Advancements in the application of MRI radiomics in meningioma

Abstract

Meningiomas are among the most common intracranial tumors, and challenges still remain in terms of tumor classification, treatment, and management. With the popularization of artificial intelligence technology, radiomics has been further developed and more extensively applied in the study of meningiomas. This objective and quantitative technique has played an important role in the identification, classification, grading, pathology, treatment, and prognosis of meningiomas, although new problems have also emerged. This review examines the application of magnetic resonance imaging (MRI) techniques in meningioma research. A database search was conducted for articles published between November 2017 and April 2025, with a total of 87 studies included after screening. These studies were summarized in detail, and the risk of bias and the certainty of the evidence were assessed using the Quality Assessment of Diagnostic Accuracy Studies version 2 (QUADAS-2) and radiomics quality scores (RQS). All the studies were retrospective, with most being single-center studies. Contrast-enhanced T1-weighted imaging (T1C) and T2-weighted imaging (T2WI) are the most commonly used MRI sequences. Current research focuses on five topics, namely, differentiation, grade and subtypes, molecular pathology, biological behavior, treatment, and complications, with 14, 32, 14, 12, and 19 studies addressing these topics (some of which are multiple topics). Combined imaging features with clinical or pathological features often outperform traditional clinical models. Most studies show a low to moderate risk of bias. Large, prospective, multicenter studies are needed to validate the performance of radiomic models in diverse patient populations before their clinical implementation can be considered.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13014-025-02679-8.

Introduction

Meningiomas are among the most common primary intracranial tumors and account for approximately one-third of central nervous system tumors in adults [1], whereas in children and adolescents, they are much less common (0.4-4.6%) [2]. According to the classification system proposed by the World Health Organization (WHO), meningiomas are divided into three grades. Benign meningiomas (BMs) are classified as WHO Grade I, are typically well defined and slow-growing and present a low recurrence risk—approximately 10% at five years. Atypical meningiomas (AMs) are classified as WHO Grade II; these tumors demonstrate increased mitotic activity, a higher recurrence rate—reaching up to 50% at five years—and a shorter survival time. Anaplastic or malignant meningiomas (MM) are classified as WHO Grade III tumors; these tumors clearly display malignant characteristics, almost inevitably recur, and have a five-year progression-free survival rate of less than 10% [3].

Meningiomas typically occur on the cerebral convexity, parasagittal sinus, or skull base. They are usually solitary and cause symptoms such as seizures and cranial nerve dysfunction due to compression of adjacent tissues [4]. Surgery is currently the primary treatment for meningiomas. Gross total resection (GTR) of the tumor significantly contributes to preventing recurrence and mortality [5–7]. AMs and MMs are more invasive and have higher recurrence rates; thus, various radiotherapy methods are often used to enhance local control, particularly when surgery alone seems insufficient for achieving complete resection.

Magnetic resonance imaging (MRI) is crucial for evaluating meningiomas. T1-weighted imaging (T1WI) provides anatomical details, whereas T2-weighted imaging (T2WI) and T2 fluid-attenuated inversion recovery (T2FLAIR) imaging highlight soft tissue contrasts. In contrast-enhanced MR images, meningiomas typically show enhancement due to dural infiltration or reactive vascularization [8]. They often have unclear boundaries with the dura mater, which helps differentiate them from other nonmeningeal brain lesions. Low signal intensity within the tumor, often due to calcification or vascular flow voids, may also be observed [9]. Clinically, contrast-enhanced T1-weighted imaging (T1C) is crucial for surgical planning. According to previous studies, peritumoral edema is also common in meningiomas and is present in 38-67% of intracranial meningioma patients. Specifically, 32-66% of low-grade meningioma patients and 79-94.4% of high-grade meningioma patients exhibit peritumoral edema [10, 11], which may be closely related to the grade and prognosis of meningiomas [11].

Diagnosing and treating meningiomas is challenging due to their complexity and the need for accurate preoperative classification, surgical planning, postoperative therapy selection, and prognosis assessment. Imaging data are crucial, but traditional methods are limited by machine capabilities, human resolution, researcher subjectivity, and basic calculation methods. Advanced radiomic technology provides practical help in solving these problems, and its application for the diagnosis and treatment of meningioma is consequently very important [12]. However, several questions remain: What is the current application of radiomics in meningiomas? How can radiomics improve the ability to predict meningioma diagnosis and guide treatment decisions? What are the limitations of the current study and the directions of future research? To address these questions, we conducted a review of the relevant studies focusing on these issues.

Methods

Research retrieval and selection

We searched PubMed, Embase, and Web of Science using the terms ‘meningioma’ AND ‘radiomics’ from November 2017 to May 2025. We included original English-language research articles that applied radiomic techniques to meningioma imaging data. Case reports, conference abstracts, comments and review articles were excluded. Following the title/abstract screening, the full texts of potentially relevant studies were evaluated. For this study, the inclusion criteria were as follows: (1) the research focused on meningioma, (2) at least one MRI radiomic feature was utilized for model development, (3) a clearly defined model was presented, and (4) the research had a well-defined objective and outcome. These criteria were designed to ensure the relevance and quality of the included studies, allowing for a comprehensive analysis of MRI radiomics models in meningioma research. Two individuals completed the selection and evaluation process; disputes were negotiated, and consensus was reached through discussion. The data collection and screening process is shown in Fig. 1. After thorough study-by-study screening, 87 original studies on the topic were ultimately included.

Fig. 1.

