Heterogeneity phenotypes in recurrent glioblastoma: a multimodal MRI-based spatial mapping framework for precision treatment

Abstract

Background

To develop a multimodal magnetic resonance imaging (MRI)-based spatial mapping framework for quantitatively characterizing intratumoral heterogeneity in recurrent glioblastoma (rGBM), identifying distinct imaging subregions, and classifying heterogeneity phenotypes predictive of treatment response and survival outcomes.

Methods

A total of 140 rGBM patients were recruited and underwent standardized diffusion-weighted imaging (DWI) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Pixel-wise colocalization of apparent diffusion coefficient (ADC) and DCE-MRI features identified four Multimodal Imaging Subregions (MIS). Entropy and Moran’s I quantified heterogeneity, and hierarchical clustering defined imaging phenotypes. Treatment response to 1-(2-chloroethyl)-3-cyclohexyl-1-nitrosourea (CCNU), bevacizumab (Bev) + stereotactic radiotherapy (SRT), and Bev + CCNU was assessed by volumetric and component-level changes. Survival analyses were performed using Kaplan–Meier and multivariate Cox models.

Results

MIS4, defined by low ADC and slow-rising enhancement, was consistently treatment-resistant. Three imaging phenotypes with distinct heterogeneity patterns demonstrated significant prognostic stratification across regimens. Phenotype A showed the best outcomes under Bev-based regimens, while Phenotype B responded better to CCNU. Imaging phenotypes independently predicted progression-free survival (PFS) and overall survival (OS).

Conclusion

This framework enables spatially resolved, phenotype-based analysis of rGBM heterogeneity using routine MRI. Imaging phenotypes serve as non-invasive biomarkers to guide personalized treatment planning and outcome prediction in recurrent glioblastoma.

Clinical trial registration number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-01929-1.

Background

Recurrent glioblastoma (rGBM) remains one of the most formidable challenges in neuro-oncology, characterized by aggressive invasion and marked therapeutic resistance, resulting in limited treatment benefits, poor survival outcomes, and high mortality rates [1, 2]. Although current guidelines, such as those from the National Comprehensive Cancer Network (NCCN) [3] and European Society for Medical Oncology (ESMO) [4], provide recommendations for the management of rGBM, treatment strategies remain highly individualized, with no universally accepted standard of care [5, 6]. A deeper understanding of tumor heterogeneity in rGBM patients may facilitate the optimization of therapeutic approaches and ultimately improve overall patient outcomes [7, 8].

Heterogeneity represents the most critical biological hallmark of recurrent rGBM, underlying the pronounced inter-patient variability in tumor behavior [9, 10]. rGBM exhibits substantial molecular heterogeneity in the expression of key genes such as IDH, MGMT, EGFR, TERT, and TP53, often diverging from their profiles at initial diagnosis [11]. Furthermore, individual tumors demonstrate spatiotemporal variability in gene expression. This genetic heterogeneity contributes to a highly complex tumor microenvironment [12, 13], characterized by heterogeneous interplay among tumor cells, cancer stem cells, stromal components, and immune infiltrates, as well as diversified angiogenic pathways [14]. These molecular and cellular alterations collectively manifest at the macroscopic level as inconsistent tumor texture, variable vascularization, and markedly heterogeneous spatial distribution of biologically active tumor components [15].

Previous studies have consistently demonstrated the pivotal role of tumor heterogeneity in guiding therapeutic decision-making for rGBM. Lin [16] compared primary and recurrent GBM specimens and reported significantly increased heterogeneity at the cellular, molecular, and genetic levels in rGBM, highlighting its biological complexity and therapeutic challenges. Dadario [17] investigated the major molecular subtypes of rGBM including classical, neural, proneural, and mesenchymal subtypes and their associated treatment strategies, underscoring the importance of tailoring therapies based on molecular heterogeneity. Bagley [12] provided a comprehensive overview of recent advances in rGBM immunotherapy, encompassing CAR-T cell therapy, immune checkpoint inhibitors, cancer vaccines, and oncolytic virotherapy, and emphasized the profound impact of intratumoral heterogeneity on immunotherapeutic efficacy. In recent years, several research centers have evaluated the potential of imaging-derived biomarkers in rGBM treatment stratification [18–20]. Despite the inherent risk of bias in such studies [21], this emerging field holds substantial promise for clinical translation and plays a critical role in optimizing treatment strategies for rGBM.

Tumor proliferative cells and tumor microvasculature constitute the principal structural components of the tumor microenvironment [22], exerting direct influence on tumor biological behavior and macroscopic imaging features. Elucidating their spatial organization is critical for advancing the clinical utility of rGBM heterogeneity in guiding and optimizing therapeutic strategies, ultimately enhancing patient outcomes. While previous habitat imaging and radiomic clustering studies stratified glioblastoma using multiparametric MRI features [23, 24], our framework advances this approach by integrating pixel-wise colocalization of cellular diffusion and vascular status information to derive biologically interpretable multimodal imaging subregions (MIS). We further quantify their compositional and spatial heterogeneity using entropy and Moran’s I, and link these imaging-derived phenotypes to treatment-specific responses and survival across multiple centers, emphasizing both methodological transparency and clinical translatability. In this study, we established a heterogeneity analysis framework based on diffusion-weighted imaging (DWI) and dynamic contrast-enhanced (DCE) MRI, aiming to quantitatively characterize the spatial heterogeneity of rGBM. Concurrently, we evaluated the treatment responsiveness of distinct heterogeneous components and explored the identification of heterogeneity-driven phenotypes to further inform personalized therapeutic decision-making.

