Abstract
Objective
To explore how molecular glioblastoma (mol-GBM) differs from histological glioblastoma (hist-GBM) in tumor heterogeneity using multi-parametric physiologic MRI-based tumor habitat analysis.
Materials and Methods
In this multi-institutional retrospective study, imaging data were collected from two tertiary centers: 13 mol-GBMs and 39 hist-GBMs from institution 1 (2007–2024) for habitat definition, model development, and internal validation and nine mol-GBMs were obtained from institution 2 (2020–2024) for external validation. The apparent diffusion coefficient (ADC; cellularity), relative cerebral blood volume (rCBV; vascularity), and volume transfer constant (Ktrans; permeability) were binarized into high/low categories, which yielded eight spatial habitat clusters to visualize tumor heterogeneity. Habitat proportions and intratumoral heterogeneity features were compared between groups. Multivariable logistic regression models incorporating habitat-derived features were developed to distinguish mol-GBM from hist-GBM, and their performance was evaluated.
Results
Fifty-two patients (hist-GBM, n = 39; mol-GBM, n = 13; mean age, 60.0 ± 11.1 years; 22 male) were evaluated. Habitat analysis revealed that mol-GBM had a significantly lower proportion of the most malignant habitat (cluster 3: low-ADC, high rCBV, high Ktrans; 0.80% vs. 7.0%, P < 0.001). The total proportion of high Ktrans habitats was markedly lower in the mol-GBM group (5.7% vs. 21.3%, P < 0.001). A habitat-based multivariable model incorporating tumor volume, cluster 3 proportion, high Ktrans proportion, and Shannon entropy achieved an area under receiver operating characteristic curve (AUC) of 0.87 (95% confidence interval [CI]: 0.73–0.97) at internal validation using a leave-one-out cross-validation, which was substantially greater, albeit nonsignificant, than the AUC for tumor size alone (0.70, 95% CI: 0.51–0.85, P = 0.057), with an external validation accuracy of 88.9% (8/9).
Conclusion
Multiparametric physiologic MRI habitat analysis demonstrated differences in the tumor heterogeneity between mol-GBM and hist-GBM. Tumor permeability (Ktrans) was the most discriminating parameter. Incorporating habitat-derived features alongside tumor volume may improve the discrimination of the two subtypes on MRI analysis. Further studies, including more rigorous external validation, are warranted.
Keywords: Glioblastoma, Molecular glioblastoma, Tumor habitat, Diffusion MRI, Perfusion MRI
INTRODUCTION
The 2021 WHO classification of tumors of the central nervous system reclassified histologically low-grade, IDH-wildtype diffuse astrocytoma (WHO grades 2–3) harboring specific molecular alterations (TERT promoter mutation, EGFR amplification, or chromosome 7 gain/10 loss) as molecular glioblastoma (mol-GBM), WHO grade 4 [1,2]. While hist-GBM is notoriously aggressive [3], mol-GBM demonstrates a similarly poor prognosis [4]. However, longer progression-free survival has been reported for mol-GBM, which may represent an early infiltrative stage of GBM progression [4,5].
Currently, the most established finding is that mol-GBMs demonstrate contrast enhancement less frequently than hist-GBMs [5,6,7]. Nevertheless, a recent study reported that among non-enhancing IDH-wildtype GBMs, 45% were hist-GBM and 55% were mol-GBM [8], underscoring the need for advanced MRI to distinguish the two types, given the limitations of conventional imaging.
Physiologic MRI, including diffusion-weighted imaging (DWI), dynamic susceptibility contrast (DSC) perfusion imaging, and dynamic contrast-enhanced (DCE) MRI, provides quantitative insight into tumor cellularity, vascularity, and permeability [9,10,11]. Because mol-GBM lacks pseudopalisading necrosis or microvascular proliferation, these sequences may provide insights into microstructural and vascular heterogeneity that differ between mol-GBM and hist-GBM. Such heterogeneity is associated with biological aggressiveness [12] and can be quantified using the apparent diffusion coefficient (ADC; cellularity), relative cerebral blood volume (rCBV; vascularity), and volume transfer constant (Ktrans; permeability). Previous studies have shown the value of these physiologic MRI parameters in characterizing the biological behavior of GBM [13,14,15,16,17]. Moreover, tumor habitats based on ADC and rCBV have stratified IDH-wildtype GBMs into prognostic subgroups [17,18,19], highlighting the value of physiologic MRI parameters in noninvasive prediction of outcomes.
This study evaluated physiologic MRI differences between mol-GBM and hist-GBM. To further explore these differences, ADC, rCBV, and Ktrans parameters were integrated and a higher-dimensional habitat analysis was performed to visualize and quantify tumor heterogeneity. This multiparametric approach aimed to provide a deeper radiologic insight into mol-GBM.
MATERIALS AND METHODS
Study Population
This retrospective study was approved by the Institutional Review Board of institution 1 (internal dataset, IRB No. 2212-077-1385) and institution 2 (external dataset IRB No. 2023-0525) with a waiver of informed consent. Between November 2007 and October 2024, 689 patients with IDH-wildtype GBM (2021 WHO criteria) were identified at institution 1.
The eligible patients were adults (≥18 years) who had undergone conventional and advanced MRI (DWI, DSC, and DCE). Patients were classified according to the 2021 WHO criteria into hist-GBM, defined as an IDH-wildtype diffuse glioma with microvascular proliferation and/or intratumoral necrosis, and mol-GBM, defined as a histological grade 2–3 IDH-wildtype diffuse glioma carrying the TERT promoter mutation, EGFR amplification, or +7/-10 chromosome variations. Patients with incomplete MRI or MRI acquired more than 6 months before surgery were excluded. After applying the exclusion criteria, 13 patients with mol-GBM and 476 with hist-GBM were identified. Following 1:3 propensity score matching by age and sex, 52 patients (13 mol-GBM, 39 hist-GBM) were included.
For the external dataset, 18 patients from institution 2 enrolled between February 2020 and February 2024 met identical inclusion criteria, producing 9 mol-GBM (mol-GBM [ext]). The patient enrollment flowchart is shown in Figure 1.
