert neurosurgeons. Classifiers trained on encoder-derived features achieved the highest F1-scores of 0.97 ± 0.01, demonstrating strong predictive performance particularly when edema content was included. Ensemble learning on encoder- and PyRadiomics-extracted features did not improve accuracy compared to the individual approaches. Prediction performance varied across scanner vendor, field strength, tumor location, and volume quintiles, with 3 T scanners yielding superior results, notably higher accuracy for smaller tumors under 1.73  cm³, and comparatively reduced performance in foramen magnum and intraventricular regions. This represents an important advance with clear clinical relevance, as smaller tumors are more difficult to classify regarding future growth. Our findings establish the feasibility of predicting meningioma growth risk from a single MRI scan, offering a non-invasive approach for early risk stratification and personalized surveillance strategies. By reducing reliance on serial imaging, the approach has the potential to support informed clinical decisions while improving resource allocation and ensuring timely intervention.

1. Introduction

Meningiomas are the most common primary intracranial tumors in adults that arise from the meningothelial cells of the arachnoid layer of the meninges, accounting for over one-third of central nervous system (CNS) tumors and are typically slow growing (Maniar et al., 2023, Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025). According to the World Health Organization (WHO), meningiomas are classified into three grades: Grade I being benign and minimally aggressive; Grade II being associated with elevated risks of recurrence; and Grade III being the most aggressive form and associated with poor clinical outcomes (Yang et al., 2024, Mostafa et al., 2024). While typically benign, approximately 20 % exhibit atypical characteristics and rarely malignant features (Yang et al., 2024, Wang, et al., 2024). This necessitates close monitoring for tumors that are not operated upon at the outset (Yang et al., 2024, Wang, et al., 2024).

Standard imaging modalities, including magnetic resonance imaging (MRI) and computed tomography (CT), are used for diagnosis and monitoring, with contrast enhancement aiding in tumor characterization (Hanna et al., 2023, Goldbrunner et al., 2021, Nagai Yamaki et al., 2021, Farajzadeh et al., 2023, Sadeghzadeh et al., 2025). However, despite the widespread use of these imaging modalities, clinical decision-making at the time of diagnosis remains largely reactive with limited capacity to prospectively predict individual tumor growth behavior (Hanna et al., 2023, Goldbrunner et al., 2021, Nagai Yamaki et al., 2021, Farajzadeh et al., 2023). Despite their generally slow growth behavior, meningiomas display highly variable progression within each tumor grade, with reported annual volume changes (AVC) ranging from complete stability (∼0.1  cm³) to substantial increases (∼1.8  cm³), in some cases exceeding 90 % per year (Sughrue et al., 2010, Olivero et al., 1995, Davis et al., 2022).

Clinically, these tumors may present with seizures, progressive neurological deficits (such as hemiparesis, aphasia, or visual disturbances), and cognitive impairments, depending on their size, location, and severity (Wiemels et al., 2010, Whittle et al., 2004). While surgical resection remains the primary treatment for symptomatic patients, the unpredictable growth pattern complicates clinical decision making, particularly in asymptomatic and/or conservatively managed cases where immediate surgery is not indicated (Alexiou et al., 2010, Opalak et al., 2023). In such cases, clinicians lack reliable tools to distinguish tumors that will remain indolent from those likely to progress and exhibit clinically significant growth (Alexiou et al., 2010, Opalak et al., 2023). This highlights the importance of early indicators of tumor behavior to support optimal treatment planning.

Current monitoring strategies rely principally on sequential imaging, typically through semiannual follow-ups, to assess tumor progression (Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025, Hanna et al., 2023, Goldbrunner et al., 2021, Nagai Yamaki et al., 2021). However, this standardized approach does not account for patient-specific growth trajectories, as tumor progression may follow heterogeneous and non-linear patterns, leading to potential overtreatment or delayed intervention (Davis et al., 2022). On the other hand, the use of CT imaging exposes patients to ionizing radiation and may pose contrast-related risks, particularly for those requiring repeated imaging, along with the associated inconvenience of attending multiple sessions (Wang, et al., 2024, Nieto Alvarez et al., 2024, Bhatia et al., 2024). Importantly, current predictive approaches do not provide individualized growth risk estimation at baseline and therefore fall short in supporting personalized surveillance strategies without reliance on follow-up imaging (Hashiba et al., 2009), leaving an important gap in clinical management. This highlights the need for more advanced strategies that can help reduce imaging frequency, mitigate the risks associated with variable growth rates, and offer a structured framework to guide risk prediction beyond routine follow-up scheduling.

Despite notable progress in the application of deep learning (DL) across neuro-oncology (Farajzadeh et al., 2023, Ahsan et al., 2024), the specific challenge of predicting meningioma growth using these methods remains largely unexplored. Most existing DL-based studies focus on tumor detection, segmentation, or retrospective volumetric change analysis across multiple imaging time points, rather than prospection of tumor growth risk at the time of diagnosis (Farajzadeh et al., 2023, Ahsan et al., 2024, Abbasi Sureshjani et al., 2024, Roy et al., 2020, Maizer and Alhijawi, 2024, Goswami, 2021, Sadeghzadeh et al., 2025). Previous research has primarily focused on automating volumetric comparisons of intracranial tumors (i.e., glioma, meningioma, and pituitary tumors) and eye tumors across multiple imaging rounds to support clinical decision-making (Maizer and Alhijawi, 2024, Goswami, 2021). Other studies have employed invasive techniques, such as biopsies, to classify tumor growth in lymphomas and breast cancer through repeated tissue sampling (Abbasi Sureshjani et al., 2024, Roy et al., 2020). However, these methods still require frequent invasive procedures which fail to reduce the dependence on multiple imaging/tissue sampling rounds, maintaining a significant burden on the patient and clinical resources.

In parallel, molecular and survival-based studies have examined genetic markers (Librizzi et al., 2024) associated with meningioma growth and prognosis, as well as overall and progression-free survival (Wach et al., 2025). Beyond growth-focused analyses, imaging-based research has applied radiomics and machine learning (ML) to meningioma characterization tasks such as tumor grading, consistency prediction, recurrence risk estimation, and outcome-related stratification using baseline MRI features (Park et al., 2019, Zhai et al., 2021). Similar approaches have also been extended more broadly across neuro-oncology to predict tumor aggressiveness, treatment response, or survival from single or limited imaging time points (Zhai et al., 2021, Pendem et al., 2025). While these studies demonstrate image-derived features can capture clinically meaningful tumor phenotypes, their predictive aims differ fundamentally from longitudinal growth behavior, and they do not directly address predictive growth risk classification at the time of diagnosis.

