Radiomics in glioblastoma recurrence: advances in prediction, localization, and differentiation from treatment-related effects

Abstract

Background

Glioblastoma multiforme (GBM), the most aggressive primary brain tumor, is characterized by rapid recurrence and poor prognosis despite multimodal therapy. Accurate differentiation of GBM recurrence from treatment-related effects (TrE), including pseudoprogression (PsP) and radiation necrosis (RN), remains a major clinical challenge due to overlapping imaging features on conventional Magnetic Resonance Imaging (MRI).

Main body

Radiomics has emerged as a noninvasive quantitative imaging approach that extracts high-dimensional features from medical images and integrates them with machine learning algorithms. This review summarizes recent advances in radiomics for recurrent GBM, including characterization of recurrence patterns, preoperative and postoperative recurrence risk prediction, spatial localization of recurrent lesions, and differentiation of recurrence from TrE. Key technical and clinical challenges are also discussed, including data heterogeneity, limited external validation, model generalizability, and biological interpretability.

Conclusions

By linking imaging biomarkers with clinical and biological insights, radiomics demonstrates significant translational potential for improving recurrence assessment in GBM. Future efforts should focus on multicenter validation, standardized imaging protocols, and enhanced interpretability to facilitate reliable clinical implementation.

Introduction

Gliomas, originating from glial cells, are the most common primary brain tumors in adults [1]. According to the 2021 World Health Organization (WHO) grading system, gliomas are classified into grades I-IV. Specifically, glioblastoma multiforme (GBM), a grade IV astrocytoma, has an incidence rate of 3.19 per 100,000 populations and a median survival of approximately 15 months, with the latter varying by age, sex, and geographic location [2, 3]. The standard treatment for newly diagnosed GBM involves maximal surgical resection followed by radiotherapy and temozolomide chemotherapy [4, 5]. Despite undergoing standard treatment, almost all patients experience tumor recurrence within 6–9 months due to the aggressive and invasive nature of the GBM cells [6–8]. Moreover, recurrent GBM often develops resistance to chemotherapy [9, 10]. These factors collectively contribute to a dismal prognosis for patients. Moreover, distinguishing GBM recurrence from pseudoprogression (PsP) is clinically challenging, as they often share overlapping radiological features like Magnetic Resonance Imaging (MRI) contrast enhancement [11]. This diagnostic challenge easily leads to either unnecessary intervention for PsP or delayed treatment for recurrence, worsening the difficulty of managing GBM [12, 13]. Therefore, precise prediction and identification of GBM recurrence is crucial for selecting appropriate treatment strategies and improving prognosis.

However, conventional MRI encounters substantial hurdles in differentiating true tumor progression from treatment-related effects (TrE) [14]. Approximately 20% of patients develop pseudoprogression, which arises from inflammation and blood brain barrier disruption induced by treatment [15, 16]. It manifests as temporary enlargement of contrast enhancing lesions within the first 3 months after chemoradiation [16]. Moreover, radiation necrosis, which typically develops 3–12 months after treatment, also gives rise to contrast enhancing lesions [17, 18]. Both PsP and radiation necrosis (RN) simulate true tumor progression on conventional MRI, thereby complicating the accurate assessment of treatment response.

Owing to the heterogeneous nature of GBM, multiple imaging modalities have been employed to achieve comprehensive tumor characterization [19]. In recent years, radiomics has emerged as a framework for extracting quantitative imaging biomarkers and correlating them with clinical outcomes across various cancers [20–22]. By enabling high-throughput analysis of radiological datasets, radiomics not only enhances tumor phenotyping but also offers deeper insights to guide personalized treatment strategies in neuro-oncology [23].

Building on this framework, radiomics has been increasingly applied to predict the prognosis, gene mutation, pathogenesis, and therapeutic responses of patients with GBM [24–28]. Notably, it has shown considerable utility in predicting GBM recurrence. For example, Du et al. developed a decision tree model that integrates the preoperative MRI-derived radiomics score (Rad-score) with clinicopathological factors. This model demonstrated robust performance in predicting the risk of GBM recurrence within one year after gross total resection [29].

In this review, we provide an integrative overview of current radiomics-based approaches for predicting and discriminating GBM recurrence across the disease continuum. We summarize applications in both preoperative and postoperative settings, including recurrence risk stratification, early surveillance, spatial localization of recurrent lesions, and differentiation of tumor recurrence from treatment-related imaging changes such as PsP and RN. Particular attention is given to emerging models that combine multimodal MRI features with molecular biomarkers (e.g., isocitrate dehydrogenase [IDH] mutation and O6-methylguanine-DNA methyltransferase [MGMT] promoter methylation) and clinical variables (e.g., age and Karnofsky Performance Status). By synthesizing current evidence, identifying methodological limitations, and outlining future research directions, this review aims to inform ongoing translational efforts to bridge radiomics research and clinically meaningful recurrence assessment in glioblastoma.

Review methodology

This narrative review was conducted through a structured literature search aimed at identifying studies investigating radiomics-based approaches for predicting and discriminating glioblastoma recurrence. The search strategy covered major biomedical databases, including PubMed, Web of Science, and Scopus, and encompassed studies published between January 2018 and December 2024. Search terms combined keywords and Medical Subject Headings (MeSH) related to glioblastoma recurrence and imaging-based modeling, including “glioblastoma,” “recurrence,” “radiomics,” “machine learning,” “deep learning,” “MRI,” and “treatment-related effects.” In addition, reference lists of relevant reviews and key original articles were manually screened to identify further pertinent studies.

Relevant studies included original research articles involving human subjects that addressed clinically relevant aspects of glioblastoma recurrence, such as recurrence risk prediction, spatial localization of recurrent lesions, or differentiation between tumor recurrence and treatment-related imaging changes. Studies employing radiomics and machine learning approaches based on medical imaging data were considered. Conference abstracts, review articles, and non–imaging-based investigations were excluded.

All included studies were assessed for relevance, and data were extracted with emphasis on study objectives, patient cohorts, imaging modalities and acquisition protocols, feature extraction or modeling strategies, validation approaches, and reported clinical implications. Given the substantial methodological heterogeneity across available studies, including variations in imaging protocols, feature definitions, and validation strategies, a narrative synthesis approach was adopted.

The aim of this review was to provide a clinically oriented overview of current radiomics-based methods for glioblastoma recurrence assessment, highlight methodological strengths and limitations, and identify key gaps that warrant further investigation to facilitate future translational research.

GBM recurrence

Recurrence is one of the leading causes of mortality in patients with GBM [30]. Notably, the timing of recurrence exhibits substantial interpatient variability. The time to recurrence of GBM is not only associated with tumor grade but also closely linked to its molecular subtypes (Proneural, Neural, Classical, Mesenchymal) [31–34]. For instance, Proneural subtype gains no benefit from aggressive therapy, which may impact its recurrence time. Independent risk factors for tumor recurrence in GBM patients include partial resection (PR), tumor contact with the subventricular zone (SVZ), and telomerase reverse transcriptase (TERT) C228T wild-type status [35, 36]. Furthermore, KPS remains a significant independent prognostic factor for survival in patients with recurrent GBM [36, 37].

