Artificial Intelligence and Meningioma: How Computers Are Learning to Help Diagnose and Predict Brain Tumors

Table of Contents

• Key Points
• Why This Research Matters
• What Is a Meningioma?
• Key AI Terms Explained Simply
• How the Researchers Conducted This Review
• AI in Meningioma Image Analysis
• AI Predicting Tumor Pathology
• Predicting the WHO Grade Without Surgery
• Predicting Molecular Biomarkers: Ki-67 and PR
• AI in Prognostic Prediction
• Major Limitations and Gaps
• What This Means for Patients
• Practical Recommendations
• Frequently Asked Questions
• Source Information

Key Points

• AI models can outline meningiomas on MRI with high accuracy, but most evidence comes from retrospective, single-center studies with inconsistent imaging protocols.
• AI can predict tumor grade and molecular markers like Ki-67 and PR from scans, yet these tools are not validated for routine clinical use.
• The central bottleneck is the lack of large, prospective, multi-center validation showing impact on clinical decision-making.
• Patients should ask whether AI tools are used at their center as part of a research protocol or established care.

Why This Research Matters

Meningiomas are the most common primary intracranial tumor — meaning they start inside the skull rather than spreading from elsewhere in the body. They account for approximately 39% of all central nervous system tumors, which is roughly 4 in every 10 brain and spinal cord tumors. That makes them a substantial burden for neurosurgeons worldwide.

Most meningiomas are benign and classified as WHO Grade I (World Health Organization Grade I, the least aggressive category). But this is not true for everyone. Between 20% and 30% of meningiomas — about 1 in 5 to 1 in 3 — are higher-grade tumors (WHO Grade II or III) that behave aggressively and carry a poorer prognosis.

Current treatment decisions rest heavily on the WHO classification. Doctors may simply monitor asymptomatic low-grade tumors, while symptomatic or high-grade tumors usually require surgical removal, with radiation therapy considered as an add-on treatment.

The problem is that this system has real limits. Before surgery, standard imaging cannot reliably tell doctors the tumor's grade, its molecular features, or how fast its cells are dividing. Prognosis is also imprecise — two patients with the same tumor grade can have very different outcomes. Personalizing treatment is further complicated because doctors cannot easily predict in advance how firm the tumor is, whether it is stuck to nearby tissue, or how well it will respond to radiation.

This is where artificial intelligence comes in. The authors argue that AI could help fill these gaps by extracting information from routine scans that the human eye cannot detect.

What Is a Meningioma?

A meningioma is a tumor that grows from the meninges — the protective layers of tissue that surround the brain and spinal cord. Because of where they grow, even "benign" meningiomas can cause serious symptoms by pressing on brain tissue, nerves, or blood vessels.

The WHO grading system sorts meningiomas into three grades based on how the cells look under a microscope and how aggressively they behave. Grade I tumors are the slowest growing. Grade II (sometimes called atypical) and Grade III (anaplastic, meaning highly abnormal and fast-growing) tumors are more likely to return after treatment.

Grade matters enormously. It shapes whether a patient needs surgery right away, how much tissue the surgeon tries to remove, and whether radiation therapy is added afterward.

Key AI Terms Explained Simply

Before diving into the findings, it helps to understand the vocabulary. The authors define these terms carefully, and they are worth knowing because they appear throughout the research.

• Artificial intelligence (AI): the broad field of making computers perform tasks that normally require human intelligence.
• Machine learning (ML): a subset of AI in which computers learn patterns from data rather than being explicitly programmed.
• Deep learning (DL): a powerful type of machine learning that uses multi-layered neural networks (computer systems loosely inspired by the brain) to learn complex patterns.
• Radiomics: the high-throughput extraction of quantitative features — hundreds of measurable numerical characteristics — from medical images, which AI models then analyze.
• Imaging biomarkers: measurable features from scans that reveal biological information about a tumor, such as its grade or molecular status.