Motivated by evidence indicating that early treatment of meningiomas can reduce mortality by up to 65 % (Aizer et al., 2015), this study seeks to address a critical gap in neuro-oncology: the absence of non-invasive, image-based methods capable of predicting meningioma growth risk (i.e., into growing, stable, and shrinking) at the time of diagnosis from a single imaging session, without reliance on serial follow-up imaging. Previously, we have demonstrated that neurosurgeon-derived MRI and CT features can meaningfully support meningioma growth risk prediction (Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025). Building on this, here we ask a more fundamental question: can the growth trajectory of a meningioma be predicted at the time of diagnosis using only the imaging information contained in a single baseline contrast-enhanced T1-weighted (T1W + C) MRI scan? Specifically, we investigate whether automated computer vision and ML approaches can classify tumors as growing, stable, or shrinking from a single imaging session (examples of tumor behavior are shown in Fig. 1). We hypothesize that a single baseline MRI, with at least one meningioma, contains sufficient latent tumor-specific and peritumoral imaging features that are sufficient for reliable growth risk classification. To address this question, the study uses T1W + C data of 192 patients with single meningioma tumors and pursues four core objectives:

• 1.We evaluate whether 1) deep-learning encoder-based feature extraction at the tumor and peritumoral levels and 2) PyRadiomics-derived features from the entire MRI, fed into ML classifiers, can predict meningioma growth risk across the categories of growing, stable, or shrinking.
• 2.We examine how predictive performance is influenced by tumor size and location, with particular emphasis on smaller tumors, where early risk stratification is most clinically valuable.
• 3.We assess the robustness of predictions across scanner vendor, machine type, and magnetic field strength, to understand potential hardware reflected effects on performance and generalizability.
• 4.Finally, we quantify the contribution of peritumoral information, including edema region, by systematically expanding the tumor margin (i.e., bounding boxes used for predictions) from 0 % to 400 % and whole-brain, to identify how much surrounding tissue meaningfully supports growth prediction and where additional context begins to degrade performance.

Fig. 1.

Examples of growth trajectories from three patients at three subsequent scan rounds. Yellow arrows indicate tumor location. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

2. Methods

2.1. Dataset

The dataset used in this study comprises T1W + C MRI scans from 336 meningioma-confirmed patients (265 females and 71 males, with one or more tumors), collected between 2002 and 2020 across multiple Auckland City hospitals (Davis et al., 2022). The cohort includes patients of diverse ethnic backgrounds, including Māori (n = 61), Pasifika (n = 32), European (n = 187), and other ethnicities (n = 56). Depending on clinical and surgical requirements, imaging was performed over multiple scan rounds, with each patient undergoing a minimum of two and a maximum of five scans. The data reflect substantial real-world acquisition heterogeneity, encompassing 21 distinct scanner configurations across Siemens, Philips, and GE platforms, using both 1.5 T and 3 T magnetic field strengths. Acquisition parameters differed modestly across sites and vendors, reflecting routine clinical practice, with examples including Siemens Verio 3 T, Philips Achieva 3 T, and GE Signa HDxt 1.5 T. This diversity reflects the inherent variability found in clinical imaging settings. In the dataset, the thresholds for growing and shrinking tumors were defined as volume changes of + 15 % and −15 %, respectively, and annotated by a neurosurgeon (BD), as detailed in (Davis et al., 2022). All MRI scans were pre-processed through a standardized skull stripping step designed and outlined by BraTS 2024 pipeline (Kazerooni et al., 2023).

2.2. Ethics

The study was approved by the Auckland Health Research Ethics Committee (Reference number: AH24358) at the University of Auckland, New Zealand.

2.3. Proposed method

We used only the MRI scans in the database, as our primary objective was to evaluate the predictive value of this modality for meningioma growth risk, rather than that of CT images. Our approach, depicted in Fig. 2, consists of two main steps:

• -Step 1: Since no dedicated meningioma-specific feature extractor was available, we trained a deep neural network, namely a fully convolutional network (FCN), for meningioma segmentation to enable its encoder to learn image features from tumor regions (Fig. 2, Fig. 3, Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11, Fig. 12, Fig. 13, Fig. 14). Post-training, the decoder block of this network was discarded, and the encoder block was then retained for use as a ‘deep-learning-based feature extractor’ in the subsequent step (Fig. 2, Fig. 3, Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11, Fig. 12, Fig. 13, Fig. 14).
• -Step 2: To integrate the growth risk prediction pipeline, this step uses features extracted through two approaches: the encoder block of the FCN; and the PyRadiomics (Van Griethuysen et al., 2017) feature extraction module, processed for the classification task (Fig. 2).

Fig. 2.

The schematic of the proposed method for meningioma growth risk prediction. Note that the RoIs used in both steps contain a tumor which was manually delineated by a neurosurgeon.

Fig. 3.

Flowchart of data preparation.

Fig. 4.

Distribution and longitudinal trends of meningioma growth among the 192 patients. (A), Distribution of tumor volumes. Quintiles are illustrated with dashed red lines, including: [0.15, 0.84), [0.84, 1.73), [1.73, 3.06), [3.06, 6.87), and [6.87, 57.16] cm³. (B) Volumetric trajectories across scan rounds categorized as growing, stable, and shrinking tumors. Blue lines indicate medians and orange lines indicate outliers determined by third quartile (Q3) + 1.5 × interquartile range (IQR). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5.

Examples of accurate vs. inaccurate segmentations across growth classes. Red: ground truth, green: FCN segmentation. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 6.

Normalized confusion matrices of the winner algorithms in Table 2, Table 3 averaged across all folds of cross-validation. (A) encoder-extracted features, SVM, mix-up; (B) PyRadiomics features, SVM, feature perturbation; (C) ensemble learning on both feature extractors, LR. Italic numbers are average of total cases in 5-folds.

Fig. 7.