The management of recurrent GBM is highly dependent on accurate radiographic identification [38–40]. However, the imaging manifestations of recurrent GBM are heterogeneous and can present in four distinct patterns: local recurrence, distant recurrence, multifocal recurrence, and diffuse recurrence (Fig. 1).

Fig. 1.

Schematic Representation of GBM Recurrence Patterns. Created with BioRender.com

Local recurrence is the most common pattern, accounting for over 80% of cases. It typically occurs within 2 cm of the original tumor margin or resection cavity and is characterized by nodular or thick, irregular enhancement on post-contrast T1-weighted MRI [41–43]. Distant recurrence is defined as the occurrence of contrast-enhancing lesions located more than 2 cm away from the resection cavity [44, 45]. Morphologically, these lesions usually exhibit well-defined, linear, or nodular enhancement [46]. Multifocal recurrence is identified by the presence of two or more lesion sites, each with mostly or completely well-defined borders, separated by regions of normal brain signal [47]. This pattern is suggestive of extensive tumor infiltration or subependymal spread [48]. Diffuse recurrence refers to recurrent lesions that are either centered at the primary tumor site or extended beyond 3 cm from the primary site or resection cavity margin. A key feature of this pattern is that at least 50% of the lesion margin is qualitatively evaluated as poorly defined [47].

Clinical management of recurrent GBM

Management of GBM remains one of the most challenging issues in neuro-oncology. Despite advances in surgical, radiotherapeutic, and systemic approaches, no universally accepted standard of care has been established for recurrent disease [49, 50]. Contemporary management has therefore shifted toward individualized, context-dependent decision-making that reflects the biological heterogeneity of recurrence and the limited durability of available therapies.

Historically, repeat surgical resection and re-irradiation have been considered for selected patients with localized recurrence, favorable performance status, and surgically accessible lesions [51, 52]. While these approaches may offer cytoreduction or symptomatic relief in carefully selected patients, their applicability is constrained by patient selection, treatment-related risks, and the limited durability of survival benefit reported across studies [53, 54]. Systemic therapies, including alkylating agents, nitrosoureas, and anti-angiogenic treatments such as bevacizumab, have expanded therapeutic options for recurrent GBM [55]. Nevertheless, survival benefits remain limited and highly variable, underscoring the need for improved patient stratification [56].

Importantly, treatment strategies differ substantially according to recurrence patterns. Localized recurrence may prompt consideration of repeat surgery or focal re-irradiation, whereas distant or multifocal recurrence generally favors systemic or palliative approaches [57, 58]. Diffuse recurrence, characterized by infiltrative growth and poorly defined margins, poses particular challenges for local therapies and is therefore more commonly managed with systemic or supportive strategies [59]. In contrast, treatment-related effects such as pseudoprogression or radiation necrosis typically warrant conservative management rather than therapeutic escalation. Misclassification of true tumor recurrence and treatment-related changes can therefore directly lead to inappropriate treatment selection and suboptimal patient outcomes [60, 61].

As a result, accurate radiographic assessment is central to therapeutic decision-making during GBM recurrence [62]. Yet, post-treatment imaging findings are often complex and heterogeneous, reflecting a mixture of tumor progression, therapy-induced changes, and diverse recurrence patterns, particularly after radiotherapy or emerging treatments [63, 64]. These challenges have motivated the exploration of advanced quantitative imaging approaches that aim to provide more objective and reproducible descriptors of recurrent disease and to support individualized treatment stratification.

Within this clinical context, quantitative imaging approaches have gained increasing attention as potential decision-support tools. By capturing tumor heterogeneity, spatial growth characteristics, and biological aggressiveness, such approaches may refine recurrence classification, guide patient selection for salvage therapies, and ultimately improve management strategies for recurrent GBM [65–68]. Beyond improving diagnostic accuracy, emerging radiomics-based biomarkers may also influence clinical management. Imaging-derived recurrence risk estimates could help identify patients who may benefit from intensified salvage treatments versus those more suitable for conservative strategies. Quantitative risk stratification may further support tailored surveillance intervals, while spatially resolved risk maps have the potential to assist biopsy targeting or focal treatment planning in heterogeneous post-treatment lesions. Together, these roles suggest that radiomics may complement conventional imaging by providing decision-support information rather than serving solely as a diagnostic adjunct.

The role of radiomics in recurrent GBM

Given the heterogeneity of recurrence patterns and the reliance of clinical management on accurate imaging assessment, radiomics has emerged as an advanced computational approach that extracts high-dimensional quantitative features from medical images, enabling the noninvasive characterization of tumor heterogeneity, microenvironment, and biological behavior [19, 69]. At its core, a standardized radiomics workflow typically encompasses several key stages: (1) acquisition of high-quality, standardized medical images (e.g., MRI, CT); (2) delineation of the region of interest (ROI), often the tumor, via manual or increasingly automated segmentation; (3) high-throughput extraction of quantitative features capturing lesion morphology, intensity statistics, textural patterns, and spatial relationships; (4) feature selection and dimensionality reduction to identify the most informative biomarkers; and (5) construction and validation of predictive or diagnostic models using machine learning algorithms [70, 71]. These steps are summarized in Fig. 2, which provides a schematic overview of the standardized radiomics pipeline.

Fig. 2.

Schematic overview of a standardized radiomics workflow

In several solid tumors, radiomics-based models have been successfully applied to clinically meaningful tasks, including lesion characterization, treatment response assessment, and risk stratification. In lung cancer, radiomics has shown robust performance in distinguishing benign from malignant pulmonary nodules, providing support for noninvasive diagnostic decision-making [72, 73]. In breast cancer, MRI-derived radiomic signatures have been associated with pathological complete response to neoadjuvant chemotherapy, facilitating pre-treatment stratification and individualized therapy planning [74–76]. Similarly, in cervical cancer, deep learning–based imaging models have been applied to predict recurrence and metastatic risk [77]. Collectively, these applications illustrate that, when aligned with well-defined clinical questions and standardized imaging protocols, radiomics can meaningfully inform clinical decision-making.

In parallel with imaging-based radiomics, computational pathology has emerged as a complementary AI approach using whole-slide histopathology images to characterize tumor biology. Recent transformer-based weakly supervised models, such as that proposed by Jiang et al., achieve diagnostic-level glioma classification and molecular marker discovery, underscoring the growing potential of integrated radiology–pathology AI frameworks.

Biological interpretability remains a critical consideration for the clinical translation of radiomics-based models. Texture-based features, including wavelet-derived metrics, are thought to capture multi-scale intratumoral heterogeneity related to variations in cellularity, necrosis, and microvascular architecture [78]. First-order intensity features may reflect tissue composition and vascular permeability, particularly on contrast-enhanced and perfusion-sensitive MRI, whereas shape-related features can indicate infiltrative growth and poorly defined tumor margins [79–81]. Complementary evidence from AI-based computational pathology further supports this interpretability framework, as recent weakly supervised models applied to whole-slide histopathology images have demonstrated the feasibility of inferring glioma classification and molecular characteristics directly from tissue morphology [82]. Although these associations do not imply direct histopathological correspondence, they provide biological plausibility for linking quantitative imaging features with underlying tumor aggressiveness and microenvironmental complexity [80].

