A scoping review of artificial intelligence applications in meningioma from image analysis to prognostic prediction

Abstract

Background

Meningiomas constitute the most prevalent primary intracranial tumors, accounting for approximately 39% of all central nervous system tumors and representing a substantial neurosurgical challenge.

Objective

This review aims to examine and summarize the current applications of artificial intelligence (AI) technologies throughout the diagnosis and treatment processes of meningiomas.

Methods

A search was conducted in the Web of Science core collection and Scopus and PubMed, databases on November 9, 2025, utilizing a search strategy that incorporated the term “meningioma” along with related AI terminologies in the title. Literature was screened based on pre-defined inclusion and exclusion criteria, resulting in 52 articles being selected for this review.

Results

AI technologies have demonstrated considerable promise and added value in the management of meningiomas. In image analysis, deep learning models have facilitated automatic and highly precise tumor segmentation, significantly outperforming traditional manual methods. Regarding pathological prediction, AI models have successfully non-invasively predicted crucial biomarkers, such as WHO classification and the Ki-67 index, from preoperative MRI scans. In prognostic prediction, AI models have exhibited robust capabilities in forecasting overall survival, progression-free survival, and recurrence risk.

Conclusion

AI technology represents a formidable new instrument for the precise diagnosis and treatment of meningiomas, showing notable potential for clinical translation.

Introduction

Meningioma is the most common primary intracranial tumor, representing a major neurosurgical disease burden. While most are benign (WHO Grade I), 20%–30% are higher-grade (WHO Grade II/III) tumors with aggressive behavior and poor prognosis [1–4]. Current management relies on the WHO classification. Asyptomatic low-grade tumors may be monitored, while symptomatic or high-grade lesions typically require surgical resection, with radiotherapy considered adjunctively. However, this paradigm faces precision medicine challenges. Preoperatively, conventional imaging cannot reliably predict tumor grade, molecular features, or proliferation activity. Prognostication is imprecise, as outcomes vary significantly even within the same grade. Furthermore, personalizing treatment is hindered by difficulties in preoperatively predicting tumor consistency, adhesions, and radiotherapy responses [5, 6].

Artificial intelligence (AI), particularly machine and deep learning, offers transformative potential to address these gaps. For clarity, AI is the overarching field, with machine learning (ML) as a subset that enables computers to learn from data. Deep learning (DL) is a powerful ML technique using multi-layered neural networks. Radiomics is an application domain that involves the high-throughput extraction of quantitative features from medical images, which can then be analyzed by AI/ML models. AI can automatically analyze medical images to perform precise tumor segmentation and extract high-dimensional radiomic features. By integrating this data with clinical information, AI models show promise for non-invasively predicting tumor grade, molecular status, and patient prognosis. This capability to generate “imaging biomarkers” supports more precise diagnosis, outcome forecasting, and personalized treatment planning [7–9].

While previous reviews have often focused on specific AI applications, such as tumor grading or a single treatment modality [10–13], this review provides a comprehensive and systematic overview of AI across the entire meningioma care continuum. Beyond summarizing technical performance, we critically examine key translational challenges. Furthermore, we compare the maturity and readiness of AI applications in segmentation, diagnosis, and prognosis prediction.

Given this context, this review provides a overview of AI applications across the meningioma care continuum. We focus on advancements in image analysis, pathological prediction, prognostic evaluation, and decision support, examining current performance, limitations, and future pathways for clinical integration.

Literature search and inclusion/exclusion methods

To ensure the rigor and thoroughness of this review, we established a clear strategy for literature search, screening, and data extraction, adhering to the preferred reporting project guidelines. Literature Search Strategy: A search was conducted across three principal databases, Web of Science Core Collection, Scopus and PubMed on November 9, 2025. We employed a title-based search strategy to precisely identify pertinent research. The search query was formulated as follows: Meningioma* and ((Machine Learning) OR (artificial intelligence) OR (artificial-intelligence) OR (Intelligence*) OR (deep learning)). This approach was designed to exhaustively capture all original research literature featuring both meningioma and associated AI terms (such as ML, AI, and DL). Literature Screening and Inclusion/Exclusion Criteria: Subsequent to the initial search, all bibliographic entries were imported into the EndNote reference management system, and duplicates were eliminated. Selection proceeded according to predefined inclusion and exclusion criteria: Inclusion criteria: (1) Original research articles; (2) Research pertaining to meningioma; (3) Studies focusing on the development or validation of AI models (including ML and DL) applied at any stage of meningioma imaging analysis, diagnosis, grading, prognosis prediction, or treatment decision-making; (4) Availability of full text. Exclusion criteria: (1) Non-English publications; (2) Editorials, case reports, conference abstracts, and book chapters; (3) Studies not involving the application or validation of AI models; (4) Inaccessible or substandard data. Screening was performed independently by two researchers, who initially assessed titles and abstracts for relevance, then accessed full texts of potentially eligible studies for deeper evaluation. Any disagreements were resolved through discussion between the initial researchers or by arbitration from a third researcher. Ultimately, 52 articles were selected for inclusion in the review. The detailed schematic diagram of the literature screening process is shown in Fig. 1.

