Machine learning-based integration of dosiomics and pre-radiotherapy multimodal MRI radiomics for survival stratification in patients with glioblastoma multiforme

Abstract

Background

Tumor heterogeneity is a significant factor contributing to the marked differences in survival rates among glioblastoma multiforme (GBM) patients, who face a poor prognosis. To improve personalized treatment, it is essential to identify specific tumor characteristics that capture this variability and aid in predicting survival. This study aimed to evaluate the utility of dosiomics and radiomics in predicting overall survival (OS). The central hypothesis was that integrating dosiomics and radiomics could improve survival outcome predictions.

Methods

A total of 74 GBM patients from The Cancer Imaging Archive were retrospectively included. Dosiomic features from the gross tumor volume (GTV) of planned dose distributions, along with radiomic features from the contrast-enhanced tumor (CET) and edema/non-contrast-enhanced tumor (ED/nCET) subregions across various pre-radiation MRI modalities, were extracted and optimized using L1-based feature selection. Logistic Regression (LR) models were built utilizing different feature configurations to assess the discriminative power of dosiomic and radiomic features, considering the impact of heterogeneous subregions. Model performance was assessed through stratified 10-fold cross-validation (CV).

Results

The dosiomic model exhibited a mean area under the receiver operating characteristic (ROC) curve (AUC) of 0.80 0.12. The subregion-based models demonstrated mean AUC values of 0.90 0.09 for the CET subregion and 0.76 0.10 for the ED/nCET subregion, indicating that the CET subregion significantly outperformed the ED/nCET subregion (p-value < 0.05). The mean AUC values for modality-based models were as follows: 0.86 0.12 for T1-weighted contrast-enhanced (T1CE), 0.84 0.18 for T1-weighted (T1), 0.85 0.14 for T2-weighted (T2), and 0.76 0.21 for fluid-attenuated inversion recovery (FLAIR) sequences. There was no significant difference in discrimination power among the four modalities (p-value > 0.05). The combined dosiomic and CET model improved performance to 0.96 0.07 (p < 0.05).

Conclusions

Dosiomic and pre-radiotherapy MRI-derived radiomic features are capable of stratifying GBM patients into two long-term and short-term groups. Notably, the integration of dosiomics and radiomics significantly enhances survival prediction in GBM patients.

Introduction

Glioblastoma multiforme (GBM), the most devastating and threatening primary brain neoplasm, develops from glial or neural progenitor cells that undergo genetic and epigenetic alterations. It is characterized by an almost inevitable tendency to relapse after rigorous therapy and has a dismal prognosis. It can impact individuals of any age; however, it predominantly affects the older adult population. Maximum safe surgical resection followed by chemoradiotherapy is the current standard treatment of GBM tumors. However, the aggressive and infiltrative nature of GBM prevents the possibility of achieving a curative treatment, which lowers the median survival time for the majority of patients to under two years [1–7].

Despite the remarkable advances in treatment, the inconsistent response remains a significant challenge. This poor prognosis could be associated with the high degree of intrinsic intra-tumoral heterogeneity. Such heterogeneity is characterized by the presence of both clonal and subclonal differentiated tumor cell populations, glioma stem cells, and various elements of the tumor microenvironment [8–11]. The broad spectrum of the overall survival (OS) of GBM patients despite a similar treatment and poor prognosis highlights the need for precision medicine and personalized treatment [12]. Based on the precision medicine definition of the National Institute of Health (NIH), cancer treatment should be customized based on the biological heterogeneity of each individual. Therefore, determining tumor-specific prognostic markers and predicting patients’ survival outcomes is pivotal for ideal and personalized treatment [6, 13].

Identifying the molecular profile of GBM (genomics) is a widely utilized approach for assessing intra-tumor heterogeneity. This genomic information is typically captured through high-throughput sequencing techniques such as whole-exome sequencing, targeted gene panels, or RNA sequencing, which examine tumor DNA and RNA to reveal genetic variations and expression patterns. Despite genomics offering enhanced accuracy in predicting treatment responses and survival outcomes, its invasive nature and limitations in capturing the full heterogeneity of tumors underscore the necessity for a non-invasive and robust technique to assess tumor characteristics [8, 14, 15]. Although tumor heterogeneity complicates the use of biopsy-based genomics, it provides the capacity for the application of medical imaging-based radiomics as a non-invasive and repeatable alternative [9, 12]. Radiomics is an imaging algorithm that extracts a large number of high-throughput quantitative imaging features related to shapes, intensities, and textures, enabling the assessment of intra-tumor heterogeneity for predicting clinical outcomes such as survival [2, 9, 11, 16–20].

Numerous studies have been undertaken to predict the survival of GBM patients using preoperative magnetic resonance imaging (MRI)-based radiomics, employing either machine learning models or conventional statistical methods [2, 8, 16, 21–26]. Preoperative MRI offers a superior evaluation of tumor integrity and heterogeneity. In contrast, postoperative and pre-radiotherapy MRI, which reveal residual enhanced tumor and rapid early progression (REP) on pre-radiotherapy MRI—with nearly a 50% prevalence rate—[27–29] exhibit significant prognostic value associated with survival. The potential of pre-radiotherapy MRI, which accounts for tumor progression between surgery and the initiation of radiotherapy, has been demonstrated to predict the survival of GBM patients [30–32]. In addition to examining the contrast-enhanced tumor, the non-enhanced component or edema has also been evaluated as a potential predictor of survival outcomes in patients with GBM [31]. The mentioned studies on postoperative or pre-radiotherapy MRI lack an assessment of tumor subregions utilizing multimodal MRI radiomics, including T1-weighted (T1), T1-weighted contrast-enhanced (T1CE), T2-weighted (T2), and fluid-attenuated inversion recovery (FLAIR) imaging, as well as machine learning model evaluation.

Regarding the primary role of radiotherapy in the treatment course of GBM, studies indicated that dose-volume histogram (DVH) parameters can serve as additional prognostic factors for survival in GBM patients [33, 34]. The main drawback of DVH-based metrics is their reliance on partial information from the three-dimensional planned dose distribution, leading to a loss of spatial information. In contrast, the dosiomic concept has introduced innovative dose distribution metrics that extract quantitative features encoding the spatial three-dimensional dose map [35–38]. Notwithstanding the promising results from radiomic- and dosiomic-based models individually, combining the two has been shown to enhance the accuracy of prognosis prediction for other tumors [39, 40].

To the best of our knowledge, no research has evaluated the association between dosiomic features and survival in GBM patients. Therefore, this study was designed to investigate the prognostic capacity of dosiomic features derived from the planned dose distribution for GBM patients. Notably, it was hypothesized that integrating dosiomics and radiomics into a single model would provide an enhancement effect for survival prediction. To achieve this, machine learning models were initially constructed that utilized dosiomic and radiomic variables extracted from dose distributions and pre-radiotherapy MRI scans to differentiate between short- and long-term OS. The predictive performance of each subregion and modality, as well as their combination, was specifically evaluated. Subsequently, a combined model was developed that integrated both dosiomics and radiomics, aiming to enhance the accuracy of the prediction model.