In the context of GBM, radiomics has shown considerable potential in prognostic stratification, molecular subtyping, and recurrence identification, thereby supporting the formulation of personalized therapeutic strategies [80].

Predicting GBM recurrence

Radiomics also holds substantial promise in predicting and discriminating GBM recurrence. By analyzing preoperative tumor, postoperative residual tumor, or peritumoral imaging features, radiomics can detect subtle changes in the tumor microenvironment, identify high-risk regions for recurrence, and differentiate GBM recurrence from TrE [28, 83–85] (Fig. 3). This capability facilitates the development of personalized follow-up protocols and therapeutic strategies, thereby expanding the clinical utility of radiomics to cover the entire continuum of GBM management [86].

Fig. 3.

Schematic overview of the role of radiomics in GBM. Created with BioRender.com

Preoperative radiomics for predicting GBM recurrence

Identifying patients at risk of early GBM recurrence prior to treatment initiation is clinically crucial, as it enables proactive therapeutic planning and timely intervention during disease progression. However, as an invasive procedure, tissue biopsy carries inherent risks of misdiagnosis, primarily due to tumor heterogeneity and sampling errors [87]. Consequently, noninvasive imaging biomarkers capable of capturing spatially heterogeneous tumor biology have become an area of growing interest.

In recent years, radiomics has advanced significantly in the preoperative diagnosis and recurrence prediction of GBM. The representative radiomics models were summarized in Table 1.

Table 1.

Models for predicting GBM recurrence

Author	Time	Radiomics Type	Model Name	SD	Software Used for Feature Extraction	Modalities Used for Feature Extraction	TSS	VSS	Training Set AUC	Validation Set AUC	Validation Strategy	Key Limitations	Ref.
Lundemann M et al.,2018	N/A	Hand-crafted	Binomial logistic regression model	P	ITK-SNAP, Mirada Workstation	18F-FET PET/CT, 18F-FDG PET/MRI	15	N/A	0.77	N/A	Internal	Small single-center cohort; voxel-wise modeling with limited patient-level validation; no independent external validation	[83]
WANG et al.,2021	2010–2018	Hand-crafted	Nomogram	R	N/A	Contrast enhanced sequence, FLAIR, T1WI, T2WI	86	36	0.85	0.84	Internal	Single-center retrospective; modest cohort size; 2D feature extraction; no external validation; limited survival analysis	[88]
Lao et al.,2021	2012,2019	Hand-crafted	Proximity Estimation-coupled Support Vector Machine SVMPE​	R	N/A	T1WI, CE-T1WI, T2WI, FLAIR, ADC	20	30	0.73 (F1 score)	0.73 (F1 score)	Internal	Single-center retrospective; imaging protocol heterogeneity; no independent external validation; selection bias	[89]
Du et al.,2023	2017–2020	Hand-crafted	Decision tree model	R	ITK-SNAP	T1WI, T2WI, T2-FLAIR, DWI, CE-T1WI	98	36(test set) + 37(Wuhan Union Hospital) + 26(The Second Affiliated Hospital of Xuzhou Medical University)	0.85	0.719 0.81 0.702	Internal+ External	Single-center retrospective; small sample size; manual segmentation variability; no external validation	[29]
Salari et al.,2024	2014–2020	Hand-crafted	RFC, SVC	R	3D Slicer	CE-T1WI, T2- FLAIR	95	N/A	0.829(RFC)	N/A	Internal	Small sample size; limited MRI sequences; incomplete survival data; lack of multicenter and molecular integration	[90]
Rathore et al.,2018	2006–2013	Hand-crafted	SVM with Gaussian kernel	R	GLISTR algorithm	T1WI, CE-T1WI, T2WI, T2-FLAIR, DTI, DSC	31	59	0.83	0.91	Internal+ External	Single-center cohort; no independent external validation; limited feature complexity; no histopathological confirmation of infiltration	[91]
Shim et al.,2021	2010–2019	Hand-crafted	Neural network model, Cox-LASSO	R	NordicICE	CE-T1WI, T2-FLAIR, DSC	154	38	0.995(LR) 0.986(DR)	0.969(LR) 0.864(DR)	Internal	Single-center retrospective; limited sample size; binary endpoint without time-to-event modeling; no multimodal integration; no external validation	[66]
Cepeda et al.,2023	N/A	Hand-crafted	XGBoost, RF, LightGBM	R	ITK-SNAP, LifEx	T1WI, CE-T1WI, T2WI, FLAIR, ADC	40	15	N/A	0.81	Internal+ External	Retrospective; limited cohort relative to feature dimensionality; internal validation only; overfitting risk	[92]
Luzzi et al.,2024	N/A	DL-assisted hand-crafted radiomics	Radiomics-based multifactorial in silico model	PoC	FMRIB’s Automated Segmentation Tool	T1WI, CE-T1WI, T2WI, FLAIR, DWI, DSC	1	N/A	Jaccard index: 0.69(short-term tumor growth) 0.26(long-term recurrence site)	N/A	N/A	Single-patient proof-of-concept; low spatial prediction accuracy; no postoperative biomechanical modeling; no cohort-level validation	[93]

Abbreviations: SD Study Design, DL Deep Learning, R Retrospective, P Prospective, PoC Proof-of-concept, RFC Random Forest Classifier, SVC Support Vector Classifier, SVM Support Vector Machine, CatBoost Categorical Boosting, XGBoost Extreme Gradient Boosting, RF Random Forest, LightGBM Light Gradient Boosting Machine, LR Local Recurrence, DR Distant Recurrence, TSS Training Set Size, VSS Validation Set

Representative work by Lundemann et al. demonstrated the feasibility of voxel-wise recurrence prediction using multiparametric imaging [83]. By integrating ¹⁸F-fluoroethyl L-tyrosine (FET) Positron Emission Tomography (PET)/Computed Tomography (CT), ¹⁸F-fluorodeoxyglucose (FDG) PET/MRI, diffusion-weighted imaging (DWI), and dynamic contrast-enhanced (DCE) perfusion MRI, the study showed that regions with higher preoperative metabolic and hemodynamic activity were more likely to develop recurrence, with ¹⁸F-FET uptake emerging as the strongest predictor. Their logistic regression model achieved a voxel-level Area Under the Curve (AUC) of approximately 0.77, illustrating the potential of metabolic–perfusion imaging to delineate infiltrative tumor margins beyond conventional structural MRI.

Building on this work, Wang et al. developed a preoperative predictive nomogram integrating multi-dimensional data to stratify early recurrence risk in GBM, which exhibited superior predictive accuracy [88]. This integrative approach achieved AUCs in the mid-0.80 range in validation, outperforming models based on radiomics or clinical features alone. Their findings highlight the added value of combining quantitative imaging heterogeneity metrics with established radiological and systemic indicators of tumor aggressiveness.

Across studies, several common feature types are summarized in Table 1. Metabolic PET parameters and perfusion-derived metrics are frequently associated with future recurrence, likely reflecting hypermetabolic and hypervascular tumor subregions that are resistant to therapy. Diffusion features, indicative of tumor cellularity, also appear repeatedly. These observations support the biological premise that preoperative imaging can identify aggressive tumor habitats prone to early regrowth. Despite methodological heterogeneity, these approaches are largely complementary and underscore the growing value of imaging biomarkers in precision neuro-oncology.