Flow chart illustrating the literature search and selection process

Risk of bias applicability

The risk of bias in individual studies was assessed by two investigators. We used different tools to assess the risk of bias depending on the characteristics of the studies. Data were extracted from these studies, tabulated, and then examined for bias risk and suitability using the Quality Assessment of Diagnosis Accuracy Studies version 2 (QUADAS-2) tool [13]. The tool covers four sources of bias: (1) patient selection, (2) index test, (3) standard domain, and (4) flow and timing bias. On the basis of the information provided by the study, each bias risk was assessed as high, unclear, or low risk. If all signaling questions in a field are answered “yes,” then the risk of bias is considered “low.” An answer of “no” to any of the signaling questions indicates potential bias. The “unclear” category is used only when the reported data are insufficient to make a judgment. We also included the radiomics quality score (RQS) in this review [14].

Result synthesis

Owing to differences in study design, model algorithms, patient cohorts, assessment strategies, and performance measures, narrative synthesis was selected over meta-analysis. Meta-analysis is not recommended for studies of diagnostic test accuracy that have significant differences in patient cohorts and test settings because it would produce biased results [15].

Results

Study characteristics

The research process of radiomics, which includes image acquisition and preprocessing, region of interest (ROI) delineation, feature extraction and screening, and model construction, is shown in Fig. 2. The 87 studies included are summarized in detail in Table 1. In general, the number of studies applying radiomics to meningioma is increasing annually, and the scope of research is gradually expanding. The sample size of most studies was typically between 100 and 400 patients, and 66.7% (58/87) of the studies had single-center designs. The researchers preferred to use T1C to determine tumor boundaries and T2WI to assist in assessing the nature of the tumor and the presence of peritumoral edema. The combination of the two imaging modalities can simplify the search for key features. In terms of feature extraction and model construction, machine learning algorithms are commonly employed in current research (Fig. 3).

Fig. 2.

Process of radiomic study of meningioma

Table 1.