Material and methods

Study population

A total of 74 patients newly diagnosed with glioblastoma treated by radiotherapy at the Burdenko National Medical Research Center of Neurosurgery between 2014 and 2020 from The Cancer Imaging Archive (TCIA, http://www.cancerimagingarchive.net/collection/burdenko-gbm-progression/) [41] were retrospectively included in this study. The dataset encompasses 180 patients; however, information regarding the date of death is only available for 78 of them. Thus, 78 patients were eligible based on the following inclusion criteria: 1) complete pre-radiotherapy MRI sequences, including T1, T1CE, T2, and FLAIR; 2) associated radiotherapy planning files comprising RTSTRUCT and RTDOSE files; and 3) known dates of surgery and death. Following exclusion criteria that eliminated patients with poor image quality, making GBM subregions segmentation and texture analysis inaccurate, and a patient who exhibited no enhancement after surgery, a total of 74 patients were ultimately analyzed. OS time was defined as the interval between the date of surgery and the date of death. Accordingly, patients were stratified into two binary classes based on their median survival time: long-term survival (OS ≥ 18.8 months, n = 39, label:1) and short-term survival (OS < 18.8 months, n = 35, label:0). The summary of patient characteristics for both long-term and short-term classes is presented in Table 1. As the TCIA datasets do not include patient identifiers, hence, ethical approval from an institutional review board was not required.

Table 1.

Characteristics of GBM patients in long-term and short-term classes

Characteristic	Class		P- value
	Long-term	Short-term
Patient No. (%)	39 (52.7%)	35 (47.3%)
Age (years), mean SD	52.44 14.47	60.94 11.70	0.017
Gender, No. (%)			0.94
Female	17 (44%)	14 (40%)
Male	22 (56%)	21 (60%)
Tumor location, No. (%)			0.73
Frontal lobe	10 (25%)	8 (23.5%)
Temporal lobe	8 (20%)	5 (14.7%)
Parietal lobe	3 (7.5%)	6 (17.6%)
Occipital lobe	1 (2.5%)	- (0%)
Frontotemporal lobe	2 (5%)	1 (2.9%)
Frontoparietal lobe	2 (5%)	3 (8.8%)
Parietotemporal lobe	9 (22.5%)	6 (17.6%)
Parietooccipital lobe	3 (7.5%)	1 (2.9%)
others	2 (5%)	4 (11.8%)
Hemisphere, No. (%)			0.24
Left	27 (69.2%)	20 (57.1%)
Right	12 (30.8%)	13 (37.1%)
Bilateral	- (0%)	2 (5.7%)
OS (months), mean SD	30.86 12.47	12.56 3.96	NA

SD:standard deviation; OS: overall survival; NA: not applicable

Methodology

The workflow proposed for overall OS prediction, illustrated in Fig. 1, includes the following steps: 1) dataset and target definition, 2) preprocessing of multimodal MR images and planned dose distribution files, 3) delineation and labeling of regions of interest (ROIs), 4) extraction of dosiomic and radiomic features, 5) selection of relevant features using a supervised method, and 6) development of both single and combined predictive models.

Fig. 1.

Overview of the study pipeline for survival prediction in GBM patients. The workflow includes dataset definition, image preprocessing, registration, ROI delineation, feature extraction from MRI and RT dose data, L1-based feature selection, and machine learning model development, such as logistic regression with cross-validation. Models were built using individual and combined radiomic and dosiomic features, followed by performance evaluation through ROC analysis

Image preprocessing

The MRI studies sourced from the TCIA dataset were collected from various sites, utilizing different scanner vendors and diverse scanning protocols. These variations can influence radiomic features; therefore, image standardization is essential to normalize intensity and spatial resolution before conducting three-dimensional radiomic analysis [11, 17]. The entire preprocessing pipeline was conducted using the FMRIB Software Library (FSL) (version 6.0.6.5) [42] and the 3D Slicer software (version 5.2.2) (http://www.slicer.org) [43]. Initially, MRI scans were converted to Neuroimaging Informatics Technology Initiative (NifTI) format. Prior to image registration, the scans were reoriented to the RAS (neurological) format, which is the standard orientation in FSL. The registration process involved several steps: 1) T1, T2, and FALIR scans were registered to T1CE as a reference using the FLIRT algorithm in FSL; 2) T1CE scans were aligned with the SRI-24 atlas; and 3) the registered T1, T2, and FLAIR scans were registered to the atlas by applying the transformation matrix obtained during the T1CE to atlas registration, resulting in co-registered, resampled volumes with isotropic voxel dimensions of 1 mm³ [2] (Fig. 2). In most instances, affine registration was successfully completed using FLIRT in FSL; however, in some cases where this algorithm led to distortions, the’SlicerElastix’ extension within 3D Slicer was employed to achieve better registration results. The accuracy of the registration process was evaluated through both visual and quantitative approaches. Visual assessment involved a meticulous slice-by-slice alignment of the registered volumes to confirm precise anatomical correspondence. Quantitative validation was conducted by calculating the Dice Similarity Coefficient (DSC) between the registered masks and the reference mask, yielding a DSC value of 1, which indicates complete overlap. Finally, voxel intensities were normalized to a range of [0, 1] using min-max normalization. To evaluate the impact of intensity normalization on variability arising from multicenter MRI data, the coefficient of variation (CV), defined as the standard deviation divided by the mean, was calculated for the selected features across all subjects, both before and after intensity normalization. The results are provided in the Supplementary Information (SI).

Fig. 2.

Registration workflow of multimodal MRI images. T1, T2, and FLAIR were co-registered to T1CE, which was then aligned to an atlas. The transformation matrix was applied to T1, T2, and FLAIR for standardized spatial alignment

To standardize the dose distribution images prior to dosiomic analysis, the CT scans were first resampled to a uniform resolution of 1 mm³. Subsequently, the dose distribution images and structure delineations from the CT scans, stored in the RTDOSE and RTSTUCT file formats, respectively, were resampled to match the isotropic resolution of the CT scans. Since RTSTRUCT files do not contain volumetric information that can be resampled, they were converted to binary label maps before the resampling process. The preprocessing of the RTDOSE and RTSTRUCT files were carried out using 3D Slicer.