Materials and methods

Patient cohort

This prospective study was approved by the institutional ethics review board. Written informed consent was obtained from all participants. Patients with rGBM were prospectively recruited from three independent medical centers between May 2020 and February 2023. The discovery cohort comprised patients from the Affiliated Hospital of Yangzhou University and the Affiliated Hospital of Jiangsu University, while the validation cohort included patients from the Affiliated People’s Hospital of Jiangsu University. Inclusion criteria were as follows: (1) age > 18 years; (2) histopathologically confirmed glioblastoma (IDH-1 wild-type) after surgical resection; (3) adjuvant chemoradiotherapy administered according to the Stupp protocol; (4) first recurrence confirmed by serial MRI follow-up and evaluated by the multidisciplinary team based on the Response Assessment in Neuro-Oncology (RANO) criteria [25]; (5) ineligibility for or refusal to undergo repeat tumor resection; (6) treatment with one of the following regimens based on individual clinical characteristics and molecular profiles: 1-(2-chloroethyl)-3-cyclohexyl-1-nitrosourea (CCNU) monotherapy, bevacizumab combined with stereotactic radiotherapy (Bev + SRT), or bevacizumab combined with CCNU (Bev + CCNU). Patients were excluded according to the following criteria: (1) presence of other intracranial malignancies; (2) comorbid severe neurological disorders such as stroke or epilepsy; (3) significant dysfunction of the cardiovascular, pulmonary, hepatic, renal, or coagulation systems, or unstable vital signs; (4) history of neurosurgical procedures other than GBM resection; (5) incomplete clinical follow-up data; (6) absence of paired pre- and post-treatment MRI scans for the first treatment cycle or failure to meet MRI quality control standards required by the study. All enrolled patients underwent survival follow-up for 12 months starting from the time of first recurrence. Progression-free survival (PFS) was recorded from the date of recurrence to the date of radiological progression or loss to follow-up. Overall survival (OS) was defined from the date of recurrence to death or loss to follow-up. MRI examinations were performed within one week before treatment initiation and upon completion of the first therapeutic cycle, serving as the primary imaging data for comparative analysis. Thereafter, MRI follow-up was conducted every two weeks until the end of follow-up or death.

This study was approved by the Institutional Ethics Committees of all participating centers. Written informed consent was obtained from all participants.

MRI data acquisition

All patients underwent conventional magnetic resonance imaging (MRI) on a 3.0-T scanner, including axial T1-weighted imaging (T1WI), sagittal T1WI, axial T2WI, fluid-attenuated inversion recovery (FLAIR), diffusion-weighted imaging (DWI), as well as dynamic contrast-enhanced (DCE)-MRI. Detailed information regarding the imaging platforms and acquisition parameters is provided in the Supplementary Materials (Doc S1, Table S1).

MRI post-processing

Image preprocessing

Axial T1WI, DWI, and DCE-MRI scans obtained before and after treatment were manually imported into 3D Slicer (version 5.6.1; https://www.slicer.org). To address potential spatial misalignment between different imaging sequences, ADC images (calculated from DWI sequences) were rigidly or affinely registered to the corresponding DCE phases to achieve spatial correspondence. To ensure consistency in image dimensions across participating institutions, all registered images were resampled to a unified matrix size with isotropic voxel spacing (1 × 1 × 1 mm³). Subsequently, intensity normalization was performed, ADC maps were standardized using Z-score normalization within a brain parenchyma mask, while DCE images were normalized using a local maximum intensity-based method.

Tumor segmentation and extraction

Tumor regions of interest prior to treatment (VOI_pre) were automatically segmented using the deep learning-based model provided by Oncohabitats (https://www.oncohabitats.upv.es) [26]. The generated masks encompassed both the contrast-enhancing tumor core and the central necrotic region. These VOI_pre masks were subsequently mapped onto the standardized pre-treatment ADC and DCE images to define VOI_pre−ADC and VOI_{pre−Enhancement}, respectively. Using the same approach, tumor regions of interest following treatment were delineated to generate VOI_post−ADC and VOI_{post−Enhancement}.

Generation of tumor imaging subregions

For each individual, an unsupervised kernel density estimation (KDE) approach was applied to the ADC maps to construct a continuous probability density function (PDF) of voxel intensities within the tumor region. The ADC value corresponding to the peak of the distribution was defined as the mode. Voxels with ADC values greater than the mode were classified as ADC-high, whereas those below the mode were classified as ADC-low. For the DCE images, a temporal intensity curve (TIC) pattern analysis was performed across multiple dynamic phases to automatically categorize each voxel within the tumor region into distinct enhancement patterns (rapid wash-in and wash-out, plateau, or slow-rising enhancement types). Following spatial co-registration of the VOI_pre−ADC and VOI_{pre−Enhancement} volumes, voxel-wise fusion was performed based on the binary classification of ADC and the pattern classification of DCE. This fusion generated a set of new multimodal imaging subregions, termed Multimodal Imaging Subregions (MIS), representing distinct tumor subregions. The corresponding MIS-based masks were then projected back onto the individual patient’s VOI_pre to construct a spatial heterogeneity map based on MIS distribution. Using the same methodology, post-treatment MIS was derived and mapped onto the VOI_post to generate corresponding heterogeneity maps. All image normalization, KDE estimation, classification, and fusion procedures were implemented using Python 3.12.

Quantification of tumor imaging subregions and components

Quantitative analysis of MIS and their constituent components was performed using the 3D Slicer. All measurements were based on pixel-wise volumetric calculations. Each MIS was assigned a unique index label, and its total volume was denoted as V_MIS, subsequently, tumor entropy was calculated to quantify the compositional complexity of the intratumoral MIS [27], whereas Moran’s I was applied to evaluate the spatial distribution homogeneity of the MIS within the tumor [28]. Given that each MIS results from the spatial colocalization of ADC and DCE derived voxels, the respective volumes within each MIS were further decomposed into ADC-based volume (V_ADC) and DCE-based volume (V_Enhancement), calculated via decoupling of the colocalized pixel contributions. To evaluate changes in MIS and their components before and after treatment, the difference in MIS volume (ΔV_MIS) was computed based on index label correspondence. Boolean operations (intersection and difference) on spatial masks were used to localize and quantify changes in individual components, yielding ΔV_ADC and ΔV_Enhancement. Within these, changes involving simultaneously altered colocalized voxels were defined as ΔV_overlap, while component-specific changes were recorded as ΔV_ADC−only and ΔV_{Enhancement−only}, respectively.