Fig. 1. Flow diagram of the patient selection process. IDH = isocitrate dehydrogenase, GBM = glioblastoma, TERT = telomerase reverse transcriptase, EGFR = epidermal growth factor receptor, DSC = dynamic susceptibility contrast, DCE = dynamic contrast-enhanced, Mol-GBM = molecular glioblastoma, Hist-GBM = histological glioblastoma, Mol-GBM (ext) = molecular glioblastoma from the external dataset.
The internal dataset of institution 1 (n = 52) was used for habitat definition, model development, and internal validation. The external dataset of institution 2 (n = 9, all mol-GBM) was used for the independent evaluation.
MRI Acquisitions
For institution 1, MRI studies of all patients were performed using 3T MRI units (Magnetom Verio or Magnetom Skyra; Siemens Healthineers, Erlangen, Germany) with a 32-channel head coil. The MRI protocol included T2-weighted imaging (T2WI), T2 fluid-attenuated inversion recovery (FLAIR), T1-weighted imaging (T1WI), DWI, DCE-MRI, and DSC-MRI, and contrast enhanced (CE) T1WI. For institution 2, the MRI scan was acquired using a 3T MRI unit (Ingenia 3.0 CX, Philips Healthcare, Amsterdam, the Netherlands) including the same sequences applied in institution 1. The detailed MRI acquisition parameters are shown in the Supplement (Magnetic Resonance Imaging Acquisition).
Image Processing
The three-dimensional (3D) CE-T1WI of each patient were resampled to 1 mm isometric voxels. Skull stripping was performed on all MRI sequences using the SynthStrip algorithm [20]. The skull-stripped CE-T1WI served as the reference image, and all other modalities (T2WI, FLAIR, DWI, DSC, and DCE) were co-registered using rigid transformations with the Advanced Normalization Tools [21]. After registration, tumor masks were generated using HD-GLIO [22], and refined by an experienced neuroradiologist (K.S.C, 7 years) and a well-trained research assistant (10 years).
To mitigate inter-institutional batch effects, the ADC, rCBV, and Ktrans values were normalized using normal white matter reference distributions. The ADC map was calculated using b values of 0 and 1,000 s/mm2, based on a two-point estimate of signal decay. For DSC and DCE analyses, the pharmacokinetic map calculations were performed using commercial software (Nordic ICE, v4.1.2; Nordic-NeuroLab, Bergen, Norway). For the DSC analysis, the rCBV map was obtained using the Weisskoff–Boxerman method for leakage correction [23]. For DCE analysis, the Ktrans was calculated using the two-compartment pharmacokinetic model (Extended Tofts Model, eTofts) [24]. Conventional histogram features of ADC, rCBV, and Ktrans were extracted. Further details are provided in the Supplement (Parametric Mapping for Advanced Image, Batch Effect Correction, Histogram-Based Analysis, and Supplementary Fig. 1).
Multiparametric Physiologic MRI-Based Tumor Habitats
Tumor habitats were defined using voxel-wise ADC, rCBV, and Ktrans values. For each parameter, Gaussian mixture modeling was applied to derive a single threshold, dichotomizing values into low and high categories. Based on the resulting 2 × 2 × 2 combination, each voxel was classified into one of eight physiologic habitats (e.g., cluster 3: low ADC, high rCBV, high Ktrans), each representing a distinct tumor microenvironment. Habitat definitions were fixed using the internal dataset and were applied consistently in all subsequent analyses.
For each tumor, the proportion of the eight habitat clusters was calculated relative to the total tumor volume. The overall tumor size was represented by log-transformed total tumor volume. Clusters were also grouped by shared physiologic trait, all clusters with high Ktrans (clusters 1, 3, 5, 7), high rCBV (clusters 2, 3, 6, 7), or low ADC (clusters 0, 1, 2, 3). Cluster 3 (low ADC, high rCBV, high Ktrans) was prespecified as a marker of the most malignant-appearing tumor habitat.
Intratumoral heterogeneity was quantified using Shannon entropy, the Simpson diversity index, and Pielou evenness derived from the habitat proportions of the eight habitat clusters. The external cohort, consisting exclusively of mol-GBM cases, was used to assess the consistency of the habitat distribution patterns.
Model Development and Validation
Multivariable logistic regression models were developed using the internal cohort (institution 1; n = 52). Features were standardized according to the z-score using the parameters of the internal cohort and were applied identically to the external cohort.
Univariable and multivariable logistic regression analysis was performed to evaluate the association between habitat-derived features and GBM subtypes (mol-GBM vs. hist-GBM). Univariable analysis included all prespecified habitat-derived features: individual cluster proportions, composite habitat proportions (high Ktrans, high rCBV, and low ADC), Shannon entropy, and log-transformed tumor volume. Variables for the multivariable analysis were selected a priori based on biological relevance and univariable significance, without stepwise selection. As a result, three logistic regression models were defined using variables selected for multivariable modeling to assess the added discriminatory value of habitat-derived features beyond tumor volume. Model V incorporated log-transformed tumor volume only, serving as a clinically accessible reference. Model H added the cluster 3 proportion and the high Ktrans proportion, selected based on their biological relevance as markers of the malignant microenvironment and the overall permeability burden, respectively. Model H+ further incorporated Shannon entropy to evaluate the incremental value of intratumoral heterogeneity. All models were configured to predict hist-GBM. Further details are provided in the Supplement (Univariable Logistic Regression Analysis and Multivariable Logistic Regression Analysis).
Internal validation was performed using leave-one-out cross-validation (LOOCV) within the internal cohort (n = 52). For each fold, the z-score normalization and an L2-regularized logistic regression (C = 1.0) were fit to the n - 1 training subjects and were applied to the held-out subject. The 52 aggregated out-of-fold probabilities were used to construct receiver operating characteristic (ROC) curves, compute area under the curve (AUC) values, and (95% confidence intervals [CIs] by 1,000-iteration bootstrap), and to define a Youden index-derived threshold, which was held fixed for all subsequent evaluations. After LOOCV, each model was refit on the full internal dataset for external evaluation; this step did not contribute to internal validation metrics. The fixed coefficients were applied to the external cohort using the same threshold. Further details are provided in the Supplement (Model Validation).