Putting these together, AI can automatically analyze medical images, precisely outline tumors, and pull out high-dimensional radiomic features. When combined with patient clinical information, these models show promise for non-invasively predicting tumor grade, molecular status, and prognosis — all without a single incision.

The authors also note how their review differs from earlier ones. Previous reviews often focused on narrow topics, such as tumor grading alone or a single treatment modality. This review instead covers the entire meningioma care continuum and, importantly, compares how mature and "ready for prime time" AI applications are in different tasks.

How the Researchers Conducted This Review

This is a scoping review, a type of study designed to map out all the available evidence on a topic rather than answer one narrow question. The authors followed the Preferred Reporting Items for Systematic Reviews guideline (a widely used checklist that keeps reviews transparent and rigorous) for their search, screening, and data extraction.

The team searched three major databases: Web of Science Core Collection, Scopus, and PubMed. The search was run on November 9, 2025. To keep results precise, they used a title-based search strategy, looking for the term "meningioma" combined with AI-related terms such as machine learning, artificial intelligence, deep learning, and intelligence.

All references were imported into EndNote reference management software, and duplicates were removed. Studies were then screened against pre-set rules.

Studies were included if they:

1. Were original research articles
2. Concerned meningioma
3. Focused on developing or validating AI models (including machine and deep learning) applied at any stage of meningioma imaging analysis, diagnosis, grading, prognosis prediction, or treatment decision-making
4. Had full text available

Studies were excluded if they:

1. Were not published in English
2. Were editorials, case reports, conference abstracts, or book chapters
3. Did not involve the application or validation of AI models
4. Had inaccessible or substandard data

Two researchers screened the studies independently. They first judged titles and abstracts for relevance, then read the full text of potentially eligible papers. Any disagreements were settled by discussion between the two researchers or by a third researcher acting as arbitrator.

In the end, 52 articles met all criteria and were included in the review.

AI in Meningioma Image Analysis

AI is fundamentally changing how doctors analyze meningioma images. The core idea is to replace a time-consuming, manual process — one that depends heavily on the subjective experience of each physician — with an automated, quantitative, and intelligent tool.

This covers the entire workflow: finding the tumor, outlining it precisely, measuring its volume, and recognizing and classifying its features.

One striking finding: research in this field only began in 2018, and the number of publications has increased every year since. This tells us that applying AI to meningioma is a relatively young but rapidly growing area of interest.

Why segmentation matters

Segmentation means drawing the exact boundary of a tumor on an image. Accurate segmentation and volume measurement are pivotal for tracking how fast a meningioma grows, planning surgery or radiation, and monitoring whether treatment is working.

Traditional methods rely on manual outlining and simple linear measurements, often using geometric assumptions such as the ABC/2 method (a formula that estimates volume from three diameter measurements). These old methods are slow, poorly repeatable, and tend to overestimate volume because meningiomas are often irregularly shaped.

AI models, especially deep learning architectures based on convolutional neural networks (CNNs) — a type of network especially good at processing images — learn complex features directly from image pixels in an end-to-end process.

How well do they work? Performance is typically measured with the Dice similarity coefficient (DSC), a score from 0 to 1 that compares the computer's outline with the correct outline (1.0 means perfect overlap).

• Kang and colleagues ran a multicenter study of 659 patients and reported a DSC of 0.893 on an external validation set using an nnU-Net model — showing real potential for generalization.
• Other studies reported DSCs ranging from 0.81 to 0.92 on internal patient groups.

These numbers deserve context, however. High Dice scores are promising, but their clinical meaning depends on tumor size and location. Segmenting small tumors (less than 1 cm³) or those at the skull base remains more challenging. Also, the reported "superior" performance is mainly measured against manual methods or traditional geometric approximations inside controlled research settings — not against the messy reality of diverse hospitals.