After quantifying the MIS subregions, we further evaluated the associations between MIS composition and key biological/clinical indicators, including O6-methylguanine-DNA methyltransferase (MGMT) status, Karnofsky Performance Status (KPS), tumor volume, edema severity, postoperative to recurrence interval (PRI), and survival outcomes (OS and PFS). The proportions of MIS1-4 were extracted from the discovery cohort, and correlation coefficients (rho) were computed. The results were visualized using a heatmap to illustrate both the direction and magnitude of the correlations.

Treatment protocols

Stereotactic radiotherapy (SRT) planning and dosimetry

Stereotactic radiotherapy (SRT) was delivered using volumetric-modulated arc therapy (VMAT) with image guidance. Patients were immobilized with a thermoplastic mask, and planning CT was rigidly registered with contrast-enhanced MRI. The gross tumor volume (GTV) was defined as the enhancing lesion, expanded by 2–3 mm to create the clinical target volume (CTV), with an additional 1–2 mm for the planning target volume (PTV). The prescribed dose was 35–40 Gy in 5 fractions (7–8 Gy per fraction), respecting OAR constraints [29] (Table S2). Dose calculations were performed using the RayStation 12 A treatment planning system with Monte Carlo algorithms (dose-to-medium). End-to-end dosimetric accuracy was verified via EPID/film validation, with point dose errors < 3%. For translational analysis, mean dose (D_mean) and coverage (D₉₅) were recalculated for each multimodal imaging subregion (MIS). Dose mapping to heterogeneous subregions followed an established voxel/region-wise distribution analysis framework [30].

Chemotherapy regimens

For patients treated with CCNU monotherapy, lomustine (CCNU) was administered orally at a standard dose of 110 mg/m² every 6 weeks, consistent with the EORTC 26,101 phase III trial [31].

For the Bev + CCNU combination regimen, treatment followed the CeTeG/NOA-09 [32] and BELOB [33] trial protocols, with lomustine given at 90 mg/m² every 6 weeks in combination with bevacizumab 10 mg/kg intravenously every 2 weeks. Dose modifications, treatment delays, or discontinuations were implemented according to hematologic and non-hematologic toxicities, in alignment with established clinical trial guidelines.

Statistical analysis

Statistical analyses and data visualization were performed using IBM SPSS Statistics version 27.0 and GraphPad Prism version 10.1.2. A two-sided P value ≤ 0.05 was considered statistically significant. Data with normal distribution and homogeneity of variance were expressed as mean ± standard deviation (SD), whereas non-normally distributed data were presented as median with range. Categorical variables were summarized as counts and percentages. Intergroup comparisons were conducted using analysis of variance (ANOVA). Correlation analyses were performed using either the Pearson or Spearman method, depending on data distribution.

Hierarchical clustering for rGBM phenotypic classification was performed in Python 3.12. To assess robustness, alternative clustering algorithms (k-means, Gaussian mixture, and spectral clustering) were applied with cluster numbers ranging from 2 to 5. Internal validity indices (Silhouette, Calinski–Harabasz, Dunn, and Davies–Bouldin) were calculated, and stability was further evaluated using 1,000 bootstrap resampling as well as leave-one-center-out (LOCO) and leave-one-feature-out (LOFO) analyses.

Kaplan–Meier survival analysis with log-rank tests was employed to evaluate the prognostic stratification capacity of imaging phenotypes across different treatment regimens. After adjusting for relevant clinical covariates, the independent prognostic value of imaging phenotypes was assessed using a Cox proportional hazards model, with results reported as hazard ratios (HR) and 95% confidence intervals (CI). The incremental contribution of imaging phenotypes to prognostic models was quantified by changes in the concordance index (C-index), likelihood ratio χ² statistics, Akaike/Bayesian information criteria (AIC/BIC), and net reclassification improvement (NRI). Time-dependent AUCs and calibration curves were also generated.

Results

Patient characteristics

A total of 140 eligible patients with rGBM were enrolled in the discovery cohort, including 31 patients in the CCNU monotherapy group (19 males and 12 females; 66.5 ± 9.3 years), 49 patients in the Bev + SRT group (31 males and 18 females; 68.2 ± 7.4 years), and 60 patients in the Bev + CCNU group (36 males and 24 females; 67.3 ± 6.9 years). Significant intergroup differences were observed across three key parameters—peritumoral edema, MGMT promoter methylation status, and postoperative to recurrence interval (PRI) features (p < 0.01), (Table 1) summarizes the clinical characteristics of patients across the three different treatment groups. In addition, an independent validation cohort comprising 77 eligible rGBM patients was established, in which the three treatment groups similarly exhibited significant differences in these parameters (p < 0.01).

Table 1.

Sociodemographic and clinical characteristics of patients with rGBM

Characteristic	CCNU (n = 31)	Bev + SRT (n = 49)	Bev + CCNU (n = 60)	p value
Sex
Male, n (%)	19 (61.29)	31 (63.27)	36 (60.00)	0.962
Female, n (%)	12 (38.71)	18 (36.73)	24 (40.00)
Age(y), mean (SD)	66.5 ± 9.3	68.2 ± 7.4	67.3 ± 6.9	0.064
Tumor locationa
Non-midline-crossing, n (%)	25 (80.65)	45 (91.84)	49 (81.67)	0.249
Midline-crossing, n (%)	6 (19.35)	4 (8.16)	11 (18.33)
Tumor volume(mm3)b, median (range)	29457.31 (1456.27-77245.64)	30884.52 (987.34-80752.02)	31332.55 (987.34-74722.70)	0.844
Edema volume(mm3)c, median (range)	20665.20 (8456.53-30245.61)	48766.23 (24567.07-87245.44)	51432.64 (27666.25-94205.64)	0.000**
KPS score
≥ 60, n (%)	22 (70.97)	33 (67.34)	43 (71.67)	0.879
< 60, n (%)	9 (29.03)	16 (32.65)	17 (28.33)
MGMT (%)
> 30, n (%)	24 (77.42)	21 (42.86)	44 (73.33)	0.001**
≤ 30, n (%)	7 (22.58)	28 (57.14)	16 (26.67)
PRI (m), mean (SD)	4.4 ± 0.9	6.5 ± 1.7	4.2 ± 1.2	0.000**