Fig. 1.

Schematic diagram of the literature screening process

The core role of AI in meningioma image analysis

AI technologies, especially DL and ML, are fundamentally transforming the paradigm of meningioma image analysis. The primary function of AI is to convert the traditionally time-consuming and labor-intensive process of manual image analysis, which depends heavily on the subjective experience of physicians, into an automated, quantitative, and intelligent tool for precision medicine. This transformation includes the entire spectrum of processes, from the initial detection of the tumor, through precise segmentation and volume measurement, to advanced feature recognition and classification. Consequently, AI provides unprecedented objective evidence, thereby enhancing support for clinical decision-making [14]. Figure 2 shows the annual publications related to AI in meningioma research. Research in this field began in 2018 and has been increasing year by year since then. This shows that the application of AI in meningioma is an area that has gradually gained attention in recent years. However, it is crucial to critically assess the translational gap between promising research and clinical deployment. A significant limitation across many studies, including those cited herein, is the frequent reliance on retrospective, single-institution, and relatively small datasets, along with a general lack of standardization in imaging protocols and tumor annotations. These factors raise concerns about data leakage, overfitting, and, most importantly, poor generalizability to external, real-world patient populations. The current scarcity of rigorous, prospective multi-center external validations represents the central bottleneck hindering the reliable clinical integration of these models, a challenge that must be prioritized in future research to realize the field’s potential.

Fig. 2.

Annual publications related to AI in meningioma research

Accurate tumor segmentation and volume quantification are pivotal for evaluating meningioma growth rates, devising surgical or radiotherapy strategies, and monitoring the efficacy of treatments. Traditional methods of manual delineation and linear measurements, which typically rely on simplistic geometric assumptions (e.g., the ABC/2 method), are characterized by low efficiency, poor repeatability, and a propensity for volume overestimation due to irregular tumor shapes [15]. In contrast, AI models, particularly those employing DL architectures based on CNNs, are capable of extracting complex features directly from image pixels via end-to-end learning processes. Technical reports have demonstrated the feasibility of this approach, with various convolutional neural network (CNN)-based models achieving high segmentation accuracy on retrospective datasets, as measured by metrics like the Dice similarity coefficient (DSC). For instance, Kang et al., in a multicenter study (n = 659), reported a DSC of 0.893 on an external validation set using an nnU-Net model, showcasing potential for generalization [16]. Other studies have reported DSCs ranging from 0.81 to 0.92 on internal cohorts [17–20]. It is important to contextualize these performance figures. While high Dice scores are promising, their clinical interpretation depends on factors like tumor size and location; for example, segmenting small (< 1 cm³) or skull-base meningiomas remains more challenging [16]. Furthermore, the reported “superior” performance is primarily in comparison to manual methods or traditional geometric approximations within controlled research settings. The true test of robustness—consistent performance across diverse, real-world clinical imaging protocols from unseen institutions—has been less extensively validated. Comparative analyses suggest that AI-based volumetric methods are more resilient to tumor shape variation than traditional linear techniques [15]. However, the direct clinical impact of this improved technical accuracy—such as on surgical planning efficiency or long-term patient outcomes—requires prospective evaluation.

Beyond volumetric assessment, AI has been explored for the automated detection of meningiomas and specific imaging features like the dural tail sign. Proof-of-concept studies report high detection sensitivities and classification accuracies on specific test sets [21–24]. These results, while encouraging, are often derived from relatively small, curated datasets. Challenges such as false positives from mimicking structures and the models’ performance on truly independent, prospectively collected screening data are not yet fully characterized, indicating this application remains in a developmental phase rather than being clinically routine.

The integration of segmentation with diagnostic tasks represents an advanced research direction. Several studies have developed end-to-end systems that perform both segmentation and WHO grade prediction. Zhang et al. reported a grading accuracy of 81.52% alongside segmentation performance [25], while Chen et al. and Jun et al. achieved AUCs of 0.918 and 0.770 for grade prediction, respectively, in their validation cohorts [19, 26]. These integrated models represent a significant step towards comprehensive AI-assisted workflows. Crucially, some studies have shown that radiomic features extracted from AI-generated segmentations can be comparable to those from manual outlines, supporting the use of automation for downstream analysis [27]. The development of interpretability methods, like attention mapping, also helps build trust by highlighting the image regions influential for the model’s prediction [26].

In summary, the reviewed literature strongly supports the technical feasibility of AI for automating and enhancing various aspects of meningioma image analysis. Demonstrated capabilities include precise segmentation, automated detection, and non-invasive grade prediction with promising metrics. However, a critical gap remains between technical performance in retrospective studies and proven clinical readiness.​ Most evidence is derived from retrospective, single-center or limited multi-center studies. Key limitations that hinder generalizability and immediate clinical deployment include variability in sample sizes, class imbalance favoring low-grade tumors, and heterogeneity in MRI acquisition protocols across institutions. Therefore, while AI establishes a methodological​ foundation for intelligent preoperative assessment, its transformative impact on clinical neuro-oncology is still emerging. Future work must prioritize large-scale, prospective, multi-center trials with standardized imaging protocols to rigorously assess clinical utility, impact on decision-making, and ultimately, patient outcomes.