That said, comparative analyses suggest AI-based volumetric methods are more resilient to tumor shape variation than traditional linear techniques. But whether this improved technical accuracy actually improves surgical planning efficiency or long-term patient outcomes still needs prospective evaluation (studies that follow patients forward in time).

Detection and feature recognition

AI has also been explored for automatically detecting meningiomas and specific imaging features, such as the dural tail sign (a tapering enhancement along the brain lining often seen with meningiomas).

Proof-of-concept studies report high detection sensitivities and classification accuracies on specific test sets. However, these results often come from relatively small, curated datasets. Two problems are not yet fully characterized: false positives caused by structures that mimic tumors, and how the models perform on truly independent, prospectively collected screening data. The authors conclude this application remains in a developmental phase, not a clinical routine.

Combining segmentation with diagnosis

Some of the most advanced work combines segmentation with diagnostic tasks in a single system. These end-to-end systems both outline the tumor and predict its WHO grade.

• Zhang and colleagues reported a grading accuracy of 81.52% alongside their segmentation performance.
• Chen and colleagues achieved an AUC of 0.918 for grade prediction in their validation group.
• Jun and colleagues achieved an AUC of 0.770 in their validation group.

(The AUC, or area under the curve, is a score from 0.5 to 1.0, where 0.5 means the model is guessing and 1.0 means perfect discrimination.)

Notably, some studies show that radiomic features extracted from AI-generated segmentations are comparable to those from manual outlines. That supports using automation for downstream analysis. Interpretability tools, such as attention mapping — which highlights which parts of an image influenced the model's decision — also help build trust.

The bottom line on image analysis: the technical feasibility is strongly supported. But most evidence comes from retrospective, single-center or limited multi-center studies, with variability in sample sizes, class imbalance (too few high-grade tumors) and inconsistent MRI protocols across institutions.

AI Predicting Tumor Pathology

Accurate pathological assessment is essential for treatment planning and prognosis. The WHO grading system is the foundation, and it directly shapes treatment strategy. Traditionally, grading has relied solely on examining tissue removed during surgery — an invasive and time-consuming process.

AI offers a different path: extracting quantitative features from routine preoperative MRI that the human eye cannot discern, enabling non-invasive, precise prediction of grade before any surgery happens.

Predicting the WHO Grade Without Surgery

Distinguishing low-grade (Grade I) from high-grade (Grade II/III) tumors before surgery is a major focus of AI research, because this directly affects surgical planning and whether added therapy like radiation is needed.

The earliest efforts used machine learning models built on manually defined radiomic features:

• Hale and colleagues reported an artificial neural network (ANN) model achieving an AUC of 0.89 using MRI features interpreted by neuroradiologists.
• Park and colleagues showed that combining features from multiparametric MRI — conventional MRI (cMRI), apparent diffusion coefficient (ADC) maps, and susceptibility-weighted imaging (SWI) — could improve performance, highlighting the value of multi-sequence data.

These studies proved the concept, but their reliance on subjective feature definitions and often modest sample sizes limits how well they apply to new patients.

With deep learning, the field shifted toward CNNs, which learn features autonomously from raw images and avoid the subjectivity and information loss of manual feature extraction.

• Zhu and colleagues developed a deep learning radiomics (DLR) model using conventional contrast-enhanced T1-weighted MRI. It achieved an AUC of 0.811 in an independent validation group — significantly outperforming handcrafted-feature radiomics models.
• Banzato and colleagues compared the Inception-V3 and AlexNet models, pre-trained on ADC maps and enhanced T1-weighted images. Inception-V3, based on ADC maps, achieved the best diagnostic accuracy. Both models performed poorly on enhanced T1-weighted images, showing that model architecture and image sequence choice matter. The ADC value, which reflects tissue cell density, proved crucial for grading.
• Jun and colleagues built an interpretable multiparametric deep learning model combining T1-enhanced and T2-weighted images. It achieved fully automatic tumor segmentation (Dice coefficient 0.910) and tumor grading with an external validation AUC of 0.770 — surpassing human physicians. Using correlation-weighted class activation mapping (RCAM), they found the model focused mainly on the surface region of meningiomas, suggesting it uses features at tumor margins to grade tumors. This insight strengthened clinician trust in the model's reasoning.
• Saadh and colleagues conducted a large multicenter study using a fusion strategy that combined handcrafted radiomic features with deep learning features from 3D autoencoders. On T1-enhanced images, this hybrid model achieved state-of-the-art performance. A reproducibility analysis showed that 127 out of 215 features had high reliability (intraclass correlation coefficient, ICC, greater than 0.75 — meaning repeated measurements agreed closely).