Regions of interest delineation

Heterogeneous subregions were manually delineated slice by slice on multimodal MR images by an experienced radiologist. The two identified subregions were the contrast-enhanced tumor (CET) and the FLAIR hyperintensity, which encompasses edema and non-contrast-enhanced tumor (ED/nCET). These areas were delineated on T1CE and FLAIR sequences, respectively. Precise delineation of the enhanced tumor can be challenging in post-operative MRIs due to the presence of surgical blood products. To address this, a digital subtraction map of pre- and post-contrast T1 images was created, facilitating the differentiation between true contrast-enhancement and confounding postoperative changes, in cases with hyperintense signals on T1 scans [28, 29]. Ultimately, the subregions delineated on the T1CE and FLAIR images were mapped to other sequences based on shared atlas space alignment achieved via the T1CE-to-atlas transformation matrix. ROI labeling was conducted by 3D Slicer.

Radiomic and dosiomic-based feature extraction and heterogeneity measurements

High-throughput radiomic-based features were extracted using SlicerRadiomics, a tool within 3D Slicer that incorporates the PyRadiomics library. Features were calculated from two labeled subregions, CET and ED/nCET, across four MRI sequences to characterize the local, regional, and global heterogeneity of GBM [15]. Each radiomic set comprised 14 shape-based features, 18 first-order statistical or histogram-based features, and 75 higher-order statistical or texture-based features derived from both the original and wavelet-filtered images. The wavelet transform was employed to decompose the original image into various frequency components along the x, y, and z axes, resulting in images labeled as LLH, LHL, LHH, HLL, HLH, HHL, HHH, and LLL, where’L’ and’H’ signify low and high frequencies, respectively. In total, 6808 radiomic-based features were extracted for each patient.

Dosiomic features were extracted from the dose distribution images based on the ROI corresponding to the Gross Tumor Volume (GTV_FLAIR) delineated on FLAIR images. This ROI encompassed the surgical cavity, residual contrast-enhanced tumor, non-contrast-enhanced tumor, and the peritumoral edema. The types of features extracted for dosiomic analysis were identical to those utilized in the radiomic analysis.

Feature selection

Extremely high-dimensional data introduces an overfitting problem, where a machine learning model performs exceptionally well on training data but poorly on new, unseen data. To address this issue, feature selection can effectively overcome overfitting by identifying a smaller subset of features that maintain a high level of integrity and relevance to the target variable, rather than relying on the complete original set of features. In classification tasks, the goal of feature selection is to select a subset of highly discriminative features capable of distinguishing between samples from different classes. Features strongly associated with OS were identified using a sequential selection methodology designed to enhance model robustness. This involved an initial univariate analysis to reduce dimensionality, followed by a multivariate analysis to identify the most relevant features.

Univariate analysis

Before conducting multivariate analysis, the Mann-Whitney U test was employed to identify radiomic features with p-values less than 0.05 and area under the receiver operating characteristic (ROC) curve (AUC) values greater than 0.65. The AUC is calculated using the formula

where U is the Mann-Whitney statistic, and and are the number of patients with long and short survival, respectively.

Multivariate analysis

To determine the most appropriate feature selection method for the dataset, multiple approaches were systematically evaluated. Each method was assessed by training machine learning models using the features it selected. Model performance was measured based on the AUC, the standard deviation of AUC across 10-fold cross-validation (CV), and the degree of overfitting. Three feature selection techniques were compared: L1-based Logistic Regression (LR), also known as LASSO, Support Vector Machine (SVM) with Recursive Feature Elimination and Cross-Validation (SVM-RFECV), and LR with RFECV (LR-RFECV). Although ensemble-based models were also explored, they were excluded from the final comparison due to excessive selected features and pronounced overfitting, ultimately, the L1-based feature selection demonstrated the highest AUC values and the lowest standard deviation across the 10 folds, and was therefore employed as the most suitable method for the given dataset. Before conducting feature selection, the extracted features were standardized using the’StandardScaler’ from the’scikit-learn’ library. This standardization is recognized as a crucial step, especially for algorithms that are sensitive to the scale of features. L1-based feature selection is particularly effective for high-dimensional datasets, especially when the number of predictors exceeds the number of samples. In this approach, feature selection is achieved by minimizing a regularized loss function of the form:min

where L( denotes the loss function, are the model coefficients, and is the L1 norm that penalizes certain coefficients, shrinking them to zero and promoting sparsity in the model. The strength of this regularization is governed by the hyperparameter C, which is the inverse of the regularization coefficient [6, 44–47]. The optimal value of C was searched through 10-fold CV, aimed at minimizing classification error. Feature selection was conducted on various feature sets, including both individual and combined features, and the features with non-zero coefficients were ultimately retained as the final signature for use in the models. For combined models, the optimal preselected features were combined for additional round of selection. This hierarchical strategy facilitated early dimensionality reduction and enhanced the robustness of the final feature set.

Classification and model evaluation

To identify the most appropriate machine learning model for predict survival outcomes of patients with GBM, several classifiers were evaluated, including LR, SVM, k-nearest neighbors (KNN), and Naive Bayes (NB). Model performance was measured based on the AUC, the standard deviation of AUC across 10-fold CV, and the degree of overfitting.

Among the classifiers trained using features selected through L1-based feature selection, both LR and SVM achieved the highest mean AUC values across the 10 folds. Although their AUCs were nearly identical, LR exhibited a lower standard deviation compared to SVM. Thus, LR was implemented as the most suitable model for further analyses.

A variety of classification models were trained using different feature classes, both individually and in combination, as outlined below:

Single Model:

1. A model using dosiomic features (M_D).

Combined Models:

1. Models employing radiomic features to evaluate each subregion across different modalities:

○ CET (T1CE + T1 + T2 + FLAIR) (M_R-CET)

○ ED/nCET (T1CE + T1 + T2 + FLAIR) (M_{R- ED/nCET})

2. Models utilizing radiomic features to assess each modality acrossvarious subregions:

○ T1CE (CET + ED/nCET) (M_R-T1CE)

○ T1 (CET + ED/nCET) (M_R-T1)

○ T2 (CET + ED/nCET) (M_R-T2)

○ FLAIR (CET + ED/nCET) (M_R-FLAIR)

3. A combination of single radiomic models across all subregions and modalities: CET (T1CE + T1 + T2 + FLAIR) + ED/nCET (T1CE + T1 + T2 + FLAIR) (M_{R-CET + R- ED/nCET})

4. A combination of the top-performing radiomic model with the dosiomic model (M_{R-CET + D})

5. A combination of the age variable with all the aforementioned models:

○ (M_D + Age),

○ (M_{R-CET + Age}),

○ (M_{R- ED/nCET + Age}),

○ (M_{R-T1CE + Age}),

○ (M_{R-T1+ Age}),

○ (M_{R-T2 + Age}),

○ (M_{R-FLAIR + Age}),

○ (M_{R-CET + R- ED/nCET + Age}),

○ (M_{R-CET + D + Age}).

