Correlation of contrast-enhancing tumor region on pre-radiotherapy MRI with overall survival in glioblastoma

Abstract

Aim

Heterogeneity assessment of contrast-enhancing tumor regions on pre-radiotherapy magnetic resonance imaging (MRI) holds prognostic significance in glioblastoma (GBM), given its association with tumor progression and overall survival (OS). This study aimed to evaluate the prognostic value of radiomic features extracted from contrast-enhancing tumor regions on pre-radiotherapy MRI across multiple imaging sequences.

Methods

A total of 74 GBM patients from The Cancer Imaging Archive were analyzed and categorized into short- and long-term survival groups based on their median survival of 18.8 months. Radiomic features were extracted from original, wavelet-filtered images, and Laplacian of Gaussian (LoG)-filtered images both individually and in combination across four MRI sequences: T1-weighted contrast-enhanced (T1WCE), T1W, T2-weighted (T2W), and fluid-attenuated inversion recovery (FLAIR). Feature selection was conducted using least absolute shrinkage and selection operator (Lasso) and radiomic scores (Rad-scores) were constructed. The capability of Rad-scores in stratifying OS was evaluated using receiver operating characteristic (ROC) and calibration curve analyses, along with Kaplan–Meier survival analysis. Furthermore, clinical characteristics were integrated to determine whether this combination enhanced the stratification performance.

Results

ROC curve analysis of the Rad-scores demonstrated that wavelet features markedly improved performance compared to the combination of original and LoG features (AUC: 0.90 [0.81–0.96] vs. 0.83 [0.72–0.90]; BH-adjusted p = 0.0183) in differentiating between short- and long-term survivors. Combining wavelet and LoG-filtered features significantly improved survival stratification, outperforming the combination of original and LoG features (AUC: 0.92 [0.84–0.97] vs. 0.83 [0.70–0.90]; BH-adjusted p = 0.0229). Additionally, Kaplan–Meier survival analyses and log-rank tests further confirmed significant correlations between all radiomic feature groups and OS. Integrating clinical characteristics did not notably enhance the discriminatory power (p > 0.05). All the Rad-scores and the combined Rad-sores with clinical data, revealed good agreement between the predicted and actual OS probabilities (p > 0.05), except for the combination of original and LoG features and wavelet features with clinical characteristics (p $<$ 0.05).

Conclusion

Radiomic analysis of contrast-enhancing tumors on pre-radiotherapy MRI sequences provides strong prognostic insights in GBMs. These findings demonstrate the potential of integrating wavelet-filtered radiomic features for risk stratification; prospective validation is required before clinical implementation.

Introduction

Glioblastoma multiforme (GBM), a subtype of high-grade gliomas, is widely recognized as one of the most virulent, lethal, and recurrent primary brain tumors, posing substantial challenges to the medical community. GBMs are histologically classified as high-grade astrocytic tumors and stratified into two primary molecular subtypes: isocitrate dehydrogenase (IDH)-wild type and IDH-mutant. This malignancy accounts for approximately 57% of all gliomas and 48% of primary malignant central nervous system tumors, exhibiting both the highest incidence and the poorest survival outcomes among brain neoplasms. Despite adherence to standard treatment protocols, the prognosis remains dismal, largely due to the pronounced intra-tumoral heterogeneity of GBM, which underlies its therapeutic resistance and frequent recurrence [1–4].

Magnetic resonance imaging (MRI) offers significant contributions to the clinical management of GBM, supporting diagnosis, treatment planning, and monitoring of therapeutic response. High-resolution MRI enables the identification of diverse tumor patterns and facilitates a nuanced characterization of tumor biology through the extraction of quantitative imaging features associated with patient prognosis and survival. The active or contrast-enhanced tumor region, visualized using gadolinium on T1-weighted contrast-enhanced (T1WCE) images, is typically regarded as a hypercellular rich in viable tumor cells, making it particularly informative for evaluating spatial heterogeneity. Complementary evaluation of this contrast-enhancing tumor region using additional MRI sequences, such as T1W, T2-weighted (T2W), and fluid-attenuated inversion recovery (FLAIR), may offer further insights into tumor heterogeneity and improve prognosis assessments related to patients’ survival [5–7].

The standard postoperative MRI protocol for patients with GBM includes three key time points: an early postoperative scan, a pre-radiotherapy scan, and a response assessment scan typically conducted 2–6 weeks following completion of chemo-radiotherapy. The early postoperative MRI is recommended within 48–72 h after surgery to minimize the confounding effects of benign postoperative enhancement, thereby an accurate evaluation of the remaining tumor tissue burden [8–11]. The presence of residual active tumor on postoperative MRI—quantified through volumetric analysis or radiomic features—has been shown to be a significant prognostic indicator of overall survival (OS) in GBM patients [11–13].

While postoperative MRI is widely recognized as a useful tool for assessing residual enhancement and guiding radiotherapy planning, the interval of 4–6 weeks between surgery and the initiation of radiotherapy—considering a tumor doubling time of 21–30 days—may lead to GBM tumor regrowth, a phenomenon referred to as rapid early progression. This has been reported to occur in 19–75% of cases. Failure to account for such progression may lead to insufficient radiation coverage, potentially compromising patient outcomes. To address this, a pre-radiotherapy MRI is recommended to capture early tumor progression and improve delineation of the contrast-enhancing tumor burden, which may, in turn, enhance prognostic accuracy [9, 10, 14–16]. Several studies have evaluated the prognostic value of contrast-enhanced tumor volume on pre-radiotherapy MRI and its correlation with OS in GBM patients [9, 10, 17, 18]. Notably, residual tumor volume measured on pre-radiotherapy scans has been identified as a significantly more reliable predictor of OS compared to tumor volume assessed on postoperative images [9]. However, these investigations have primarily focused on contrast-enhanced volumes in isolation, without integrating information from multimodal MRI sequences for comprehensive tumor assessment.