Notes: Data of age and PRI are expressed as the mean ± standard deviation; Data of tumor volume and peritumoral edema volume are expressed as the median(numbers in parenthesis are range); Data of other characteristics are expressed as the numbers (numbers in parenthesis are percentage)

Abbreviations: CCNU = Cyclohexyl-Chloro-Nitrosourea; Bev = Bevacizumab; SRT = Stereotactic Radiotherapy; KPS = karnofsky performance status; MGMT = O6-methylguanine-DNA methyltransferese; PRI = postoperative to recurrence interval

^a Indicated as the location of the main bulk of tumor

^b Measured based on the DCE sequence

^c Measured based on the T2WI and FLAIR sequences

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Quantification of tumor heterogeneity in rGBM

Pixel-wise colocalization analysis identified four distinct imaging subregions within rGBM tumors, designated as MIS1 (blue), MIS2 (green), MIS3 (yellow), and MIS4 (red) (Fig. 1A). Among them, MIS1 exhibited a relatively stable proportion across patients (21.5–29%), whereas the proportions of MIS2, MIS3, and MIS4 showed substantial variability (MIS2: 4.8–61%; MIS3: 1.1–56.6%; MIS4: 5.3–54.5%) (Fig. 1B). The identification of these subregions demonstrated robust reproducibility and high consistency in the independent validation cohort. These four imaging-derived subregions collectively characterize the spatial heterogeneity of rGBM and exhibit distinct entropy values and Moran’s I indices. Notably, entropy was not significantly correlated with tumor volume (ρ = 0.008, p = 0.928) (Fig. 1C), nor was Moran’s I (ρ = 0.030, p = 0.723) (Fig. 1D), indicating that both metrics serve as volume-independent quantitative descriptors of rGBM heterogeneity.

Fig. 1.

Quantification of Intratumoral Heterogeneity in rGBM. (A) Identification of imaging subregions in rGBM based on pixel-wise colocalization of ADC and DCE modalities, with corresponding habitat maps. (B) Distribution of MIS subregion proportions at the population level. (C) Correlation between MIS entropy values and total tumor volume. (D) Correlation between MIS Moran’s I values and total tumor volume

The correlation heatmap revealed distinct biological and clinical associations across MIS subregions. The MIS4 proportion showed negative correlations with MGMT, KPS, PRI, OS, and PFS, while demonstrating positive correlations with tumor volume and edema severity. In contrast, MIS1 and MIS2 exhibited opposite correlations with these features, whereas MIS3 demonstrated intermediate correlations between MIS4 and MIS1/2 (Fig. S1).

Therapeutic response of rGBM heterogeneity components to different treatment regimens

Following one treatment cycle, tumor volumes exhibited varying degrees of reduction across all three therapeutic regimens, with no statistically significant differences in the extent of shrinkage among groups (p > 0.05) (Fig. 2A). Under CCNU monotherapy, the MIS1 subregion showed the greatest reduction (22%), whereas MIS2 and MIS3 demonstrated comparable decreases (16% and 17%, respectively). In contrast, the Bev + SRT combination resulted in the most pronounced reduction in MIS3 (30%), followed by MIS2 (22%), while MIS1 exhibited only a modest decrease (7%). For the Bev + CCNU combination, both MIS2 and MIS3 displayed the highest reduction rates (24% and 27%, respectively), with no significant difference between them, and MIS1 showed a moderate reduction (15%). However, regardless of treatment modality, all interventions had a minimal effect on reducing MIS4 (CCNU: 5%; Bev + SRT: 9%; Bev + CCNU: 11%) (Fig. 2B-D). Further analysis of intraregional component alterations revealed that CCNU primarily reduced regions characterized by high ADC values. Bev + SRT mainly targeted contrast-enhancing regions, particularly plateau-type enhancement areas. Bev + CCNU exhibited a dual effect, significantly reducing both high ADC regions and contrast-enhancing components. Nevertheless, none of the treatment strategies demonstrated a meaningful impact on reducing slow-rising enhancement regions, which remained largely resistant to intervention (Fig. 2B-D).

Fig. 2.

Therapeutic Response Efficiency of rGBM Tumors and Their Heterogeneous Components. (A) Comparison of tumor volume reduction rates among three treatment regimens. (B–D) Changes in imaging subregions and their respective component proportions under each treatment strategy

Dosimetric validation and MIS dose-response

Dosimetric validation confirmed robust treatment delivery, with median point dose error < 3% and consistent agreement between planned and measured distributions. Across clinical sampling, the median deviation in MIS D_mean and D₉₅ was < 2%, indicating high reproducibility. Exploratory analysis showed no significant correlation between MIS-level dose metrics and volumetric regression (ΔV%) (Table S3).

Imaging phenotype identification and heterogeneity characterization in rGBM

Hierarchical clustering analysis identified three distinct imaging phenotypes of rGBM (optimal cluster number = 3; silhouette coefficient = 0.535) (Fig. 3A-B). These phenotypes exhibited markedly divergent heterogeneity profiles, as reflected by both entropy (p < 0.01) and Moran’s I values (p < 0.01). Specifically, Phenotype A was characterized by a high entropy (range: 0.55–0.73; median: 0.65) and a low Moran’s I (range: -1.00-0.70; median: -0.85), indicating substantial compositional diversity with a spatially dispersed distribution of subregions. Phenotype B exhibited a low entropy (range: 0.46–0.65; median: 0.54) coupled with a high Moran’s I (range: 0.35–0.85; median: 0.65), representing tumors with homogeneous subregional composition but spatially clustered distribution. In contrast, Phenotype C displayed both high entropy (range: 0.65–0.84; median: 0.75) and high Moran’s I (range: 0.55-1.00; median: 0.75), indicative of tumors with pronounced compositional heterogeneity and spatially concentrated subregions (Fig. 3C).