Other Statistical Analyses
Patient characteristics were compared between hist-GBM and mol-GBM in the internal dataset, and between internal mol-GBM and mol-GBM (ext), using the independent t-test or Mann–Whitney U test for continuous variables, depending on the normality of distribution assessed by the Shapiro–Wilk test, and the chi-squared test for categorical variables. Differences in voxel-wise habitat distributions were analyzed by pairwise permutational multivariate analysis of variance (PERMANOVA) with post-hoc Mann–Whitney U tests. The DeLong test was used for ROC AUC comparison. No formal adjustment for multiple comparisons was applied. A two-sided P-value <0.05 was considered statistically significant. Statistical analyses were performed using Python (version 3.10.13; Python Software Foundation, Wilmington, DE, USA) with the SciPy (version 1.15.3), scikit-learn (version 1.7.0), Matplotlib (version 3.10.3), and Seaborn (version 0.13.2) libraries. Propensity score matching was conducted using R (version 3.4.0; R Foundation for Statistical Computing, Vienna, Austria).
RESULTS
Patient Characteristics
The baseline characteristics of the patients are summarized in Table 1. Age and sex did not differ between hist-GBM and mol-GBM subtypes nor between the internal and external mol-GBM cohorts.
Table 1. Baseline characteristics of the internal and external dataset.
| Characteristics | Internal dataset | External dataset | P * | P † | ||
|---|---|---|---|---|---|---|
| Hist-GBM (n = 39) | Mol-GBM (n = 13) | Mol-GBM (ext) (n = 9) | ||||
| Age at diagnosis, yrs | 59.6 ± 11.6 | 61.1 ± 10.0 | 64.6 ± 10.9 | 0.683 | 0.453 | |
| Sex | 0.746 | 1.000 | ||||
| Male | 17 (43.6) | 5 (38.5) | 4 (44.4) | |||
| Female | 22 (56.4) | 8 (61.5) | 5 (55.6) | |||
| Baseline KPS | 82.6 ± 11.6 | 86.7 ± 13.7 | 95.0 ± 7.6 | 0.234 | 0.163 | |
| WHO grade | <0.001 | 0.178 | ||||
| Grade 2 | 0 (0) | 2 (15.4) | 4 (44.4) | |||
| Grade 3 | 0 (0) | 11 (84.6) | 5 (55.6) | |||
| Grade 4 | 39 (100) | 0 (0) | 0 (0) | |||
| Period between MRI and biopsy/surgery, day | 3.7 ± 10.2 | 14.2 ± 32.3 | 33.1 ± 27.1 | 0.239 | 0.004 | |
| Contrast Enhancement on T1 | <0.001 | 1.000 | ||||
| No | 2 (2.6) | 9 (69.2) | 6 (66.7) | |||
| Yes | 37 (94.9) | 4 (30.8) | 3 (33.3) | |||
| Tumor volumes, mm3 | 76,148 [74,763] | 38,843 [27,009] | 29,026 [18,674] | 0.009 | 0.229 | |
| IDH-wildtype | 39 (100) | 13 (100) | 9 (100) | - | - | |
| TERT promoter mutation | 0.034 | 0.206 | ||||
| No | 11 (28.2) | 0 (0) | 3 (33.3) | |||
| Yes | 15 (38.5) | 8 (61.5) | 6 (66.7) | |||
| NA | 13 (33.3) | 5 (38.5) | 0 (0) | |||
| EGFR amplification | 0.164 | 0.017 | ||||
| No | 9 (23.1) | 6 (46.2) | 9 (100) | |||
| Yes | 29 (74.4) | 7 (53.8) | 0 (0) | |||
| NA | 1 (2.5) | 0 (0) | 0 (0) | |||
| +7/-10 chromosome variations | 1.000 | 0.080 | ||||
| No | 20 (51.3) | 5 (38.5) | 1 (11.1) | |||
| Yes | 5 (12.8) | 1 (7.7) | 5 (55.6) | |||
| NA | 14 (35.9) | 7 (53.8) | 3 (33.3) | |||
| PTEN mutation | 0.589 | 0.494 | ||||
| No | 36 (92.3) | 11 (84.6) | 9 (100) | |||
| Yes | 3 (7.7) | 2 (15.4) | 0 (0) | |||
| MGMT methylation | 0.502 | 0.164 | ||||
| No | 24 (61.5) | 10 (76.9) | 3 (33.3) | |||
| Yes | 14 (35.9) | 3 (23.1) | 5 (55.6) | |||
| NA | 1 (2.5) | 0 (0) | 1 (11.1) | |||
| Ki-67, % | 38.3 ± 21.0 | 14.8 ± 13.0 | NA | <0.001 | - | |
Data are presented as mean ± standard deviation, median [interquartile range], and number (%).
*P-values for pairwise comparison between Mol-GBM and Hist-GBM, †P-values for pairwise comparison between Mol-GBM and Mol-GBM (ext).
Mol-GBM = molecular glioblastoma, Hist-GBM = histological glioblastoma, Mol-GBM (ext) = molecular glioblastoma from the external dataset, KPS = Karnofsky performance score, IDH = isocitrate dehydrogenase, TERT = telomerase reverse transcriptase, EGFR = epidermal growth factor receptor, +7/−10 = gain of chromosome 7 and loss of chromosome 10, PTEN = phosphatase and tensin homolog, MGMT = O6-methylguanine-DNA methyltransferase, NA = not available, - = not applicable
Among the mol-GBM group (n = 13), eight patients harbored a TERT promoter mutation (8/13, 61.5%), seven patients EGFR amplification (7/13, 53.8%), and one patient +7/-10 chromosome variations (1/13, 7.7%). For the mol-GBM (ext) group (n = 9), six patients had TERT promoter mutation (6/9, 66.7%) and five patients had +7/-10 chromosome variations (5/9, 55.6%).