Confusion matrices of the winner classifiers stratified across tumor volume quintiles; (A) encoder-extracted features, SVM, mix-up; (B) PyRadiomics features, SVM, feature perturbation; (C) ensemble learning on both feature extractors, LR. Italic numbers are average of total cases in 5-folds.

Fig. 8.

Confusion matrices of the winner classifiers stratified across scanner brands; (A) encoder-extracted features, SVM, mix-up; (B) PyRadiomics features, SVM, feature perturbation; (C) ensemble learning on both feature extractors, LR. Numbers in parentheses are average of total cases in 5-folds.

Fig. 9.

Confusion matrices of the winner classifiers stratified across scanner magnetic field strengths; (A) encoder-extracted features, SVM, mix-up; (B) PyRadiomics features, SVM, feature perturbation; (C) ensemble learning on both feature extractors, LR. Italic numbers are average of total cases in 5-folds.

Fig. 10.

Confusion matrices of the winner classifiers stratified across scanner type; (A,D) encoder-extracted features, SVM, mix-up; (B,E) PyRadiomics features, SVM, feature perturbation; (C,F) ensemble learning on both feature extractors, LR. Italic numbers are average of total cases in 5-folds.

Fig. 11.

Confusion matrices of the winner classifiers stratified across location in brain; (A,D) encoder-extracted features, SVM, mix-up; (B,E) PyRadiomics features, SVM, feature perturbation; (C,F) ensemble learning on both feature extractors, LR. Italic numbers are average of total cases in 5-folds.

Fig. 12.

Growth risk prediction performance stratified by meningioma location. Values represent normalized F1-scores for each region, reported for the best performing model in each configuration: (1) encoder-extracted features with SVM and mix-up augmentation, (2) PyRadiomics features with SVM and feature perturbation, and (3) ensemble learning combining both feature extractors with LR, shown in order from left to right. The tumor shapes and sizes shown are for illustrative purposes and do not reflect actual anatomical dimensions or information in the cohort of the sub-regions. The skull-based image (left) was adapted from “https://www.anatomystandard.com” under a CC BY-NC-SA 4.0 license. The cortical brain section (right) was adapted from ‘Slagter − Drawing Coronal section of the brain − no labels’ at AnatomyTOOL.org by Ron Slagter under a CC BY-NC-SA 4.0 license.

Fig. 13.

Visualization of accurate and inaccurate segmentation which resulted in misclassification. Red: ground truth, green: segmentation. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 14.

Impact of peritumoral margin size and tumor volume on meningioma growth prediction performance. F1-scores are shown as a function of progressively expanded tumor margins (0 % to 400  %), evaluated with and without inclusion of neurosurgeon-identified tumor volume using the best-performing classifier.

The following sections provide the technical methodologies employed across data preparation, feature extraction, modeling, and classification stages.

2.4. Step 1: Encoder-based feature learning

Initially, all 408 meningiomas were manually segmented by a neurosurgeon registrar (SMJ) and refined by a consultant neurosurgeon (JAC). In part I, the sub-volumes (i.e., regions of interest: RoIs) containing a tumor were extracted from the full brain MRI scans with respect to the manual segmentations. These RoIs, rather than the complete MRI, were used to train an FCN with an encoder-decoder architecture for tumor segmentation.

To do so, all MRIs were used without exclusion and initially pre-processed using the contrast-limited adaptive histogram equalization (CLAHE) algorithm (Pizer et al., 1990) to enhance scan contrast while limiting amplification to avoid noise over-amplification. Isotropic voxel resampling was implemented to standardize the scan resolution to 1 × 1 × 1 mm³, with each millimeter cube corresponding to one voxel. Subsequently, MRI patches of 64 × 64 × 64 voxels (justification for size is available in Section 2.5.2) were generated with a 25 % overlap. A patch was retained if more than 50 % of its pixels were non-zero (indicating brain tissue). However, an additional 20 % of the patches with fewer than 50 % brain pixels (in each patch) were randomly chosen and included to improve model generalizability in handling blank and non-related regions. Patches with fewer than 0.02 % tumor pixels (equivalent to 52 voxels, determined empirically) were excluded without exception. These patches were augmented through random intensity shifts, scaling, and flipping during the training of the segmentator. Lastly, all patches were randomly shuffled for each training epoch to ensure data diversity.

The encoder of the FCN segmentator in Fig. 2-I consisted of four blocks, each with two 3D convolutional layers employing ReLU activation, “same” padding, and “He-normal (He et al., 2015)” kernel initialization, followed by a 3D max pooling layer. The number of filters in the blocks increased progressively from 16 to 32, 64, and 128. The expansive path used up-sampling layers to incrementally restore spatial resolution, with each step followed by two 3D convolutional layers configured similarly to the contracting path. The number of filters decreased sequentially from 128 to 4 through six blocks. The final layer produced a 1-channel segmentation mask through a 1 × 1 × 1 convolutional layer with a sigmoid activation function. The training was performed using 5-fold cross-validation with a 70:10:20 split, with the constraint that data from any single patient was included in only one set to prevent data leakage. This step aimed to make the encoder learn detailed, meningioma-specific features from the RoIs.

Post-training, in part II, the expansive part of the network (i.e., its decoder) was discarded and only the encoder was retained for use as a feature extractor in Step 2. The encoder contains latent space features from the bottleneck which were reformatted by flattening them into a 1D vector, resulting in a 1 × 8192 vector per tumor for subsequent processing. From here on, we refer to these features as ‘encoder-extracted’, ‘encoder-derived’, or ‘encoder-based’ features where appropriate. To re-emphasize, the FCN was trained solely to enable its encoder to learn image features within meningioma-containing RoIs, and its segmentation outputs were not used anywhere in the pipeline. This focused design was aimed at capturing meningioma-specific features to support growth predictions in the second step below.

2.5. Step 2: Meningioma growth prediction

Step 2 consists of two pathways for predicting growth. In path A, an RoI containing a tumor (with respect to the manual segmentation) was initially passed to the pre-trained encoder to extract features. Note that automatic segmentations of the FCN were not used to crop the RoIs in this step, therefore, incorrect segmentations of FCN do not affect downstream tasks. In path B, the full, uncropped MRIs and their corresponding manual segmentations were passed to the PyRadiomics (Van Griethuysen et al., 2017) to extract radiomic features. In the following, sequential steps including data preparation, encoder- and PyRadiomics-based feature extraction, data augmentation approach, and tumor growth prediction (i.e., classification) procedures are explained.