Fig. 3.

Identification and Characterization of rGBM Imaging Phenotypes and Heterogeneity Features. (A–B) Optimal phenotype clustering of rGBM tumors using hierarchical clustering (optimal cluster number = 3). (C) Comparison of heterogeneity metrics across the three identified phenotypes. (D) Habitat distribution and phenotypic characteristics of rGBM imaging subregions across the three phenotypes

Robustness analyses confirmed that the three-phenotype structure was consistently reproduced across alternative clustering algorithms, with stable validity indices (Silhouette 0.51–0.54, CH > 400, DB < 0.65) and high reproducibility on bootstrap and sensitivity tests (median ARI = 0.73, NMI = 0.70) (Table S4). Importantly, the same phenotype classification and heterogeneity patterns were further validated in the independent cohort, with representative maps shown in (Fig. 3D).

Independent prognostic value of imaging phenotypes for survival outcomes across different treatment regimens

In the CCNU group, patients with Phenotype B exhibited the longest PFS and OS, followed by those with Phenotype A, while Phenotype C had the poorest outcomes (Fig. 4A). Multivariate survival analysis demonstrated that, in addition to MGMT promoter methylation serving as an independent predictor for both PFS and OS, imaging phenotype classification was also an independent prognostic factor (Tables 2 and 3). In the Bev + SRT group, patients with Phenotype A achieved the best PFS and OS, followed by Phenotype C, with Phenotype B showing the worst outcomes (Fig. 4B). Notably, multivariate analysis indicated that imaging phenotype alone was an independent predictor of both PFS and OS in this group (Tables 4 and 5). In the Bev + CCNU group, Phenotype A was associated with the most favorable PFS and OS, followed by Phenotype B, and Phenotype C had the poorest prognosis (Fig. 4C). Consistent with the CCNU group, multivariate analysis revealed that both MGMT promoter methylation and imaging phenotype were independent prognostic factors for PFS and OS (Tables 6 and 7). Beyond serving as independent prognostic factors, imaging phenotypes also provided incremental predictive value when added to clinical baseline models. Across endpoints and treatment regimens, incorporation of phenotype classification improved model performance, with C-index gains of 0.06–0.10 and corresponding increases in likelihood ratio χ² (all p ≤ 0.015). Additional metrics, including changes in AIC/BIC, net reclassification improvement, and time-dependent AUCs, consistently supported the added prognostic contribution (Table S5).

Fig. 4.

Survival Differences Among rGBM Imaging Phenotypes Across Treatment Regimens. (A) Kaplan–Meier curves of PFS and OS across the three phenotypes under CCNU treatment. (B) Kaplan–Meier curves of PFS and OS across the three phenotypes under Bev + SRT treatment. (C) Kaplan–Meier curves of PFS and OS across the three phenotypes under Bev + CCNU treatment

Table 2.

Univariate and multivariate analysis of PFS outcomes for rGBM patients in CCNU group (n = 31)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype B	Reference
Phenotype A	0.031 *	1.585 (1.044–2.408)	0.048 *	1.519 (1.004–2.297)
Phenotype C	0.002 **	2.703 (1.812–4.035)	0.001 **	2.846 (1.841–4.399)
Sex (Male)	0.722	1.107 (0.813–1.503)	—	—
Age (≥ 65y)	0.621	1.203 (0.830–1.744)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.799	0.934 (0.715–1.220)	—	—
Tumor volume (≥ 30147.71mm3)	0.439	1.190 (0.867–1.632)	—	—
Edema volume (≥ 28774.38mm3)	0.023*	1.742 (1.091–2.358)	0.181	1.350 (0.896–2.030)
KPS (< 60)	0.208	1.351 (0.914–1.992)	—	—
MGMT (≤ 30)	0.019 *	1.804 (1.124–2.402)	0.031*	1.468 (1.038–2.265)
PRI (≤ 4.1 m)	0.351	1.168 (0.829–1.648)	—	—
Treatment duration (≥ 4.7 w)	0.008 **	1.860 (1.370–2.520)	0.274	1.345 (0.850–2.128)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Table 3.

Univariate and multivariate analysis of OS outcomes for rGBM patients in CCNU group (n = 31)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype B	Reference
Phenotype A	0.027 *	1.672 (1.082–2.585)	0.042 *	1.437 (0.981–2.106)
Phenotype C	0.001 **	2.925 (1.968–4.345)	0.002 **	2.713 (1.706–4.309)
Sex (Male)	0.768	1.083 (0.781–1.501)	—	—
Age (≥ 65y)	0.588	1.214 (0.823–1.791)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.841	0.926 (0.701–1.222)	—	—
Tumor volume (≥ 30147.71mm3)	0.384	1.238 (0.891–1.721)	—	—
Edema volume (≥ 28774.38mm3)	0.017 *	1.684 (1.098–2.313)	0.193	1.304 (0.872–1.951)
KPS (< 60)	0.255	1.322 (0.894–1.952)	—	—
MGMT (≤ 30)	0.013 *	1.745 (1.107–2.369)	0.028 *	1.427 (1.021–2.198)
PRI (≤ 4.1 m)	0.406	1.145 (0.808–1.623)	—	—
Treatment duration (≥ 4.7 w)	0.006 **	1.798 (1.327–2.476)	0.263	1.312 (0.847–2.115)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Table 4.