Summary of the application of radiomics in meningioma

	Author	Year	Number of Patients	Multicenter	Special grade or position	MR Sequences	Radiomics Analysis*	Outcome	Aim and theme
1	Yuan et al.[23]	2025	97	No	Cerebellopontine angle	T1C, T2WI, DWI, ADC	Machine learning	Differentiation	Differentiation
2	Li et al.[80]	2025	442	Yes	/	T1C	Deep learning	Grade	Grade and subtypes
3	Qu et al.[81]	2025	259	No	Sellar	T1C, T1WI, T2WI	Machine learning	Differentiation	Differentiation
4	Zhang et al.[82]	2025	115	No	/	T1C, T1WI, T2WI, T2FLAIR, DWI, ADC	Deep learning	Grade	Grade and subtypes
5	Song et al.[78]	2025	451	Yes	AM	T1C, T1WI, T2WI, T2FLAIR, ADC	Cox regression	Recurrence	Treatment and prognosis
6	Gui et al.[59]	2025	601	Yes	/	T1C, T2WI, DWI	Deep learning	Sinus invasion	Biological behavior
7	Bo et al.[67]	2025	151	No	/	T1C, T1WI, T2WI	Logistic regression	Postoperative cerebral edema	Treatment and prognosis
8	Chen et al.[77]	2025	250	No	High-Grade	T1C, T2WI	Cox regression	Recurrence	Treatment and prognosis
9	Yin et al.[83]	2024	3064	No	/	T1C	Deep learning	Differentiation	Differentiation
10	Lin et al.[51]	2024	739	Yes	/	T1C, T2WI	Machine learning	PR	Molecular pathology
11	Fan et al.[37]	2024	227	No	/	T1C	Logistic regression	Grade	Grade and subtypes
12	Hu et al.[68]	2024	148	No	/	T1C, T1WI, T2WI, ADC	Machine learning	Postoperative cerebral edema and hemorrhage	Treatment and prognosis
13	Karabacak et al.[38]	2024	698	Yes	/	T1C, T1WI, T2WI, T2FLAIR	Machine learning	Grade	Grade and subtypes
14	Liang et al.[25]	2024	1221	Yes	/	T1C, T1WI, T2WI	Deep learning	Differentiation	Differentiation
15	Li et al.[24]	2024	171	Yes	/	T1C, T2WI	Logistic regression	Differentiation	Differentiation
16	Kalasauskas et al.[84]	2024	226	No	/	T1C	Machine learning	Grade, recurrence	Grade and subtypes, Treatment and prognosis
17	Xiong et al.[85]	2024	99	No		T1C, T1WI, T2FLAIR	Machine learning	Differentiation	Differentiation
18	Zeng et al.[79]	2024	164	No	Grade II	T1C, T1WI, T2WI	Machine learning	Recurrence	Treatment and prognosis
19	Kertels et al.[86]	2024	160	Yes	High-Grade	T1C, T1WI, T2WI, T2FLAIR	Machine learning	Grade	Grade and subtypes
20	Yu et al.[55]	2024	469	Yes	AM	T1C, T2WI, T2FLAIR	Logistic regression	Brain invasion	Biological behavior
21	Wang et al.[87]	2024	230	Yes	/	T1C, T1WI, T2WI, T2FLAIR, DWI	Deep learning	Differentiation	Differentiation
22	Azamat et al.[52]	2024	132	No	/	T1C, T1WI, T2WI, T2FLAIR, SWI	Deep learning	NF2, S-100	Molecular pathology
23	Zhang et al.[88]	2024	98	No	Peritumoral edema	T1C, T2WI, T2FLAIR	Deep learning	Grade	Grade and subtypes
24	Yang et al.[89]	2024	326	No	/	T1C	Deep learning	Grade	Grade and subtypes
25	Gui et al.[58]	2024	331	Yes	/	T1C, T2WI, DWI	Logistic regression	Sinus invasion	Biological behavior
26	Duan et al.[90]	2024	318	No	/	T1C	Deep learning	Ki-67	Molecular pathology
27	Ren et al.[18]	2024	224	Yes	AM	T1C, T2FLAIR	Cox regression	Recurrence	Treatment and prognosis
28	Zhang et al.[91]	2024	202	No	/	T2FLAIR	Deep learning	Consistency	Biological behavior
29	He et al.[92]	2024	169	No	/	T1C, T1WI, T2WI	Cox regression	Recurrence	Treatment and prognosis
30	Park et al.[93]	2023	212	Yes	/	T1C, T1WI, T2WI, T2FLAIR	Machine learning	Grade	Grade and subtypes
31	Cai et al.[94]	2023	160	No	/	T1C, T2WI	Machine learning	Grade	Grade and subtypes
32	Jo et al.[95]	2023	162	No	/	T1C, T1WI, T2WI, T2FLAIR	Machine learning	Radiotherapy	Treatment and prognosis
33	Chen et al.[49]	2023	1192	Yes	/	T1C	Deep learning	Grade, Ki-67, H3K27me3, PR	Grade and subtypes, Molecular pathology
34	Han et al.[96]	2023	423	No	/	T1C, T2WI	Machine learning	Subtype	Grade and subtypes
35	Zhao et al.[97]	2023	568	No	/	T1C, T2WI	Logistic regression	Grade	Grade and subtypes
36	Li et al.[98]	2023	313	Yes	/	T1C, T1WI, T2WI	Machine learning	Grade, Ki-67	Grade and subtypes, Molecular pathology
37	Han et al.[35]	2023	544	No	/	T1C, T2WI	Machine learning	Grade	Grade and subtypes
38	Akkurt et al.[48]	2023	117	No	High-Grade	T1C	Machine learning	TERT Promotor Mutations	Molecular pathology
39	Khanna et al.[99]	2023	343	No	/	T1C, T1WI, T2WI, T2FLAIR, DWI, ADC	Machine learning	Ki-67	Molecular pathology
40	Li et al.[73]	2023	595	Yes	/	T1C, T2WI, ADC	Machine learning	Post-GKS complications	Treatment and prognosis
41	Duan et al.[50]	2023	157	No	High-Grade	T1C	Machine learning	PR	Molecular pathology
42	Wang et al.[61]	2023	599	Yes	/	T1C, T2WI	Logistic regression	Sinus invasion	Biological behavior
43	Moon et al.[46]	2023	576	Yes	/	T1C, T1WI, T2WI, T2FLAIR, ADC	Machine learning	Ki-67, p53	Molecular pathology
44	Duan et al.[44]	2023	340	No	/	T1C	Machine learning	Grade	Grade and subtypes
45	Ouyang et al.[100]	2023	280	Yes	/	T1C	Machine learning	Ki-67	Molecular pathology
46	Krähling et al.[101]	2023	167	No	/	T1C	Machine learning	Mitosis cycles	Molecular pathology
47	Sun et al.[47]	2022	135	Yes	/	T1C, T1WI, T2WI	Machine learning	NF2	Molecular pathology
48	Guo et al.[11]	2022	444	No	/	T1C, T1WI, T2WI	Machine learning	Grade	Grade and subtypes
49	Wang et al.[20]	2022	96	Yes	Parasellar	T1C, T1WI, T2WI	Machine learning	Differentiation	Differentiation
50	Park et al.[71]	2022	155	No	Grade II	T1C, T2WI	Logistic regression	Radiotherapy	Treatment and prognosis
51	Musigmann et al.[102]	2022	138	No	/	T1C	Machine learning	Risk of STR	Treatment and prognosis
52	Zhao et al.[45]	2022	371	Yes	/	T1C	Machine learning	Ki-67	Molecular pathology
53	Duan et al.[103]	2022	188	No	/	T1C	Logistic regression	Grade	Grade and subtypes
54	Speckter et al.[74]	2022	93	No	/	T1C, T1WI, T2WI, T2FLAIR, DWI, DTI	Linear regression	Post-GKS complications	Treatment and prognosis
55	Stadlbauer et al.[26]	2022	167	No	/	advMRI, phyMRI	Machine learning	Differentiation	Differentiation
56	Park et al.[104]	2022	318	Yes	/	T1C, T2WI	Deep learning	Grade	Grade and subtypes
57	Hsieh et al.[16]	2022	57	No	Parasagittal and Parafalcine	T1C, T1WI, T2WI, T2FLAIR	Machine learning	Recurrence	Treatment and prognosis
58	Chen et al.[34]	2022	609	Yes	/	T1C, T1WI, T2WI	Deep learning	Differentiation	Differentiation
59	Sun et al.[60]	2022	1048	Yes	/	T1C, T2WI	Deep learning	Sinus invasion	Biological behavior
60	Duan et al.[31]	2022	188	No	/	T1C	Machine learning	Grade	Grade and subtypes
61	Zhang et al.[105]	2022	242	No	/	T1C, T2WI	Logistic regression	Grade	Grade and subtypes
62	Fan et al.[106]	2022	220	No	/	T1C, T2WI	Logistic regression	Differentiation	Differentiation
63	Yang et al.[32]	2022	132	No	/	T1C	Deep learning	Grade	Grade and subtypes
64	Li et al.[107]	2021	284	No	Grade II	T1C, T1WI, T2WI	Logistic regression	Brain invasion	Biological behavior
65	Xu et al.[70]	2021	322	No	/	T1C	Logistic regression	Postoperative seizure	Treatment and prognosis
66	Khanna et al.[43]	2021	306	No	Grade I	T1C, T1WI, T2WI, T2FLAIR, DWI, ADC	Machine learning	Ki-67	Molecular pathology
67	Zhai et al.[63]	2021	172	No	/	T1C, T2WI, T2FLAIR, ADC	Logistic regression	Consistency	Biological behavior
68	Ko et al.[75]	2021	128	No	Grade I	T1C, T2WI	Machine learning	Recurrence	Treatment and prognosis
69	Xiao et al.[69]	2021	136	No	/	T1C, T2WI	Machine learning	Postoperative cerebral edema	Treatment and prognosis
70	Han et al.[108]	2021	131	No	/	T1C, T2WI, T1WI	Machine learning	Grade	Grade and subtypes
71	Zhang et al.[57]	2021	490	No	/	T1C, T2WI	Logistic regression	Bone invasion	Biological behavior
72	Kandemirli et al.[109]	2021	108	No	/	T1C	Machine learning	Brain invasion	Biological behavior
73	Joo et al.[56]	2021	454	No	/	T1C, T2WI	Machine learning	Brain invasion	Biological behavior
74	Zhang et al.[19]	2020	1728	Yes	/	T1C, T2WI	Machine learning	Brain invasion	Biological behavior
75	Zhang et al.[22]	2020	235	No	Anterior skull base	T1C	Machine learning	Differentiation	Differentiation
76	Wei et al.[110]	2020	292	No	/	T1C, T1WI, T2WI	Machine learning	Differentiation	Differentiation
77	Chu et al.[30]	2020	98	No	/	T1C	Logistic regression	Grade	Grade and subtypes
78	Tian et al.[21]	2020	127	No	Sellar/Parasellar	T1C, T2WI	Logistic regression	Differentiation	Differentiation
79	Chen et al.[111]	2019	150	No	/	T1C	Machine learning	Grade	Grade and subtypes
80	Morin et al.[76]	2019	303	Yes	/	T1C, T1WI, T2WI, T2FLAIR, DWI	Machine learning	Grade, survival	Grade and subtypes, Treatment and prognosis
81	Laukamp et al.[112]	2019	71	No	/	T1C, T1WI, T2WI, T2FLAIR, DWI, ADC	Logistic regression	Grade	Grade and subtypes
82	Hamerla et al.[33]	2019	138	Yes	/	T1C, T1WI, T2WI, T2FLAIR, ADC	Machine learning	Grade	Grade and subtypes
83	Zhang et al.[17]	2019	60	No	Skull base	T1C, T2WI, DWI, ADC	Machine learning	Recurrence	Treatment and prognosis
84	Zhu et al.[113]	2019	181	Yes	/	T1C	Deep learning	Grade	Grade and subtypes
85	Niu et al.[114]	2019	241	No	/	T1C	Fisher discriminant	Subtype	Grade and subtypes
86	Park et al.[115]	2018	136	No	/	T1C, ADC, DTI	Machine learning	Grade, subtype	Grade and subtypes
87	Coroller et al.[116]	2017	175	No	/	T1C	Machine learning	Grade	Grade and subtypes