2.5.1. Data preparation for feature extraction

The sequential data preparation steps are depicted in Fig. 3. The original dataset, with a combination of 1028 MRI and CT scans from 336 patients, did not include unique identifiers for individual tumors and, with some scans containing multiple tumors, it was impossible to include such patients in this study. This is a particularly important aspect as the growth status of each individual tumor serves as the model’s ground truth for growth prediction, which requires prior identification of every tumor. To address this, only records with a single tumor from the original dataset were included, resulting in a total of 864 scans from 295 patients (including both MRIs and CTs). Since our primary objective was to evaluate the predictive value of MRI, CT images were excluded, resulting in a dataset that contained 524 scans from 195 patients. Every scan of these patients was included, except their final round of imaging (to prevent bias, as growth status beyond that round was not recorded), resulting in 334 scans. In addition, scans without tumors (i.e., blank segmentations where skull stripping caused total tumor removal), or those with errors (e.g., scans showing incomplete brain), were excluded resulting in 329 scans from 192 patients.

Amongst the cohort, tumor behavior was consistent across all imaging time points, with 137 patients exhibiting progressive growth, 47 remained stable, and 8 cases exhibiting progressive shrinkage. The resulting scans of these patients contained 329 total tumors, with a distribution of 239 growing, 77 remaining stable, and 13 shrinking. Tumor volumes within the resulting scans ranged from 0.15 cm³ to 57.16 cm³ with a size distribution shown in Fig. 4A and per-class volumetric behavior shown in (Fig. 4B), indicating linear growth (Davis et al., 2022). Patient demographics and clinical distribution are available in Table 1. Notably, the resulting dataset contained no instances in which a new meningioma (i.e., an additional tumor) appeared on subsequent scans, nor contained a case where one or more previously-existing tumors shrank to disappearance, leaving only single meningiomas in subsequent scans. Thus, each scan contained only one tumor from the time of initial diagnosis onward. The median scan interval was 750 days, with an interquartile range (IQR) of 548.5 days. Although this indicates variability in follow-up times, it does not affect the study reliability because tumor behavior was consistent and linear within each growth class (Davis et al., 2022); and that the models ultimately are to classify a growth risk per tumor and do not depend on the scan interval.

Table 1.

Demographic and clinical distribution of the 192 patients in the filtered dataset.

Variable	Category	Count	Percentage (%)
Sex	Female	151	78.6
	Male	41	21.4
Ethnicity	European	107	55.7
	Māori	32	16.7
	Not Stated	2	1.0
	Other	34	17.7
	Pasifika	17	8.9
Age at diagnosis	[30, 35)	1	0.5
	[35, 40)	3	1.6
	[40, 45)	12	6.2
	[45, 50)	15	7.8
	[50, 55)	31	16.1
	[55, 60)	34	17.7
	[60, 65)	24	12.5
	[65, 70)	22	11.5
	[70, 75)	25	13.0
	[75, 80)	15	7.8
	[80, 85)	9	4.7
	[85, 90]	1	0.5
Underwent operation	Yes	31	16.1
	No	161	83.9
WHO grade	I	27	14.1
	II	4	2.1
	Not available for non-operated patients	161	83.9
Histopathology confirmation	Yes	31	16.1
	No (not available for non-operated patients)	161	83.9

2.5.2. Encoder-based feature extraction

Initially, RoIs were cropped to include the manually segmented tumor along with a 100 % margin on each side (i.e., 200  % along each axis) to capture both the tumor and its surrounding peritumoral environment for feature representation. This margin was chosen based on indications from our prior work suggesting that the peritumoral region may contain additional contextual information that can be biologically relevant for predicting tumor growth (Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025). The RoIs were zero-padded to match the largest cropped RoI and were then rescaled to fit a 64 × 64 × 64 voxel.

In our dataset, the mean tumor dimensions were 18.2 × 19.7 × 18.4 ± 8.4 × 9.0 × 8.6  mm³. Accordingly, the chosen input size of 64 × 64 × 64 voxel provided, on average, approximately a 200 % margin of surrounding tissue, offering sufficient peritumoral context (Dincer et al., 2020, Larson et al., 1995, Hutzelmann et al., 1998). This further assists in maintaining a balance between anatomical coverage and computational efficiency.

The resulting RoIs were passed through the trained encoder for feature extraction, resulting in a 1 × 8192 feature vector per tumor. No image/data augmentation was applied to the MRIs during this phase. Within the encoder-extracted feature, columns with constant values (i.e., zero standard deviation) were removed to reduce excessive dimensionality for the ML task, resulting in a 1 × 7981 feature vector per tumor. Note that although the large number of features may capture noise and increase the risk of overfitting in downstream tasks if not carefully refined, the adopted train–test splitting strategy and the use of cross-validation (see following sections) avoid overfitting to noise.

2.5.3. PyRadiomics feature extraction

As an alternative feature extraction approach, we used PyRadiomics (Van Griethuysen et al., 2017), an open-source Python library developed for extracting a wide range of quantitative features from medical imaging modalities such as CT, MRI, and PET. Here, we employed PyRadiomics to extract high-throughput radiomic features such as first-order statistics, shape descriptors, and texture features from tumor regions (with respect to the manual segmentation masks). In contrast to path A, no cropping was applied in this path to comply with the input requirements of the PyRadiomics implementation.

2.5.4. Data augmentation

Both sets of extracted features from the encoder and PyRadiomics blocks were augmented class-wise by a 100 % augmentation factor, via four different augmentation approaches: 1) Gaussian noise injection, 2) mix-up, 3) feature perturbation, and 4) SMOTE (Liu et al., 2022). To avoid bias, augmentation was applied within each growth class rather than across the entire dataset. This is particularly important as oversampling the underrepresented classes to match the dominant class would result in ground truth samples becoming a subset of the oversampled data, which could skew the model evaluation. For instance, the ‘shrink’ class, with only 13 original cases, would increase to 239 if oversampled. Lastly, the non-augmented data was preserved for testing, which also makes sure that the results are not inflated by the majority class due to the extreme class imbalance.