Univariate and multivariate analysis of PFS outcomes for rGBM patients in Bev + SRT group (n = 49)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype A	Reference
Phenotype C	0.001 **	2.931 (1.952–4.399)	0.002 **	2.778 (1.451–4.375)
Phenotype B	0.021 *	1.706 (1.082–2.689)	0.039 *	1.598 (1.023–2.495)
Sex (Male)	0.688	1.158 (0.814–1.647)	—	—
Age (≥ 65y)	0.575	1.263 (0.823–1.938)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.613	0.912 (0.626–1.330)	—	—
Tumor volume (≥ 29654.22mm3)	0.472	1.137 (0.788–1.641)	—	—
Edema volume (≥ 45672.73mm3)	0.008**	1.956 (1.196–2.501)	0.079	1.382 (0.964–2.147)
KPS (< 60)	0.186	1.429 (0.846–2.413)	—	—
MGMT (≤ 30)	0.158	1.321 (0.903–1.933)	—	—
PRI (≤ 4.4 m)	0.227	1.318 (0.843–2.062)	—	—
Treatment duration (≥ 4.2 w)	0.005 **	1.951 (1.295–2.938)	0.077	1.267 (0.938–1.758)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Table 5.

Univariate and multivariate analysis of OS outcomes for rGBM patients in Bev + SRT group (n = 49)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype A	Reference
Phenotype C	0.002 **	2.842 (1.936–4.220)	0.030*	2.416 (1.311–3.869)
Phenotype B	0.035 *	1.652 (1.072–2.602)	0.046 *	1.573 (1.016–2.438)
Sex (Male)	0.703	1.138 (0.798–1.623)	—	—
Age (≥ 65y)	0.598	1.244 (0.816–1.896)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.632	0.929 (0.636–1.320)	—	—
Tumor volume (≥ 29654.22mm3)	0.488	1.163 (0.798–1.693)	—	—
Edema volume (≥ 45672.73mm3)	0.010 **	1.874 (1.162–2.429)	0.082	1.368 (0.933–2.072)
KPS (< 60)	0.202	1.397 (0.826–2.365)	—	—
MGMT (≤ 30)	0.167	1.342 (0.897–1.941)	—	—
PRI (≤ 4.4 m)	0.239	1.305 (0.831–2.060)	—	—
Treatment duration (≥ 4.2 w)	0.007 **	1.899 (1.278–2.815)	0.085	1.253 (0.921–1.742)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Table 6.

Univariate and multivariate analysis of PFS outcomes for rGBM patients in Bev + CCNU group (n = 60)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype A	Reference
Phenotype B	0.009 **	2.103 (1.206–3.666)	0.019 *	1.908 (1.112–3.273)
Phenotype C	0.001 **	2.749 (1.587–4.763)	0.004 **	2.621 (1.347–5.100)
Sex (Male)	0.642	1.198 (0.762–1.884)	—	—
Age (≥ 65y)	0.514	1.332 (0.807–2.199)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.591	0.921 (0.609–1.394)	—	—
Tumor volume (≥ 35231.65mm3)	0.427	1.187 (0.775–1.818)	—	—
Edema volume (≥ 49864.58mm3)	0.024*	1.736 (1.087–2.398)	0.072	1.354 (0.962–2.081)
KPS (< 60)	0.197	1.389 (0.813–2.375)	—	—
MGMT (≤ 30)	0.013 *	1.747 (1.127–2.710)	0.042 *	1.532 (1.017–2.309)
PRI (≤ 4.5 m)	0.255	1.287 (0.837–1.978)	—	—
Treatment duration (≥ 4.4 w)	0.036 *	1.648 (1.034–2.628)	0.087	1.429 (0.950–2.148)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

Table 7.

Univariate and multivariate analysis of OS outcomes for rGBM patients in Bev + CCNU group (n = 60)

Characteristic	Univariate		Multivariate
	p value	HR (95%CI)	p value	HR (95%CI)
Phenotype A	Reference
Phenotype B	0.014 *	1.976 (1.158–3.372)	0.027 *	1.809 (1.095–3.189)
Phenotype C	0.008 **	2.521 (1.505–4.223)	0.026**	2.344 (1.271–4.509)
Sex (Male)	0.661	1.174 (0.744–1.852)	—	—
Age (≥ 65y)	0.532	1.305 (0.791–2.153)	—	—
Tumor location (Non-midline-crossing / Midline-crossing)	0.609	0.936 (0.620–1.415)	—	—
Tumor volume (≥ 35231.65mm3)	0.445	1.164 (0.763–1.777)	—	—
Edema volume (≥ 49864.58mm3)	0.031 *	1.688 (1.051–2.371)	0.069	1.327 (0.945–2.005)
KPS (< 60)	0.211	1.354 (0.790–2.322)	—	—
MGMT (≤ 30)	0.017 *	1.702 (1.101–2.651)	0.039 *	1.489 (1.008–2.256)
PRI (≤ 4.5 m)	0.272	1.264 (0.824–1.961)	—	—
Treatment duration (≥ 4.4 w)	0.042 *	1.602 (1.014–2.532)	0.094	1.389 (0.931–2.107)

* p < 0.05 or ** p < 0.01 was regarded as statistically significant

No statistically significant phenotype × treatment interaction was detected for PFS or OS, indicating that the incremental prognostic contribution of the imaging phenotype was broadly consistent across treatment regimens (Table S6-S7). The robustness and generalizability of the prognostic value of imaging phenotypes were successfully validated in the independent validation cohort.