AI-driven prediction of meningioma pathological features

Accurate pathological assessment of meningiomas is essential for treatment planning and prognosis. The WHO grading system serves as the fundamental method for the pathological classification of meningiomas, directly influencing the treatment strategy and prognosis for patients. Traditionally, this grading has relied solely on postoperative pathological examinations, which are invasive and time-consuming. AI technology, particularly ML and DL, has the potential to revolutionize this process by extracting quantitative features from routine preoperative MRI that are difficult to discern by the human eye, thus enabling non-invasive and precise preoperative grading.

Classification and grading

The WHO grading of meningiomas is fundamental for assessing biological behavior and clinical prognosis. Preoperative, non-invasive prediction of this grade, particularly distinguishing low-grade (Grade I) from high-grade (Grade II/III) tumors, is a major focus of AI research due to its direct implications for surgical planning and adjuvant therapy. Initial efforts employed machine learning (ML) models on manually defined radiomic features. For example, Hale et al. reported an ANN model achieving an AUC of 0.89 using neuroradiologist-interpreted MRI features [27]. Park et al. later demonstrated that models integrating features from multiparametric MRI (cMRI, ADC, SWI) could improve performance, suggesting the value of multi-sequence data [28]. While these studies established proof-of-concept, their reliance on subjective feature definitions and often modest sample sizes limits generalizability.

With the advent of DL, the focus has shifted towards CNNs, which are capable of autonomously learning hierarchical features from raw images, thereby circumventing the subjectivity and potential information loss associated with manual feature extraction. Zhu et al. developed a deep learning radiomics (DLR) model using conventional enhanced T1-weighted MRI. This model achieved an AUC of 0.811 in an independent validation cohort, distinguishing itself in meningioma-level MRI classification and significantly outperforming handcrafted feature-based radiomics models [29]. Banzato et al. provided further insights by comparing the Inception-V3 and AlexNet models, pre-trained on ADC maps and enhanced T1-weighted images. Their findings indicated that the Inception-V3 model, which was based on ADC maps, achieved the best diagnostic accuracy. However, both models performed poorly on enhanced T1-weighted images, highlighting the importance of model architecture and the selection of image sequences. Additionally, the tissue cell density information reflected by the ADC value proved to be crucial for grading [30]. In recent years, research has shifted towards adopting more advanced network architectures and focusing on model interpretability. Jun et al. constructed an interpretable multiparametric DL model that combined T1-enhanced and T2-weighted images. This model not only achieved fully automatic tumor segmentation (Dice coefficient = 0.910) but also performed tumor grading. In external validation, its grading AUC was 0.770, which surpassed the performance of human physicians. Through correlation-weighted class activation mapping (RCAM), it was discovered that the model primarily focused on the surface region of meningiomas, indicating its recognition of the features at tumor margins for grading purposes, thereby significantly enhancing clinicians’ trust in the model’s decision-making process [26]. Saadh et al. conducted a large-scale multicenter study that explored a fusion strategy combining handcrafted radiomic features with DL features extracted using 3D autoencoders. On T1-enhanced images, this hybrid model achieved state-of-the-art performance. Reproducibility analysis revealed that 127 out of 215 features exhibited high reliability (intraclass correlation coefficient, ICC > 0.75) [31]. These robust data suggest that AI models, particularly DL models, exhibit nearly perfect and robust discriminative capabilities in the preoperative grading of meningiomas. Despite these promising results, several challenges remain. Primarily, the generalization capability of these models is a critical issue. Many models exhibit strong performance in internal validation but experience significant performance declines during external multicenter validation [32]. This decline is largely attributable to the variability in MRI scanning protocols, equipment, and parameters across different medical institutions. Implementing batch effect correction algorithms and conducting rigorous external validation are essential strategies to address this issue. Secondly, data imbalance presents a frequent challenge, especially with a limited sample size for high-grade meningiomas. Techniques such as the Synthetic Minority Oversampling Technique (SMOTE) serve as effective countermeasures [33]. In summary, AI models, particularly DL-based approaches, have demonstrated compelling technical performance in the preclinical, retrospective task of non-invasive meningioma grade prediction. Current trends involve leveraging multiparametric data, developing interpretable systems, and conducting more rigorous external validations. However, the clinical readiness of these tools remains limited.​ Key challenges include the persistent variability in MRI protocols across institutions, the common issue of class imbalance in training data, and the scarcity of prospective trials measuring impact on clinical decision-making or patient outcomes.​ Therefore, while AI offers a powerful investigative tool​ with the potential to inform preoperative planning, its role in routine clinical decision-support for precision medicine is not yet established.​ Translation will require standardized imaging protocols, large-scale prospective validation, and explicit demonstration of clinical utility beyond retrospective diagnostic accuracy.