This structured approach allowed for a comprehensive analysis of the factors influencing survival in GBM patients. A stratified 10-fold CV approach was employed to identify the optimal model and assess its predictive performance on a validation set. The performance of the models during the CV was evaluated using the mean and standard deviation of the AUC from the ROC analysis across the 10 folds of the training and validation sets. The feature selection and model training processes were carried out using the’scikit-learn’ library (version 1.2.2) [48] in Python.

Model interpretation

It is of great importance to interpret the process of machine learning models to understand the main features influencing the model’s output. In linear models, such as LR, the coefficients indicate the overall importance of each feature. However, since these coefficients are scale-dependent, their interpretation can be complicated unless the features are standardized. Even with standardization, features with higher coefficients may not always be the most important if they exhibit low variance. Besides, coefficients do not allow for the assessment of local feature importance or how this importance varies with different feature values. A more effective approach than relying solely on the coefficients is to implement the SHAP (Shapley Additive exPlanations) method. SHAP provides insights into both the contribution of each feature for individual samples (local explanations) and the overall significance of each feature across all instances (global explanations).

To calculate SHAP values across the 10 folds, the important features that appeared most frequently across the iterations were identified as the most significant predictors for survival outcomes. The SHAP values were computed on the training sets using the SHAP (version 0.42.1) package [49] in Python.

Statistical analysis

In this study, the Mann-Whitney U test and the chi-square test were employed to analyze patient characteristics, including the continuous variable (age) and categorical variables (gender, tumor location, and hemisphere), in order to identify significant intergroup differences between the short-term and long-term survival groups. To evaluate the importance of the final features in each model signature, the radiomics score (rad-score) for each patient was computed by multiplying each feature’s coefficient by its corresponding feature value and summing the results. Rad-score is a single numeric value produced for a given patient that summarizes the information contained in a set of radiomic features. This is commonly used in radiomic studies with a linear relationship between features and coefficients. Subsequently, a univariate LR was fitted for each rad-score, and Wald-test p-values were calculated to assess the association of the rad-score coefficients. Kaplan-Meier survival analysis was subsequently carried out to evaluate the association between the different signatures and OS. Optimal threshold values of the rad-scores were identified using ROC curve analysis in MedCalc Statistical software (version 23.2.1) (MedCalc Software Ltd, Ostend, Belgium; https://www.medcalc.org; 2024) [50]. Patients were then categorized into high- and low-risk groups according to these thresholds, and group differences were assessed using the log-rank test.

All statistical tests were two-sided, with a significance level set at 0.05. Statistical analyses were conducted using the’stats’ module (version 0.14.0) and the’lifelines’ package (version 0.30.0) [51] in Python.

Results

Characteristics of GBM patients

In this study, a total of 74 patients with a confirmed diagnosis of GBM were classified into two groups based on survival duration: a long-survival group consisting of 17 female and 22 male patients, and a short-survival group comprising 14 female and 21 male patients. The average ages for the long-survival and short-survival groups were 52.44 years and 60.94 years, respectively (refer to Table 1). Statistical analysis revealed a significant difference in age between the two groups (p-value < 0.05), while gender did not demonstrate any significant discriminative power in two classes (p-value > 0.05). Additionally, there were no significant differences in tumor location and hemisphere variables between the long and short overall survival categories (p-value > 0.05). The highest incidence of tumors was observed in the frontal lobe as well as in the left hemisphere.

Feature selection

A summary of the performance of each feature selection method and its corresponding machine learning models for dosiomic features is presented in Table 2.

Table 2.

Training and validation AUC values from 10-fold CV for various feature selection methods and corresponding machine learning models for dosiomic features

Model	Feature Selection Algorithm
	L1-based		SVM-RFECV		LR-RFECV
	Training AUC (Mean SD)	validation AUC (Mean SD)	validation AUC (Mean SD)	validation AUC (Mean SD)	Training AUC (Mean SD)	validation AUC (Mean SD)
LR(L2)	0.87 0.01	0.80 0.12	0.67 0.03	0.65 0.23	0.71 0.02	0.69 0.20
SVM	0.88 0.01	0.82 0.14	0.67 0.03	0.65 0.23	0.70 0.03	0.68 0.20
KNN	0.86 0.01	0.73 0.15	0.80 0.02	0.67 0.20	0.75 0.02	0.61 0.20
NB	0.84 0.02	0.79 0.16	0.66 0.03	0.63 0.22	0.71 0.03	0.69 0.20

The results of L1-based feature selection are summarized in Table 3. The optimal regularization hyperparameter (C) was identified at the minimum classification error for each feature class. The selected features with non-zero coefficients are listed in order of their coefficients. Notably, the high-performing dosiomic features consist entirely of wavelet types, whereas the radiomic signatures include both original and wavelet features.

Table 3.

The signatures chosen through L1-based feature selection for different feature classes with their corresponding coefficients, arranged in descending order based on their contribution to the signature, optimal c hyperparameter values, and the p-values obtained from the Wald-test