Discussion

rGBM remains one of the most challenging malignant central nervous system tumors for clinicians, primarily due to its poor prognosis and limited sensitivity to current therapeutic strategies. Treatment selection for rGBM is highly individualized and largely dependent on the expertise of multidisciplinary clinical teams. Despite recommendations from large-scale clinical trials and established treatment guidelines, the overall therapeutic benefit for rGBM patients remains suboptimal [34, 35]. Conventional clinical indicators for individualized treatment decision-making, such as Karnofsky Performance Status (KPS), molecular biomarkers, tumor volume, contrast enhancement patterns, and peritumoral edema, provide limited guidance for therapeutic optimization [2, 36]. Therefore, there is an urgent need for novel, biologically informative biomarkers with stronger predictive capabilities to guide treatment planning and adaptation in rGBM. Tumor habitat imaging and spatial phenotype mapping have been validated as effective tools for risk stratification and prognostic evaluation across various tumor types [24]. The integration of multi-parametric MRI data offers a novel framework for identifying risk subgroups and optimizing treatment strategies in rGBM, while simultaneously providing a powerful platform for mechanistic investigations into treatment responses. In this study, we employed a pixel-wise MRI colocalization technique combined with heterogeneity phenotype characterization to systematically investigate the structured relationship between intratumoral composition and therapeutic response. Compared with conventional habitat or radiomic clustering approaches, our framework emphasizes voxel-level fusion of complementary biophysical processes and explicit modeling of spatial organization, thereby generating biologically interpretable imaging biomarkers with potential clinical utility for supporting treatment decision-making and optimization in rGBM patients.

In this study, we, for the first time, identified four distinct imaging-based heterogeneity subregions in rGBM using a pixel-wise colocalization approach that integrates diffusion properties with contrast enhancement characteristics. These subregions exhibit markedly divergent biological signatures, likely representing distinct composite structures of proliferating tumor cells and tumor-associated microvasculature within the tumor microenvironment. Specifically, MIS1 is characterized by high ADC values combined with rapid wash-in and wash-out enhancement patterns, potentially corresponding to tumor regions with low cellular proliferation and well-preserved microvascular function [37]. MIS2, defined by high ADC values coupled with plateau-type enhancement, may represent areas of low tumor cell density accompanied by moderately increased vascular permeability [38]. In contrast, MIS3 displays imaging features indicative of hypercellular regions with abundant abnormal microvasculature [38]. MIS4 represents a biologically distinct subregion characterized by both low ADC and atypical enhancement patterns. This subregion likely reflects regions with not only highly proliferative tumor cells but also abnormally high vascular resistance, reminiscent of the chaotic, ring-like, or reniform vascular structures recently described in glioblastoma [39]. These aberrant vascular morphologies are prone to impaired perfusion and are often associated with enhanced invasive potential [40]. Analysis of the proportional variability of these subregions across patients revealed substantial inter-individual heterogeneity, underscoring the highly personalized pathological landscape of rGBM. Importantly, heterogeneity metrics-specifically entropy (reflecting compositional complexity) and Moran’s I (reflecting spatial autocorrelation), were not significantly correlated with tumor volume, indicating that imaging-based heterogeneity serves as an independent biological property rather than a mere reflection of tumor size. This suggests that the heterogeneity signature identified in this study provides complementary biological information beyond traditional clinical and imaging parameters, offering a novel imaging biomarker framework for rGBM characterization. Exploratory correlations reinforce the biological plausibility of MIS subregions. The distinct patterns, with MIS4 associated with adverse profiles, MIS1/2 with favorable associations, and MIS3 representing an intermediate state, suggest that MIS reflects not only imaging heterogeneity but also biologically and clinically meaningful tumor characteristics.

Despite achieving comparable degrees of overall tumor volume reduction across the three treatment regimens (CCNU, Bev + SRT, and Bev + CCNU), the response patterns of the intratumoral heterogeneity subregions differed significantly, offering novel mechanistic insights into therapeutic response in rGBM. The CCNU group exhibited the most pronounced reduction in MIS1, which is likely attributable to the alkylating mechanism of CCNU exerting stronger cytotoxic effects within this subregion. This suggests that CCNU has greater efficacy in targeting proliferating tumor cells compared to tumor-associated microvasculature [41]. In the Bev + SRT group, MIS3 exhibited the greatest reduction, indicating a synergistic ablative effect on hypervascular regions with dysfunctional microvasculature. This response likely results from the VEGF pathway inhibition by bevacizumab combined with enhanced radiosensitivity in hypoxic [42–44], poorly perfused tumor compartments. The Bev + CCNU group demonstrated effective suppression of both MIS2 and MIS3, reflecting a broader-spectrum antitumor effect. This may be attributed to the dual mechanism of vascular normalization or anti-angiogenesis from bevacizumab and direct cytotoxicity from the alkylating agent [45]. Notably, MIS4 consistently exhibited poor response across all regimens, underscoring its role as a core niche of treatment resistance. This phenotype appears driven by VEGF-independent angiogenesis [46], hypoxia-induced stemness [47], immune evasion, and vessel co-option [48], compounded by dense cellularity with perfusion-diffusion mismatch that limits drug delivery. Dosimetric validation further confirmed that MIS4’s resistance cannot be attributed to underdosage but reflects intrinsic biological refractoriness. Together, these findings position MIS4 as a candidate imaging biomarker of therapeutic resistance and highlight the need for integrated multimodal strategies: targeting hypoxia, stemness, immune escape, or metabolism and in combination with radiation and anti-angiogenic therapy, beyond dose escalation alone.