In the realm of precision medicine, radiomics has emerged as a rapidly advancing research discipline centered on the large-scale extraction of quantitative features from medical images. This technique enables the transformation of complex multimodal imaging data into high-dimensional, mineable information. Radiomics has demonstrated significant potential across a range of clinical applications, including treatment response evaluation, OS and progression-free survival (PFS) prediction, tumor segmentation, disease restaging, lesion classification, and dose prediction [19–23].

This study aimed to investigate the prognostic value of contrast-enhancing tumor regions on pre-radiotherapy MRI by employing radiomics and heterogeneity-based assessment approaches across multiple MRI sequences. Unlike prior studies that predominantly focused on a single modality and volumetric assessment, this work uniquely leverages multimodal imaging to extract a broad spectrum of radiomic features. These features, derived from original images, as well as from wavelet-transformed and Laplacian of Gaussian (LoG)-filtered images, were analyzed independently. The association between these features and patient OS was assessed through receiver operating characteristic (ROC) curve analysis, calibration curve analysis, and Kaplan–Meier survival analysis. This approach underscores the potential of integrating multi-sequence MRI data to enhance prognostic modeling in GBM patients, representing a novel step forward in precision imaging. Furthermore, clinical characteristics were incorporated to assess whether this integration could further enhance stratification accuracy.

Materials and methods

Patients

The patient cohort comprised 74 individuals newly diagnosed with GBM, sourced from The Cancer Imaging Archive (TCIA, http://www.cancerimagingarchive.net/collection/burdenko-gbm-progression/) [24]. As shown in Table 1, the median age was 57.5 years (range 8–85), with 58.11% male patients. The median OS was 18.8 months (range 4.3–68.3). This retrospective study included patients who underwent four pre-radiotherapy MRI sequences: T1W, T2W, and FLAIR. OS was defined as the time elapsed between surgical intervention and death, with a censored status of 1 indicating that the patient was deceased. Patients were stratified into two groups based on survival outcomes: a short-term survival group (OS < 18.8 months) and a long-term survival group (OS ≥ 18.8 months). The cutoff point of 18.8 months corresponds to the median survival time of the study cohort, making it an appropriate method for distinguishing between patient groups. Since median survival represents the time point at which half of the patients are expected to survive, it naturally creates two balanced groups: those with survival times below the median and those above. Moreover, median survival is resistant to outliers, which is advantageous in survival analyses where some patients may have markedly longer or shorter survival times. As the data were obtained from the TCIA, a publicly accessible and anonymized dataset, institutional review board approval was not required for this study.

Table 1.

Clinical and demographic characteristics of patients with GBM

Characteristic	Value
Number of patients	74
Age (years)
Median (range)	57.5 (8–85)
Gender, n (%)
Male	43 (58.11)
Female	31 (41.89)
Overall survival (months)
Median (range)	18.8 (4.3–68.3)
Tumor location, n (%)
Frontal lobe	18 (24)
Temporal lobe	13 (18)
Parietal lobe	9 (12)
Occipital lobe	1 (1.35)
Frontotemporal lobe	3 (4)
Frontoparietal lobe	5 (6.76)
Parietotemporal lobe	15 (20)
Parietooccipital lobe	4 (5.41)
Others	6 (8.11)

Study workflow

The overall workflow of the current study consisted of four main stages, as depicted in Fig. 1, comprises several steps: (1) image preprocessing and registration; (2) image segmentation; (3) feature extraction and reduction; and (4) heterogeneity-survival analysis.

Fig. 1.

Radiomic workflow showing image registration, tumor segmentation, feature extraction, harmonization and reduction, and survival analysis. MRI sequences were registered to an atlas, segmented, and used to extract radiomic features. Selected features informed Rad-score calculation and were evaluated using ROC curve, calibration curve and Kaplan–Meier analyses

First, in the image registration step, pre-radiotherapy MRI sequences, including T1WCE, T1W, T2W, and FLAIR images, were spatially aligned to a standard anatomical atlas to ensure anatomical structures and voxel-wise correspondence across modalities. Following registration, the image segmentation phase was performed to delineate the contrast-enhancing tumor compartment on all co-registered sequences. The segmented region of interest (ROI) was used for subsequent radiomic analysis. Next, in the feature extraction and reduction step, a comprehensive set of radiomic features was extracted from the ROI, including original features, wavelet-transformed features, and LoG-filtered features. Feature reduction was conducted using Lasso with 100 bootstrap iterations to select the most informative and non-redundant features. These selected features were then used to construct radiomic signatures and compute radiomic scores (Rad-scores) for each patient.

In the final stage, heterogeneity-survival analysis, the prognostic performance of the radiomic signatures was evaluated. ROC curve analysis was performed to assess the discriminatory power of the Rad-scores with survival status. Kaplan–Meier analysis with log-rank tests was used to stratify patients into risk groups based on Rad-scores and to investigate the association between imaging-derived heterogeneity and OS.