Despite these encouraging results, serious challenges remain. The key one is generalization: many models perform strongly in internal validation but decline significantly during external multicenter validation. This happens largely because MRI scanning protocols, equipment, and parameters differ across institutions.

Data imbalance is another frequent problem, especially when high-grade meningiomas are rare in a dataset. Techniques such as the Synthetic Minority Oversampling Technique (SMOTE), which generates synthetic examples of the underrepresented group, serve as effective countermeasures.

The authors make their position clear: AI offers a powerful investigative tool that could inform preoperative planning, but its role in routine clinical decision support is not yet established.

Predicting Molecular Biomarkers: Ki-67 and PR

Beyond grade, certain molecular markers matter for assessing how aggressively a meningioma behaves. Two are especially important.

The Ki-67 proliferation index measures how many tumor cells are actively dividing. It is considered the gold standard for cell proliferation activity, and its level correlates closely with how fast a meningioma grows and how likely it is to return.

Progesterone receptor (PR) expression status matters too. Meningiomas often carry progesterone receptors, and negative PR expression typically indicates a more aggressive tumor and lower sensitivity to hormone-based (endocrine) therapy.

Multiple studies have tested whether AI can predict these markers from scans:

• A machine learning model using radiomic features from multiparametric MRI, combined with LASSO regression (a technique that selects the most useful features) and support vector machine (SVM) algorithms, achieved AUCs of 0.84 and 0.83 in the discovery and independent validation groups. It performed consistently in both skull-base and non-skull-base tumors.
• A ResNet50 model refined through transfer learning achieved an excellent AUC of 0.905 on the internal test set. Remarkably, it predicted Ki-67, H3K27me3 (a protein modification linked to tumor behavior), and PR status simultaneously — showing one model can handle multiple predictive tasks.
• Duan and colleagues designed a nomogram (a visual scoring tool) integrating clinical features, radiomics scores, and deep transfer learning (DTL) scores. Its predicted AUC of 0.779 for Ki-67 was not statistically significantly different from its sub-models, but it delivered greater net clinical benefit across a broad range of threshold probabilities in decision curve analysis (DCA) — proving useful as a comprehensive assessment tool.
• Zhao and colleagues systematically compared clinical imaging features versus radiomic features. They identified intratumoral necrosis (dead tissue inside the tumor) and maximum tumor diameter as independent predictors of high Ki-67 expression. A clinical-radiomics model incorporating these features achieved an AUC of 0.837 in internal testing.
• For PR status, Gao and colleagues extracted deep transfer learning features from MRI using a finely tuned ResNet50 model. Comparing classifiers — logistic regression (LR), SVM, and Naive Bayes — they found the SVM model had the most robust predictive performance, and DCA confirmed its high clinical applicability.
• Lin and colleagues ran a comprehensive multicenter study extracting radiomic features from T2-weighted and contrast-enhanced T1-weighted images. They integrated these with clinical predictors such as tumor location and enhancement patterns, building a joint model using the Extreme Gradient Boosting (XGBoost) algorithm. It performed robustly in the training set and multiple external validation sets, achieving AUC values of 0.907, 0.827, 0.846, and 0.807.