Biomarkers

Aside from the WHO classification, key molecular pathological biomarkers for meningiomas, such as the Ki-67 proliferation index and progesterone receptor (PR) expression status, play a crucial role in assessing tumor proliferative activity, invasiveness, recurrence risk, and sensitivity to endocrine therapy. Advances in AI technology show potential for the non-invasive preoperative prediction of these critical biomarkers, representing a promising research direction for detailed molecular subtyping and personalized treatment in meningioma cases. The Ki-67 index serves as the gold standard for evaluating cell proliferation activity, with its levels being closely associated with the growth rate and recurrence risk of meningiomas. Numerous studies have explored the feasibility of using AI to predict the Ki-67 index. Specific studies have adopted diverse methodologies. One ML model that utilized radiomic features from multiparametric MRI, combined with LASSO regression and support vector machine (SVM) algorithms, achieved AUCs of 0.84 and 0.83 in the discovery and independent validation cohorts, respectively. This model showed consistent performance in both skull base and non-skull base tumors within the study cohorts [34]. The ResNet50 model, refined through transfer learning, not only attained an excellent AUC of 0.905 in the internal test set but also simultaneously predicted Ki-67, H3K27me3, and PR statuses, thereby showcasing the potential of a single model for multiple predictive tasks [32]. To improve model performance and potential utility, researchers have explored fusion models. Duan et al. designed a nomogram that integrates clinical features, radiomics scores, and deep transfer learning (DTL) scores. Although this model’s predicted AUC of 0.779 for Ki-67 was not statistically significantly distinct from its sub-models, it demonstrated a greater net clinical benefit across a broad range of threshold probabilities in its decision curve analysis (DCA), thus proving its utility as a comprehensive assessment tool [35]. Zhao et al.’s research systematically evaluated the worth of clinical imaging features versus radiomic features, identifying intratumoral necrosis and maximum tumor diameter as independent predictors of high Ki-67 expression. A clinical-radiomics model that incorporated these features achieved an AUC of 0.837 in internal testing [36]. PR expression status is another pivotal biomarker; negative PR expression often indicates more aggressive tumors. Gao et al. specifically investigated the prediction of PR expression, extracting deep transfer learning features from MRI using a finely tuned ResNet50 model. Upon comparing classifiers such as logistic regression (LR), SVM, and Naive Bayes, they discovered that the SVM model exhibited the most robust predictive performance, and DCA confirmed its high clinical applicability [37]. Lin et al. conducted a comprehensive multicenter study in which radiomic features were extracted from T2-weighted and contrast-enhanced T1-weighted images. They integrated these features with clinical predictive factors, such as tumor location and enhancement patterns, to construct a joint model utilizing the Extreme Gradient Boosting (XGBoost) algorithm. This model demonstrated robust performance in both the training set and multiple external validation sets, achieving AUC values of 0.907, 0.827, 0.846, and 0.807, respectively. Moreover, the model showed an association with recurrence-free survival in the analysis, suggesting its potential prognostic value [38]. The capabilities of AI extend beyond the prediction of single markers. Chen et al. developed a DL model that is capable of simultaneously predicting the WHO classification, Ki-67 index, H3K27me3, and PR status. This model demonstrated high predictive accuracy for all categories on the internal test set, illustrating a multi-task learning approach for preoperative assessment [32]. An attention-based DL network was also developed to automatically segment tumors and simultaneously predict the WHO grade and Ki-67 expression. In predicting Ki-67 expression, the 2.5D Swin Transformer architecture of the network achieved an AUC of 0.89 on the external test set, significantly surpassing the performance of models based solely on clinical and radiomics data. This approach demonstrates the technical capability of end-to-end DL architectures to handle multiple prediction tasks [39]. Beyond imaging data, the integration of AI with metabolomics offers a novel approach to understanding the biology of meningiomas. An untargeted metabolomics and lipidomics analysis was conducted on 85 meningioma samples of varying grades using liquid chromatography-high resolution mass spectrometry (LC-HRMS). Subsequently, ML techniques were employed for feature selection, successfully identifying three potential biomarkers that distinguish between low-grade and high-grade meningiomas [40]. Halder et al. expanded on this approach by combining Raman spectroscopy, ATR-FTIR, and LC-MS/MS for metabolomics analysis of patient serum. They utilized ML to identify metabolites, such as glycochenodeoxycholic acid and indole-3-acetic acid, achieving an accuracy of approximately 90% in differentiating meningiomas from healthy controls. This research lays the groundwork for developing fluid-based, minimally invasive diagnostic and prognostic tools [41]. AI-driven biomarker prediction is an active area of research in meningioma molecular pathology. Whether through precise predictions of Ki-67 index and PR expression status, the combined analysis of multiple markers, or the exploration of metabolomics features, AI models have shown promising performance in retrospective studies, indicating potential for future clinical translation. These technologies aim to facilitate non-invasive molecular subtyping and the identification of high-risk patients. If successfully validated and integrated, they could contribute to a more precise molecular pathological assessment in the future.