*In studies employing multiple analytical methods, only the modeling approach of the optimal model, as reported in the Results section, is presented

Diffusion tensor imaging (DTI)

Fig. 3.

Statistical analysis of the radiomic study of meningioma. A: Research count by year. B: Research count by number of patients. C: Research count by MRI sequence. D: Number of multicenter or single-center studies. E: Radiomic analysis method statistics

Aim and outcome

Most current research on the application of radiomics to meningioma has focused mainly on the following issues: differentiating meningioma from other intracranial tumors, classification and subtype determination, auxiliary molecular pathological diagnosis, identification of biological behavior, treatment selection and efficacy, and associations with postoperative complications (Fig. 4). More than one-third of these studies (32/87) focused on the preoperative noninvasive grading of meningiomas, which reflects the prevalence of and high concern with this problem in clinical practice. Similarly, the differentiation of meningiomas from other intracranial tumors was a common research topic and was the focus of 14 studies. In terms of molecular pathology, Ki-67 has drawn significant attention, with eight studies applying radiomics to predict Ki-67 levels in meningiomas. Brain invasion is also common in meningiomas and is closely related to treatment, with 5 studies considering it the subject of study. In addition, the prognosis of meningiomas has been a challenge of widespread concern, and accurate prognostic prediction can be of great help to patients. As a result, nine and one current studies used radiomic techniques to predict meningioma recurrence and survival, respectively.

Fig. 4.

Research aims and outcome statistics; a single study can have multiple research aims and outcomes

Risk of bias assessment

Given that radiomics methodologies in the context of meningioma have become relatively well established and consistent, related studies are generally subject to less methodological bias. A total of 14 studies were considered to have significant bias according to the QUADAS-2 assessment (Supplementary Table 1). Temporally, most of these studies were published during the early phase of radiomics development, with methodological bias primarily concentrated in two key areas. First, the sample sizes were relatively small. For example, in the studies by Hsieh et al. [16] and Zhang et al. [17], Only 57 and 60 patients were included, respectively, which inevitably led to bias. Second, some studies included too few MRI sequences, such as only T1C sequences. T1C is generally considered to provide the most information in the preoperative diagnosis and treatment of meningiomas, but other sequences are equally important, such as T2WI, which can be of great help in revealing the nature of meningiomas and peritumoural edema. The mean RQS of the 87 studies was 12.94 (Table 2). All studies were retrospective (Supplementary Table 1). Furthermore, 80.5% of the studies described the image scanning process in detail, and 52.9% used multiple segmentation to screen for stable features. Feature dimensionality reduction was performed in all studies, with 58.6% combining imaging features with other clinical or pathological features to build models and 39.1% comparing imaging models with commonly used clinical models. The use of calibration curves and clinical decision curves was described in only 35.6% and 31.0% of the studies, respectively. Ren et al. [18] and Zhang et al. [19] had a maximum RQS of 19.