Postoperative radiomics for predicting GBM recurrence

Despite maximal safe resection followed by radiotherapy and temozolomide, glioblastoma remains characterized by a high rate of local recurrence and poor long-term survival [94]. Early identification of recurrent disease is therefore critical for timely therapeutic intervention. However, histopathological confirmation is often impractical in the early post-treatment period due to surgical risks and the frequent occurrence of pseudoprogression (PsP), which limits the reliability of conventional imaging-based assessment [95, 96]. These diagnostic uncertainties have driven growing interest in radiomics as a noninvasive tool for postoperative surveillance and early recurrence prediction [97].

Current postoperative radiomics studies are predominantly retrospective and single-center, with cohort sizes typically below 100 patients, as shown in Table 1. Most models are built on multiparametric MRI, especially Contrast-Enhanced T1-Weighted Imaging (CE-T1WI) and T2-Fluid-Attenuated Inversion Recovery (FLAIR), frequently combined with diffusion-derived metrics such as Apparent Diffusion Coefficient (ADC). Extracted features are largely hand-crafted texture, intensity, and wavelet descriptors that may capture residual tumor infiltration, microvascular proliferation, and therapy-induced microstructural heterogeneity [94, 98]. Reported predictive performance generally falls within a moderate-to-high range, with AUC values commonly between approximately 0.75 and 0.85, although external validation remains limited.

Several representative studies illustrate different methodological directions within this field. Lao et al. proposed a voxel-wise framework combining support vector machine (SVM) framework with stem cell niche (SCN) proximity estimation (denoted as SVMPE) [89]. Using postoperative multiparametric MRI, their model aimed to localize subclinical high-risk regions around the resection cavity and achieved an F1-score of about 0.73 in the test set, demonstrating the potential of spatially resolved prediction to identify recurrence-prone areas months before radiological progression.

Furthermore, Du’s research team conducted a multicenter prospective cohort study that improved the generalization of the model [29]. To mitigate inter-site variability related to MRI acquisition, all centers employed 3.0 T scanners with standardized imaging protocols, and image preprocessing steps such as resampling, intensity normalization, and inter-sequence registration were applied to reduce potential batch effects arising from differences in scanner parameters (e.g., TR and TE). Based on this harmonized imaging framework, they developed an interpretable decision tree model integrating radiomic signatures with key clinical and molecular variables, including age, MGMT promoter methylation, KPS, and TERT mutation status. The model achieved an AUC of 0.85 in the training cohort and approximately 0.72 in both internal and external validation cohorts, underscoring the added value of combining imaging heterogeneity metrics with established prognostic biomarkers for postoperative risk stratification.

In addition, Salari et al. established a multimodal MRI-driven predictive framework for GBM recurrence, which achieved higher prognostic precision than Du’s study [90]. Utilizing contrast-enhanced T1-weighted (CE-T1WI) and T2 fluid-attenuated inversion recovery (T2_FLAIR) sequences from the Burdenko Glioblastoma Progression Dataset, their random forest–based model achieved an AUC of approximately 0.83, outperforming single-sequence models. The study underscores the clinical potential of multimodal radiomics, which provides a comprehensive representation of tumor heterogeneity and biological behavior.

Across studies, several consistent patterns emerge. CE-T1WI–derived features frequently contribute to performance, likely reflecting neovascularization and blood–brain barrier (BBB) disruption, while T2-FLAIR and diffusion features capture non-enhancing infiltrative tumor and treatment-induced tissue alterations. Models that incorporate clinical or molecular variables generally outperform imaging-only approaches, supporting the concept that postoperative recurrence risk is shaped by both tumor biology and host factors.

Predicting the site of GBM recurrence

In the clinical management of GBM, achieving early and precise localization of recurrence remains a critical challenge. Recurrence frequently arises from infiltrative tumor cells extending beyond the visibly abnormal regions on routine imaging, leading to spatially heterogeneous regrowth patterns and complicating timely intervention. Notably, over 80% of recurrences occur near the resection cavity, highlighting the spatial preference of infiltrative tumor cells and the difficulty of adequately covering high-risk regions using uniform radiotherapy margins [99]. In this context, radiomics, through quantitative feature extraction from multimodal imaging and machine learning applications, offers a noninvasive and reproducible method to predict GBM recurrence sites, thereby optimizing personalized treatment planning.

Current studies on recurrence site prediction are typically voxel-wise or region-based and rely heavily on multiparametric MRI, especially CE-T1WI, T2-FLAIR, diffusion metrics, and perfusion imaging such as Dynamic Susceptibility Contrast (DSC)-derived Cerebral Blood Volume (CBV). Extracted features often include texture heterogeneity, intensity distributions, and perfusion dynamics, which are thought to reflect tumor cellularity, angiogenesis, and microstructural invasion. Reported performance is generally moderate to high, with region-based or voxel-level AUCs frequently around 0.80–0.90, although most evidence remains derived from retrospective single-center cohorts.

One representative line of work focuses on biology-driven spatial modeling of recurrence risk. Rathore et al. demonstrated that multiparametric radiomic features from the peritumoral edema region could distinguish areas destined for recurrence from those that remained stable [91]. Perfusion and diffusion features, including elevated relative Cerebral Blood Volume (rCBV) and reduced ADC, were particularly informative, and the multiparametric model significantly outperformed conventional MRI alone (AUC around 0.9). Their spatial mapping revealed 97% of pathologically confirmed recurrences originated within 20 mm of the original tumor core, which was consistent with preoperative infiltration hotspots identified by the model. This finding challenges the clinical practice of uniform radiotherapy margins, instead advocating for personalized dose escalation guided by radiomic risk maps. The study underscores the potential of ML to decode tumor biology from routine imaging, enabling precision targeting of infiltrative GBM niches.

Other studies have emphasized advanced machine learning integration of perfusion-driven radiomics. Shim et al. developed a neural network–based model using radiomic features derived from CBV maps to differentiate local versus distant recurrence patterns [66]. The model achieved high AUCs in validation (approximately 0.97 for local recurrence and 0.86 for distant recurrence), suggesting that vascular heterogeneity in both enhancing tumor and peritumoral edema regions may encode spatial relapse patterns. Importantly, their radiomics-based risk scores were also associated with progression-free survival, indicating that spatial prediction and temporal prognosis may be biologically linked.

To improve generalizability, Cepeda et al. evaluated voxel-based radiomics across multicenter datasets and applied ensemble machine learning algorithms for recurrence region prediction [92]. Their model achieved region-level AUCs around 0.8 in external validation and demonstrated spatial concordance between predicted high-risk regions and actual recurrence sites on follow-up MRI. The model demonstrated robustness across diverse MRI protocols and scanner manufacturers, validated through multi-institutional cohorts. Notably, preoperative MRI analysis generated recurrence probability maps that aligned with subsequent tumor regrowth patterns, indicating clinical potential for guiding personalized surgical margins. This approach addresses the challenge of distinguishing infiltrative tumor from edema in peritumoral regions, offering a tool to optimize adjuvant therapies and improve patient outcomes. However, the model still has room for improvement in terms of incorporated variables.