Together, these results point toward detailed molecular subtyping and personalized treatment planning for meningioma patients — all from imaging alone. But the same caveat applies: promising retrospective results do not equal clinical readiness.

AI in Prognostic Prediction

Beyond diagnosing and grading, AI models have been developed to forecast what will happen to patients over time. According to the review's findings, AI models have demonstrated robust capabilities in predicting overall survival (how long patients live after diagnosis), progression-free survival (how long patients survive without the tumor growing or returning), and recurrence risk (the chance the tumor comes back).

This is a meaningful advance because conventional prognosis is imprecise. Outcomes vary significantly even among patients with the same tumor grade. A tool that could better stratify patients could help doctors decide who needs closer monitoring and who might benefit from more aggressive treatment from the start.

The link between biomarkers and prognosis reinforces why biomarker prediction matters. Because the Ki-67 index is closely tied to tumor growth rate and recurrence risk, an AI model that predicts Ki-67 non-invasively is, indirectly, also supplying prognostic information. The same logic applies to PR status: negative receptor expression signals a more aggressive tumor.

The authors emphasize that this potential must be tested rigorously before it changes practice. Retrospective accuracy is one thing; proving that a model's predictions improve real patient outcomes is another entirely.

Major Limitations and Gaps

The authors are notably candid about the gap between research promise and clinical deployment. Their central criticism applies to nearly every study they reviewed.

• Retrospective, single-institution data: Most studies look backward at existing records from one hospital, rather than following patients forward.
• Small datasets: Sample sizes are often modest, raising concerns about overfitting (models memorizing quirks of one dataset rather than learning generalizable patterns).
• Lack of standardization: Imaging protocols and tumor annotations vary widely, creating risks of data leakage (when information the model should not have access to sneaks in and inflates performance) and poor generalizability.
• Performance drop in external validation: Many models that excel internally decline significantly when tested at other hospitals, largely due to differences in MRI scanners, settings, and protocols.
• Class imbalance: High-grade meningiomas are less common, so training datasets are skewed toward low-grade tumors.
• Scarcity of prospective trials: There is a near-total absence of prospective, multi-center external validation showing impact on clinical decision-making or patient outcomes.

The authors call this lack of rigorous, prospective, multi-center external validation "the central bottleneck" hindering reliable clinical integration. Addressing it, they argue, must be the priority for future research.

They also note that even the technical wins need scrutiny. The direct clinical impact of more accurate segmentation — on surgical planning efficiency or long-term outcomes — has not yet been demonstrated.

What This Means for Patients

If you or a loved one has been diagnosed with a meningioma, here is what this review suggests about the current state of AI.

AI is not yet part of routine care. Despite impressive research results, no AI tool described here has been validated well enough to replace or reliably supplement standard clinical decision-making. You should be cautious of any clinic marketing AI-based diagnosis as a proven standard.

AI's near-term value is in research and investigation. AI can generate objective, quantitative measurements that support clinical decisions — as an investigative tool, not an established one.

Several potential benefits are on the horizon:

• Faster, more consistent tumor volume measurements to track growth over time
• Better preoperative estimates of tumor grade, which could refine surgical and radiation planning
• Non-invasive estimates of Ki-67 and PR status, potentially reducing the need for some invasive testing
• Improved prognostic information to guide how closely a tumor is monitored

Interpretability matters for trust. Some models now reveal which image regions drove their predictions — for example, focusing on the tumor's surface regions when grading. That transparency helps clinicians judge whether a model's reasoning is biologically sensible.

Practical Recommendations

Based on this review, here is what the findings suggest for patients, clinicians, and researchers.

1. For patients: Ask whether AI tools are used at your center as part of a research protocol or established care. Standard care remains the WHO classification of surgically removed tissue.
2. For clinicians: Treat AI predictions as supporting information, not definitive answers, until prospective multi-center validation is available.
3. For researchers: Prioritize large-scale, prospective, multi-center trials with standardized imaging protocols.
4. For study design: Apply batch effect correction algorithms to handle differences between institutions, and use techniques such as SMOTE to address class imbalance.
5. For measurement: Demonstrate clinical utility explicitly — not just diagnostic accuracy on retrospective data.