Table 2.

Radiomics quality scores

RQS	Rate	Average score
Image protocol quality	80.5%	0.80
Multiple segmentations	52.9%	0.53
Phantom study on all scanners	0.0%	0
Imaging at multiple time points	0.0%	0
Feature reduction or adjustment for multiple testing	100.0%	3
Multivariable analysis with nonradiomic features	58.6%	0.59
Detect and discuss biological correlates	2.3%	0.02
Cutoff analyses	67.8%	0.68
Discrimination statistics	100.0%	2
Calibration statistic	35.6%	0.72
Prospective study registered in a trial database	0.0%	0
Validation	100.0%	2.40
Comparison to “gold standard”	39.1%	0.78
Potential clinical utility	31.0%	0.62
Cost-effectiveness analysis	0.0%	0
Open science and data	79.3%	0.82
Total score (Maximum:36)	/	12.94

Discussion

Differentiating meningiomas from other tumors

Radiomics technology is vital for early disease recognition, particularly in the preoperative assessment of space-occupying lesions in certain areas, such as the skull base. Currently, MR images tend to be employed in radiomic studies for tumor differentiation.

The differential diagnosis of sellar region tumors remains a clinical challenge. Wang et al. applied radiomics to distinguish parasellar cavernous hemangiomas from meningiomas by extracting radiomic features from five MRI sequences and evaluating three feature selection methods and six classifiers. Radiomics models incorporating support vector machine (SVM) and K-nearest neighbors (KNN) classifiers based on T2WI features and ADC images have demonstrated good predictive performance [20]. Tian et al. revealed significant differences in four qualitative MR image features between craniopharyngiomas and meningiomas, and cystic changes in particular were identified as independent predictors of the diagnosis. Image features and texture features from T1WI and T2WI are significantly associated with cystic changes, aiding in the comparative analysis of craniopharyngiomas and meningiomas [21]. In a study focusing on anterior skull base lesions, Zhang et al. attempted to distinguish multiple types of space-occupying lesions using radiomic and machine learning classifiers. The combination of least absolute shrinkage and selection operator (LASSO) regression for feature selection and linear discriminant analysis for the classification algorithm showed the best overall performance in differentiating among pituitary adenomas, meningiomas, craniopharyngiomas, and Rathke’s cysts [22]. Recently, Yuan et al. developed a differential diagnostic model for cerebellopontine angle schwannomas and meningiomas using multimodal MRI and an SVM approach [23]. In addition, studies by Li et al. and Liang et al. demonstrated that radiomic techniques outperform clinical models in the preoperative differentiation of solitary fibrous tumors and angiomatous meningiomas, which may contribute to improved patient outcomes [24, 25].

Stadlbauer et al. investigated whether multiple machine learning algorithms built from high-dimensional radiomic features from advanced MRI (AdvMRI) and physiological MRI (PhyMRI) could reliably classify contrast-enhanced brain tumors. PhyMRI enables the quantitative assessment of microvascular structures, angiogenesis, oxygen metabolism, and tissue hypoxia in tumors. The study included five common types of solid brain cancers: glioblastomas, anaplastic gliomas, meningiomas, primary central nervous system lymphomas, and brain metastases. The classifiers developed in the present study outperformed neuroradiologists in terms of accuracy, precision, and classification error but not in terms of sensitivity or specificity [26]. The QUADAS-2 and RQS scores revealed less research bias in this area. However, as noted by An et al., radiomic and machine learning studies involving small sample sizes may yield unreliable results if they are based on single random splits of the data into training and test datasets. This highlights the need for larger datasets and robust validation methods in radiomic research [27].

WHO classification of meningiomas

According to the Central Brain Tumor Registry of the United States (CBTRUS) 2021 report, high-grade meningiomas (WHO grades II and III) accounted for 19.9% of all newly diagnosed meningiomas from 2014–2018 [28]. High-grade meningiomas typically have 7- to 8-fold higher recurrence rates and poorer prognoses than low-grade meningiomas do [6]. The surgical approaches and postoperative treatments for high- and low-grade meningiomas differ substantially, making the preoperative grading of meningiomas clinically important. Studies have shown that high- and low-grade meningiomas also present differences in conventional MRI features, such as tumor location, shape, and size; areas of necrosis; heterogeneous enhancement; and peritumoral edema volume [8]. Radiomics can be used to amplify these differences, indicating that the evaluation of relevant clinical and radiomic features is essential for grading meningiomas preoperatively [29]. Table 3 shows the model effect of the application of radiomics in the classification of meningioma.

Table 3.