Beyond data-driven radiomics, hybrid mechanistic–radiomic frameworks have also been explored. Luzzi et al. proposed an in silico multifactorial model integrating radiomic inputs with tumor growth dynamics, vascular parameters, and tissue biomechanics [93]. Notably, the results demonstrated high short-term predictive accuracy (Jaccard index 0.69), supporting its feasibility for intraoperative margin guidance. However, long-term recurrence predictions showed limited reliability (Jaccard index 0.26), a limitation attributed to unmodeled postoperative tissue remodeling and immune-microenvironment interactions. Despite this, the study represents an exploratory methodological advance by pioneering the integration of Vascular Endothelial Growth Factor (VEGF)-mediated angiogenesis and Diffusion Tensor Imaging (DTI)-derived white matter tractography into growth simulations, thereby providing mechanistic insights into recurrence patterns. While constrained by its single-patient proof-of-concept design, the work nonetheless highlights the critical need for hybrid models that unify cellular kinetics, vascular dynamics, and tissue biomechanics to improve prognostic accuracy. This approach not only advances computational neuro-oncology but also establishes a paradigm for personalized therapeutic planning in GBM management.

Despite promising advances, recurrence site prediction remains an emerging field constrained by small retrospective cohorts, heterogeneous recurrence definitions, and limited histopathologic validation of infiltrative margins, as reflected in Table 1. Most models depend on manual or semi-automated segmentation and lack prospective validation in radiotherapy planning workflows. Future progress will require standardized imaging protocols, multicenter harmonization, and integration of molecular and microenvironmental biomarkers to achieve clinically actionable spatial risk maps for personalized treatment targeting.

Radiomics for differentiating GBM recurrence from TrE

Post-treatment MRI of GBM patients frequently reveals three distinct yet radiologically overlapping conditions—PsP, RN, and true tumor recurrence—each requiring drastically different clinical management. This overlap poses a major diagnostic challenge, as conventional MRI’s qualitative assessment fails to reliably distinguish them, risking either over intervention or delayed treatment that worsens patient outcomes [14, 100].

PsP is transient treatment-induced abnormal signals/contrast enhancement mimicking tumor progression, usually occurring within 6 months (mostly first 3 months) of chemoradiotherapy [101]. Pathologically, it stems from treatment-related local tissue reactions (inflammation, vasogenic edema, BBB disruption, etc.) [15]. It is more common in MGMT-promoter-methylated GBM (Temozolomide [TMZ]-treated patients), with stable neurological function and longer overall survival (OS) [102]. Mild PsP resolves spontaneously, severe cases may turn to irreversible necrosis [15].

RN is a late, irreversible complication of radiation-induced neurotoxicity in healthy brain tissue, typically developing 3–12 months post-RT [102]. Its mechanism involves RT-triggered pro-inflammatory cytokines (e.g., TNF-α) causing endothelial cell death, BBB breakdown, etc [103, 104]. Histologically, it is characterized by coagulative necrosis. However, its imaging appearance often resembles tumor progression, which complicates noninvasive diagnosis [105].

Recurrence is the top cause of GBM mortality, with over 90% occurring within 6–9 months of initial therapy [106]. Pathologically, it is more aggressive (e.g., TMZ resistance) with unique molecular alterations (e.g., MGMT promoter methylation) [107]. The substantial imaging overlap with treatment-related changes highlights the need for quantitative tools that can support earlier and more accurate identification. Relevant studies on radiomics for differentiating GBM recurrence from TrE are summarized in Table 2.

Table 2.

Models for differentiating GBM recurrence from post-treatment

Author	Time	Differentiation Target	Radiomics Type	Model Name	SD	Software Used for Feature Extraction	Modalities Used for Feature Extraction	TSS	VSS	Training Set AUC	Validation Set AUC	Validation Strategy	Key Limitations	Ref.
Wang et al.,2020	2015–2018	Recurrence vs. RN	Hand-crafted	Integrated radiomics-based model	R	ITK-SNAP	¹⁸F-FDG PET, ¹¹C-MET PET, FLAIR, T1WI, T2WI, CE-T1WI	112	48	0.988	0.914	Internal	Single-center retrospective; manual segmentation; limited PET availability; no multicenter validation	[108]
Park et al.,2021	2016–2019	Recurrence vs. RN	Hand-crafted	Radiomics model	R	3D Slicer	T2WI, CE-T1WI, ADC	86	41	0.93	0.8	External	Retrospective; small sample size; limited external validation; protocol variability sensitivity	[109]
Sadique et al.,2024	N/A	Recurrence vs. RN	DL-assisted hand-crafted radiomics	MRF, NFRF	R	trained 3D deep learning mode	FLAIR, T1WI, T2WI, CE-T1WI	158	N/A	0.892	N/A	N/A	Severe class imbalance; no independent external validation; potential resampling overfitting; imaging heterogeneity	[110]
Kim et al.,2019	2011–2018	Recurrence vs. PsP	Hand-crafted	Multiparametric radiomics model	R	MITK	CE-T1WI, ADC, FLAIR, CBV	61	34 (external) + 23 (internal prospective)	0.9	0.85(external) 0.96(internal)	Internal+ External	Small external validation set; moderate external performance; selection bias; protocol variability effects	[111]
Elshafeey et al.,2019	N/A	Recurrence vs. PsP	Hand-crafted	Radiomic models based on Ktrans and rCBV maps	R + P	3D Slicer	DSC, DCE, Post-contrast T1WI	98	7	0.891(SVM) 1(C5.0)	0.89	Internal	Retrospective design; small pseudoprogression cohort; semi-automated segmentation; no prospective external validation	[112]
Ren et al.,2023	2018–2022	Recurrence vs. TrE	Hand-crafted	SVM, KNN	R	3D Slicer	T2-FLAIR, T1WI, T2WI, CE-T1WI	90	41	0.994	0.965	Internal	Single-center retrospective; imaging-based labels; small cohort; no external validation	[28]

Abbreviations: SD Study Design, DL Deep Learning, R Retrospective, P Prospective, MRF Multiresolution Radiomic Feature Model, NFRF Non-Fractal Radiomic Feature Model, SVM Support Vector Machine, KNN K-Nearest Neighbors, TSS Training Set Size, VSS Validation Set Size

GBM recurrence vs. radiation necrosis

The differentiation between postoperative recurrent brain tumors (recurrent brain tumors [rBT], specifically recurrent GBM) and RN remains a core challenge in neuro-oncology diagnosis and clinical management. Although both entities appear as contrast-enhancing lesions on conventional MRI, their required management strategies differ substantially. Aggressive therapeutic interventions (e.g., reoperation, targeted therapy) are indicated for rBT, while RN generally warrants conservative management (e.g., observation, symptomatic treatment) [98, 113]. In this context, radiomics—a paradigm leveraging high-throughput quantitative feature extraction coupled with ML—has emerged as a promising quantitative approach.