Tumor Habitat Analysis
Eight tumor habitat clusters were defined by combining binarized ADC, rCBV, and Ktrans values using GMM. Cutoff values derived from the hist-GBM and mol-GBM were ADC = 1.326, rCBV = 4.420, and Ktrans = 0.026. The pairwise PERMANOVA showed significant differences in overall habitat distributions between hist-GBM and mol-GBM (pseudo-F = 4.45, P = 0.003) and between hist-GBM and mol-GBM (ext) (pseudo-F = 4.38, P = 0.003), but not between mol-GBM and mol-GBM (ext) (pseudo-F = 0.63, P = 0.765).
Median proportion (%) in cluster 0 (low ADC, low rCBV, low Ktrans) was significantly higher in the mol-GBM (40.0% vs. 71.4%, P < 0.001). In contrast, the proportion of cluster 3 (low ADC, high rCBV, high Ktrans) was significantly higher in the hist-GBM than in mol-GBM (7.0% vs. 0.8%, P ≤ 0.001). Similarly, cluster 7 (high ADC, high rCBV, high Ktrans) was also more prevalent in the hist-GBM (1.4% vs. 0.3%, P = 0.036) (Table 2). External validation using the mol-GBM (ext) dataset showed consistent results with hist-GBM in clusters 0 and 3. No significant differences were found between mol-GBM and mol-GBM (ext) across all clusters (Supplementary Table 1).
Table 2. Comparison of tumor habitat cluster proportions.
| Cluster | ADC | rCBV | Ktrans | Cluster proportions (%) | P | ||
|---|---|---|---|---|---|---|---|
| Hist-GBM (n = 39) | Mol-GBM (n = 13) | ||||||
| 0 | ↓ | ↓ | ↓ | 40.0 [28.9–60.1] | 71.4 [64.0–79.7] | <0.001* | |
| 1 | ↓ | ↓ | ↑ | 6.1 [1.8–10.6] | 3.0 [1.6–3.3] | 0.059 | |
| 2 | ↓ | ↑ | ↓ | 6.3 [4.7–8.6] | 5.2 [3.1–7.8] | 0.612 | |
| 3 | ↓ | ↑ | ↑ | 7.0 [3.5–15.3] | 0.8 [0.6–2.5] | <0.001* | |
| 4 | ↑ | ↓ | ↓ | 15.4 [7.2–29.6] | 11.4 [5.9–16.5] | 0.263 | |
| 5 | ↑ | ↓ | ↑ | 1.3 [0.4–3.9] | 0.4 [0.4–1.2] | 0.099 | |
| 6 | ↑ | ↑ | ↓ | 1.2 [0.5–2.2] | 0.7 [0.4–1.6] | 0.352 | |
| 7 | ↑ | ↑ | ↑ | 1.4 [0.3–3.4] | 0.3 [0.2–0.9] | 0.036* | |
| Composite habitat proportions | |||||||
| Low ADC (clusters 0 + 1 + 2 + 3) | 78.1 [54.2–87.7] | 87.8 [78.6–93.7] | 0.163 | ||||
| High rCBV (clusters 2 + 3 + 6 + 7) | 22.7 [12.7–35.0] | 10.1 [6.9–11.4] | 0.001* | ||||
| High Ktrans (clusters 1 + 3 + 5 + 7) | 21.3 [10.8–34.2] | 5.7 [3.6–7.9] | <0.001* | ||||
Data are median [interquartile range].
*P < 0.05.
ADC = apparent diffusion coefficient, rCBV = relative cerebral blood volume, Ktrans = volume transfer constant, Hist-GBM = histological glioblastoma, Mol-GBM = molecular glioblastoma
The total proportion of high-permeability habitats (high Ktrans; clusters 1, 3, 5, and 7) was markedly higher in hist-GBM compared with mol-GBM (21.3% vs. 5.7%, P < 0.001). Likewise, the proportion of high-vascularity habitats (high rCBV; clusters 2, 3, 6, and 7) was significantly higher in hist-GBM (22.7% vs. 10.1%, P = 0.001). In contrast, no significant difference was observed in the overall proportion of high-cellularity habitats (low ADC; clusters 0, 1, 2, and 3) between the two groups (78.1% vs. 87.8%, P = 0.163).
The 3D habitat distributions and the corresponding 2D projections (Fig. 2, Supplementary Fig. 2) further confirmed these distinctions. Tumor volumes for clusters 3 (red; low ADC, high rCBV, high Ktrans) and 7 (gray; high ADC, high rCBV, high Ktrans) were greater in the hist-GBM than the in mol-GBM and mol-GBM (ext). The clustering patterns of mol-GBM and mol-GBM (ext) appeared to be identical (Fig. 2B, C), visualizing the absence of significant differences noted in Supplementary Table 1.
Fig. 2. Voxel-wise habitat clustering of Hist-GBM, Mol-GBM, and Mol-GBM (ext) based on multiparametric MRI. Tumor voxels were classified into eight clusters based on 2 × 2 × 2 combinations of binary high/low grouping of ADC, rCBV, and Ktrans values using a GMM. GMM-determined cutoff values were as follows: ADC, 1.326; rCBV, 4.420; and Ktrans, 0.026. Shown are the three-dimensional cluster distributions with their respective two-dimensional projections onto each parameter axis. A: Voxel data from patients in the Hist-GBM group (n = 39). For improved visualization and fair comparison, 10% of all voxels are displayed. Voxel data for Hist-GBM patients are provided in Supplementary Figure 2. B: Voxel data from all patients in the Mol-GBM group (n = 13). C: External validation of the established clustering criteria using voxel data from the Mol-GBM (ext) group (n = 9). Hist-GBM = histological glioblastoma, Mol-GBM = molecular glioblastoma, Mol-GBM (ext) = molecular glioblastoma from external dataset, ADC = apparent diffusion coefficient, rCBV = relative cerebral blood volume, Ktrans = volume transfer constant, GMM = Gaussian mixture model.