Feature class	Signature	Optimum C		Wald-test p-value
Dosiomics	D-wavelet-HLH-glszm- SizeZoneNonUniformityNormalized −0.188121 D-wavelet-HHH-gldm- DependenceNonUniformityNormalized −0.168252 D-wavelet-LLH-glrlm-RunLengthNonUniformityNormalized 0.164690 D-wavelet-LLL-glszm- GrayLevelNonUniformityNormalized −0.106006 D-wavelet-LLL-glszm- LargeAreaLowGrayLevelEmphasis 0.035056 D-wavelet-HHH -glszm-ZoneEntropy 0.034892 D-wavelet-HLL-firstorder-Skewness −0.023983 D-wavelet-HLL-gldm-LargeDependenceHighGrayLevelEmphasis −0.024342	0.15	< 0.001
Radiomics (CET subregion)	ET-T1-wavelet-LLL-glszm ZoneVariance −0.664250 ET-T1-original-shape-MinorAxisLength −0.588929 ET-FLAIR-wavelet-LHL-glcm-ClusterShade 0.521030 ET-T2-original-firstorder-90%ile 0.502047 ET-CE-wavelet-HHH-glszm-HighGrayLevelZoneEmphasis 0.489112 ET-FLAIR-wavelet-LHH-glszm-GrayLevelVariance 0.461872 ET-T1-wavelet-LHL-glszm-SmallAreaLowGrayLevelEmphasis −0.320021 ET-CE-wavelet-LHH-glszm-ZoneEntropy 0.301751	0.5	< 0.001
Radiomics (ED/nCET subregion)	ED-CE-wavelet-LLH-glszm-GrayLevelVariance 0.694279 ED-T2-original-firstorder-Range 0.491939 ED-T2-wavelet-LLH-glszm-LowGrayLevelZoneEmphasis 0.446590 ED-T1-wavelet-LLL-firstorder-10%ile 0.346523 ED-FLAIR-wavelet-HLL-glszm-LowGrayLevelZoneEmphasis 0.327698 ED-CE-wavelet-LHH-firstorder-Mean 0.282761 ED-CE-wavelet-LHH-glcm-MaximumProbability 0.282156 ED-T2-wavelet-HLL-firstorder- 90%ile 0.226168 ED-T2-original-firstorder-Skewness 0.152979	1	< 0.001
Radiomics (T1CE modality)	CE- ED-wavelet-LLH-glszm-GrayLevelVariance 1.338741 CE- ET-wavelet-HHH-glszm-HighGrayLevelZoneEmphasis 0.927724 CE- ET-wavelet-HHL-glcm-MaximumProbability 0.925756 CE- ET-wavelet-HHH-firstorder-Uniformity 0.800589 CE- ED-wavelet-HLH-firstorder-Mean 0.403815 CE -ET-wavelet-LLH-glszm-LowGrayLevelZoneEmphasis 0.377762 CE- ET-wavelet-LHL-glcm-Autocorrelation 0.360801 CE- ET-wavelet-LHH-glszm-ZoneEntropy 0.347177	2	< 0.001
Radiomics (T1 modality)	T1- ET-wavelet-LLL-glszm-ZoneVariance −0.726593 T1- ET-wavelet-LHL-glszm-SmallAreaLowGrayLevelEmphasis −0.597261 T1- ET-original-firstorder-10%ile 0.484738 T1- ED-wavelet-HLH-firstorder-Median 0.364657 T1- ET-original-shape-MinorAxisLength −0.303637 T1- ED-wavelet-LLL-firstorder-10%ile 0.132799	0.5	< 0.001
Radiomics (T2 modality)	T2- ET-wavelet-HHH-firstorder-Uniformity 1.106373 T2- ED-original-firstorder-Range 0.785276 T2- ET-wavelet-LLH-glrlm-LowGrayLevelRunEmphasis 0.521436 T2- ET-wavelet-HLL-firstorder-RootMeanSquared 0.464775 T2- ED-wavelet-LLH-glszm-LowGrayLevelZoneEmphasis 0.457707 T2- ET-original-glrlm-ShortRunEmphasis 0.409775 T2- ET-wavelet-LHL-glrlm-ShortRunHighGrayLevelEmphasis 0.172776 T2- ED-original-firstorder-Skewness 0.127432	1	< 0.001
Radiomics (FLAIR modality)	FLAIR-ET -wavelet-LHH-glszm-GrayLevelVariance 1.094150 FLAIR-ET -wavelet-LHL-glcm-ClusterShade 0.787417 FLAIR-ET-original-glrlm-ShortRunLowGrayLevelEmphasis 0.580911 FLAIR-ED-wavelet-HLL- glszm GrayLevelVariance 0.398520 FLAIR-ET -wavelet-LLH-gldm- SmallDependenceLowGrayLevelEmphasis 0.387383 FLAIR-ET -wavelet-LLL-glszm-ZonePercentage −0.224954 FLAIR-ED-wavelet-HLL-glszm-LowGrayLevelZoneEmphasis 0.169161 FLAIR-ET-original-glszm-ZonePercentage −0.104111	3	< 0.001
Radiomics (Integration of subregions and modalities)	ET-T1-original-shape-MinorAxisLength −1.077103 ED-CE-wavelet-LLH-glszm-GrayLevelVariance 1.013329 ET-T2-wavelet-HHH-firstorder-Uniformity 0.656449 ET-T1-wavelet-LLL-glszm-ZoneVariance −0.610914 ET-CE-wavelet-HHH-glszm-HighGrayLevelZoneEmphasis 0.540799 ET-T2-wavelet-LLH-glrlm-LowGrayLevelRunEmphasis 0.498013 ET-FLAIR-wavelet-LHH-glszm-GrayLevelVariance 0.497398	1	< 0.001
Radiomics (CET subregion) + Dosiomics	ET-T1-original-shape-MinorAxisLength −1.723343 ET-FLAIR-wavelet-LHL-glcm-ClusterShade 1.281707 D-wavelet-LLL-glszm-LargeAreaLowGrayLevelEmphasis 1.200515 ET-CE-wavelet-HHH-glszm-HighGrayLevelZoneEmphasis 1.045803 D-wavelet-HLL-firstorder-Skewness −0.992684 ET-T2-original-firstorder-90%ile 0.967957 D-wavelet-LLL-glszm-GrayLevelNonUniformityNormalized −0.940228 D-wavelet-LLH-glrlm-RunLengthNonUniformityNormalized 0.861047 ET-CE-wavelet-LHH-glszm ZoneEntropy 0.750654 ET-T1-wavelet-LLL-glszm ZoneVariance −0.714647 ET-FLAIR-wavelet-LHH-glszm GrayLevelVariance 0.651517	2	< 0.001

In the combined analysis of radiomic and dosiomic features, four dosiomic and seven radiomic variables demonstrated higher coefficients than those observed in the dosiomic-only and radiomic-only analysis (refer to Table 3). A nuanced observation was that the top features from the dosiomic-only and radiomic-only analyses did not retain the highest coefficients in the combined model. This can be attributed to the influence of feature interactions and multicollinearity. In the combined setting, redundant or correlated features may have their contributions redistributed, while features offering complementary information across domains can become more prominent. For instance, ET-T1-original-shape-MinorAxisLength and D-wavelet-LLL-glszm-LargeAreaLowGrayLevelEmphasis showed substantially increased coefficients in the combined model, highlighting complementary contributions that enhance predictive value.

Figure 3 illustrates the results of rad-score calculations for each patient across various feature groups. The purple and blue bars represent the rad-scores for GBM patients with long and short OS, respectively. It is evident that the majority of the rad-scores for the long-term survival group are positive, while those for the short-term survival group are predominantly negative. This indicates that as rad-scores become more negative, the likelihood of prolonged survival diminishes. The statistical significance of these relationships was confirmed using Wald-test p-values by fitting univariate LR models to each rad-score, revealing that the rad-scores from all feature categories significantly discriminated between long-term and short-term survival outcomes (p-value < 0.001).

Fig. 3.

Bar plots demonstrating the rad-scores for each patient categorized based on long-term and short-term OS concerning various feature classes of A) dosiomics, B) Radiomics (CET subregion), C) Radiomics (ED/ nCET subregion), D) Radiomics (T1CE modality), E) Radiomics (T1 modality), F) Radiomics (T2 modality), G) Radiomics (FLAIR modality), H) Radiomics (integration of subregions and modalities), and I) Radiomics (CET subregion) + dosiomics

The Kaplan-Meier analysis results for rad-scores derived from dosiomic and radiomic signatures are presented in Fig. 4. All rad-scores significantly stratified patients into high- and low-risk groups, as evidenced by the clear separation of the survival curves. Among the evaluated models, the signature integrating CET radiomic features with dosiomic features achieved the greatest separation between risk groups, highlighting its superior prognostic value compared to other feature sets.