Across existing studies, several common methodological patterns are evident, as reflected in Table 2. Most investigations are retrospective and single-center, with training cohort sizes typically ranging from about 80 to 160 patients, while RN usually accounts for less than one-third of cases, introducing class imbalance that may bias model performance. External validation is relatively uncommon, and when performed, validation cohorts are often small (< 50 cases). Reported validation AUCs generally fall within the 0.80–0.91 range, whereas training AUCs frequently exceed 0.90, indicating possible overfitting. In terms of imaging inputs, models most often combine conventional MRI (CE-T1WI, T2WI/FLAIR) with quantitative sequences such as diffusion (ADC) and perfusion metrics (CBV or Ktrans). A smaller subset incorporates metabolic PET, which provides complementary biological information beyond structural MRI.

Representative studies illustrate these broader trends. Wang et al. developed an integrated radiomics model combining multimodal PET (¹⁸F-FDG and Carbon-11 Methionine [¹¹C-MET]), MRI, and clinical variables, achieving high diagnostic performance (AUC 0.988 in training and 0.914 in validation) [108]. This study exemplifies the added value of metabolic imaging, as PET-derived tumor-to-background ratios helped capture proliferative activity that conventional MRI cannot reliably depict. Similarly, Park et al. demonstrated that diffusion-based radiomics could maintain diagnostic performance in an external validation cohort (validation AUC ≈ 0.80) [109]. Their findings support the notion that ADC-derived features are relatively robust across institutions, likely because diffusion metrics are quantitative and less dependent on scanner-specific intensity normalization than conventional MRI sequences.

To enhance model generalization and bridge the gap between radiomic biomarkers and clinical outcomes, Sadique et al. introduced a different methodological angle by combining multiresolution fractal (MRF) features with conventional radiomics and incorporating survival analysis [110]. Their model achieved competitive classification performance (AUC ≈ 0.89) and suggested potential prognostic relevance of fractal features. This approach reflects a growing interest in extending radiomics beyond classification toward outcome-oriented modeling.

Despite variations in feature engineering and modeling strategies, several feature categories recur across studies. Diffusion-derived texture features are frequently selected, likely reflecting differences in cellular density between viable tumor and necrotic tissue. Perfusion heterogeneity metrics (e.g., CBV, Ktrans variability) also show consistent discriminatory value, aligning with the contrast between tumor neovascularization and radiation-induced vascular damage. When available, PET metabolic parameters provide additive information, supporting multimodal integration as a key performance driver.

Overall, current evidence suggests that quantitative diffusion and perfusion features are among the more reproducible imaging biomarkers for differentiating rBT from RN. However, broader clinical translation will depend on improved standardization of radiomics workflows, automated segmentation, multicenter validation, and closer integration with molecular and clinical data to enhance robustness and biological interpretability [94].

GBM recurrence vs. pseudoprogression

The standard treatment regimen of GBM typically consists of maximal safe surgical resection followed by concurrent chemoradiotherapy. However, within 3 months of treatment, new or enlarging contrast-enhancing lesions frequently appear on follow-up imaging, creating substantial diagnostic uncertainty between true tumor recurrence and PsP. PsP typically reflects transient inflammatory or treatment-related vascular injury and may resolve without intervention, whereas tumor recurrence represents active neoplastic proliferation requiring prompt therapeutic escalation [114]. This distinction has become even more complex in the era of immunotherapy, as immune-mediated inflammatory responses may further mimic radiographic progression, while immune-related pseudoprogression can reflect treatment-associated immune activation rather than true tumor recurrence [95–97, 115, 116]. Because biopsy is invasive, subject to sampling error, and not feasible in all patients, reliable noninvasive imaging biomarkers are critically needed.

Radiomics has therefore been explored as a quantitative approach for differentiating PsP from recurrence. Similar to studies on radiation necrosis, most PsP-focused radiomics investigations are retrospective and single-center, with training cohorts typically ranging from about 60 to 100 patients, as summarized in Table 2. PsP cases often represent a minority subgroup, again introducing class imbalance. External validation is performed in only a subset of studies, and validation cohorts are generally small. Reported validation AUCs commonly fall within the 0.85–0.90 range, suggesting moderate-to-good diagnostic performance but also reflecting limited large-scale confirmation.

Methodologically, diffusion and perfusion imaging form the backbone of most PsP models. Quantitative ADC features capture cellular density differences, while perfusion parameters (e.g., CBV, Ktrans) reflect vascular permeability and neoangiogenesis, which differ between inflammatory treatment effects and active tumor growth. Multimodal integration consistently outperforms single-sequence approaches, indicating that PsP and recurrence differ along multiple biological dimensions rather than a single imaging parameter.

Representative studies illustrate these trends. Kim et al. developed a multiparametric radiomics model combining diffusion and perfusion MRI features, achieving strong diagnostic performance (AUC ≈ 0.90 in training and ≈ 0.85 in external validation) [111]. Their work highlights the added value of integrating DWI and perfusion-weighted imaging (PWI) features, as integrated models outperformed individual parameters such as ADC or CBV alone. Similarly, Elshafeey et al. constructed radiomics models using perfusion-derived parameters (Ktrans and rCBV) from DSC and DCE imaging, reporting an AUC of approximately 0.89 [112]. Together, these studies emphasize the central role of diffusion and perfusion heterogeneity in distinguishing PsP from true tumor recurrence.

Across studies, diffusion-derived texture features and perfusion heterogeneity metrics are the most consistently retained radiomic predictors, aligning with the biological contrast between inflammatory tissue injury and proliferative tumor vasculature. However, integration of clinical and molecular variables remains uncommon, despite known associations between MGMT promoter methylation and the likelihood of PsP. As a result, many current models provide limited biological contextualization. Overall, while radiomics demonstrates reproducible promise for differentiating PsP from recurrence, broader clinical translation will require standardized imaging and preprocessing pipelines, larger multicenter cohorts, and closer integration of imaging features with molecular and clinical data to enhance robustness and interpretability.

GBM recurrence vs. treatment-related effects

TrE typically refer to local tissue inflammation, necrosis, or changes in vascular permeability induced by RT or chemotherapy [117]. These therapy-associated changes may manifest as new or enlarging enhancing lesions and surrounding edema on follow-up imaging, often resembling the radiographic appearance of tumor recurrence. Misdiagnosis may lead to overtreatment (e.g., unnecessary reoperation) or delayed intervention, directly affecting patient quality of life and prognosis. Therefore, there is a strong clinical need for reliable noninvasive approaches to improve differentiation between TrE and true tumor recurrence [118].

Radiomics-based modeling has been explored to address this challenge, with emerging evidence supporting the value of multiregional and multimodal feature integration. Ren et al. developed a machine learning–based radiomics model incorporating features from postoperative contrast enhancement and peritumoral edema across conventional MRI sequences, demonstrating that combined regional analysis outperformed single-region or single-sequence approaches [28]. The optimal model achieved a test AUC of approximately 0.96, highlighting the potential diagnostic advantage of integrating complementary imaging habitats.

Importantly, feature-level analyses suggested that radiomic descriptors derived from postoperative enhancement were associated with vascular-related alterations, whereas edema-based texture features provided additional discriminatory information, potentially reflecting infiltrative tumor growth and heterogeneous treatment-induced tissue responses. These findings support the concept that TrE and tumor recurrence differ not only within enhancing tumor cores but also in the surrounding peritumoral microenvironment.