2.5.5. Classification

To capture the underlying distributions of the extracted features, the augmented features from each path were independently fed into a Gaussian mixture model (GMM) (Pearson, 1894) with 35 components (chosen empirically based on experimental investigations). To enhance feature representation, Fisher vectors (Sanchez et al., 2013) were subsequently computed from the GMMs to encode higher-order statistics of the feature distributions. These representations were then used as inputs to four ML classifiers: multi-layer logistic regression (LR), support vector machine (SVM), naïve Bayes, and decision tree. The task involved predicting and classifying tumor growth into one of the three states of growing, stable, or shrinking. Each model was selected for its distinct methodological strengths explained below.

LR was included for its backpropagation mechanism and its ability to model complex feature interactions. SVM was chosen for its ability to perform convex optimization. Decision tree was used for its interpretability and capacity to maximize information gain. Naïve Bayes was selected for its assumption of feature independence, which allows the exploration of diverse features without reliance on spatial adjacency (Tangirala, 2020).

To maximize overall predictive performance, we used ensemble learning to integrate the outputs of the individual classifiers. This strategy was applied across three scenarios: 1) using encoder-extracted features, 2) using features extracted via PyRadiomics, and 3) using the combination of both feature types. Training of the classifiers and the ensemble models was performed using stratified 5-fold cross-validation via 80:20 train-to-test split scheme (applied before augmentation) on the same underlying train/test data used for the FCN to avoid data leakage and bias. This way, no data were shared across folds between the FCN and the classifier training processes. To emphasize, the original non-augmented encoder-extracted features were preserved for testing the classifiers across all five folds to ensure unbiased evaluation and statistical validity of the training procedure. Moreover, samples from the same patients were not distributed across the training and testing sets, avoiding data leakage.

2.6. Experimental setup

We used Python 3.6 and TensorFlow 2.6.2 to train the FCN on high-performance computers (HPCs) provided by The University of Auckland, equipped with Nvidia V100 GPUs (32 GB memory) and a total of 1 TB shared RAM space, of which 10 GB was used. The growth classification models were implemented using the classifiers available in the SciKit-Learn Python library on CPU (due to the SciKit-Learn implementation) on an Apple MacBook Pro M3 Max with 36 GB RAM. All classifiers were trained within each cross-validation fold using standard SciKit-Learn implementations, with model fitting performed exclusively on training data and evaluation on held-out, non-augmented test samples.

2.7. Data analysis

2.7.1. Evaluation metrics

The Dice similarity score was used to evaluate the segmentation performance of the FCN, serving as an implicit indicator of the reliability of the encoder for downstream feature extraction. Tumor growth prediction performance was evaluated using the weighted F1-score, which for simplicity, we refer to it only as F1-score hereafter. All results are reported as mean ± standard deviation (SD) across cross-validation folds. Also, performance estimates for the shrinking class have higher uncertainty due to small sample size; we therefore present class-wise confusion matrices and interpret performance of each class cautiously.

2.7.2. Tumor size, location, machine type, and upstream FCN assessments

To assess the influence of tumor size on growth prediction performance, tumors were stratified into five volumetric quintiles spanning the full observed range (0.15–57.16 cm³), mentioned earlier: [0.15, 0.84), [0.84, 1.73), [1.73, 3.06), [3.06, 6.87), and [6.87, 57.16] cm³. In addition, given that the MRI scans were acquired across a range of vendors, machine type, and magnetic field strengths, we further examined the impact of these factors on prediction performance, along with potential associations between tumor location and growth risk classification.

Furthermore, to investigate whether misclassifications could be attributed to upstream feature extraction limitations, we quantitatively examined the FCN’s segmentations corresponding to the incorrectly classified cases. This analysis was aimed at determining whether encoder-based inferences contributed to suboptimal feature extraction, which potentially led to incorrect predictions. The findings from this assessment are presented in the Results section.

2.7.3. Peritumoral margin and volume ablation

To investigate the contribution of peritumoral context to growth prediction, we performed an ablation analysis by progressively expanding the tumor margin from 0 % to 400 % as well as whole-brain, and evaluating performance using the best-performing classifier identified in prior experiments. This analysis was aimed at quantifying how varying amounts of surrounding tissue, including edema, influence predictive accuracy and to identify the margin at which additional context no longer provides benefit.

In parallel, the model was trained both with and without the inclusion of neurosurgeon-identified tumor volume value to assess its incremental contribution to prediction performance. As the primary objective of this study was to evaluate MRI-based growth risk prediction value, tumor size and scanner-derived features were not included in these ablative analyses.

3. Results

In the following, we first report the performance of the FCN segmentator (presented solely as an indicator of its encoder’s robustness for feature extraction), followed by the evaluation results of the predictive classifiers for tumor growth prediction/categorization. We also include illustrative examples of FCN segmentations associated with inaccurately classified cases to provide insights into how the encoder-derived features may have influenced these prediction errors.

3.1. Segmentator reliability

Table 2 summarizes the segmentation performance of the FCN model (prior to discarding its decoder) to check its reliability for the downstream classification tasks. Overall, the model achieved a Dice score of 0.74 ± 0.27 and a Hausdorff 95 distance of 11.3 ± 6.1  mm on 64 × 64 × 64 voxels (RoIs). To provide greater transparency, the results in this table are stratified by tumor growth categories. The model demonstrated relatively consistent segmentation performance on growing and stable meningiomas with lower variance, whereas performance variance on shrinking tumors was slightly higher. Fig. 5 demonstrates examples of segmentation performance per growth class, with further visual examinations of the inferences confirming that the model consistently delineated meningioma regions with occasional influences from false positives and false negatives. These results collectively suggest that the model was effective in learning image features within the RoIs, justifying the use of its encoder as a feature extractor, while its decoder/segmentation outputs were not used in any part of the study.

Table 2.

Mean FCN segmentation performance.

Class	Dice	HD95 (mm)
Global	0.74 ± 0.27	11.3 ± 6.1
Growing	0.73 ± 0.09	13.1 ± 5.4
Stable	0.77 ± 0.16	9.5 ± 3.3
Shrinking	0.71 ± 0.29	16.7 ± 7.2

3.2. Tumor growth prediction

The augmentation methods applied to both encoder- and PyRadiomics-based features resulted in diverse new data points with negligible correlation to the original features (cosine similarity < 0.05; absolute feature-wise Pearson correlation < 0.001).