Fig. 4.

Kaplan-meier plots illustrating the association between rad-scores and OS concerning various feature classes of A) dosiomics, B) Radiomics (CET subregion), C) Radiomics (ED/ nCET subregion), D) Radiomics (T1CE modality), E) Radiomics (T1 modality), F) Radiomics (T2 modality), G) Radiomics (FLAIR modality), H) Radiomics (integration of subregions and modalities), and I) Radiomics (CET subregion) + dosiomics

Comparison of models’ performance

The report on OS classification obtained from LR models, based on three-dimensional dose maps and multimodal MRI studies, is summarized in Table 4. Both training and validation AUC values are presented, as the distinction between these values is important for evaluating model performance and identifying potential overfitting. The AUC value reflects the average of 10 AUC values across 10-fold CV, accompanied by the standard deviation of these AUCs around the mean, which indicates the stability of the models.

Table 4.

Training and validation AUC values from 10-fold CV for various models, including dosiomics, radiomics, a combination of both, as well as each of the models including the age variable, along with the number of optimal features used in each case

Feature class	Model name	Training AUC (Mean SD)	Validation AUC (Mean SD)	Number of optimal features
Dosiomics	MD	0.87 0.01	0.80 0.12	8
Radiomics (Subregion-based)	MR-CET	0.93 0.01	0.90 0.09	8
	MR-ED/nCET	0.85 0.02	0.76 0.10	9
Radiomics (MRI modality-based)	MR-T1CE	0.88 0.01	0.86 0.12	8
	MR-T1	0.86 0.02	0.84 0.18	6
	MR-T2	0.86 0.02	0.85 0.14	8
	MR-FLAIR	0.81 0.01	0.76 0.21	8
Radiomics (Integration of subregions and modalities)	MR-CET + R- ED/nCET	0.90 0.01	0.90 0.11	7
Radiomics (CET subregion) + Dosiomics	MR-CET + D	0.97 0.01	0.960.07	11
Dosiomics + Age	MD + Age	0.90 0.01	0.86 0.13	9
Radiomics (Subregion-based) + Age	MR-CET + Age	0.93 0.01	0.90 0.09	9
	MR- ED/nCET + Age	0.87 0.01	0.81 0.14	10
Radiomics (MRI modality-based) + Age	MR-T1CE + Age	0.90 0.01	0.88 0.08	9
	MR-T1+ Age	0.87 0.02	0.83 0.19	8
	MR-T2 + Age	0.87 0.02	0.87 0.15	9
	MR-FLAIR + Age	0.84 0.01	0.79 0.13	9
Radiomics (Integration of subregions and modalities) + Age	MR-CET + R- ED/nCET + Age	0.93 0.01	0.91 0.13	8
Radiomics (CET subregion) + Dosiomics + Age	MR-CET + D + Age	0.98 0.00	0.97 0.05	12

SD:standard deviation

Figure 5 depicts the ROC curves for each of the 10 folds, as well as the averaged ROC curve across all folds, for the different models evaluated on the validation sets

Fig. 5.

ROC curves illustrating each of the 10 folds, along with the average ROC curve across all folds for various models of A) dosiomics, B) Radiomics (CET subregion), C) Radiomics (ED/ nCET subregion), D) Radiomics (T1CE modality), E) Radiomics (T1 modality), F) Radiomics (T2 modality), G) Radiomics (FLAIR modality), H) Radiomics (integration of subregions and modalities), and I) Radiomics (CET subregion) + dosiomics on the validation sets

The dosiomic-based model (denoted as M_D), trained on extracted features from GTV volumes and incorporating eight selected features, showed a prognostic performance characterized by a mean AUC of 0.87 0.01 on the training sets and 0.80 0.12 on the validation sets.

To evaluate the heterogeneity of GBM in relation to OS using pre-radiotherapy MRI, two heterogeneous subregions, including CET and ED/nCET, were considered across four MRI modalities. The subregion-based models, named M_R-CET and M_R-ED/nCET, achieved mean AUC values of 0.93 0.01 (training) and 0.90 0.09 (validation) for the CET subregion, and 0.85 0.02 (training) and 0.76 0.10 (validation) for the ED/nCET subregion. The results indicate that the CET subregion significantly outperformed the ED/nCET subregion in stratifying the overall survival of GBM patients (p-value < 0.05). This discrepancy is further illustrated in the box plots comparing the two subregions (Fig. 6).

Fig. 6.

AUC values plotted in the box plots for different models. ALL: integration of subregions and modalities

In addition to assessing each subregion individually, the performance of each MRI modality (T1, T1CE, T2, and FLAIR) with two defined subregions was separately analyzed. This resulted in the creation of four MRI modality-based models: M_R-T1CE, M_R-T1, M_R-T2, and M_R-FLAIR. The mean AUC values for these radiomic models were as follows:

M_R-T1CE: training AUC = 0.88 0.01; validation AUC = 0.86 0.12

M_R-T1: training AUC = 0.86 0.02; validation AUC = 0.84 0.18

M_R-T2: training AUC = 0.86 0.02; validation AUC = 0.85 0.14

M_R-FLAIR: training AUC = 0.81 0.01; validation AUC = 0.76 0.21

Based on these results, the CE and FLAIR sequences demonstrated the highest and lowest AUC values, respectively. However, there was no significant difference in discriminative power among the four sequences concerning OS (p-value > 0.05). The distribution of AUCs for the models is depicted in the box plots in Fig. 6.

The combined model, which integrated data from all four modalities and two subregions (M_{R-CET + R-ED/nCET}), yielded mean AUC values of 0.90 0.01 for training and 0.90 0.11 for validation. Although this model exhibited higher training and validation AUC values compared to the other radiomic models, it only demonstrated a statistically significant improvement when compared to the M_R-ED/nCET and M_R-FLAIR models (p-value < 0.05).

To evaluate the enhancement effect of dosiomic features on radiomics, the radiomic model selected for further analysis was the M_R-CET model, which demonstrated a high AUC as well as the lowest variance among the other trained radiomic models (validation AUC: 0.90 0.09). The best features from the M_D and M_R-CET models were combined to train and assess the M_{R-CET + D} model, which exhibited significantly improved performance compared to the individual M_R-CET and M_D models (training AUC: 0.97 0.01, validation AUC: 0.96 0.07; p-value < 0.05). It is evident that the variance of the combined model was considerably lower than that of the individual models.