Following this overview, each stage is described in more detail in the subsequent subsections.

Image preprocessing and segmentation

Image standardization is regarded as a pivotal stage prior to conducting radiomic analysis. MR images acquired across different centers often exhibit variability due to differences in scanner vendors, imaging protocols, magnetic field strengths, and acquisition parameters. These factors can significantly affect image intensity values and impose non-biological variability unrelated to the underlying tumor characteristics [25, 26]. To address these issues, both intensity normalization and voxel size resampling were applied to the multi-center pre-radiotherapy MR images utilized in this study.

To ensure spatial consistency among MRI sequences and alignment with a standard anatomical space, as well as to standardize their voxel size, a series of image registration steps were carried out: (1) affine or rigid registration of T1W, T2W, and FLAIR to T1WCE, using FSL’s flirt function or 'SlicerElastix' extension with 12 or 6 degrees of freedom and a trilinear interpolation; (2) affine or rigid registration of T1WCE to the SRI-24 anatomical atlas; (3) applying the transformation matrix produced from the T1WCE-to-atlas registration to the transformed T1W, T2W, and FLAIR, aligning them with the transformed T1WCE and the same atlas space.

All images were visually inspected after registration for quality assurance. This process resulted in co-registered volumes with slice-to-slice correspondence and a normalized voxel size of 1 mm³ [27, 28]. This registration preserves the accurate anatomical locations, preventing any misalignment in anatomical structures and ensuring that extracted radiomic features are derived from properly aligned images. Following spatial normalization, intensity values were scaled between 0 and 1 using the min–max normalization [28–34]. All preprocessing steps were conducted utilizing the FMRIB Software Library (FSL, version 6.0.6.5) and the 3D Slicer software (version 5.2.2).

Contrast-enhancing tumor regions were manually segmented on pre-radiotherapy T1WCE images by an experienced radiologist. Due to post-surgical changes, hyperintense signals emerge on T1W images, which can obscure the distinction between surgical artifacts and true tumor enhancement. To improve delineation accuracy in such cases, the T1WCE–T1W subtraction images were generated using the FSL to isolate contrast-enhancing tumor regions more clearly. The resulting tumor segmentations were subsequently overlaid onto co-registered MR sequences for downstream radiomic analysis. All manual segmentations were conducted using the 3D Slicer software (version 5.2.2).

Feature extraction and harmonization

Radiomic features were extracted from three types of images: the original MR images, wavelet-transformed images, and LoG-filtered images, across four MRI sequences. For LoG filtering, kernel sizes ranged from 2 to 5 mm to capture image characteristics at varying spatial scales. From the original images, extracted features included shape-based descriptors, first-order statistics, and higher-order texture features. The wavelet-transformed images encompassed both first- and high-order features across multiple scales and frequency components. Similarly, the LoG-filtered images provided additional first- and high-order features sensitive to edge and intensity variations.

In total, for each patient, 107 features were extracted from the original images, 744 from the wavelet-filtered images, and 372 from the LoG-filtered images, all computed within the manually segmented contrast-enhancing tumor regions. The features were subsequently grouped into distinct categories—original, wavelet-based, and LoG-based—for downstream analysis, as well as combined feature sets to assess their complementary value. Radiomic feature extraction was conducted using the 'Radiomics' extension within the 3D Slicer software (version 5.2.2).

Although image preprocessing steps such as intensity normalization and voxel resampling were implemented, inherent multi-center effects persisted in the extracted radiomic features. Therefore, to mitigate the site-related variability, the ComBat harmonization method was applied to all extracted features using the neuroCombat package (version 0.2.12) in Python.

Feature reduction

To preserve radiomic features highly correlated with OS, a two-step selection strategy was implemented. Initially, the Mann–Whitney U test was used to identify radiomic features with AUCs greater than 0.65 and p-values less than 0.05. This step prioritized features demonstrating both discriminative power and statistical significance. Subsequently, Lasso feature selection was conducted over 100 bootstrap iterations to further reduce feature dimensionality. Within each iteration, tenfold cross-validation was employed to optimize the regularization hyperparameter (C), ensuring model generalizability.

Features were ultimately selected based on their frequency of inclusion across all bootstrap iterations, with group-specific thresholds applied to finalize the most robust and relevant features. The following frequency thresholds were used: original features with frequencies exceeding 60, wavelet features with frequencies over 75, LoG features with frequencies above 80, combined original and wavelet features with frequencies greater than 77, combined original and LoG features with frequencies exceeding 87, and combined wavelet and LoG features with frequencies over 70. The combination of original, wavelet, and LoG features was excluded from further analysis, as no original features met the selection criteria within this group. These thresholds were determined empirically by evaluating a range of selection frequency cutoffs across each feature group and selecting the lowest threshold that preserved high AUC performance while reducing redundancy. Due to differences in dimensionality and feature stability across groups—particularly the larger and more redundant wavelet and LoG feature sets—higher thresholds (e.g., 75 and 80) were applied to ensure model sparsity and avoid overfitting. In contrast, a lower threshold (e.g., 60) was sufficient for the original feature group, which contained fewer and more stable features. This group-specific strategy allowed us to optimize model performance and interpretability for each configuration. Finally, radiomic signatures were constructed for each feature group based on the selected features.