Table 3 presents a detailed comparison of the F1-score averages from all classifiers trained on the augmented features extracted by both the encoder and PyRadiomics approaches. Across all models, the SVM trained on encoder-extracted features, augmented with mix-up, achieved the highest average F1-score of 0.97 ± 0.01, closely followed by LR at 0.96 ± 0.01. For PyRadiomics-based features, the highest performance was also achieved by the SVM, reaching an average F1-score of 0.94 ± 0.01, when trained on data augmented using feature perturbation.

Table 3.

F1-scores of the encoder- and PyRadiomics-based approaches across different data augmentation methods. The highest performance with the lowest SD in each row is shown in bold.

Feature extractor	Model	Mix-up	Gaussian noise	Feature perturbation	SMOTE
Encoder-based	Naïve Bayes	0.59 ± 0.03	0.82 ± 0.03	0.30 ± 0.03	0.59 ± 0.02
	Decision tree	0.73 ± 0.04	0.84 ± 0.03	0.73 ± 0.02	0.79 ± 0.02
	Logistic regression	0.96 ± 0.01	0.90 ± 0.02	0.95 ± 0.03	0.97 ± 0.02
	SVM	0.97 ± 0.01	0.89 ± 0.04	0.96 ± 0.02	0.97 ± 0.02
PyRadiomics	Naïve Bayes	0.23 ± 0.04	0.62 ± 0.06	0.20 ± 0.06	0.25 ± 0.03
	Decision tree	0.72 ± 0.01	0.83 ± 0.01	0.74 ± 0.04	0.79 ± 0.05
	Logistic regression	0.86 ± 0.03	0.86 ± 0.02	0.94 ± 0.02	0.90 ± 0.01
	SVM	0.85 ± 0.02	0.86 ± 0.01	0.94 ± 0.01	0.91 ± 0.02

Table 4 presents the F1-score averages for the ensemble learning approach across the three scenarios mentioned in Section 2.5.5. Ensemble learning did not improve the performance of classifiers trained on encoder-extracted features and resulted in marginal performance reduction for models trained on PyRadiomics features, decreasing the average by 0.02 to 0.92 ± 0.01. When combining both feature types, the highest achieved performance was 0.97 ± 0.01, which matched the best result obtained without ensemble learning.

Table 4.

Mean F1-scores achieved through ensemble learning across the three scenarios. The highest performance with the lowest SD in each row is shown in bold.

Model	Encoder-based	PyRadiomics	Both
Naïve Bayes	0.97 ± 0.01	0.92 ± 0.01	0.97 ± 0.01
Decision tree	0.89 ± 0.04	0.84 ± 0.02	0.88 ± 0.04
Logistic regression	0.97 ± 0.02	0.92 ± 0.01	0.97 ± 0.01
SVM	0.97 ± 0.01	0.92 ± 0.02	0.97 ± 0.01

Fig. 6 demonstrates the normalized confusion matrices of the top-performing combinations identified in Table 3, Table 4: the SVM trained on encoder-extracted features augmented with mix-up, the SVM trained on PyRadiomics features augmented with feature perturbation, and the LR model through an ensemble learning approach on data from both feature extractors. From these results, ensemble learning improved classification performance for the ‘growing’ and ‘stable’ classes, but reduced accuracy for the ‘shrinking’ class.

Fig. 7 stratifies these confusion matrices by five tumor volume quintiles, demonstrating how the encoder-based and PyRadiomics features contributed to tumor growth prediction outcomes. While all approaches succeeded in predicting growth for larger meningiomas (>3 cm³), they faced challenges with the medium-sized tumors (1.73–3.06  cm³). Notably, both the encoder-based and PyRadiomics approaches performed particularly well for smaller tumors (<1.73  cm³).

Fig. 8, Fig. 9, Fig. 10 present stratified confusion matrices from Fig. 6 by scanner brand, machine type, and magnetic field strength, respectively. These results demonstrate consistent predictive reliability across vendors, with particularly strong performance for the ‘growing’ class and modest challenges in the ‘stable’ and ‘shrinking’ classes.

Fig. 11 further stratifies the predictions by tumor location while Fig. 12 maps these results onto a brain atlas. It is evident that growth of tumors on tentorial, falcine, foramen magnum, and intraventricular are challenging to classify.

3.3. FCN influence on classification

Representative examples of both accurate and inaccurate segmentations that were associated with incorrect predictions by the top-performing model (i.e., an SVM trained with mix-up augmentation achieving an F1-score of 0.97 – see Table 3) are provided in Fig. 13. For comparison, Fig. 5 presents examples of both accurate and inaccurate segmentations associated with correct classifications.

The portion of the cases (per growth class) that benefited from rich feature extraction (such as those in Fig. 5) can be inferred from the main diagonal elements of the confusion matrices available in Fig. 6, Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11. Similarly, the portion of the cases impacted by poor feature extraction (such as those in Fig. 13) is reflected in the off-diagonal elements of those matrices.

3.4. Peritumoral and volumetric information results

As shown in Fig. 14 prediction performance improved as peritumoral context was included, with F1-scores peaking at around 0.95–0.98 when tumor margins of 150–250 % were used. Smaller margins still produced reasonably good performance, with F1-scores of approximately 0.93–0.95. Beyond 250 %, however, performance declined sharply as larger regions of surrounding tissue were incorporated. Whole-brain context reduced performance and increased majority-class bias; therefore, subsequent analyses focus on tumor-centered regions. Inclusion of tumor volume provided a marginally better performance (<0.01 F1-score) for up to 225 % margin, with similarly competitive performance (<0.03 F1-score) at larger margins.

4. Discussion

This study marks a significant advancement in neuro-oncology, demonstrating for the first time that meningioma growth risk can be predicted from tumor-level image features derived from a single MRI scan. Our approach achieved high predictive performance, with average F1-scores exceeding 0.90 across 192 patients with single meningiomas, enabling reasonably high classification into growing, stable, and shrinking tumors, which establishes a robust benchmark for non-invasive prediction of tumor behavior. The consistent performance across folds further supports that the models did not overfit despite the high feature dimensionality.