The authors conclude that AI is a formidable new instrument for the precise diagnosis and treatment of meningiomas, with notable potential for clinical translation. Realizing that potential will require closing the gap between promising research and proven real-world benefit.

Frequently Asked Questions

Can AI tell me the grade of my meningioma without surgery?

Research models can predict whether a meningioma is low-grade or high-grade from MRI scans. For example, one deep learning model using contrast-enhanced T1-weighted MRI achieved an AUC of 0.811 in an independent validation group. However, these tools are not yet ready for routine clinical use and require large, prospective, multi-center validation before they can replace surgical biopsy for grading.

What does an AUC of 0.8 mean for a meningioma AI test?

AUC stands for area under the curve, a score from 0.5 to 1.0. A value of 0.5 means the model is guessing, while 1.0 means perfect discrimination. So an AUC of 0.8 means the model correctly distinguishes between two groups, such as low-grade and high-grade tumors, about 80% of the time. Higher values indicate better performance.

Can AI predict if my meningioma will come back?

AI models have shown potential to forecast recurrence risk, overall survival, and progression-free survival from imaging and clinical data. However, these predictions come from retrospective studies and have not been proven to improve real patient outcomes. They are not yet reliable enough for routine use in guiding monitoring or treatment decisions. More rigorous prospective trials are needed before they can change practice.

What is the Ki-67 index and can AI measure it from my scan?

The Ki-67 proliferation index measures how many tumor cells are actively dividing. It is considered the gold standard for cell proliferation activity and correlates with how fast a meningioma grows and how likely it is to return. Several AI models have predicted Ki-67 from MRI scans with AUCs ranging from 0.779 to 0.905 in internal tests, but these are not yet validated for routine clinical use.

Why do AI models for meningioma often fail when tested at other hospitals?

Many models perform well in internal validation but decline significantly during external multicenter validation. This happens largely because MRI scanning protocols, equipment, and parameters differ across institutions. Other issues include small datasets, class imbalance, and lack of standardization. The authors call the lack of rigorous, prospective, multi-center external validation the central bottleneck hindering reliable clinical integration.

How accurate is AI at outlining a meningioma on an MRI?

AI models can automatically outline tumors with high accuracy. In a multicenter study of 659 patients, an nnU-Net model achieved a Dice similarity coefficient of 0.893 on an external validation set. Other studies reported scores from 0.81 to 0.92 on internal patient groups. However, segmenting small tumors or those at the skull base remains more challenging, and clinical benefit is not yet proven.

I was just diagnosed with a meningioma — when should I get a second opinion on my scans and pathology?

A meningioma diagnosis rests on the WHO grade of surgically removed tissue, and grade shapes whether surgery is needed, how much tissue is removed, and whether radiation is added. Standard imaging cannot reliably reveal grade, molecular features, or cell division rate before surgery, and prognosis is imprecise even among patients with the same grade. Because treatment decisions depend so heavily on accurate grading, an independent review of your imaging and pathology before committing to a plan is reasonable. Diagnostic Detectives Network provides independent expert second opinions.

Source Information

Original article title: A scoping review of artificial intelligence applications in meningioma from image analysis to prognostic prediction.

Authors: Xu N, Ma W, Dong W, Yin B.

Publication details: Discover Oncology, 2026; volume 17, article 319. DOI: 10.1007/s12672-026-04476-5. Published open access under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Study type: Scoping review of 52 original research articles identified from Web of Science Core Collection, Scopus, and PubMed as of November 9, 2025.

This patient-friendly article is based on peer-reviewed research. It is intended for educational purposes and does not replace consultation with a qualified healthcare professional.