Machine learning based radiomics approach for outcome prediction of meningioma – a systematic review

Meningioma is the most prevalent type of cancer identified in the brain which is mostly affecting adults. Approximately 7.86 individuals per 100,000 are diagnosed annually with meningioma globally, which is significantly attributed to the credibility of advancements in imaging technology for tumour diagnosis in the last three decades.^1–2 The World Health Organization (WHO) has classified meningiomas into three distinct grades of which the Grade I meningiomas are represented by their slow growth, while Grade II meningiomas are categorised as atypical and Grade III meningiomas represent the most fatal form, being malignant in nature. Thus, both Grade II and Grade III meningiomas display a high rate of recurrence following routine treatment, which may obviously end up with a reduced life expectancy. These facts shed light on the early detection and subsequent treatments which are quite mandatory for the precise diagnosis of meningiomas.^3–5 After meningioma surgery, overall survival at 5 and 10 years respectively were 92.6% and 85%. Survival is prolonged in women, younger adults and those with convexity and benign tumor.⁶

Magnetic Resonance Imaging (MRI) is an emerging and most effective method for detecting meningiomas and monitoring specific medical conditions. Conventional MR imaging findings such as tumor size, bone invasion, and parasagittal location have all been identified as important imaging parameters related to progression/recurrence (P/R) in meningiomas.^7–9

Radiomics incorporates a range of advanced techniques used to conduct a comprehensive analysis of medical images. Radiomics is primarily intended to analyse the characteristics of tissues and lesions, with respect to their shape and heterogeneity, and monitor changes over a period through serial imaging, during treatment period.^10–12 Examining tissue heterogeneity is crucial, as genomic studies highlight its role as prognostic factor for survival and a challenge in cancer management. Radiomic features strongly correlate with cellular heterogeneity indices. Unlike traditional biopsies, which sample limited tumour areas, radiomics enables heterogeneity assessment across the entire tumour. Leveraging large datasets, radiomics identifies unrecognized markers and patterns related to disease progression and treatment. Additionally, integrating radiomic data with clinical, histological, genomic, and other datasets through machine learning enhances its diagnostic and prognostic utility.^13–18 There are no reviews conducted on evaluating the performance of machine learning (ML) based radiomic models in the prediction of recurrence and overall survival for meningioma. Hence, the aim of the study is to summarize the performance of the ML based radiomics models in the prediction of outcomes such as P/R and overall survival analysis for meningioma.

Methods

PRISMA¹⁹ (Preferred reporting items for systematic reviews and meta-analysis) standards were followed in conducting the review. (PRISMA checklist- Extended data)

Results

A total of eight articles were included ( Figure 1).

Figure 1. Flow chart for study selection.

Discussion

Meningioma is the most prevalent brain tumor, primarily affecting adults, and it occurs with incidence of 7.86 per 100,000 people worldwide each year.^1–2 MRI is the recommended imaging modality, and recurrence is associated with traditional imaging criteria such as tumor size and bone invasion.^7–9 Radiomics, provides a non-invasive substitute for biopsies by evaluating tumor heterogeneity and identifying prognostic markers. Using machine learning to integrate radiomics with clinical and pathological data may improve the accuracy of diagnosis and prognosis of tumors.^10–12 However there is lack of systematic review of ML-base radiomics models for meningioma survival and recurrence prediction.

Zhang et al.²¹ and Kalasauskas et al.²⁷ reported the effectiveness of radiomics-based ML models in predicting tumor recurrence and aggressiveness in meningiomas. Using radiomic features from contrast enhanced T1Weighted image and ADC maps as primary predictors, Zhang et al.²¹ created a random forest-based radiomics model that achieved 90% accuracy in predicting P/R in SBM. Their research showed that radiomics performed better than traditional MR imaging in objective and quantitative tumor categorization, offering useful prognostic information for decisions about surgery and treatment. Kalasauskas et al.²⁷ reported that radiomic analysis is more accurate than traditional neurological evaluation in predicting tumor recurrence and CNS WHO grade. The ML model performed well across heterogeneous MRI datasets, emphasizing the importance of the data preprocessing in lowering variability across varied MRI datasets. According to their study, preoperative evaluation based on radiomics may help with treatment stratification by directing regarding the extent of resection and follow-up imaging. Both the studies support the clinical promise of radiomics for better surgical planning, recurrence prediction, and tailored meningioma care.