Performance indicators of different radiomic models in grading studies of meningiomas

	Author	AUC*	Sensitivity*	Specificity*	Accuracy*
1	Li et al.[80]	0.87	0.75	0.86	0.80
2	Zhang et al.[82]	0.83	0.50	0.91	/
3	Fan et al.[37]	0.85	/	/	/
4	Karabacak et al.[38]	0.84	/	/	0.73
5	Kalasauskas et al.[84]	0.93	/	/	/
6	Kertels et al.[86]	0.88	/	/	0.90
7	Zhang et al.[88]	/	0.95	0.89	0.93
8	Yang et al.[89]	0.96	0.94	0.87	0.88
9	Park et al.[93]	0.86	0.73	0.73	0.73
10	Cai et al.[94]	0.74	0.72	0.69	/
11	Chen et al.[49]	0.97	0.96	0.96	0.92
12	Zhao et al.[97]	0.91	/	/	/
13	Li et al.[98]	0.86	0.82	0.66	/
14	Han et al.[35]	0.95	0.98	0.79	0.83
15	Duan et al.[44]	0.77	/	/	/
16	Guo et al.[11]	0.88	0.82	0.84	0.83
17	Duan et al.[103]	0.95	0.85	0.87	0.86
18	Park et al.[104]	0.83	0.85	0.70	0.73
19	Duan et al.[31]	0.88	0.85	0.83	0.79
20	Zhang et al.[105]	0.81	0.74	0.79	0.77
21	Yang et al.[32]	0.99	0.90	0.92	0.93
22	Han et al.[108]	0.96	0.87	0.92	/
23	Chu et al.[30]	0.96	0.95	0.92	0.94
24	Chen et al.[111]	0.93	/	/	0.76
25	Morin et al.[76]	0.78	/	/	0.71
26	Laukamp et al.[112]	0.91	0.78	0.89	/
27	Hamerla et al.[33]	0.97	1.00	0.97	/
28	Zhu et al.[113]	0.81	0.77	0.90	/
29	Park et al.[115]	0.86	0.90	0.75	0.94
30	Coroller et al.[116]	0.86	0.82	0.63	/

*All values are based on the optimal radiomic model as reported in the Results section of the corresponding study, prioritizing those obtained from the training cohort

Chu et al., Yang et al., and Duan et al. constructed models for preoperatively grading meningiomas on the basis of T1C image features [30–32]. T2WI, T2FLAIR, DWI, and apparent diffusion coefficient (ADC) images were subsequently gradually incorporated into these models. Although different centers may adopt different approaches, models using radiomic features derived from multiparametric MRI often achieve higher area under the curve (AUC) values and exhibit high sensitivity and specificity in distinguishing low- from high-grade meningiomas [33]. Chen et al. introduced an external validation set al.ongside internal training and validation sets in the development of radiomic-based models. They developed a deep learning-based segmentation method capable of accurately extracting meningioma features from multiparametric MR images, facilitating the deployment of radiomics in clinical practice for preoperative meningioma grading [34]. Nomograms constructed on the basis of constructed radiomic models can more clearly and intuitively display the contributions of relevant features to meningioma grading [35].

Traditional radiomic studies on meningioma grading are often based on features extracted from images of the tumor mass. However, over time, the use of radiomics in these studies was no longer limited to the tumor itself. Previous studies have suggested that the potential pathogenic mechanisms underlying the development of peritumoral edema might differ among meningiomas of different grades because of the secretion of different angiogenic factors, leading to potentially heterogeneous imaging characteristics of edema [10, 36]. Consequently, Guo’s study was the first to separately delineate peritumoral edema as an ROI and incorporate it into a radiomic model. They subsequently demonstrated that radiomic features of the peritumoral edema region could aid in the preoperative grading of WHO Grade I and Grade II tumors [11]. Although earlier studies on this topic were of poor quality with a large potential bias, the quality of recent studies has significantly increased, mainly due to larger sample sizes and more detailed image descriptions [37, 38].

Molecular pathology of meningiomas

Molecular pathology plays a crucial role in managing meningiomas through the analysis of the molecular characteristics of tumor cells, which can help determine the biological behavior, prognosis, and treatment response of the tumor. Conducting molecular pathology studies on meningiomas can assist physicians in more accurately classifying meningioma subtypes, thereby providing more precise treatment strategies and establishing a foundation for personalizing therapy. The latest 2021 WHO Classification of Central Nervous System Tumors (5th edition) combines certain molecular genetic features with a histological diagnosis to establish an integrated diagnostic and stratified reporting system [39]. This new approach defines multiple tumor types and subtypes, reflecting a deeper understanding of the genetic backgrounds and clinical features of these diseases [39]. In recent studies, researchers have used genomic and molecular methods to identify many molecular markers of meningioma recurrence and prognosis, including Ki-67, NF2, AKT1, TRAF7, H3K27me3, and the telomerase reverse transcriptase (TERT) promoter. In addition to classic meningioma driver gene mutations such as biallelic NF2 inactivation, specific molecular features have been identified as key factors for classifying meningioma. For example, comutations in TRAF7 and KLF4 are molecular markers of secretory meningiomas; SMARCE1 mutations are characteristic genetic changes in clear cell meningiomas; and BAP1 mutations are associated with rhabdoid meningiomas [40]. In terms of grading, regardless of the histological features of the tumor, TERT promoter mutations and/or CDKN2A/B homozygous deletions can be used to diagnose WHO grade III meningiomas [41]. In addition to assisting with meningioma classification and grading, molecular pathology indicators have been increasingly shown to be correlated with prognosis; for example, loss of H3K27me3 often indicates a poorer prognosis [42].