Hierarchical clustering analysis identified three distinct imaging-based heterogeneity phenotypes in rGBM, each characterized by markedly different subregional compositions. Phenotype A exhibited an unbalanced subregion distribution with a predominance of MIS4, yet the spatial arrangement of subregions was highly dispersed. Phenotype B demonstrated relatively balanced proportions across all subregions, but with a clustered spatial distribution. In contrast, Phenotype C was characterized by both a high proportion of MIS4 and a tightly aggregated spatial pattern. These distinct spatial configurations likely contribute to differential therapeutic responses across treatment regimens [49, 50]. Survival analysis revealed pronounced prognostic differences among phenotypes under different treatments. In the CCNU group, patients with Phenotype B achieved the longest PFS and OS, while those with Phenotype C had the poorest outcomes. This suggests that CCNU is more effective in tumors with spatially clustered but compositionally balanced subregions, potentially due to the presence of functional microvascular clusters that facilitate efficient drug delivery to proliferative tumor cells when MIS4 is not dominant [51]. In the Bev + SRT group, Phenotype A was associated with the most favorable survival outcomes, whereas Phenotype B exhibited the worst prognosis. This indicates that a dispersed subregional distribution may enhance the therapeutic efficacy of combined anti-angiogenic and radiotherapy approaches, possibly by improving the suppression of neovascularization and enhancing radiation targeting efficiency in spatially scattered vascular compartments. Under the Bev + CCNU regimen, Phenotype A again conferred the best survival outcomes, while Phenotype C was associated with the poorest prognosis. This pattern highlights the treatment-resistant nature of clustered MIS4, which may possess inherent resistance to alkylating agents and reduced sensitivity to bevacizumab due to vascular endothelial growth factor (VEGF)-independent angiogenic mechanisms [48, 52]. Multivariate Cox regression analysis confirmed that imaging heterogeneity phenotypes serve as robust, independent prognostic factors for both PFS and OS across all treatment regimens. These findings collectively demonstrate that imaging-derived phenotypes not only reflect the intrinsic biological architecture of rGBM but also correlate closely with therapeutic sensitivity. The consistent performance of this classification in the validation cohort underscores its strong generalizability. Incorporating heterogeneity phenotyping into clinical decision-making frameworks has the potential to significantly enhance the precision of personalized treatment strategies for rGBM, ultimately improving patient outcomes.

In this study, tumor segmentation was performed automatically using the Oncohabitats deep-learning model with manual quality control, which ensured accuracy while maintaining efficiency; in future applications, this step could be fully automated with integrated deep learning models to further streamline workflow and facilitate routine clinical integration. (Doc S2).

This study has several limitations. First, the sample size was relatively small, particularly within individual treatment subgroups, which may introduce selection bias and limit statistical power. Second, the absence of direct correlation with histopathological or molecular subtypes constrains mechanistic interpretation, leaving some aspects of tumor heterogeneity inferred solely from imaging features. Third, heterogeneity in subsequent therapies and supportive care after progression could not be fully captured. Although survival analyses were stratified by primary regimens and adjusted for major covariates, variations in second-line treatments and supportive measures may have influenced outcomes and partially confounded the prognostic impact of MIS phenotypes. Future studies incorporating larger multicenter cohorts, standardized treatment protocols, detailed longitudinal clinical records, and comprehensive molecular profiling will be essential to validate robustness, clarify biological underpinnings, and accelerate clinical translation of this imaging-based heterogeneity framework.

In conclusion, this study innovatively integrates multimodal MRI data with pixel-wise colocalization and spatial clustering to construct comprehensive heterogeneity maps and phenotypic classifications of rGBM. This approach not only elucidates treatment-specific response patterns but also establishes a reproducible and generalizable imaging-based phenotypic classification system with substantial clinical applicability. Notably, within the context of multi-line chemoradiotherapy, this heterogeneity-driven phenotypic framework holds significant potential as a practical tool for guiding personalized treatment selection and survival prediction. Our findings provide novel imaging-based evidence to support precision medicine strategies in the management of rGBM.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADC

Apparent diffusion coefficient

ANOVA

Analysis of variance

Bev

Bevacizumab

CCNU

1-(2-chloroethyl)-3-cyclohexyl-1-nitrosourea (lomustine)

CI

Confidence interval

DCE-MRI

Dynamic contrast-enhanced magnetic resonance imaging

DWI

Diffusion-weighted imaging

EGFR

Epidermal growth factor receptor

ESMO

European Society for Medical Oncology

FLAIR

Fluid-attenuated inversion recovery

HR

Hazard ratio

IDH-1

Isocitrate dehydrogenase 1

KDE

Kernel density estimation

KPS

Karnofsky performance status

MGMT

O6-methylguanine-DNA methyltransferase

NCCN

National Comprehensive Cancer Network

OS

Overall survival

PDF

Probability density function

PFS

Progression-free survival

PRI

Postoperative to recurrence interval

RANO

Response Assessment in Neuro-Oncology

rGBM

Recurrent glioblastoma

SRT

Stereotactic radiotherapy

T1WI

T1-weighted imaging

T2WI

T2-weighted imaging

TERT

Telomerase reverse transcriptase

TIC

Time-intensity curve

TP53

Tumor protein p53

VEGF

Vascular endothelial growth factor

VOI

Volume of interest

Author contributions

Y.Z, YF.L, and W.W designed and supervised the study. D.H and RC.W collected and organized the clinical and imaging data. Y.J and Z.L performed the MRI post-processing and feature extraction. Y.L and YH.X conducted the statistical analysis. Y.Z validated the results. Y.Z was the major contributor in writing the original draft. All authors contributed to the manuscript revision and approved the final version for submission.

Funding

National Natural Science Foundation of China (81871343); Key Research and Development Program of Jiangsu Province (Social Development, BE2021693);Key Medical Research Project of the Jiangsu Provincial Health Commission (K2024006).

Data availability

Raw data were generated at the Department of Radiology, Affiliated Hospital of Yangzhou University, Department of Radiology, Affiliated Hospital of Jiangsu University, and Department of Radiology, Affiliated People’s Hospital of Jiangsu University. Derived data supporting the findings of this study are available from the corresponding author on request.

Declarations

Ethics approval and consent to participate

This study was conducted according to the tenets of the Declaration of Helsinki. The ethics committee of Affiliated Hospital of Yangzhou University (Approval No. YDFY20200021); /Affiliated Hospital of Jiangsu University (Approval No. SWYXLL20200136)/ Affiliated People’s Hospital of Jiangsu University (Approval No. ZJYRY2020341) approved this prospective study, and all participants provided written informed consent.

Consent for publication

Consent for publication was given by the Ethic’s Review Board of Affiliated Hospital of Yangzhou University, the Ethic’s Review Board of the Affiliated Hospital of Jiangsu University, and the Ethic’s Review Board of the Affiliated People’s Hospital of Jiangsu University. All patients included in this study consented in the publication of the data.

Competing interests

The authors declare no competing interests.