Consistency

Intraoperative consistency of meningioma is a crucial factor that influences the surgical strategy and the extent of resection. Soft tumors are more amenable to aspiration and removal, whereas firm or fibrotic tumors adhere more tightly to brain tissue. This adhesion increases the difficulty of achieving complete resection and elevates the risk of neurovascular damage. Consequently, the preoperative, non-invasive, and accurate prediction of meningioma consistency holds substantial clinical value. It guides neurosurgeons in crafting individualized surgical plans, preparing specific instruments, and estimating the duration of the surgery. Advanced AI technologies, particularly radiomics and DL, are being actively explored for this application. Initial research primarily utilized traditional ML algorithms to analyze radiomic features extracted from multiparametric MRI scans. Zhai et al. were pioneers in creating a predictive model based on radiomics nomograms. They extracted a comprehensive set of features from multiparametric MRI, employed variance selection and LASSO regression for dimensionality reduction, and ultimately utilized a logistic regression classifier to develop the model. This model reported AUCs of 0.861 and 0.960 on the training and internal validation datasets, respectively, suggesting an association between radiomic features and tumor texture in that study cohort [42]. With further technological advancements, the focus of research has shifted towards integrating radiomics and DL features to enhance model performance. The study conducted by Zhang et al. exemplifies this trend. Their approach involved not only extracting hand-designed radiomic features but also learning deep features automatically using CNNs and then combining these features. The integrated model, established using a logistic regression classifier, showed improved performance on the validation set compared to single-feature models, with precision, recall, and F1 score all above 0.85 in that study [43]. A multi-center study led by Lin et al. corroborated similar findings. They compared a model based solely on radiomics (Rad_Model), a model using purely DL (DL_Model), and a model that integrated both methodologies (DLR_Model). The results consistently demonstrated that the DLR_Model, which combined both feature types, performed best. It achieved an AUC of 0.854 in the external test cohort, which was higher than the performance of models using either feature type alone in that analysis [44]. Beyond leveraging image data alone, research has explored the integration of images with other intraoperative modal data to define tumor consistency. Cepeda et al. conducted an innovative pilot study in which tumor elasticity, measured by intraoperative ultrasound elastography (IOUS-E), served as the reference standard to classify tumors as “hard” or “soft.” Subsequently, they utilized an ML classifier to analyze radiomic features from preoperative MRI to predict this elasticity standard. In a pilot study with a small sample size (n = 18), the model reported a classification accuracy of 94%. This approach provides a novel pathway for the future multimodal definition and prediction of tumor texture [45]. The predictive value of AI in determining meningioma consistency extends beyond surgical applications, including early warnings of treatment-related complications. Li et al.’s study concentrated on peritumoral edema, a frequent complication of Gamma Knife radiosurgery (GKS). They developed a predictive model using ML that incorporated clinical, semantic, and ADC map radiomic features, which achieved a C-index of 0.780 during external validation. This research underscores the intrinsic connection between the radiomic features of tumors and their biological behaviors, such as their responses to treatments, thereby broadening the scope of clinical predictions from preoperative planning to prognostic assessments [46]. The utility of AI models extends beyond the prediction of single indicators. Gui et al. devised a model that integrates radiomics and DL features to diagnose preoperative venous sinus invasion in meningiomas. Although primarily focused on sinus invasion, the model also reflects the often tough, fibrotic texture of invasive tumors. This integrated model demonstrated AUCs exceeding 0.76 across various validation sets, illustrating the efficacy of multi-task learning in this context [47]. Enhanced predictive performance is not solely dependent on the algorithms used but also on precise identification of critical regions. The significance of the peritumoral area has thus garnered considerable attention. A DL-based radiomics model that analyzes both the tumor and its surrounding peritumoral tissue has shown that incorporating the peritumoral region significantly improves performance over models that consider only the tumor itself. This suggests that peritumoral tissue harbors essential information for predicting the biological behavior of tumors [48]. There has been notable advancement in AI-driven prediction of meningioma texture, evolving from early radiomics-based models to sophisticated algorithms that integrate both radiomics and DL features. These studies report progressively improved predictive accuracy on retrospective datasets, demonstrating the technical feasibility of non-invasive tumor texture differentiation in a research context. If prospectively validated, such tools could potentially aid surgical planning by providing preoperative insights into tumor consistency.

Application of AI in meningioma prognosis and treatment response prediction

The accurate prediction of long-term prognosis and response to specific treatments in meningioma patients is crucial for personalized clinical management, which aims to improve quality of life and extend survival. Research is exploring the use of AI to integrate multimodal data, including clinical, imaging, pathological, and molecular information, with the goal of developing enhanced predictive tools. These tools are increasingly applied in various critical areas, such as predicting survival, assessing recurrence risk, issuing warnings about treatment complications, and enhancing patient quality of life.