Representative Cases for Tumor Habitat Clustering
Representative cases (Fig. 3) illustrate these patterns. Patient A, a case of hist-GBM with marked contrast enhancement, showed a large proportion of cluster 3 (red) within the enhancing lesion. Patient B, despite minimal enhancement, also showed a marked proportion of cluster 3 in the tumor area. In contrast, patient C, a mol-GBM case with a non-enhancing lesion, showed a predominance of cluster 0 (blue; low ADC, low rCBV, low Ktrans) within the FLAIR hyperintense region. The stacked voxel count plot (Fig. 3D) showed a higher cluster 3 proportion in hist-GBM than in mol-GBM. Moreover, high Ktrans clusters (cluster 1, 3, 5, and 7) were more frequent in the hist-GBM (Fig. 3E). These patterns were most evident in tumors with strong contrast enhancement.
Fig. 3. Visualization of tumor clusters in Hist-GBM and Mol-GBM patients. A-C: Multimodal MRI of three patients with IDH-wildtype glioblastoma, presented as the T1CE image, FLAIR image, a magnified (×4) tumor region, and the corresponding cluster map overlaid on the magnified area: (A) Hist-GBM, CE (+); a 58-year-old male with aphasia presenting an enhancing lesion in the left temporal lobe. (B) Hist-GBM, CE (±); a 52-year-old female with right upper extremity paresthesia presenting a right temporal mass with minimal enhancement. (C) Mol-GBM, CE (-); a 56-year-old female with headache shows a non-enhancing mass in the left frontal and medial parietal lobes. D: Stacked bar plots demonstrate cluster-wise tumor volume distribution, highlighting increased cluster 3 volume in Hist-GBM compared with Mol-GBM. E: Ktrans values were highest in the enhancing Hist-GBM, case A, followed by the minimally enhancing Hist-GBM, case B, and lowest in the non-enhancing Mol-GBM, case C. Hist-GBM = histological glioblastoma, Mol-GBM = molecular glioblastoma, IDH = isocitrate dehydrogenase, T1CE = T1-weighted contrast-enhanced, FLAIR = fluid-attenuated inversion recovery, CE = contrast enhancement, Ktrans = volume transfer constant.
Discrimination Between Mol-GBM and Hist-GBM
Univariable logistic regression was performed to evaluate the association between habitat-derived features and the GBM subtype, the results are summarized in Supplementary Figure 3A. Features with strong univariable associations, cluster 3 proportion, high Ktrans proportion, and Shannon entropy, were incorporated into multivariable models as described above.
Multivariable logistic regression demonstrated that the incorporation of habitat-derived features improved discrimination compared with tumor volume alone (Fig. 4, Table 3). In the internal validation cohort, Model V achieved an AUC of 0.70 (95% CI: 0.51–0.85). Model H showed improved discrimination compared to Model V with a marginal P-value (AUC: 0.85, 95% CI: 0.70–0.96; P = 0.053). Model H+ achieved the highest AUC (0.87, 95% CI: 0.73–0.97), whereas differences from Model V and Model H were not significant (P = 0.057 and 0.434, respectively). Binary classification performance based on the Youden index-derived thresholds is summarized in Table 4. In the external cohort (mol-GBM only), classification accuracies were 77.8% (7/9) for Model H and 88.9% (8/9) for Model H+. As a sensitivity analysis, a multivariable model incorporating all eight habitat clusters using a centered log-ratio transformation also demonstrated robust discrimination (AUC: 0.82; Supplementary Fig. 3B), supporting the consistency of habitat-based modeling.
Fig. 4. Multivariable logistic regression analysis for identifying features associated with Hist-GBM versus Mol-GBM. Receiver operating characteristic curves for multivariable models predicting Hist-GBM from internal validation: Model V (volume only; AUC = 0.70 [95% CI: 0.51–0.85]), Model H (volume, cluster 3 proportion, and high Ktrans; AUC = 0.85 [95% CI: 0.70–0.96]), and Model H+ (volume, cluster 3 proportion, high Ktrans, and Shannon entropy; AUC = 0.87 [95% CI: 0.73–0.97]). Hist-GBM = histological glioblastoma, Mol-GBM = molecular glioblastoma, AUC = area under the curve, CI = confidence interval, Ktrans = volume transfer constant.
Table 3. Factors associated with histological glioblastoma.
| Variable | Model V | Model H | Model H+ | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| β | aOR | 95% CI | P | β | aOR | 95% CI | P | β | aOR | 95% CI | P | |
| Tumor volume (log) | 0.643 | 1.90 | 1.15–3.72 | 0.020 | 0.525 | 1.69 | 0.94–3.62 | 0.078 | 0.415 | 1.51 | 0.78–3.14 | 0.186 |
| Cluster 3 proportion | - | - | - | - | 0.264 | 1.30 | 0.71–2.78 | 0.450 | 0.174 | 1.19 | 0.66–2.38 | 0.568 |
| High Ktrans proportion | - | - | - | - | 1.082 | 2.95 | 1.82–5.23 | <0.001 | 0.642 | 1.90 | 1.23–3.14 | 0.002 |
| Shannon entropy | - | - | - | - | - | - | - | - | 0.708 | 2.03 | 1.10–4.70 | 0.050 |
β = standardized regression coefficient, aOR = adjusted odds ratio, CI = confidential interval, Ktrans = volume transfer constant
Table 4. Binary classification performance using Youden index–derived threshold.
| Model | Youden threshold | Internal validation | External evaluation | ||
|---|---|---|---|---|---|
| AUC (95% CI) | Mol-GBM accuracy (n = 13, %) | Hist-GBM accuracy (n = 39, %) | Mol-GBM (ext) accuracy (n = 9, %) | ||
| Model V | 0.772 | 0.70 (0.51–0.85) | 11/13 (84.6) | 23/39 (59.0) | 9/9 (100)* |
| Model H | 0.612 | 0.85 (0.70–0.96) | 11/13 (84.6) | 33/39 (84.6) | 7/9 (77.8) |
| Model H+ | 0.652 | 0.87 (0.73–0.97) | 12/13 (92.3) | 32/39 (82.1) | 8/9 (88.9) |
Decision thresholds are Youden index-derived and held fixed for external evaluation.
*The 100% external accuracy for Model V reflects its high Youden threshold (0.772), driven by the smaller tumor volumes characteristic of Mol-GBM, resulting in all 9 external cases being assigned low predicted probabilities.