Rad-scores were then calculated for each patient based on the selected features included in each radiomic signature. These scores were computed as linear combinations of the selected features, weighted by their respective coefficients, thereby enabling quantitative assessment of the predictive performance of each radiomic signature. Furthermore, radiomic signatures were combined with selected clinical characteristics to evaluate the potential of clinical characteristics, in conjunction with Rad-scores. Age and tumor location characteristics, which demonstrated statistically significant AUC values of 0.66 and 0.63, respectively (p$<$ 0.05), were integrated with the radiomic signatures.

Feature reduction and Rad-score computation were implemented using the 'scikit-learn' library (version1.2.2) in Python.

Prognostic performance of the radiomic and clinico-radiomic models

ROC and calibration curve analysis

To evaluate the discriminative performance of each Rad-score, as well as the combined Rad-scores and clinical characteristics in differentiating between short- and long-term survivors, radiomic and clinico-radiomic models were developed using univariable and multivariable logistic regression models and ROC curve analyses were conducted. Each ROC curve was generated by plotting the true positive rate (TPR) against the false positive rate (FPR) across a range of classification thresholds. The optimal cutoff value, which serves as the best threshold for classifying the binary outcomes, was identified based on the maximum Youden index. The predictive performance of each model was assessed using the area under the ROC curve (AUC). To compare the AUCs across different models, the predicted probabilities served as the classification variable in both univariable and multivariable ROC curve analyses. The statistical significance of the differences was evaluated using the DeLong test. For pairwise comparisons in radiomic models, the Benjamin–Hochberg (BH) procedure was applied to adjust p-values to control the false discovery rate (FDR) associated with multiple hypothesis tests. After applying the BH correction at a significance level of 0.05, the adjusted significance threshold was modified from 0.05 to 0.0229.

Calibration curves serve as informative tools for assessing the reliability of predicted probabilities in classification tasks. Thus, calibration curves were plotted to examine how closely the predicted probabilities from radiomic and clinico-radiomic models align with the observed outcomes. Additionally, the Hosmer and Lemeshow test was conducted to quantify the agreement between the models' predictions and the actual outcomes, providing a measure of the models' calibration performance. The significance level was set at 0.05; p-values below this threshold indicated poor calibration (lack of fit), while those above 0.05 suggested good calibration (no significant lack of fit).

Kaplan–Meier analysis

Kaplan–Meier survival analysis was conducted to evaluate the association between radiomic-derived scores and OS. For each radiomic signature, patients were stratified into low and high-risk groups based on the optimal cutoff value, determined using Youden’s index from ROC curve analysis of the calculated Rad-scores. The optimal cutoff values for various signatures—the original, wavelet, LoG, the combined original and wavelet, original and LoG, and wavelet and LoG—were identified as 0.51, −2.45, 1.12, −1.79, −2.13, and −0.28, respectively. Survival differences between groups were evaluated using the log-rank test, with p-values less than 0.05 considered statistically significant.

All ROC and Kaplan–Meier analyses were carried out using MedCalc Statistical software (version 23.2.1) (MedCalc Software Ltd, Ostend, Belgium; https://www.medcalc.org; 2024). Calibration curves were plotted using the 'scikit-learn' library (version1.2.2) in Python.

Results

Feature reduction

Figures 2 and 3 present bar plots of the selected radiomic features with non-zero coefficients identified through Lasso feature selection across the different radiomic groups. The features are color-coded based on their origin: features extracted from original images are shown in orange, those from wavelet-transformed images in blue, and those from LoG-filtered images in green. Figure 2 consists of three bar plots (a–c) that visualize the mean coefficients of radiomic features selected via Lasso across different feature groups. Panel a displays the original features with non-zero coefficients derived from CE, T1W, T2W, and FLAIR MRI sequences. Among these, T1-original first-order 10 percentile shows the highest positive coefficient, suggesting a stronger association with the OS. In contrast, T1-original first-order minimum feature exhibits a negative coefficient, indicating an inverse relationship with the outcome.

Fig. 2.

Bar plots showing the mean coefficients of the top radiomic features selected from three different feature sets: a original features, b wavelet-transformed features, and c LoG-filtered features

Fig. 3.

Bar plots showing the mean coefficients of the top radiomic features selected from three different feature sets: a original + wavelet features, b original + Laplacian of Gaussian (LoG) features, and c wavelet + LoG features

Panel b presents wavelet-transformed features, which generally show higher positive coefficient magnitudes than original features. The feature FLAIR-wavelet-LHL glcm ClusterShade stands out with the highest positive coefficient across all features.

Panel c includes features derived from the LoG-filtered images. Several features from T1W, T2W, and CE sequences contribute meaningfully to the model, with T2-log-sigma-5–0 mm-3D ngtdm coarseness exhibiting the largest negative coefficient among all features, suggesting it is a highly influential predictor in the LoG group.

Overall, the wavelet and LoG-transformed features exhibit greater discriminative power and higher coefficient values compared to the original features, underlining the added value of multimodal and multi-scale feature extraction in enhancing predictive performance.