Survival and recurrence risk prediction

Predicting OS and PFS is the most critical application of AI in the prognostic assessment of meningiomas. A large-scale study involving 12,197 patients with WHO grade II and III meningiomas, sourced from the National Cancer Database (NCDB), utilized various ML algorithms, including TabPFN and XGBoost, to predict patient mortality at 12, 36, and 60 months. Among the models tested, the one based on the TabPFN algorithm showed the highest performance in that study, with reported AUCs of 0.805, 0.781, and 0.815 for predicting mortality at 12, 36, and 60 months, respectively. The team also integrated these high-performance models into an online web application, which significantly enhanced their clinical translation and application [49]. Song et al. concentrated on predicting atypical meningiomas (WHO grade II), known for their significant prognostic heterogeneity among high-grade meningiomas. They developed a Random Survival Forest (RSF) model that incorporated 16 clinicopathological parameters which showed promising performance in predicting recurrence and mortality risks within their cohort, with C-indexes of 0.808 and 0.889, respectively. The model also reported AUCs of 0.83, 0.82, and 0.86 for predicting 1-year, 3-year, and 5-year PFS, respectively [50]. A multicenter study in Asia conducted by Kim et al. compared the efficacy of the traditional Cox proportional hazards model with that of ML algorithms, specifically the gradient boosting machine and the random survival forest, in predicting PFS in patients with atypical meningiomas. Interestingly, the study revealed that the traditional Cox model (C-index = 0.74) slightly outperformed the more complex ML models. It successfully identified age, postoperative platelet count, performance status, Simpson resection grade, and adjuvant radiotherapy (ART) as significant prognostic factors. Furthermore, the model effectively stratified patients by risk and demonstrated that the benefit of ART for high-risk patients (a 15.9% improvement in 5-year PFS) was significantly greater than for low-risk patients (a 5.6% improvement), thus providing direct evidence supporting the precise application of adjuvant therapy [51]. The study by Nguyen et al. offered crucial clinical insights. Through a long-term follow-up (median follow-up of 10.1 years) of 823 patients with completely resected WHO grade I meningiomas, it was found that, despite complete resection, 6.8% of patients experienced recurrence. Employing ML algorithms such as gradient boosting decision trees and random forests, the researchers identified the Ki-67 proliferation index as the most significant factor for recurrence risk, in addition to follow-up time. Consequently, they recommended a minimum follow-up of 8 years for patients with a Ki-67 index below 8%, and at least 12 years for those with a Ki-67 index of 8% or higher, thereby providing valuable quantitative evidence for the development of clinical follow-up strategies [52]. Applying AI to accurately predict specific subtypes of patients represents an important research direction. By integrating clinical and MRI texture features, ML successfully predicted early progression or recurrence (P/R) of parasagittal and parafalcine (PSPF) meningiomas, achieving a model AUC as high as 0.91. The calculated radiomics score was identified as an independent high-risk factor for P/R, with a hazard ratio (HR) of 15.73 [53].

Preliminary studies suggest AI may have value in predicting treatment-related complications, thereby facilitating early interventions and adjustments to treatment strategies. Karri et al. conducted a pioneering study employing ML algorithms, based on routine demographic and perioperative data, to predict health-related quality of life (HRQoL) outcomes in patients with low-grade gliomas, meningiomas, and acoustic neuromas during a period of 12 to 60 months post-surgery. Their model showed capability in predicting a range of symptoms, such as loss of appetite, constipation, nausea, vomiting, diarrhea, dyspnea, and fatigue. This approach suggests a potential tool for early identification of patients who may require enhanced supportive care, which could aid in resource planning [54]. The future of prognostic prediction models lies in the deep fusion of multimodal data. One approach combines MRI radiomic features with clinical indicators to construct a clinical-radiomics-radiomic (CRR) model, predicting the malignant biological behavior of meningiomas using a combination of the WHO classification and Ki-67 index. Although studies have demonstrated that semantic features, such as heterogeneous enhancement and peritumoral edema, are significantly associated with adverse biological behaviors, the CRR model has not shown a significant performance improvement compared to pure radiomics models. This suggests the limitations of simple feature stacking, and indicates that deeper fusion algorithms may be the future direction [55]. Cutting-edge research even extends to the epigenetic level. Herrgott et al.’s research is dedicated to developing ML models based on DNA methylation features in liquid biopsy specimens, aiming to accurately predict the diagnosis and prognosis of meningiomas [56]. This represents a future trend in prognostic prediction towards ultra-early, non-invasive, and molecular mechanism-level approaches. AI is being investigated for application across various aspects of prognosis and treatment response prediction in meningiomas. From traditional predictions of survival and recurrence risk to early warnings of treatment complications, quality of life assessments, and even innovative explorations based on liquid biopsy and metabolomics, AI research is contributing to the development of more comprehensive prognostic prediction frameworks. Current research indicates that models based on clinical and radiomics characteristics can achieve predictive performance that merits further clinical investigation. Future development will focus more on the deep integration of multimodal data, such as the cross-integration of imaging, pathology, genomics, and metabolomics, higher levels of external validation, and prospective clinical trials to prove actual clinical efficacy. Moreover, the development of highly interpretable and user-friendly clinical decision support systems is crucial. Ultimately, successful translation of these research tools into clinical practice could contribute to more personalized management strategies for meningioma patients.