Mol-GBM = molecular glioblastoma, AUC = area under the curve, CI = confidential interval, Hist-GBM = histological glioblastoma, Mol-GBM (ext) = molecular glioblastoma from external dataset
DISCUSSION
This study demonstrated that mol-GBM and hist-GBM differ in tumor heterogeneity using a 3D habitat analysis based on multiparametric physiologic MRI. Mol-GBM showed a smaller proportion of habitats with high vascularity and high permeability, particularly cluster 3 and high Ktrans clusters. Hist-GBM also exhibited greater intratumoral diversity, as reflected by a higher Shannon entropy.
Prior imaging studies comparing mol-GBM and hist-GBM subtypes have relied largely on conventional MRI or global summary statistics. Although histogram metrics showed significant differences (Supplementary Table 2), such approaches overlook intratumoral heterogeneity [25]. The habitat analysis approach employed in this study overcomes this limitation. By partitioning the tumor into biologically distinct subregions based on a multiparametric physiological signature, this approach identifies localized malignant foci obscured by whole-tumor averaging [17,18,19]. The discovery that the most malignant-appearing habitat (cluster 3; low ADC, high rCBV, high Ktrans) was found in hist-GBM validates the utility of this method and suggests its potential for guiding targeted biopsies toward the most aggressive tumor regions.
Although prior studies have utilized ADC (cellularity) and rCBV (vascularity), this approach has limitations, particularly in minimally enhancing tumors [17,18,19]. We developed this physiological characterization by incorporating Ktrans, a quantitative parameter reflecting disruption of the blood-brain barrier and microvascular leakage [11,26,27,28]. Pathologically, aggressive GBM (particularly hist-GBM) is characterized by hypoxia-driven VEGF expression, which promotes the histologic hallmark of aberrant, immature neovasculature, and microvascular proliferation [29]. This immature vasculature is characterized by high permeability, which manifests as elevated Ktrans values on DCE-MRI [30]. The significant predominance of high Ktrans habitats in hist-GBM confirms that Ktrans sensitively captures this aggressive angiogenic activity. In multivariable analysis, high Ktrans proportion was the only habitat feature retaining independent significance between models H and H+ (odds ratio: 2.95 and 1.90, respectively), positioning it as the key differentiator between hist-GBM and mol-GBM subtypes. Previous literature has associated “cellular, hypovascular” habitats (low ADC, low rCBV) with treatment resistance and poor prognosis, often attributed to hypoxia [17,18,19]. Our study refined this concept by demonstrating that this category is not monolithic. We identified two distinct clusters within the following profile: cluster 0 (low Ktrans) and cluster 1 (high Ktrans). Cluster 0, characterized by low permeability despite low vascularity, was significantly more prevalent in mol-GBM, whereas cluster 1 did not present low permeability. This finding suggests that hypoxic regions with concurrent high permeability may reflect true regions of active, hypoxia-driven angiogenesis and vascular leakiness [30], constituting a less aggressive and proliferative microenvironment characteristic of mol-GBM [31]. This stratification, which highlights a key biological difference between the hist-GBM and mol-GBM subtypes, was detectable only when Ktrans was incorporated into the habitat analysis.
ADC alone provided limited discrimination. The proportion of low ADC clusters did not differ significantly between subtypes, whereas the high Ktrans and high rCBV clusters did. ADC values are affected by not only cellularity but also by edema, necrosis, and nuclear morphology [32,33], reducing their specificity. Additionally, mol-GBM may exhibit cellular density comparable to that of hist-GBM; Griessmair et al. [34] reported lower ADC in mol-GBM than in lower-grade astrocytoma lacking GBM-defining molecular alterations. Thus, ADC is informative but insufficient alone to distinguish mol-GBM from hist-GBM.
This study has some limitations. First, the small sample size of the mol-GBM group (n = 13 internal, n = 9 external) reflects the rarity of this entity (2.7% of IDH-wildtype GBMs in our cohort) and limits statistical power and generalizability. The findings of this exploratory study require validation in larger, prospective cohorts. Second, combining MRI data from different vendors may introduce residual batch effects despite normalization to normal white matter; future studies employing advanced harmonization techniques may further mitigate this issue. Third, the external cohort included only mol-GBM cases, limiting the evaluation of discrimination performance and generalizability. Although consistent habitat distributions were observed, external validation including both subtypes is required. Fourth, due to the retrospective study design, direct histopathological correlation of the imaging-defined habitats was not feasible. Future prospective studies integrating radiologic, histologic, and molecular data are warranted to strengthen the biological relevance and clinical applicability of these findings.
In conclusion, a 3D habitat analysis using ADC, rCBV, and Ktrans can distinguish mol-GBM from hist-GBM subtypes by capturing differences in intratumoral heterogeneity. Mol-GBM exhibited significantly smaller tumor proportions within the most malignant habitat, and differences between two subtypes were greatest for permeability (Ktrans). Incorporating habitat-derived features alongside tumor volume may improve discriminating the two subtypes on MRI. These findings provide novel radiologic insights into the recently classified but less-characterized mol-GBM and a provides a foundation for future radiogenomic investigations. Additional studies, including more rigorous external validation, are warranted.
Acknowledgments
We would like to thank Hyeryeon Son, RN, BSN (Department of Radiology, Seoul National University Hospital) for her invaluable assistance with data collection and analysis.
Footnotes
Conflicts of Interest: Seung Hong Choi, Deputy Editor, and Ji Eun Park, Editorial Board Member, of the Korean Journal of Radiology, were not involved in the editorial evaluation or decision to publish this article. The remaining authors have declared no conflicts of interest.
- Conceptualization: Minseo Choi, Yunseo Choi, Kyu Sung Choi.
- Data curation: Kyu Sung Choi.
- Formal analysis: Minseo Choi, Kyu Sung Choi, Junhyeok Lee, Seong Yun Jeong, Minchul Kim.
- Funding acquisition: Kyu Sung Choi.
- Investigation: Minseo Choi, Yunseo Choi, Kyu Sung Choi.
- Methodology: Minseo Choi, Kyu Sung Choi, Yunseo Choi, Junhyeok Lee.
- Project administration: Seung Hong Choi.