By showing that machine-inferred latent features from a single timepoint can provide sufficient information for accurate risk stratification, our study offers a promising alternative to strategies that require multiple scans or invasive procedures, which are subject to inter-operator variability. From a theoretical perspective, these findings collectively support the notion that tumor growth dynamics are implicitly encoded in baseline morphological and textural imaging signatures, reinforcing representation learning as a powerful framework for modeling latent disease processes. In the absence of comparable studies, our proposed approach not only addresses a key unmet clinical need, but also opens new avenues for integrating imaging data into early, patient-specific clinical decision-making.

Previous research has inferred tumor growth only from longitudinal imaging – such as in studies of ocular (Goswami, 2021) or intracranial lesions (Maizer and Alhijawi, 2024) – or through tissue sampling – including studies on lymphoma (Abbasi Sureshjani et al., 2024) or ductal carcinoma cohorts (Roy et al., 2020). In contrast, we have introduced a promising methodology that predicts growth trajectory at the time of diagnosis, opening avenues for early, personalized clinical decision making. Clinically, this capability has direct implications for surveillance planning, as it may help identify patients who would benefit from closer follow-up or earlier intervention, versus those suitable for conservative monitoring.

Our stratified results in Fig. 7 indicated that the models achieved high predictive accuracy for smaller meningiomas (i.e., <1.73 cm³), which is clinically valuable for personalized treatment planning and early intervention. Accuracy was reduced for medium-sized tumors, reflecting the potential impact of biological uncertainty during intermediate growth stages. Prediction accuracies improved again for larger meningiomas, which may have already plateaued growth, making classification easier. Our findings align with the growing emphasis on non-invasive diagnostic tools and illustrated how single-scan prediction models can complement established longitudinal MRI surveillance (Goldbrunner et al., 2021, Hanna et al., 2023, Nagai Yamaki et al., 2021), helping to refine imaging schedules and interventions.

The greater prediction stability observed in Fig. 8 for images acquired from Philips scanners may be attributed to vendor-specific acquisition parameters. This provides insights that leaves space for future work with richer data across all the vendors to investigate potential manufacturer-specific relations. Results in Fig. 9 highlight better model performance for 3  T compared to 1.5 T scans, likely owing to the improved signal-to-noise ratio provided by higher field strength, which enhances tumor visibility and boundary definition. Lastly, results in Fig. 11, Fig. 12 showed decreased accuracy for tumors situated in complex anatomical regions (e.g., tentorial, falcine, foramen magnum, and intraventricular regions), findings that may stem from reduced contrast or structural complexity in these locations.

While the primary aim of this study was to investigate localized tumor- and peritumoral-level features through the encoder-based approach, complemented with the full-image radiomic profiling via PyRadiomics, our further investigations from a more conventional end-to-end learning strategy (Farajzadeh et al., 2022) by directly feeding the entire original MRI into the classifiers failed to converge. This likely reflects the challenges of capturing meaningful growth-related patterns without guidance from segmentation or structural priors. As an alternative, our strategy of extracting meningioma-level features via an encoder and using them as inputs for conventional ML classifiers (selected based on their methodological strengths) proved to be effective.

While combining the two feature sets through ensemble learning was expected to enhance predictive performance, it did not improve the performance and, in contrast, reduced accuracy for the ‘shrinking’ class (see Fig. 6, Fig. 7). Since it only affected four cases in the ‘shrinking’ class, which comprised a total of 13 cases, it is difficult to draw a conclusion as to whether this is due to mild overfitting or underlying ambiguity in the polling models. Importantly, this pattern is unlikely to reflect overfitting, as overall performance remained stable across training and held-out test sets under strict cross-validation, with no evidence of degraded generalization. Rather, the limited number of affected cases suggests that this inconsistency may stem from intrinsic ambiguity within the shrinking category itself or from subtle disagreements between the polling models when class boundaries are weakly defined. Overall, encoder-derived features demonstrated superior performance, likely because they captured meningioma-specific and peritumoral imaging patterns within the RoIs, resulting in greater predictive precision compared to the PyRadiomics-based approach. From a theoretical standpoint, this finding highlights the advantage of task-oriented representation over generic handcrafted descriptors for modeling complex and heterogenous tumor phenotypes.

Importantly, the models demonstrated consistent reliability across all growth classes, even in the face of class imbalance. As shown in Fig. 6, the classifiers avoided overfitting to the majority class and achieved balanced performance across all groups, highlighting their ability to generalize to underrepresented classes. There was no clear trend linking the occasional misclassification observed in Fig. 5, Fig. 13 with weaknesses in encoder-based feature learning. The one notable exception involved a meningioma near the skull (Fig. 13F) that was largely removed during skull stripping, resulting in poor feature extraction and misclassification from ‘shrinking’ to ‘stable’. This case highlights a known challenge of automated skull stripping in tumors adjacent to the skull which indicates the need for further refinement in preprocessing steps. One might question why we developed a custom feature extractor rather than adopting established public segmentation models such as nnU-Net (Kang et al., 2023) for a similar task. While models like U-Nets excel at pixel-level segmentation by preserving spatial detail through skip connections, they produce weaker bottlenecks (Farajzadeh et al., 2023, Farajzadeh et al., 2022), which are less suited for downstream tasks that depend on stronger abstraction of features (Aqeel, 2021). In contrast, our FCN design encourages a richer latent space representation through its encoder block, making it more appropriate for a focused feature extraction task rather than segmentation accuracy.

The peritumoral ablation analysis showed that tissues surrounding the tumor contributed meaningfully to growth risk prediction, whereas inclusion of excessive surrounding information, including whole-brain context, degraded performance. This behavior is consistent with predictions being driven by local tumor and peritumoral information rather than global image characteristics. The optimal margin identified in this study, approximately 200 %, corresponds on average to a radial neighborhood of nearly 5 cm³ around the tumors in our dataset. This aligns with established pathological evidence that indicate meningiomas often extend beyond the visible tumor boundary (Qasem et al., 2025, Guenther et al., 2019, Wang et al., 2013, Huang et al., 2019, Chen et al., 2023). The adjacent dura commonly exhibits a fibrous dural tail with microscopic tumor involvement extending several centimeters from the tumor base (Chen et al., 2023), and residual tumor cells have been identified within standard surgical dural margins of approximately 1.5 cm, with invasion reported up to 3 cm in some convexity meningiomas (Chen et al., 2023). Consistent with these findings, T1W + C MRI typically demonstrates homogeneous tumor enhancement with a dural tail extending 1–3 cm beyond the primary lesion boundary (Chen et al., 2023), while, peritumoral edema may extend variably, most often extending 50–100 % of the tumor radius (approximately 1–3 cm in absolute terms), beyond the tumor margin, in some cases involving distant brain regions (Qasem et al., 2025, Guenther et al., 2019, Wang et al., 2013, Huang et al., 2019, Chen et al., 2023).