In Fig. 3, the mean coefficients of selected radiomic features derived from three different combinations of image processing methods—original + wavelet, original + LoG, and wavelet + LoG—are presented across three subpanels (a, b, and c). Panel (a) displays the top features obtained from the combination of original and wavelet-transformed images. Among these, most of the predictive features were derived from the wavelet domain, including T2-wavelet-LLH glszm HighGrayLevelZoneEmphasis, T1-wavelet-HHH first-order uniformity, and CE-wavelet-HLL glcm Imc2. Only one feature in this panel—T1-original first-order 90Percentile—originated from the original image space, and it was the one with the highest coefficient in original features signature.

Panel (b) presents the mean coefficients for features derived from original and LoG-filtered images. This combination highlights the strong predictive contribution of LoG-derived features, extracted with sigma values of 2.0, 4.0, and 5.0 mm. Notably, CE-log-sigma-2–0-mm-3D glrlm ShortRunHighGrayLevelEmphasis and T2-log-sigma-4–0-mm-3D glrlm LowGrayLevelRunEmphasis exhibit some of the highest positive coefficients, reaching values close to 5. In contrast, original feature of CE-original first-order Minimum is associated with relatively lower coefficient.

Panel (c) illustrates the features obtained from a combination of wavelet and LoG transformations. The coefficients in this group are more modest in magnitude compared to panel (b), ranging approximately from −0.5 to + 2.5, suggesting a more balanced contribution of features from both domains.

ROC and calibration curve analyses

The discriminative performance of various radiomic models in distinguishing between short- and long-term survival groups is summarized in Table 2. All Rad-scores derived from the respective feature groups demonstrated a statistically significant association with OS, effectively stratified patients into distinct survival categories, as evidenced by AUC values and p-values (p $<$ 0.0001). Notably, the wavelet features showed a markedly improved performance compared to the combination of original and LoG features (AUC: 0.90 [0.81–0.96] vs. 0.83 [0.72–0.90]; BH-adjusted p = 0.0183). Likewise, the combined wavelet and LoG features significantly surpassed the performance of the original and LoG features group (AUC: 0.92 [0.84–0.97] vs. 0.83 [0.72–0.90]; BH-adjusted p = 0.0229). No statistically significant differences were observed among the remaining comparisons (BH-adjusted p > 0.0229). These discrepancies in classification performance are further illustrated by the ROC curves presented in Fig. 4a.

Table 2.

ROC and calibration curve analyses for individual and combined radiomic models

Radiomic model	AUC (95% CI)	p-value	Youden index	Hosmer and Lemeshow test p-value
Original	0.80 (0.69–0.88)	$<$ 0.0001	0.5326	0.31
Wavelet	0.90 (0.81–0.96)	$<$ 0.0001	0.6689	0.87
LoG	0.85 (0.75–0.92)	$<$ 0.0001	0.6293	0.24
Original + wavelet	0.91 (0.81–0.96)	$<$ 0.0001	0.7260	0.92
Original + LoG	0.83 (0.72–0.90)	$<$ 0.0001	0.6176	0.04
Wavelet + LoG	0.92 (0.84–0.97)	$<$ 0.0001	0.7604	0.20

Fig. 4.

ROC curves illustrating the classification performance of different radiomic and clinic-radiomic models in distinguishing between short- and long-term survivals

Incorporating clinical variables into the various Rad-scores, as shown in Table 3, did not lead to any significant enhancement in the models' discriminative ability (p > 0.05). The ROC curve comparisons are displayed in Fig. 4b.

Table 3.

ROC and calibration curve analyses for individual and combined clinico-radiomic models

Clinico-radiomic model	AUC (95% CI)	p-value	Youden index	Hosmer and Lemeshow test p-value
Original + clinical	0.86 (0.76–0.93)	$<$ 0.0001	0.6352	0.34
Wavelet + clinical	0.94 (0.86–0.98)	$<$ 0.0001	0.8432	0.01
LoG + clinical	0.84 (0.74–0.92)	$<$ 0.0001	0.6491	0.42
Original + wavelet + clinical	0.91 (0.81–0.96)	$<$ 0.0001	0.7121	0.75
Original + LoG + clinical	0.88 (0.78–0.94)	$<$ 0.0001	0.7239	0.43
Wavelet + LoG + clinical	0.96 (0.89–0.99)	$<$ 0.0001	0.8374	0.65

Calibration curve analysis of all radiomic and clinico-radiomic models revealed non-significant results based on the Hosmer–Lemeshow test (p > 0.05), indicating good agreement between the predicted and actual probabilities of OS. However, exceptions were observed for the combination of original and LoG features in the radiomic models and the combination of wavelet features with clinical characteristics in the clinico-radiomic models (p $<$ 0.05). The calibration curves for both radiomic and clinico-radiomic models are illustrated in Fig. 5.

Fig. 5.

Calibration curves illustrating the agreement between predicted probabilities and observed outcomes in different radiomic and clinico-radiomic models for distinguishing between short- and long-term survival

Kaplan–Meier analysis

Further evaluation of the predictive capability of radiomic signatures derived from the contrast-enhancing tumor region on pre-radiotherapy MRI was conducted using Kaplan–Meier analysis (Fig. 6). For each radiomic feature class, patients were categorized into low- and high-risk groups based on the optimal cutoff values from ROC curve analysis of Rad-scores. Notably, Rad-scores derived from wavelet, LoG, and combined feature classes all demonstrated highly significant p (p < 0.0001), indicating a stronger association between Rad-scores and OS in GBM patients. In contrast, only original features, showed p-value of 0.0006, suggesting a comparatively lower prognostic strength.

Fig. 6.