AI-assisted surgical and treatment decisions

The application of AI technology in the field of meningioma has grown to encompass not only postoperative pathological prediction and prognostic analysis but also the essential elements of preoperative planning and treatment decision support. By analyzing the inherent patterns in multimodal data, AI models can assist neurosurgeons by providing quantitative predictions on surgical resectability, potential risks, short-term postoperative outcomes, and even radiotherapy responses. This research direction aims to establish a basis for more personalized surgical and treatment planning. Accurately predicting the extent of meningioma resection (GTR vs. STR) and its invasiveness (e.g., venous sinus invasion) is vital for formulating surgical strategies, evaluating surgical complexity, and managing expectations of both patients and physicians. Akkurt et al. explored the use of DL to directly predict the resection status of meningiomas, offering a novel approach to preoperative planning [57]. Utilizing preoperative enhanced T1-weighted MRI, the researchers extracted radiomic features and developed a predictive model using various ML algorithms. In a study of 138 patients, the model demonstrated consistent performance between training and an independent test set, reporting an accuracy of 88%. Studies indicate that the intracranial location and morphological characteristics of tumors are crucial for predicting complete resection feasibility, and this model shows particularly high predictive accuracy for cases with substantial residual volume post-surgery [58]. For the more complex prediction of venous sinus invasion, Gui et al. conducted a large-scale multicenter study. They creatively combined manually extracted radiomic features (3948) with over 24,000 deep features automatically derived from multiple DL architectures (VGG, ResNet, DenseNet), selecting 21 features most relevant to sinus invasion. Their radiomics-deep learning fusion model (DLR) achieved AUCs of 0.814 and 0.769 in the internal and external validation sets, respectively, showing improved performance over models using either feature type alone in this analysis. This approach shows potential as a tool for preoperative assessment of sinus invasion, a factor relevant to surgical planning [47]. Predicting short-term postoperative outcomes is essential for optimizing clinical workflows, effectively allocating healthcare resources, and ensuring adequate perioperative patient preparation. In patients with skull base meningiomas, a logistic regression algorithm employing elastic net regularization was utilized to predict three key postoperative outcomes: prolonged hospital stay (LOS > 4 days), non-standard discharge (e.g., transfer to a rehabilitation facility instead of home), and high hospitalization costs (> $47,887). The model achieved AUC scores of 0.798 and 0.752 for predicting prolonged LOS and non-standard discharge, respectively; however, it exhibited poor performance in predicting costs. This suggests that clinical factors might have a significantly greater impact on costs than radiological characteristics. This study illustrates the potential use of ML models, such as elastic net, for predicting outcomes related to healthcare resource utilization [59]. Expanding upon this predictive scope, the study by Karabacak et al. utilized data from 7000 patients registered in the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) database. They developed an ML model to predict five short-term adverse outcomes: prolonged LOS, non-home discharge, 30-day readmission, unplanned reoperation, and major complications. The LightGBM algorithm performed best in predicting prolonged LOS and non-home discharge, while the Random Forest outperformed it in predicting readmission, reoperation, and complications. This study also deployed the best-performing model via a web application, a step that facilitates accessibility for further testing [60]. Muhlestein et al. introduced a novel “guided ensemble modeling” approach by merging the most effective models from various algorithms into a single super ensemble model to predict discharge management following meningioma resection. This ensemble model showed improved predictive performance compared to traditional logistic regression in their analysis, identifying tumor size, visits to the emergency room, BMI, the convexity of the tumor, and preoperative motor deficits as key influencing factors [61]. AI has also demonstrated unique value in predicting specific neurofunctional outcomes and radiosurgical treatment responses. Iranmehr et al. concentrated on the most concerning postoperative visual outcomes in patients with suprasellar meningiomas. They developed an ML model to predict postoperative visual recovery with the highest possible sensitivity. The study found that visual complaints, optic nerve atrophy, tuberculum sellae, and olfactory groove involvement were associated with poor preoperative vision. Additionally, extension into the cavernous sinus and intraoperative vascular involvement were significant predictors of poor postoperative visual recovery. Although the study did not conclude that early intracranial optic canal decompression (IOCD) markedly improved visual outcomes, the model provides important evidence for preoperative consultation and surgical timing [62]. In the field of radiosurgery, predicting treatment response and complications is crucial for determining indications. A DL workflow based on Mask R-CNN and DeepMedic can be used to automatically segment and quantify the volume of cerebral edema around meningiomas after GKRS treatment, replacing tedious manual delineation. Its segmentation results show a similarity coefficient of up to 84.7% with the human gold standard Dice, laying the foundation for subsequent large-scale studies on the progression of edema and the development of predictive models [63]. Research into AI for predicting postoperative outcomes in meningioma is an active area, though the current evidence is primarily derived from retrospective studies. AI applications are being investigated across various stages of meningioma management, from preoperative planning to outcome prediction. From accurately predicting the extent of resection and sinus invasion preoperatively, to providing early warnings of complications, functional outcomes, and medical resource consumption during and after surgery, and to predicting responses to radiosurgery. The goal of this research is to develop comprehensive, quantitative decision-support tools for the perioperative period. Current evidence from retrospective studies suggests that ML models utilizing multimodal data can show promising predictive performance compared to traditional methods in specific settings. Future developments will likely focus on prospective, multicenter clinical validation to demonstrate their practical clinical utility, the development of real-time, interactive surgical navigation systems, and ensuring the interpretability of algorithms and fairness of models. This will ensure that different patient subgroups can benefit equitably. Future research is needed to determine if these advances can be successfully translated from the research phase into routine clinical practice to support precision treatment approaches (Fig. 3).

Fig. 3.