The broader relevance of this study lies in the growing recognition that medical scans contain latent information which may not be apparent to the human eye but is accessible to ML algorithms (Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025, Doi, 2007). Even when scans are acquired across different scanners, with different image quality and voxel sizes, or when tumor behavior is influenced by a complex interplay of biological and demographic factors, an appropriate fusion of computer vision and ML can uncover patterns that inform predictions of future growth. Our study showed that the morphological and radiomic features captured within baseline MRI are capable of reflecting the biological potential of a tumor, which can offer valuable insights into its likely progression (Doi, 2007, Goldbrunner et al., 2021, Hanna et al., 2023, Nagai Yamaki et al., 2021, Sadeghzadeh et al., 2025, Sadeghzadeh et al., 2025).

Limitations and future directions: The imbalance in class representation remains a key challenge, which we aim to address in future studies through incorporating larger multi-center datasets and additional augmentation strategies. Skull stripping, while generally effective, can occasionally remove tumors located near the skull, although we mitigated this by excluding affected cases and verifying the source of misclassifications where relevant. This limitation motivates future investigation into skull-aware preprocessing pipelines or segmentation-free feature extraction strategies that are more robust to tumors adjacent to the skull. Future studies may also explore the integration of demographic and clinical variables as additional inputs to assess their complementary impact on growth prediction, which was not examined in the present study due to the current study design. Growth categories were defined based on expert neurosurgical assessment, which may have inevitably introduced a degree of subjectivity. This highlights the potential value of models that predict continuous growth rates rather than discrete categories, which may better capture the nuance of tumor behavior and reduce interobserver variability. While radiomics could be associated with concerns around reproducibility and false positives (Moskowitz et al., 2022), our use of PyRadiomics was solely designed and intended as a comparator to establish a baseline for future work rather than as a definitive solution. In this study, we did not seek to interpret or investigate the biological association of individual features, as our focus was on validating the representational power of the extracted features, as evidenced by consistently high performance across multiple classifiers. In this study, we did not seek to interpret or investigate the biological association of individual features, as our primary focus was on validating the representational power of the extracted features, as evidenced by consistently high performance across multiple classifiers. Future work may build on this foundation to explore the relationship between learned imaging representations and underlying tumor biology, helping to bridge the gap between predictive performance and biological interpretability. Lastly, in the current study we chose to work with manual tumor segmentations performed by expert neurosurgeons to ensure that the RoIs were clinically accurate and anatomically meaningful. This provided a reliable reference for feature extraction through the encoder which resulted in a robust growth risk prediction in the downstream classification stage (as supported by the evaluation results). We acknowledge that this may not be readily available in routine clinical practice and therefore limit immediate clinical deployment. However, if reliable automated segmentation models show surgeon-level accuracy in the future, there will be opportunity for their integration into the pipeline. This limitation also highlights the importance of accurate segmentation model for downstream tasks.

In the absence of prior frameworks for single-scan meningioma growth prediction, this work lays the foundation for future developments. Our findings demonstrated the feasibility of using imaging data alone to predict tumor growth risk, which offers hope for developing computer vision-based tools that could ultimately support more personalized tumor surveillance strategies and better-informed clinical decision-making.

5. Conclusion

This study demonstrates for the first time that meningioma growth risk can be reliably predicted from tumor-level imaging features extracted from a single contrast-enhanced T1-weighted baseline MRI scan, without the need for longitudinal imaging or invasive procedures. By leveraging a meningioma-specific encoder for feature representation learning and combining it with conventional ML classifiers, we achieved consistently high predictive performance across growing, stable, and shrinking tumors in a cohort of 192 patients with single meningiomas. Our results highlight the feasibility of non-invasive and patient-specific growth risk stratification at the time of diagnosis, which complements the conventional MRI monitoring by informing healthcare. Beyond predictive performance, our findings provide evidence that baseline MRI encodes latent morphological and peritumoral signatures within MRI scans contain sufficient information to capture biologically relevant growth dynamics. We showed how the performance was influenced by scanner vendor, magnetic field strength, tumor size, and tumor location, with superior results for 3 T scanners. High accuracy was also observed for small tumors, while medium-sized tumors remained challenging to classify. This supports the use of representation learning as a principled framework for inferring future disease behavior from subtle imaging cues not readily apparent to human observers. Clinically, the ability to estimate growth risk from a single scan has the potential to inform more personalized surveillance strategies, which can enable earlier intervention for high-risk tumors while supporting conservative management for indolent cases. While further validation across larger multi-center cohorts and fully automated pipelines is warranted, this work lays the foundation for a new class of image-driven decision support tools that move meningioma management from reactive monitoring toward earlier and individualized risk-informed care.

CRediT authorship contribution statement

Nima Sadeghzadeh: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Formal analysis, Data curation, Conceptualization. Jason A. Correia: Writing – review & editing, Supervision, Resources, Methodology, Investigation. Jiantao Shen: Writing – review & editing. Sung-Min Jun: Data curation. Poul M.F. Nielsen: Writing – review & editing. Brendan Davis: Data curation. Samantha J. Holdsworth: Writing – review & editing, Supervision, Investigation, Formal analysis. Michael Dragunow: Writing – review & editing, Supervision, Funding acquisition, Resources. Richard L.M. Faull: Funding acquisition, Resources. Hamid Abbasi: Writing – review & editing, Writing – original draft, Validation, Supervision, Resources, Project administration, Investigation, Funding acquisition, Formal analysis, Conceptualization.

Funding

The work was supported by the Centre for Brain Research's Freemasons Neurosurgery Unit at the University of Auckland (No. 3718016), the Health Research Council of New Zealand (HRC 25/220), and the University of Auckland's Doctoral Scholarship.

Declaration of competing interest

The authors declare no competing interests.

Data availability

The data that has been used is confidential.