Kaplan–Meier survival curves illustrating the association between rad-score-based risk groups and overall survival in GBM patients for different radiomic feature classes: a original, b wavelet, c Laplacian of Gaussian (LoG), d combined original and wavelet, e combined original and LoG, and f combined wavelet and LoG features

Discussion

The findings of this study support the hypothesis that the contrast-enhancing tumor burden visible on multimodal pre-radiotherapy MRI holds prognostic relevance in GBM. This finding was achieved from the analysis of 74 newly diagnosed GBM patients who underwent chemo-radiotherapy. To evaluate the association between the pre-radiotherapy contrast-enhancing tumor and OS, a comprehensive radiomics-based approach was employed. Radiomic signatures (Rad-scores) were constructed using features extracted from original, wavelet-filtered, and LoG-filtered images across four MRI modalities. The stratification capability of the signatures was evaluated through ROC and calibration curve analyses, along with Kaplan–Meier survival analysis. Additionally, the combined Rad-scores and clinical characteristics were incorporated to determine whether this integration enhanced the stratification performance.

Univariable ROC analysis of the Rad-scores confirmed the predictive strength of the contrast-enhancing tumor prior to radiotherapy, with AUC values of 0.80 (95% CI 0.69–0.88), 0.90 (0.81–0.96), 0.85 (0.75–0.92), 0.91 (0.81–0.96), 0.83 (0.72–0.90), and 0.92 (0.84–0.97) for Rad-scores derived from individual classes‒original, wavelet, LoG‒as well as combinations of original with wavelet, original with LoG, and wavelet with LoG signatures, respectively. Kaplan–Meier analysis, which stratified patients into high- and low-risk groups based on optimal Rad-score thresholds, demonstrated statistically significant differences in OS across all signatures, as determined by log-rank tests. These results underscore the prognostic utility of radiomic signatures derived from the contrast-enhancing tumor on multimodal pre-treatment MRI.

Multivariable ROC analysis demonstrated no substantial improvement in discrimination power compared to univariable ROC analysis between short- and long-term OS groups (p > 0.05).

In this study, calibration curve analysis was conducted instead of solely assessing the discriminative power of the models. While AUC metric focus on the model's ability to differentiate between short- and long-term outcomes, it does not convey information about the accuracy of the predicted probabilities themselves. Calibration curve analysis of all radiomic and clinico-radiomic models indicated no significant lack of fit by the Hosmer–Lemeshow test (p > 0.05), suggesting good agreement between the predicted and observed probabilities of OS. However, the combination of original and LoG features in the radiomic models, as well as the integration of wavelet features with clinical characteristics in the clinico-radiomic models, showed significant lack of fit (p $<$ 0.05), despite rendering high AUC values of 0.83 and 0.94, respectively. This highlights that high discrimination does not necessarily equate to good calibration; a model can accurately rank patients but still over- or under-estimate their actual risk.

Although pre-radiotherapy MRI has been adopted in the treatment workflow at certain institutions, it is not yet universally mandated, as current imaging guidelines for GBM patients do not specify it as a requirement. Nonetheless, acquiring MR images prior to radiotherapy offers several potential clinical advantages. It can enhance the accuracy of treatment planning by refining the delineation of target volumes, assist in discriminating true tumor progression from radiation-induced pseudo-progression, and ultimately influence patient prognosis and therapeutic decision-making [9, 15, 36, 37].

While most prior radiomic studies have focused on preoperative MRI, our study uniquely centers on the contrast-enhancing tumor in pre-radiotherapy MRI. This imaging time point is clinically more actionable, reflecting early post-surgical dynamics that directly inform adjuvant therapy decisions. Previous research investigating the impact of pre-radiotherapy MRI-detected contrast-enhancing tumors on GBM survival has predominantly centered on volumetric analyses. These investigations have consistently identified contrast-enhancing tumor volume as a significant predictor of patient survival. For instance, Alafandi et al. [17] evaluated the prognostic relevance of contrast-enhancing tumor volume on pre-radiotherapy MRI and reported a strong correlation between contrast-enhancing tumor volume and OS in GBM patients stratified by MGMT promoter methylation status. Their study, employing Cox proportional hazards models and Kaplan–Meier survival analysis, showed that patients with larger contrast-enhancing tumor volumes experienced shorter OS. Similarly, Saraswathy et al. [18] reported a strong association between contrast-enhancing tumor volume and survival outcomes across anatomical, physiological, and metabolic pre-radiotherapy imaging modalities. Additionally, another study emphasized the prognostic superiority of pre-radiotherapy MRI over postoperative imaging. The findings indicated that contrast-enhancing tumor volume measured before radiotherapy was a significant predictor of survival outcomes (p = 0.007), whereas the post-surgery residual tumor volume did not demonstrate a statistically significant relationship with survival (p = 0.32) [9]. Building on this foundation, the present study focused on analyzing the intra-tumoral heterogeneity of the contrast-enhancing tumor on pre-radiotherapy MRI through radiomic analysis across four conventional MRI sequences. The results from both ROC curve and Kaplan–Meier survival analyses demonstrated that radiomic features extracted from the contrast-enhancing tumor serve as robust prognostic biomarkers, in alignment with previous volumetric studies.

It is important to acknowledge that contrast-enhancing granulation tissue can develop within weeks following surgery, potentially confounding interpretations of contrast enhancement. Despite this, the observed statistically significant correlation between radiomic-derived signatures and OS may imply that the imaging phenotype of the contrast-enhancing tumor exerts a predominant influence.