Current status of AI applications in meningiomas

Challenges, limitations, and future directions

Although AI has shown substantial promise in the diagnosis and treatment of meningiomas, its integration into routine clinical practice encounters several significant challenges. The foremost obstacle originates from the data used in current research, which predominantly consists of single-center, retrospective datasets. These datasets are typically characterized by limited sample sizes and a class imbalance, primarily due to the rarity of high-grade cases. A critical and often underreported technical concern within this data-centric limitation is the risk of data leakage and irreproducibility. In many studies, improper partitioning of data or the use of the same dataset for both hyperparameter tuning and final performance reporting can lead to optimistically biased, non-generalizable results. Furthermore, a lack of detailed reporting on preprocessing steps, model architectures, and code availability hampers independent verification. While some of the cited multicenter studies have employed rigorous cross-validation or held-out external test sets to mitigate these risks, the field would benefit from broader adherence to standardized reporting guidelines and public code sharing to ensure robust and reproducible findings [64, 65]. Furthermore, a critical translational gap exists in the development of predictive models, particularly for prognostication. The clinical utility of a model depends heavily on its input data and predictive targets. Comparative analysis indicates that models based purely on radiomic features often lack the contextual information provided by clinical data, while models fusing multimodal data generally demonstrate superior performance. More importantly, the clinical value of predicting an already established outcome like WHO grade is different from predicting novel, actionable endpoints such as the risk of post-operative neurological deficits or specific complications, with the latter holding greater potential to directly alter management. It is also crucial to note that predicting outcomes for low-grade (WHO I) meningiomas offers limited clinical utility, whereas accurately forecasting the recurrence or malignant transformation of high-grade (WHO II/III) tumors is paramount; future model development must reflect this hierarchy of clinical need. More critically, the variation in imaging equipment and scanning protocols across different medical institutions introduces considerable heterogeneity, which severely undermines the models’ ability to generalize. Additionally, concerns regarding privacy and security obstruct the sharing of medical data between institutions, thereby hampering the development of large-scale, high-quality datasets. From a technical standpoint, the reliability and acceptability of models are major hurdles. Many sophisticated DL models operate as “black boxes,” with decision-making processes that are opaque and difficult for clinicians to interpret and trust. This opacity limits their use in critical medical decision-making. Bridging the gap to clinical adoption requires addressing this “black box” problem head-on. The development and prioritization of explainable AI techniques are critical for fostering clinical trust. AI methods that are inherently more interpretable or that provide visual explanations are more likely to gain acceptance in hospital settings than completely opaque deep learning systems.​ Moreover, models trained on narrow datasets are susceptible to overfitting, they perform exceptionally well on training data but their performance deteriorates significantly on new, unseen data, demonstrating inadequate generalization capabilities. The clinical implementation of these models represents a crucial test, yet it is laden with challenges. Presently, the majority of research is still in the retrospective validation phase, with a lack of prospective clinical trials to ascertain the actual efficacy of AI tools in enhancing patient outcomes within real clinical settings. The pathway to regulatory approval for these models remains ambiguous. Beyond technical validation, successful integration faces significant practical hurdles. These include navigating evolving regulatory pathways, solving the pervasive challenge of model explainability as noted, and achieving seamless, low-disruption integration into existing neurosurgical workflows and hospital information systems. Furthermore, questions about how to integrate these models seamlessly and minimally into existing hospital information systems and clinical protocols, and how to train physicians to use these tools effectively, represent urgent practical issues that need resolution [66]. Looking forward, addressing these challenges necessitates a multifaceted strategy. Technologically, the adoption of privacy-preserving computing techniques, such as federated learning, offers potential for facilitating collaborative training across multiple centers while safeguarding data privacy, thereby enhancing model robustness. The development of explainable AI, which would make the decision-making processes of models transparent and comprehensible to physicians, is critical for fostering clinical trust. In terms of research directions, the immediate priority should be to undertake large-scale, multi-center prospective validation studies. Over the long term, the exploration of integrating multi-omics data, such as imaging genomics that combines macroscopic imaging with microscopic molecular data, shows promise for creating more comprehensive predictive systems. Ultimately, translating AI from research to clinical practice requires building a complete pathway encompassing robust technical validation, demonstration of clear clinical utility, regulatory compliance, and practical workflow integration. This demands close collaboration between engineers, clinicians, hospital administrators, and regulators to transform advanced technology into reliable tools that improve patient care. This advancement could significantly propel the field of meningioma diagnosis and treatment toward an era of truly precision medicine. It is important to note the limitations of this review. As a scoping review, its primary aim is to map the current landscape of AI applications in meningioma, identifying key concepts, challenges, and research gaps, rather than to provide a conclusive evaluation of comparative effectiveness. The literature search, while structured, was not exhaustive as required for a systematic review, and a formal quality assessment of included studies was not performed. Therefore, the findings should be interpreted as a descriptive synthesis that highlights trends and priorities to guide future, more definitive research.

Author contributions

Nanjian Xu and Weihu Ma and Weixin Dong and BinBin Yin wrote the main manuscript text.All authors reviewed the manuscript.

Funding

This study is supported by National Key Research and Development Program of China (2023YFC3604401) and Ningbo Clinical Research Center for Orthopedics, Sports Medicine & Rehabilitation (2024L004).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.