Radiomics has been gradually applied to this topic over the past three years, and this application has rapidly increased, yielding high-quality related studies. Ki-67, a classical molecular pathological tumor index, has been widely used in the evaluation of tumor cell proliferation. Khanna et al. first applied radiomic technology to the molecular pathological prediction of meningioma grade [43]. Multiparametric MRI-based radiomic feature analysis can be applied to stratify WHO Grade I meningiomas on the basis of Ki-67 with high accuracy. Subsequent studies have further evaluated the role of radiomics in predicting Ki-67 levels with both traditional and machine learning methods [44, 45]. P53, another classic tumor molecular pathology marker involved in cell cycle regulation, DNA repair, and apoptosis, can constitute part of a noninvasive strategy for evaluating cell proliferation through radiomic technology in meningiomas [46]. The identification of alterations to NF2 is crucial for devising surgical treatment plans and management strategies for meningiomas. Combining radiomics and machine learning has potential clinical utility in preoperatively predicting the NF2 status of meningiomas [47]. The latest classification guidelines emphasize the roles of TERT and H3K27me3 [39]. The use of radiomic technology for predicting TERT promoter mutations and H3K27me3 loss can provide important references for predicting the prognosis of high-grade meningiomas and directly influence treatment planning [48, 49]. In addition, three studies focused on the expression of progesterone receptors (PRs) in meningiomas and developed corresponding radiomic-based predictive models [49–51]. Azamat et al. further utilized susceptibility-weighted imaging (SWI) to predict the expression of the S-100 protein [52]. With advancements in molecular pathology and subsequent targeted therapy research, there is potential to change the existing treatment paradigm for meningiomas. In this process, radiomics undoubtedly plays a crucial role.

Biological behavior of meningiomas

Meningiomas can infiltrate or invade multiple structures, including brain tissue, skull bones, and venous sinuses. Brain infiltration can lead to neurological deficits, headache, and seizures, among other symptoms, and several studies have linked this process to higher tumor grades and poorer prognoses [53, 54]. The fifth edition of the WHO classification guidelines explicitly defines brain infiltration as an independent criterion for classifying meningiomas as high grade. Most preoperative imaging examinations have difficulty visualizing brain infiltration, making radiomics techniques invaluable. Zhang et al. utilized MRI T1C imaging and T2WI to identify 16 features that, alongside sex, were used to develop a clinical radiomic model for predicting the risk of brain infiltration in meningioma patients [19]. To predict brain invasion in patients with AM, Yu et al. developed a radiomic model based on three MRI sequences from 469 patients across three medical centers [55]. Additionally, Joo et al. identified peritumoral edema volume as an independent predictor of brain infiltration, integrating it into predictive models for assessing the risk of this condition [56].

Bone infiltration is also common in meningiomas and can result in skull destruction. Preoperative assessment of bone infiltration is crucial for surgical planning. A radiomic model based on T1C imaging and T2WI effectively predicts tumor bone infiltration [57]. Furthermore, meningiomas at certain locations may invade nearby specialized structures such as the nasal and venous sinuses. Venous sinus invasion can affect cerebral blood circulation, leading to increased intracranial pressure and neurological symptoms. Deep learning models employing radiomic features, particularly those extracted from both tumor and peritumoral regions, have been developed to preoperatively identify nasal and venous sinus invasion in meningioma patients [58–61]. Such models can aid in preoperatively evaluating patients, anticipating potential complications during surgery, and planning specific response strategies.

In addition to external infiltration, heterogeneity within meningioma has also been investigated in the application of radiomics to this disease. By leveraging machine learning methods, two studies separately established radiomic models and validated the corresponding nomograms for predicting meningioma consistency, aiding in personalized surgical planning [62, 63]. Overall, the biological behaviors of meningiomas substantially influence grading and surgical strategy development, but radiomics applications in this field warrant further extensive research.

Treatment options and prognosis prediction of meningiomas

Research on the application of radiomics in the treatment and prognosis of meningiomas can be divided into three categories. The first focuses on adverse reactions following treatment, such as postoperative brain edema and seizures. Postsurgical meningioma patients may experience brain tissue edema, often due to factors such as surgical trauma, local tumor destruction, or vascular damage [64]. Uncontrolled brain edema may evolve into intractable intracranial hypertension, severe irreversible neurological deficits, or even fatal brain herniation [65]. Additionally, postoperative seizures are common among meningioma patients, with approximately 20-50% experiencing at least one seizure episode after surgery. These seizures may be associated with factors such as postoperative brain edema, local inflammatory responses around the tumor, and abnormal neuronal discharge [66]. Seizures can significantly affect patients’ quality of life and may require long-term antiepileptic drug therapy. Therefore, preventing, monitoring, and treating postoperative brain edema and seizures in a timely manner are crucial for facilitating better patient recovery and reducing adverse impacts on patients’ lives. Three studies used multiparametric MRI-based radiomic features in conjunction with clinical features to assess and predict the worsening of postoperative brain edema in meningioma patients through the creation of an evaluation model [67–69]. Moreover, Hu et al. demonstrated the value of radiomics models in predicting postoperative hemorrhage [68]. Tumor-peritumoral edema and brain infiltration are major risk factors for perioperative seizures; to this end, Xu et al. developed a radiomic score derived from imaging features that demonstrated high accuracy in predicting early postoperative seizures [70].

The second category of research focuses on radiotherapy selection and its associated adverse effects. Postoperative radiotherapy selection remains contentious, particularly in the treatment of AMs and MMs. Park proposed an interpretable radiomic model for selecting patients with WHO Grade II meningiomas who may best benefit from postoperative radiotherapy [71]. Stereotactic radiosurgery (SRS) has been widely used in meningioma treatment, especially as an adjuvant therapy after subtotal resection (STR). Brain edema is a common complication following Gamma Knife radiosurgery (GKS), with a reported risk of up to 50% [72]. Among patients with post-GKS edema, 40-63% present with headaches, nausea, vomiting, ataxia, seizures, or local neurological deficits [73]. Management of these symptoms sometimes requires steroid therapy and, occasionally, surgical intervention. Thus, there is an urgent need to predict the risk of edema following GKS to assist in timely treatment decision-making. In a multicenter study, Li et al. effectively predicted the development of edema following radiotherapy for meningiomas [73]. Additionally, Speckter et al. predicted volume responses and controls following GKS for meningioma patients [74].