- Resources: Kyu Sung Choi, Inpyeong Hwang, Yae Won Park, Ji Eun Park.
- Supervision: Kyu Sung Choi, Seung Hong Choi.
- Validation: Kyu Sung Choi, Yae Won Park, Ji Eun Park.
- Visualization: Minseo Choi, Junhyeok Lee, Seong Yun Jeong.
- Writing—original draft: Minseo Choi.
- Writing—review & editing: Minseo Choi, Yunseo Choi.
Funding Statement: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00251022) (K.S.C.); the SNUH Research Fund (No. 04-2024-0600; No. 04-2025-2060) (K.S.C.); and the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) grant funded by the Ministry of Health&Welfare (No. RS-2024-00439549) (K.S.C.).
Availability of Data and Material
The data generated and/or analyzed during the current study are not publicly available due to patient privacy but are available from the corresponding author on reasonable request. The underlying code for this study is available in Multiparametric-MRI-Habitat-GBM and can be accessed via this link https://github.com/snuh-rad-aicon/Multiparametric-MRI-Habitat-GBM.
Supplement
The Supplement is available with this article at https://doi.org/10.3348/kjr.2025.1969.
References
- 1.Louis DN, Perry A, Reifenberger G, von Deimling A, Figarella-Branger D, Cavenee WK, et al. The 2016 World Health Organization classification of tumors of the central nervous system: a summary. Acta Neuropathol. 2016;131:803–820. doi: 10.1007/s00401-016-1545-1. [DOI] [PubMed] [Google Scholar]
- 2.Brat DJ, Aldape K, Colman H, Holland EC, Louis DN, Jenkins RB, et al. cIMPACT-NOW update 3: recommended diagnostic criteria for “diffuse astrocytic glioma, IDH-wildtype, with molecular features of glioblastoma, WHO grade IV. Acta Neuropathol. 2018;136:805–810. doi: 10.1007/s00401-018-1913-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ostrom QT, Bauchet L, Davis FG, Deltour I, Fisher JL, Langer CE, et al. The epidemiology of glioma in adults: a “state of the science” review. Neuro Oncol. 2014;16:896–913. doi: 10.1093/neuonc/nou087. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Ramos-Fresnedo A, Pullen MW, Perez-Vega C, Domingo RA, Akinduro OO, Almeida JP, et al. The survival outcomes of molecular glioblastoma IDH-wildtype: a multicenter study. J Neurooncol. 2022;157:177–185. doi: 10.1007/s11060-022-03960-6. [DOI] [PubMed] [Google Scholar]
- 5.Ramos-Fresnedo A, Domingo RA, Perez-Vega C, Pullen MW, Akinduro OO, Almeida JP, et al. The early infiltrative phase of GBM hypothesis: are molecular glioblastomas histological glioblastomas in the making? A preliminary multicenter study. J Neurooncol. 2022;158:497–506. doi: 10.1007/s11060-022-04040-5. [DOI] [PubMed] [Google Scholar]
- 6.Guo X, Gu L, Li Y, Zheng Z, Chen W, Wang Y, et al. Histological and molecular glioblastoma, IDH-wildtype: a real-world landscape using the 2021 WHO classification of central nervous system tumors. Front Oncol. 2023;13:1200815. doi: 10.3389/fonc.2023.1200815. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Du P, Wu X, Liu X, Chen J, Cao A, Geng D. Establishment of a prediction model based on preoperative MRI radiomics for diffuse astrocytic glioma, IDH-wildtype, with molecular features of glioblastoma. Cancers (Basel) 2023;15:5094. doi: 10.3390/cancers15205094. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Foltyn-Dumitru M, Banan R, Schell M, Mahmutoglu MA, Kessler T, Wick W, et al. Histopathological and molecular characteristics of IDH-wildtype glioblastoma without contrast enhancement: implications for clinical outcomes. Neuro Oncol. 2025;27:1878–1887. doi: 10.1093/neuonc/noaf070. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Sugahara T, Korogi Y, Kochi M, Ikushima I, Shigematu Y, Hirai T, et al. Usefulness of diffusion-weighted MRI with echo-planar technique in the evaluation of cellularity in gliomas. J Magn Reson Imaging. 1999;9:53–60. doi: 10.1002/(sici)1522-2586(199901)9:1<53::aid-jmri7>3.0.co;2-2. [DOI] [PubMed] [Google Scholar]
- 10.Law M, Yang S, Babb JS, Knopp EA, Golfinos JG, Zagzag D, et al. Comparison of cerebral blood volume and vascular permeability from dynamic susceptibility contrast-enhanced perfusion MR imaging with glioma grade. AJNR Am J Neuroradiol. 2004;25:746–755. [PMC free article] [PubMed] [Google Scholar]
- 11.Roberts HC, Roberts TP, Brasch RC, Dillon WP. Quantitative measurement of microvascular permeability in human brain tumors achieved using dynamic contrast-enhanced MR imaging: correlation with histologic grade. AJNR Am J Neuroradiol. 2000;21:891–899. [PMC free article] [PubMed] [Google Scholar]
- 12.Yabo YA, Niclou SP, Golebiewska A. Cancer cell heterogeneity and plasticity: a paradigm shift in glioblastoma. Neuro Oncol. 2022;24:669–682. doi: 10.1093/neuonc/noab269. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Yoon J, Baek N, Yoo RE, Choi SH, Kim TM, Park CK, et al. Added value of dynamic contrast-enhanced MR imaging in deep learning-based prediction of local recurrence in grade 4 adult-type diffuse gliomas patients. Sci Rep. 2024;14:2171. doi: 10.1038/s41598-024-52841-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Martucci M, Russo R, Giordano C, Schiarelli C, D’Apolito G, Tuzza L, et al. Advanced magnetic resonance imaging in the evaluation of treated glioblastoma: a pictorial essay. Cancers (Basel) 2023;15:3790. doi: 10.3390/cancers15153790. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Provenzale JM, Mukundan S, Barboriak DP. Diffusion-weighted and perfusion MR imaging for brain tumor characterization and assessment of treatment response. Radiology. 2006;239:632–649. doi: 10.1148/radiol.2393042031. [DOI] [PubMed] [Google Scholar]