Radiomic features were analyzed across multiple feature types, including those derived from original, wavelet-filtered, and LoG-filtered images, and examined both independently and in various combinations. Remarkably, features were extracted from the contrast-enhancing tumor region across four MRI sequences—T1WCE, T1W, T2W, and FLAIR—highlighting the comprehensive nature of the analysis. This multi-sequence approach enables the integration of complementary information, potentially capturing a broader spectrum of tumor characteristics and enhancing prognostic accuracy beyond that achievable by any single modality. All evaluated radiomic scores demonstrated significant discriminative potential, as reflected by statistically significant AUC values (p $<$ 0.0001).

When comparing the individual prognostic performances of the radiomic feature types, no statistically significant differences were observed among original, wavelet, and LoG features in pairwise comparisons. However, wavelet features significantly improved predictive performance compared to the combination of original and LoG features (AUC: 0.90 [0.81–0.96] vs. 0.83 [0.72–0.90]; BH-adjusted p = 0.0183). Additionally, the combined wavelet and LoG features outperformed the combination of original and LoG features (AUC: 0.92 [0.84–0.97] vs. 0.83 [0.72–0.90]; BH-adjusted p = 0.0229). These findings suggest that wavelet-derived texture features—GLCM and GLSZM classes—offer strong individual predictive power, as well as containing complementary and valuable prognostic information when integrated with LoG features.

This improvement is likely attributable to the ability of the wavelet transform to decompose MR images into multiple scales and resolutions, thereby capturing both coarse and fine textural patterns within the tumor region. Additionally, wavelet filtering enhances signal-to-noise ratio, potentially improving the robustness and discriminative capacity of the extracted features for survival prediction.

The FLAIR-wavelet-LHL glcm ClusterShade feature exhibited the greatest positive contribution among the wavelet features. This feature, which measures the asymmetry of texture heterogeneity at the LHL wavelet-filtered scale, indicates that higher values are associated with longer survival. The association between increased heterogeneity and improved prognosis suggests that certain tumor heterogeneity patterns may be linked to less aggressive tumor behavior or a better response to treatment. Although wavelet features demonstrated predictive capability, their direct connection to specific tumor biology remains uncertain. Radiomic features capture tumor characteristics at specific scales limited by the resolution of the imaging modality, and their biological basis has yet to be fully elucidated in the majority of research. To interpret these features biologically, substantial efforts in biological validation including genetic and histopathologic analyses are necessary, which is critical for establishing radiomics as a dependable and independent prognostic tool in cancer management [38].

Collectively, these findings highlight that the integration of radiomic signatures derived from pre-radiotherapy MRI into existing clinical workflows offers a promising advancement in personalized patient management. These radiomic tools serve as non-invasive, quantitative biomarkers that can augment traditional assessment methods, providing reproducible data to inform clinical decision-making. By identifying high-risk patients through specific radiomic signatures, clinicians can tailor treatment strategies more effectively, such as implementing more aggressive therapies, considering alternative interventions, or increasing the frequency of monitoring to detect early signs of progression. Compared to conventional approaches that predominantly rely on qualitative imaging assessments and clinical factors, radiomics introduces a data-driven layer of prognostic evaluation. Consequently, treatment plans can be modified promptly, potentially improving patient outcomes.

Recent developments in clinical decision support tools (CDSTs) have underscored the importance of integrating radiomics into interpretable and workflow-compatible platforms [39]. However, the practical implementation of radiomic-based tools faces several challenges, including heterogeneity in imaging protocols, variability in segmentation, and the need for external validation [40]. Furthermore, ensuring seamless integration into clinical environments requires that these tools demonstrate added value over existing nomograms and clinician-based assessments. To this end, radiomics must not only outperform traditional predictors but also offer clinically interpretable outputs that align with physician decision pathways [41]. For successful translation into practice, a realistic implementation strategy should include multi-center prospective validation, regulatory considerations for software-as-a-medical-device (SaMD) approval, user-centered interface design, and interoperability with electronic health records [42]. These steps are essential to ensure that radiomic-based decision tools are not only scientifically robust but also feasible and impactful in routine clinical care.

Nevertheless, this study provides only an initial step toward clinical translation, based on retrospective data from a limited cohort. Future studies should focus on prospective validation, standardization across institutions, and evaluation of workflow feasibility to ensure successful integration into routine clinical practice.

Although recent studies increasingly explore deep learning (DL), foundation models, and federated learning in radiomics, this study deliberately employs handcrafted radiomic features combined with Lasso-based statistical modeling to prioritize transparency, reproducibility, and clinical feasibility. Rather than using complex machine learning classifiers, we focused on statistical techniques such as logistic regression and survival analysis to evaluate the prognostic value of radiomic signatures. DL-based approaches typically require large, annotated datasets and face challenges in interpretability and clinical integration [43, 44]. In contrast, handcrafted radiomic features—paired with well-established statistical methods—are standardized, biologically interpretable, and more appropriate for implementation in data-constrained settings such as pre-radiotherapy GBM imaging [45].