The third category of research focuses on prognosis and outcome prediction, including tumor progression and recurrence and patient mortality. Prognostication has long been a challenge in meningioma research and is closely tied to the development and adjustment of clinical treatment protocols. Previous prognostic models often rely solely on clinical features, incorporating a small number of research factors into a relatively singular model. Radiomic technology captures the morphological, histological, and functional characteristics of tumors, providing clinicians with more comprehensive information. Ko et al. and Morin et al. separately established radiomic-based models for predicting progression, recurrence, and overall survival in patients with meningiomas [75, 76]. Zhang et al. and Hsieh et al. proposed predictive models for recurrence in skull base meningioma and parasagittal/falcine meningioma, respectively, which contributed to personalized treatment planning and improved patient survival rates and quality of life [16, 17]. There are significant differences in postoperative recurrence and mortality between high-grade and low-grade meningiomas; therefore, developing prognosis prediction models stratified by pathological grade is more meaningful. Currently, four studies have developed recurrence prediction models specifically for high-grade meningiomas, all of which have demonstrated superior performance compared with conventional clinical models [18, 77–79]. Future predictive models should incorporate an increasingly diverse set of factors, integrating clinical, imaging, pathological, molecular, and genetic features, among others, allowing a comprehensive understanding of disease progression and providing more precise foundations for personalized medical decision-making.

Potential applications of radiomics in surgical planning

The potential clinical application of radiomics in the planning and decision-making of meningioma surgery is an evolving field that can provide substantial assistance to surgeons. First, radiomics can help predict tumor consistency in meningiomas, which is critical for surgical planning. By analyzing the texture and signal signatures in MR images, radiomics can reveal the microscopic structure and biological behavior of tumors, which can help surgeons predict how easily a tumor will be removed during surgery. Second, radiomics can help to assess the risk of brain invasion before surgery, thereby reducing damage to normal brain tissue during surgery. Third, radiomic analysis can reveal the relative location and relationship of the tumor to important brain structures, which is critical for selecting the best surgical approach and entry route. Fourth, high-quality predictive models constructed by radiomics can also be used to predict treatment outcomes after surgery, including the likelihood of tumor recurrence and patient prognosis. In summary, radiomics can assist surgeons in specifying a personalized treatment plan, including surgery, radiation, chemotherapy, or other targeted therapies.

Limitations and future prospects of current research

Although radiomic techniques have significantly advanced our understanding of meningiomas, several limitations still persist. First, variations in imaging equipment and protocols can affect data quality and standardization, leading to reduced model accuracy and stability. Second, most current studies rely on static imaging, with limited exploration of tumor dynamics, potentially hindering insights into tumor progression. Additionally, challenges in multicenter data integration and institutional variability limit the generalizability of developed models. The RQS highlights further issues. A key shortcoming is the lack of biological validation, which hinders clinical translation. Without biological interpretability, models are often viewed as black boxes, reducing trust and applicability in clinical decision-making—especially critical in oncology. Moreover, cost-effectiveness analyses are largely lacking. These evaluations are essential for assessing whether the clinical benefits of radiomics justify their resource demands, particularly in constrained healthcare systems.

In the future, improvements in data standardization, incorporation of dynamic imaging and temporal analysis, and enhanced multicenter collaboration could increase model robustness and applicability. Integrating clinical, molecular, and biological data may also support the development of more comprehensive and accurate prognostic models for meningiomas.

Conclusion

The application of radiomic technology in meningioma research has great potential, but numerous technical and data challenges need to be overcome. Through continuous technological improvements, data standardization, and increased cooperation among multiple centers, imaging technology can be better leveraged for the early diagnosis of meningioma, the development of more effective treatment strategies, and the accurate prediction of disease prognosis.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

AdvMRI

Advanced MRI

ADC

Apparent diffusion coefficient

AUC

Area under the curve

AI

Artificial intelligence

AM

Atypical meningiomas

BMs

Benign meningiomas

CBTRUS

Central brain tumor registry of the United States

CT

Computed tomography

T1C

Contrast-enhanced T1-weighted

DTI

Diffusion tensor imaging

DWI

Diffusion-weighted imaging

GKS

Gamma knife radiosurgery

GTR

Gross total resection

KNN

K-nearest neighbors

MRI

Magnetic resonance imaging

MM

Malignant meningiomas

PhyMRI

Physiological MRI

PET

Positron emission tomography

PR

Progesterone receptor

QUADAS-2

Quality assessment of diagnosis accuracy studies version 2

ROIs

Regions of interest

RQS

Radiomics quality score

SRS

Stereotactic radiosurgery

STR

Subtotal resection

SVM

Support vector machine

SWI

Susceptibility-weighted imaging

T1WI

T1-weighted imaging

T2WI

T2-weighted imaging

T2FLAIR

T2 fluid-attenuated inversion recovery

TERT

Telomerase reverse transcriptase

WHO

World health organization

Author contributions

DS and RC: conceptualization, investigation, writing– original draft, and visualization; YL, KZ and DX: investigation, validation, writing– review and editing, and supervision; DY: writing– review and editing; FG: writing– review and editing, conceptualization and project administration. All authors have reviewed and approved the final version of this manuscript for publication.

Funding

None.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.