Collectively, these results indicate that quantitative radiomics may provide complementary information beyond conventional MRI for postoperative glioma assessment, although further multicenter validation is required before routine clinical adoption.

Other clinical applications of radiomics in GBM

Prognostic prediction

Accurate prognosis assessment is crucial for clinical management of patients with GBM. Radiomics models that integrate MRI sequences have shown robust performance in predicting OS [119]. For instance, Liu et al. developed a deep learning-based radiomics model using contrast-enhanced T1-weighted imaging (CE-T1WI) to predict OS in IDH-wildtype GBM patients after maximal safe resection [120]. This model, which integrated a ResNet-based segmentation network and the Mime machine learning framework, achieved concordance indices of 0.89, 0.81, and 0.76 in the training, internal, and external validation cohorts, respectively. These results underscore the value of automated radiomics in individualized survival prediction. Similarly, Zhang et al. established a CE-T1WI-based radiomic signature consisting of six features, which effectively predicted OS in GBM patients [121]. Notably, this model underwent externally validation across multiple centers and was found to be significantly associated with biological processes, such as DNA repair and glycolysis. Collectively, these studies highlight the clinical relevance and biological plausibility of MRI-based radiomics for multidimensional prognostic evaluation in GBM.

Prediction of molecular alterations

The 2021 WHO CNS5 classification highlights the crucial role of molecular markers in GBM diagnosis. Radiomics offers a noninvasive alternative for detecting key molecular alterations, including IDH mutation and MGMT promoter methylation [122].

For IDH mutation prediction, Li et al. developed a multiregional radiomics model based on multiparametric MRI, achieving an accuracy of 96% and an AUC of 0.90 using a random forest classifier [123]. Incorporation of clinical variables, particularly patient age, further improved model performance (accuracy 97%, AUC 0.96), highlighting the complementary value of integrating imaging-derived biomarkers with established clinical predictors. Across studies, radiomic features contributing to IDH prediction predominantly include intensity- and texture-based metrics extracted from contrast-enhanced and perfusion-sensitive sequences, which may reflect IDH-associated differences in angiogenesis, cellular density, and intratumoral heterogeneity [124, 125]. In addition, shape-related descriptors capturing tumor compactness and margin regularity have been repeatedly associated with IDH-mutant gliomas, consistent with their relatively circumscribed growth patterns compared with IDH-wildtype tumors [126, 127].

MGMT promoter methylation has likewise been extensively investigated using radiomics-based approaches. Do et al. applied a genetic algorithm–based feature selection strategy combined with machine learning classifiers, achieving an accuracy of 0.925, sensitivity of 0.894, and specificity of 0.966 [128]. From a radiologic–biological perspective, MGMT methylation status is closely linked to treatment response and therapy-induced tissue alterations, including necrosis and residual tumor burden [129, 130]. Correspondingly, predictive radiomic signatures often comprise first-order intensity features and texture metrics derived from gray-level co-occurrence and run-length matrices, which may encode imaging phenotypes associated with differential therapeutic sensitivity and post-treatment heterogeneity [131–133].

Beyond these canonical genomic and epigenetic alterations, glioblastoma progression is driven by a broader spectrum of biological programs, including metabolic reprogramming, genomic instability, and dynamic tumor–microenvironment interactions [134, 135]. Altered amino acid metabolism, for example, has been implicated in tumor growth, immune evasion, and therapy resistance, suggesting additional imaging-correlated biomarkers that may be indirectly captured by radiomic features [136, 137].

Genomic instability arising from mitotic checkpoint dysregulation further contributes to intratumoral heterogeneity. Experimental studies have demonstrated that disruption of spindle assembly checkpoint kinases such as Mps1 induces DNA damage and triggers compensatory epigenetic remodeling via SETD2 upregulation, thereby promoting tumor plasticity and molecular diversification [138]. In parallel, multi-omics analyses implicate AP-1–mediated transcriptional programs, stress-response ion channels (e.g., TRPM2), and protease signaling pathways such as ADAM15 in invasion, microenvironmental remodeling, and recurrence-associated adaptation [139–142].

Although current radiomics models typically target discrete molecular alterations, the imaging phenotypes they exploit are likely shaped by the downstream consequences of these interconnected biological processes. Collectively, these findings suggest that radiomics-based molecular prediction extends beyond purely statistical pattern recognition and may partially encode biologically meaningful variations in tumor vascularity, growth behavior, metabolic state, and treatment response, supporting its potential role as a surrogate imaging biomarker in molecularly informed GBM management [143].

Current challenges and future prospects

Radiomics has emerged as a promising quantitative methodology in GBM management, demonstrating considerable potential for noninvasive prediction, spatial localization, and discrimination of tumor recurrence. By decoding tumor phenotype and heterogeneity from routine medical images, it may augment clinical decision-making and support more personalized therapeutic strategies. However, despite these advances, radiomics should not yet be regarded as a replacement for established diagnostic or therapeutic approaches, but rather as a complementary decision-support tool. Its broader clinical adoption remains contingent upon overcoming several important technical and translational challenges. A major concern relates to the reliability and generalizability of current radiomics models, including the high risk of bias and overfitting. Most studies are retrospective and based on single-center cohorts with limited sample sizes, introducing selection bias and creating a data environment that is highly susceptible to overfitting [111]. In high-dimensional radiomics settings, where the number of extracted features far exceeds the number of patients, models may capture dataset-specific noise rather than true biological signal. This can lead to unstable feature selection, inflated internal validation performance, and marked performance degradation when applied to external cohorts, thereby limiting clinical reliability. The lack of independent external validation, particularly multicenter prospective validation, further limits confidence in real-world applicability [91]. Even when extended to multicenter settings, additional sources of variability emerge. Differences in MRI acquisition parameters, including magnetic field strength, echo time (TE), repetition time (TR), and vendor-specific protocols, introduce scanner-related batch effects that compromise feature stability and cross-center reproducibility [112]. Furthermore, variability in image preprocessing and lesion segmentation, including manual, semi-automated, and fully automated approaches, can introduce additional interobserver and inter-institutional inconsistencies, thereby affecting the robustness of extracted radiomic features. These factors underscore the need for standardized acquisition protocols and post-acquisition harmonization strategies to ensure reproducible model performance across sites. In addition, inconsistencies in clinical reference standards, recurrence definitions (e.g., Response Assessment in Neuro-Oncology [RANO] vs. iRANO), and histopathologic ground truth for treatment-related effects introduce further heterogeneity that complicates model evaluation [28, 92]. In addition to data heterogeneity and workflow variability, the limited biological interpretability of many radiomic features and the frequent omission of molecular heterogeneity also constrain the mechanistic relevance and clinical credibility of current models [29, 109]. Methodological inconsistencies also remain a nontrivial challenge. Model performance is reported using heterogeneous evaluation metrics, such as accuracy, AUC, sensitivity, and specificity, often under different validation strategies and cohort splits. The lack of standardized evaluation criteria makes direct benchmarking difficult and may obscure which models are truly robust and clinically generalizable. Furthermore, most radiomics approaches rely on single time-point imaging and therefore fail to adequately capture the dynamic evolution of post-treatment lesions, which is often essential for distinguishing recurrence from treatment-related effects in clinical practice [144].