In the comparison analysis of training and validation AUC values, the models, M_R-CET, M_R-T1CE, M_R-T1, M_R-T2, M_{R-CET + R- ED/nCET}, and M_{R-CET + D}, demonstrated consistent performance across both datasets, indicating good generalizability and a low risk of overfitting. By contrast, the M_D, M_R-ED/nCET, and M_R-FLAIR models showed greater discrepancies between training and validation AUCs, suggesting that their stability could be improved with the inclusion of additional data.

Among the patient characteristics, only the age variable exhibited significant discriminative ability between the long-term and short-term survival groups (p-value < 0.05). As a result, age was included with all the preselected feature classes to evaluate its potential impact on enhancing the predictive power of the model for overall survival. However, the findings revealed that incorporating age did not yield any significant improvements in the performance of the models (p-value > 0.05), except for the ED/nCET subregion-based model. In this instance, the M_{R- ED/nCET + Age} model achieved a mean validation AUC of 0.81 0.14, surpassing the mean validation AUC of 0.76 0.10 obtained by the M_R-ED/nCET model (p-value < 0.05).

Model interpretation

The SHAP values for the M_D, M_R-CET, and M_{R-CET + D} models are illustrated in summary plots. The gradient color bar indicates the feature value. The local contributions are displayed by the scattered dots in the plot, and the global contributions are indicated by feature importance, with the most impactful features positioned nearer to the top of the plot. Across 10 iterations, the most frequently important features for the M_D, M_R-CET, and M_{R-CET + D} models were identified as follows:

M_D: Feature 6-D-wavelet-LLL-glszm- GrayLevelNonUniformityNormalized,

Feature 0-D-wavelet-LLH-glrlm-RunLengthNonUniformityNormalized

M_R-CET: Feature 2-ET-T1-original-shape-MinorAxisLength,

Feature 5-ET-T2-original-firstorder-90%ile

M_{R-CET + D}: Feature 6-ET-T1-original-shape-MinorAxisLength,

Feature 8- ET-T2-original-firstorder-90%ile,

Feature 2-D-wavelet-LLL-glszm- GrayLevelNonUniformityNormalized.

One representative plot for each model is displayed in Fig. 7.

Fig. 7.

SHAP values for three models: A) M_D, B) M_R-CET, and C) M_{R-CET + D}. The horizontal position of each dot corresponds to the SHAP value for a feature for a particular instance, while the color of the dots represents the value of the feature. The features are arranged along the Y-axis in descending order of their absolute SHAP values

Discussion

In this study, the potential of dosiomic features derived from planned dose distributions was explored to stratify GBM patients into long-term and short-term survival groups. To the best of our knowledge, this is the first investigation into the employment of dosiomic features as prognostic factors for personalized treatment in GBM patients. A predictive model that incorporates dosiomic features was developed, achieving an average AUC of 0.80 0.12 on validation sets across a 10-fold CV strategy. More importantly, it was demonstrated that integrating MRI-based radiomic features alongside dosiomic features significantly elevated the performance of dosiomic and radiomic features, resulting in an increase in mean AUCs from 0.80 0.12 and 0.90 0.09 to 0.96 0.07.

In the dosiomic analysis, features were extracted from both the original and wavelet-transformed images. However, the final selected features using L1-based feature selection consisted solely of wavelet features. One possible explanation for this finding is that the original features did not demonstrate sufficient variation to be significant for the target outcomes. In contrast, wavelet transformations decompose an image into different frequency components, making them particularly effective for capturing subtle variations. Additionally, the original dose distribution images may contain noise, while wavelet transforms are capable of effectively distinguishing between signal and noise. This capability enhances the extraction of advanced features that are more robust and reliable for predicting survival outcomes.

In this study, pre-radiotherapy MRIs have been employed to account for tumor growth that may occur between surgical resection and the start of radiotherapy. This approach is particularly important because the post-surgical MRIs taken within the first 72 hours often do not reflect any tumor progression that could affect treatment planning and patient prognosis.

The study utilized the four commonly recommended MRI sequences for the diagnosis of GBM: T1, T1CE, T2, and FLAIR. It is generally accepted that multimodal MRIs provide unique insights into tumor heterogeneity and can enhance survival stratification [15, 23]. Two distinct assessments were conducted within the study framework. The first assessment focused on the influence of heterogeneous subregions of the tumor on the prognosis of GBM patients, specifically analyzing two CET and ED/nCET subregions across the various MRI modalities. The second assessment examined the impact of each modality, including both subregions, on patient outcomes.

The standout finding was that the rad-score of all radiomic feature classes comprising radiomics (subregion-based), radiomics (MRI modality-based), and radiomics (integration of subregions and modalities) showed a high association with survival. This finding was consistently confirmed by univariate LR and Wald-test analyses (p-value < 0.001), as well as by Kaplan-Meier survival analysis (p-value < 0.0001). The CET-based and ED/nCET-based LR models (M_R-CET and M_R-ED/nCET) achieved a mean AUC of 0.90 0.09 and 0.76 0.10, respectively. The presence of enhanced tumors on both postoperative and pre-radiotherapy MRIs has been acknowledged as a prognostic indicator for patients with glioblastoma. De Barros et al. [30] demonstrated that enhanced residual tumor volumes on pre-radiotherapy MRI were more strongly associated with OS than volumes measured on postoperative MRI. The median inter-scan interval was 27 days (range, 10–64 days). These findings underscore the potential importance of pre-radiotherapy imaging in guiding prognosis and treatment planning. The most comprehensive multicenter study, including 1,511 newly diagnosed GBM patients confirmed that residual postoperative CET volume significantly affected OS [28]. Another study indicated that pre-radiotherapy CET volume was strongly associated with OS, with stratification according to MGMT promoter methylation status [32]. Additionally, GarciaRuiz et al. [29] explored the discriminative potential of enhanced tumors for survival-based subgrouping in GBM patients using a LR model trained on radiomic features extracted from postoperative MRIs acquired at multiple time points after surgery, achieving an AUC of 0.71 (95% CI: 0.55–0.88, p-value = 0.01) in the test set. The findings of the current study align with the aforementioned literatures regarding the assessment of enhanced tumors in both postoperative and pre-radiotherapy MRI scans, whether considering volumetric measures or radiomic features.

Several factors may account for the higher AUC of our model (0.90 0.09) compared to the AUC of 0.71 reported by in GarciaRuiz et al. [29]. One plausible explanation is the employment of a distinct test set in their research. Additionally, their approach focused only T1CE images for feature extraction, whereas our study utilized four different modalities to derive the features from CET region. This comprehensive approach to feature extraction likely contributed to our model’s improved performance. Furthermore, the varying acquisition times of MRI scans in the research by GarciaRuiz et al. [29] may have influenced the performance of their model. Another possible rationale relates to the findings of De Barros et al. [30], who observed that the enhanced tumor on pre-radiotherapy MRI exhibited stronger predictive power compared to postoperative MRI.