Recent advances in the field have increasingly embraced deep radiomic approaches, which utilize DL architectures to automatically learn hierarchical imaging features directly from raw MRI data [46]. These methods have shown promise in capturing complex tumor heterogeneity and subtle imaging patterns that may be overlooked by handcrafted features, potentially improving prognostic performance and robustness across datasets [47]. However, challenges such as the need for large, well-annotated datasets and the interpretability of deep features remain areas of ongoing research [48].

Simultaneously, multi-omics integration strategies—combining imaging radiomics with genomic, transcriptomic, and proteomic data—are gaining traction in neuro-oncology [49]. Integrative models leveraging multimodal data offer a more comprehensive characterization of tumor biology, enabling refined risk stratification and personalized therapy planning. Incorporating molecular markers such as MGMT methylation status and IDH mutation alongside imaging features holds particular potential to enhance predictive accuracy and clinical relevance.

Moreover, federated learning frameworks have recently emerged as a powerful approach to enable collaborative model training across multiple institutions without sharing sensitive patient data [50, 51]. This paradigm facilitates the pooling of diverse and larger datasets, which is critical for developing generalizable and robust radiomic models, especially for rare cancers like GBM [52, 53]. Federated learning helps overcome privacy and data-sharing barriers that traditionally limit multi-center studies and may accelerate the clinical translation of AI-driven imaging biomarkers [54].

Together, these innovations represent promising directions to complement and extend the handcrafted radiomics and statistical modeling approach adopted in this study, setting the stage for future work that integrates DL, multi-omics data, and decentralized collaborative learning to optimize prognostic modeling in GBM.

This study has several limitations. First, the relatively small sample size limited the statistical power and precluded validation in an independent cohort, thereby restricting the generalizability of the findings. Due to the lack of pre-radiotherapy MRI scans in accessible local hospitals, only data from the TCIA dataset were utilized. Future studies should focus on multi-institutional collaborations, which would be more appropriate for validating established radiomic techniques. Second, the segmentation of the contrast-enhancing tumor region was performed manually by a single experienced radiologist. While expert involvement ensures high-quality delineation, manual delineation is inherently subjective and vulnerable to intra- and inter-observer variability, and future research should explore the use of semi-automated or fully automated segmentation methods to improve reproducibility and reduce operator bias [55–57]. Another limitation concerns the conversion of OS into binary categories based on the median, which can lead to the loss of detailed time-to-event information. Statistical approaches such as Cox proportional hazards regression or concordance index (C-index) analysis could offer a more comprehensive evaluation of survival outcomes by leveraging the full distribution of survival times rather than dichotomizing the outcome. Additionally, clinical factors such as MGMT methylation status and IDH mutation status were not available for all patients, and information on the extent of surgical resection was missing. Consequently, these characteristics could not be incorporated into the current analysis. Future studies should include this clinically and biologically relevant factors to improve the thoroughness of the analysis.

Although post-radiotherapy MRI data were available for some patients, the incomplete imaging sequences and challenges related to distinguishing pseudo-progression from true progression limited their inclusion in this study. Nevertheless, future research should explore the use of temporal radiomics, as it could provide deeper insights into tumor evolution, treatment response, and spatiotemporal heterogeneity in GBM.

Conclusion

This study demonstrates that radiomic features extracted from contrast-enhancing tumor regions on multimodal pre-radiotherapy MRI scans can effectively stratify overall survival in GBM patients. These findings support the clinical value of incorporating pre-radiotherapy MRI and radiomics into treatment planning to enhance survival stratification and personalize therapy for GBM patients. Future studies with larger patient cohorts are warranted to validate and further refine these prognostic models.

Abbreviations

BH

Benjamin–Hochberg

DL

Deep learning

FLAIR

Fluid-attenuated inversion recovery

FMRIB

Functional magnetic resonance imaging of the brain

FSL

FMRIB software library

FPR

False positive rate

FDR

False discovery rate

GBM

Glioblastoma multiform

IDH

Isocitrate dehydrogenase

LoG

Laplacian of Gaussian

Lasso

Least absolute shrinkage and selection operator

MRI

Magnetic resonance imaging

OS

Overall survival

PFS

Progression-free survival

ROC

Receiver operating characteristic

ROI

Region of interest

Rad-score

Radiomic score

T1WCE

T1-weighted contrast-enhanced

T2W

T2-weighted

TCIA

The Cancer Imaging Archive

TPR

True positive rate

Author contributions

Atefeh Mahmoudi: Study design, Investigation, Methodology, Conceptualization, Formal analysis, Validation, Writing—original draft, Writing—review & editing. Arash Zare Sadeghi: Methodology, Validation, Writing—review & editing. Hamed Iraji: Data curation, Visualization, Conceptualization, Writing—review & editing. Maedeh Barahman: Data curation, Visualization, Writing—review & editing. Elmira Yazdani: Formal analysis, Writing—review & editing. Seied Rabi Mahdavi: Study design, Funding acquisition, Project administration, Supervision, Writing—review & editing.

Funding

This work was supported by the School of Medicine, Iran University of Medical Sciences (IUMS), and Grant number 24094.

Data availability

This study used the publicly available dataset sourced from The Cancer Imaging Archive (TCIA, [http://www.cancerimagingarchive.net/collection/burdenko-gbm-progression/] (http:/www.cancerimagingarchive.net/collection/burdenko-gbm-progression)).

Declarations

Ethics approval and consent to participate

Since the dataset utilized in this study is publicly accessible, an ethics statement is not necessary.

Competing interests

The authors declare no competing interests.