Addressing these challenges requires coordinated methodological refinement across the radiomics workflow. Greater emphasis on prospective multicenter collaboration and harmonized imaging protocols will be essential to improve data consistency and external validity. Automated and biologically informed segmentation strategies, particularly deep learning–based approaches capable of capturing both enhancing tumor and peritumoral habitats, may reduce interobserver variability while better representing infiltrative tumor biology. Feature engineering is also expected to evolve beyond static handcrafted descriptors toward delta-radiomics, habitat imaging, and radiogenomic integration, enabling models to better reflect tumor evolution and underlying molecular heterogeneity. At the modeling stage, future efforts should prioritize clinical translatability through rigorous external validation, calibration assessment, and decision-curve analysis to ensure that predictive performance translates into meaningful clinical benefit. Multimodal integration combining imaging features with molecular biomarkers and clinical variables may further provide a more comprehensive and biologically grounded characterization of recurrent disease. In parallel, technical optimization, including improved computational efficiency, standardized preprocessing pipelines, and harmonization strategies (e.g., ComBat), will be necessary to facilitate real-world clinical implementation [29, 112].

Looking ahead, radiomics in GBM is likely to progress from retrospective image analysis toward dynamic, integrated clinical decision-support systems. Longitudinal modeling based on serial imaging may enable earlier identification of recurrence and treatment-related effects by capturing temporal evolution rather than relying solely on single time-point assessment. Advances in artificial intelligence, including deep learning–driven feature learning and self-supervised representation learning, may further enhance the robustness of imaging biomarkers while reducing dependence on handcrafted features. Integration with radiogenomics and liquid biopsy biomarkers, such as circulating tumor DNA, could facilitate noninvasive molecular stratification and real-time monitoring of tumor behavior [145–149]. Alongside advances in precision imaging and AI-driven decision support, the therapeutic landscape of GBM is also expanding toward nonthermal, nonchemical physical intervention strategies. Early studies indicate that ultrashort pulsed electric fields can disrupt glioma tumor architecture, while resonant magnetic field–based technologies are being explored for controlled bioelectromagnetic modulation [150, 151]. Ultimately, radiomics may help drive a shift from static, one-time treatment decisions toward adaptive precision neuro-oncology, enabling continuous risk stratification, therapy monitoring, and individualized intervention throughout the GBM disease course while balancing survival benefit with preservation of neurological function and quality of life.

In summary, advances in quantitative imaging and computational modeling have accelerated the development of radiomics and related imaging-based approaches, offering new opportunities to support clinical decision-making in glioblastoma management. Given the aggressive biology of GBM and the persistent difficulty in predicting and identifying tumor recurrence using conventional imaging alone, reliable noninvasive tools for recurrence assessment remain a critical unmet need. Accurate prediction and discrimination of recurrence from treatment-related effects have the potential to facilitate earlier intervention, optimize treatment selection, and improve patient stratification during post-therapy surveillance. In this review, we systematically summarized current radiomics-based methodologies applied to glioblastoma recurrence, including recurrence risk prediction, spatial localization of recurrence-prone regions, and differentiation between true tumor progression and treatment-related imaging changes. Although existing studies are limited by retrospective designs, small cohorts, and methodological heterogeneity, the accumulating evidence suggests that radiomics may provide complementary quantitative information beyond conventional imaging. Continued efforts toward standardized imaging protocols, robust validation, and integration with clinical and molecular data will be essential to advance these approaches toward reliable clinical translation.

Abbreviations

¹¹C-MET

Carbon-11 Methionine

¹⁸F-FDG

Fluorine-18 Fluorodeoxyglucose

¹⁸F-FET

Fluorine-18 Fluoroethyl-L-tyrosine

ADC

Apparent Diffusion Coefficient

AUC

Area Under the Curve

BBB

Blood–Brain Barrier

CBV

Cerebral Blood Volume

CE-T1WI

Contrast-Enhanced T1-Weighted Imaging

CT

Computed Tomography

DCE

Dynamic Contrast-Enhanced

DR

Distant Recurrence

DSC

Dynamic Susceptibility Contrast

DTI

Diffusion Tensor Imaging

DWI

Diffusion-Weighted Imaging

FET

Fluoroethyl-L-tyrosine

FLAIR

Fluid-Attenuated Inversion Recovery

GBM

Glioblastoma Multiforme

GLCM

Gray Level Co-occurrence Matrix

GLDM

Gray Level Dependence Matrix

GLRLM

Gray Level Run Length Matrix

GLSZM

Gray Level Size Zone Matrix

GTR

Gross Total Resection

IDH

Isocitrate Dehydrogenase

KNN

K-Nearest Neighbors

KPS

Karnofsky Performance Status

LASSO

Least Absolute Shrinkage and Selection Operator

LOOCV

Leave-One-Out Cross-Validation

LR

Local Recurrence

MGMT

O6-Methylguanine-DNA Methyltransferase

ML

Machine Learning

MRI

Magnetic Resonance Imaging

MRF

Multiresolution Radiomic Feature

NFRF

Non-Fractal Radiomic Feature

OS

Overall Survival

PET

Positron Emission Tomography

PoC

Proof of Concept

PsP

Pseudoprogression

Rad-score

Radiomics Score

RANO

Response Assessment in Neuro-Oncology

RFC

Random Forest Classifier

RF

Random Forest

RN

Radiation Necrosis

ROI

Region of Interest

RT

Radiotherapy

SCN

Stem Cell Niche

SD

Study Design

SVC

Support Vector Classifier

SVM

Support Vector Machine

SVMPE

Support Vector Machine with Proximity Estimation

SVZ

Subventricular Zone

TE

Echo Time

TERT

Telomerase Reverse Transcriptase

TMZ

Temozolomide

TR

Repetition Time

TrE

Treatment-related Effects

TSS

Training Set Size

VEGF

Vascular Endothelial Growth Factor

VSS

Validation Set Size

WHO

World Health Organization

XGBoost

Extreme Gradient Boosting

rCBV

relative Cerebral Blood Volume

rBT

recurrent brain tumors

PWI

Perfusion-weighted imaging

Author contributions

Zhang Tianyun, Zhu Haoliang, Sun Hangzhe, Chen Yu, Sun Xingjian, Wang Bowen, and Zhu Yang performed the literature search and drafted the manuscript. Wu Yiwen substantially contributed to the revision of the manuscript, including updating and synthesizing relevant literature, reorganizing key sections, refining the conceptual framework, and revising the manuscript in response to reviewers’ comments. Zhang Anke, Wang Kankai, and Pan Yuanbo provided critical revisions to the intellectual content and supervised the work. Pan Yuanbo conceived the study design and approved the final manuscript. All authors read and approved the final version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 82501547) and the China Postdoctoral Science Foundation (Grant No. 2023TQ0284).

Data availability

Not applicable. All data discussed in this review are taken from publicly available literature; no new datasets were generated or analysed in this work.

Declarations

Ethics approval and consent to participate

Not applicable. This manuscript is a review of previously published literature and does not involve new studies with human participants or animals.

Consent for publication

Not applicable. This manuscript does not include any individual person’s data in any form (including images or videos).

Competing interests

The authors declare that they have no competing interests.