The studies by Park et al.,²³ Hsieh et al.,²⁴ and Ko et al.²² highlighted the role of radiomics-based ML models in predicting tumor recurrence in meningiomas and refining treatment strategies. Park et al.²³ demonstrated that integrating radiomic features with clinicopathologic data significantly improved recurrence prediction in grade 2 meningiomas. Their model identified high-risk patients who benefited from ART, supporting the use of radiomics for non-invasive risk stratification and personalized treatment decisions. Hsieh et al.²⁴ introduced a radiomic score-based approach, combining clinical and MRI texture features, achieving an AUC of 0.88 in predicting P/R in parasagittal and parafalcine meningiomas. Their study confirmed that integrating clinical and imaging data enhanced predictive accuracy compared to using either alone. Ko et al.²² developed an SVM-based radiomics model, identifying adjacent bone invasion, low ADC values, and high SVM scores as key high-risk factors for recurrence. Their model outperformed traditional ADC-based assessments and emphasized radiomics’ potential in preoperative risk assessment. All three studies support radiomics as a valuable prognostic tool for identifying high-risk meningioma patients, optimizing surgical and ART decisions, and guiding personalized follow-up strategies.

The studies by Ren et al.²⁵ and He et al.²⁶ reinforce the predictive value of radiomics-based machine learning models in assessing tumor recurrence and progression-free survival (PFS) in meningiomas. Ren et al.²⁵ developed a clinicopathological-radiomics model (CPRM) incorporating radiomic features from T1Contrast and T2-FLAIR MRI alongside clinical factors, achieving an AUC of 0.858 for predicting recurrence in atypical meningiomas across multiple neurosurgical centers. Their model proved the utility of combining multimodal data for precision oncology, outperforming previous clinical and radiomics-only models.

He et al.²⁶ developed a complete clinicopathological-radiomics model (CPRM) that integrated three radiomics signatures (T1, eT1, T2), WHO grading, Simpson grading, Ki-67 index, and radiotherapy history, achieving AUCs of 0.93 for training dataset and 0.88 for testing dataset.

Their study produced a clinical normogram for personalized risk assessment and performed better than single-modality radiomics models (AUC:0.70-0.85). However, Simpson grading and Ki-67 remained the most significant predictors for progression free survival, suggesting that postoperative factors are crucial for long-term monitoring. Both studies highlighted the superior performance of integrated radiomics-clinicopathological models over traditional assessments, supporting their use in preoperative risk stratification and treatment planning for meningiomas.

The study by Morin et al.²⁸ highlighted the prognostic significance of radiologic and radiomic features in predicting meningioma grade, local failure (LF), and overall survival (OS). Their integrated model, combining preoperative MRI features with clinical variables, demonstrated significant predictive performance. ADC hypointensity and low tumor sphericity were identified as key radiomic predictors associated with high-grade meningiomas, poor local control, and reduced OS. Conventional radiological imaging features such as peritumoral edema, indistinct tumor margins, and absence of a CSF cleft sign, were associated with higher tumor grade and worse outcomes.

The review has few limitations. Firstly, meta-analysis was not conducted due to heterogeneity of data, including different MRI protocols and patient populations. Secondly, studies included had small sample sizes, limiting statistical power. Thirdly, most studies were retrospective, introducing potential biases.

Conclusion

Our review suggests that radiomics greatly improved meningioma diagnostic and prognostic accuracy when combined with clinical and pathological data through machine learning. ML models based on radiomics performed better than traditional MRI in predicting tumor aggressiveness and recurrence, offering important information for surgical planning and individualized treatment. The combination of radiomics and clinicopathological data offers a non-invasive substitute for biopsies helping in risk stratification and guiding decisions on adjuvant treatment and follow-up plans. Overall, radiomics holds great potential for improving the management and outcomes of meningioma patients.

Ethics and consent

Ethical approval and consent were not required

Reporting guidelines

Fig share: https://doi.org/10.6084/m9.figshare.28557743.v1.

Data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).