- 16.Ahn SS, Shin NY, Chang JH, Kim SH, Kim EH, Kim DW, et al. Prediction of methylguanine methyltransferase promoter methylation in glioblastoma using dynamic contrast-enhanced magnetic resonance and diffusion tensor imaging. J Neurosurg. 2014;121:367–373. doi: 10.3171/2014.5.JNS132279. [DOI] [PubMed] [Google Scholar]
- 17.Moon HH, Park JE, Kim N, Park SY, Kim YH, Song SW, et al. Prospective longitudinal analysis of physiologic MRI-based tumor habitat predicts short-term patient outcomes in IDH-wildtype glioblastoma. Neuro Oncol. 2025;27:841–853. doi: 10.1093/neuonc/noae227. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Park JE, Kim HS, Kim N, Park SY, Kim YH, Kim JH. Spatiotemporal heterogeneity in multiparametric physiologic MRI is associated with patient outcomes in IDH-wildtype glioblastoma. Clin Cancer Res. 2021;27:237–245. doi: 10.1158/1078-0432.CCR-20-2156. [DOI] [PubMed] [Google Scholar]
- 19.Park JE, Oh JY, Park DH, Lee HS, Yoon S, Kim N, et al. Mapping tumor habitats in isocitrate dehydrogenase -wild type glioblastoma: integrating MRI, pathologic, and RNA data from the Ivy Glioblastoma Atlas Project. Neuro Oncol. 2025;27:291–301. doi: 10.1093/neuonc/noae161. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Hoopes A, Mora JS, Dalca AV, Fischl B, Hoffmann M. SynthStrip: skull-stripping for any brain image. Neuroimage. 2022;260:119474. doi: 10.1016/j.neuroimage.2022.119474. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Tustison NJ, Cook PA, Holbrook AJ, Johnson HJ, Muschelli J, Devenyi GA, et al. The ANTsX ecosystem for quantitative biological and medical imaging. Sci Rep. 2021;11:9068. doi: 10.1038/s41598-021-87564-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77:e104–e107. doi: 10.1158/0008-5472.CAN-17-0339. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Bjornerud A, Sorensen AG, Mouridsen K, Emblem KE. T1- and T2*-dominant extravasation correction in DSC-MRI: part I—theoretical considerations and implications for assessment of tumor hemodynamic properties. J Cereb Blood Flow Metab. 2011;31:2041–2053. doi: 10.1038/jcbfm.2011.52. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Tofts PS, Kermode AG. Measurement of the blood-brain barrier permeability and leakage space using dynamic MR imaging. 1. Fundamental concepts. Magn Reson Med. 1991;17:357–367. doi: 10.1002/mrm.1910170208. [DOI] [PubMed] [Google Scholar]
- 25.O’Connor JP, Rose CJ, Waterton JC, Carano RA, Parker GJ, Jackson A. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res. 2015;21:249–257. doi: 10.1158/1078-0432.CCR-14-0990. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Sourbron S, Ingrisch M, Siefert A, Reiser M, Herrmann K. Quantification of cerebral blood flow, cerebral blood volume, and blood-brain-barrier leakage with DCE-MRI. Magn Reson Med. 2009;62:205–217. doi: 10.1002/mrm.22005. [DOI] [PubMed] [Google Scholar]
- 27.Cuenod CA, Balvay D. Perfusion and vascular permeability: basic concepts and measurement in DCE-CT and DCE-MRI. Diagn Interv Imaging. 2013;94:1187–1204. doi: 10.1016/j.diii.2013.10.010. [DOI] [PubMed] [Google Scholar]
- 28.Heye AK, Culling RD, Valdés Hernández Mdel C, Thrippleton MJ, Wardlaw JM. Assessment of blood-brain barrier disruption using dynamic contrast-enhanced MRI. A systematic review. Neuroimage Clin. 2014;6:262–274. doi: 10.1016/j.nicl.2014.09.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Schneider SW, Ludwig T, Tatenhorst L, Braune S, Oberleithner H, Senner V, et al. Glioblastoma cells release factors that disrupt blood-brain barrier features. Acta Neuropathol. 2004;107:272–276. doi: 10.1007/s00401-003-0810-2. [DOI] [PubMed] [Google Scholar]
- 30.Xue W, Du X, Wu H, Liu H, Xie T, Tong H, et al. Aberrant glioblastoma neovascularization patterns and their correlation with DCE-MRI-derived parameters following temozolomide and bevacizumab treatment. Sci Rep. 2017;7:13894. doi: 10.1038/s41598-017-14341-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Kaur B, Khwaja FW, Severson EA, Matheny SL, Brat DJ, Van Meir EG. Hypoxia and the hypoxia-inducible-factor pathway in glioma growth and angiogenesis. Neuro Oncol. 2005;7:134–153. doi: 10.1215/S1152851704001115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Barajas RF, Jr, Phillips JJ, Parvataneni R, Molinaro A, Essock-Burns E, Bourne G, et al. Regional variation in histopathologic features of tumor specimens from treatment-naive glioblastoma correlates with anatomic and physiologic MR Imaging. Neuro Oncol. 2012;14:942–954. doi: 10.1093/neuonc/nos128. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Eidel O, Neumann JO, Burth S, Kieslich PJ, Jungk C, Sahm F, et al. Automatic analysis of cellularity in glioblastoma and correlation with ADC using trajectory analysis and automatic nuclei counting. PLoS One. 2016;11:e0160250. doi: 10.1371/journal.pone.0160250. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Griessmair M, Delbridge C, Ziegenfeuter J, Jung K, Mueller T, Schramm S, et al. Exploring molecular glioblastoma: insights from advanced imaging for a nuanced understanding of the molecularly defined malignant biology. Neurooncol Adv. 2024;6:vdae106. doi: 10.1093/noajnl/vdae106. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data generated and/or analyzed during the current study are not publicly available due to patient privacy but are available from the corresponding author on reasonable request. The underlying code for this study is available in Multiparametric-MRI-Habitat-GBM and can be accessed via this link https://github.com/snuh-rad-aicon/Multiparametric-MRI-Habitat-GBM.