It is worth mentioning that there has been an increasing body of evidence indicating that surgically induced contrast enhancements can occur at various time points on post-surgery MRIs, exhibiting different types and frequencies. Consequently, the presence of benign enhancement on pre-radiotherapy MRIs can pose a confounding factor on model performance [52]. Nevertheless, the significant association between the rad-scores and survival, along with a predictive AUC of 0.90, suggests that the rad-score may be predominantly influenced by the characteristics of the malignant tumor. This emphasizes the importance of assessing the role of enhanced tumors in OS stratification, while also distinguishing between non-tumor-related and tumor-related contrast enhancements.

The standard surgical approach for GBM tumors consists of maximal safe resection of the tumor’s enhancing compartment. Achieving a gross total resection is associated with improved survival outcomes. However, even after a satisfactory resection, local recurrence of GBM is unavoidable owing to the infiltration of tumor cells beyond the visible boundaries of the enhancing tumor. This infiltrative region, identified as the nCET, is most effectively visualized using hyperintense T2/FLAIR sequences [32, 53, 54]. In our study, the ED/nCET subregion was evaluated by analyzing radiomic features extracted from four pre-radiotherapy MRI sequences. The rad-score for this subregion demonstrated a strong correlation with OS, as evidenced by univariate LR and Wald-test results (p-value < 0.001), along with Kaplan-Meier survival analysis (p-value < 0.0001). Furthermore, a mean AUC of 0.76 0.10 was exhibited by the ED/nCET-based model. These results are consistent with previous studies on the evaluation of the hyperintense T2/FLAIR region. Kotrotsou et al. [55] and Grabowski et al. [53] demonstrated that the volume of post-operative residual non-enhancing tissue is correlated with overall survival in GBM patients in both univariate and multivariate analyses.

Saraswathy et al. [31] highlighted the strong association between pre-radiotherapy hyperintense T2 volume and survival outcomes. However, one study presented contradictory results, indicating that while there was a significant correlation between CET and the combination of CET and ED/nCET with survival, the ED/nCET volume alone in pre-radiotherapy MRIs showed no association with survival [32]. Taken together, these findings emphasize the necessity of considering not just the CET and tumor bed but also the nCET region in radiation therapy. Therefore, imaging prior to the commencement of radiotherapy plays an important role in accurately defining this volume in treatment planning and patient’s prognosis.

In a comparative analysis of two subregion-based models, the M_R-CET significantly outperformed the M_R-ED/nCET, achieving a mean AUC of 0.90 compared to 0.76 (p-value < 0.05). The robustness of the enhancing tumor rather than the non-enhancing hyperintense T2/FLAIR has been explained due to the hypercellularity of the enhancing lesions. Research by Liu et al. [15] explored how heterogeneity in GBM affects survival outcomes, emphasizing that evaluating enhanced or active tumor regions on preoperative MRI scans offers a more accurate representation of molecular activity and intra-tumoral heterogeneity. Similarly, Wang et al. [23] and Chaddad et al. [56] indicated that features derived from the enhancing tumor were the most beneficial indicators for predicting survival compared to those from other subregions. Moreover, Grabowski et al. [53] demonstrated that the enhancing tumor region was a stronger predictor of survival than the residual volume of T2/FLAIR lesions observed in post-operative MRIs. In contrast, in some studies, the non-enhanced hyperintense T2/FLAIR regions have been identified as more predictive of outcomes than the enhanced areas.

Exemplarily, Saraswathy et al. [31] noted that the pre-radiotherapy volume of T2 hyperintensity exhibited the strongest correlation with patient survival. Likewise, Yang et al. [12] indicated that the ED-based signature prior to surgery outperformed other subregions in terms of predictive value. It has been observed that nCET regions exhibit a greater density of tumor cells compared to both enhancing tumors and necrotic compartments [54]. However, this observation is primarily applicable when evaluating nCET in isolation rather than considering the overall area of T2/FLAIR hyperintensity. In the current study, the entire residual FLAIR hyperintensity was assessed, which includes both edema and infiltrating tumor cells. The edema portion may contain persistent vasogenic edema and surgery-induced edema, which can take days to weeks to be resolved after surgery. Accordingly, while there is a significant association between the ED/nCET subregion and survival, the lower predictive power of this subregion may be raised from the sustained presence of edema rather than necessarily infiltrating nCET characterized by hypercellularity [54]. Besides, over time following resection, some non-enhancing tumor cells may proliferate, leading to the development of new contrast-enhancing lesions [32].

The combination of subregions across the four modalities (M_{R-CET + R-ED/nCET}) achieved a mean AUC of 0.90 0.11. This combined AUC demonstrated significantly better performance in contrast to the ED/nCET subregion (p-value < 0.05), while showing comparable results to the CET, with no significant differences observed (p-value 0.05). This finding can be acknowledged by the combined signature, from which only one feature was selected from the ED/nCET component, likely due to the lower cellularity present in this subregion. In total, the radiomic analysis reveals that different tumor subregions exhibit unique histogram and textural characteristics of tumor tissues, which can aid in predicting the prognosis of patients with GBM.

In addition to focusing on subregion-based analyses, the impact of multimodal MRIs (T1CE, T1, T2, and FLAIR) on the heterogeneity of GBM and its influence on survival stratification was assessed. To this end, four MRI modality-based predictive models were developed: M_R-CE, M_R-T1, M_R-T2, and M_R-FLAIR, yielding mean AUC values of 0.86 0.12, 0.84 0.18, 0.85 0.14, and 0.76 0.21, respectively. Comparable performance was exhibited by all models, with no significant discrimination found among the predictive capabilities of the various modalities (p-value 0.05). Notably, in each modality-based signature, features derived from the CET were found to contribute more to prognostic outcomes than those from the ED/nCET, highlighting the stronger prognostic value of CET variables. The combined radiomic model (M_{R-CET + R-ED/nCET}), leveraging four modalities including two subregions, with a mean AUC of 0.90 0.11, significantly outperformed only the M_R-FLAIR model, which had a mean AUC of 0.76 0.21 (p-value < 0.05) among the modality-based models. The elevated AUC value, despite the absence of significant differences, may be attributed to the relatively high variance observed in the models. To our knowledge, this represents the first study to incorporate a comprehensive analysis of all MRI modalities in pre-radiotherapy assessments for GBM by machine learning approaches. However, the impact of preoperative multiple MRI sequences on GBM survival has been explored in previous research. A study conducted by Yin et al. [4] showed that preoperative FLAR images capturing four heterogeneous subregions were found to outperform other sequences for OS classification, achieving an AUC of 0.91. In contrast, Liu et al. [15] declared that the T1CE and T2 sequences showed better performances with AUCs of 0.79 and 0.74, respectively. Plus, the combined analysis of all four sequences revealed a performance equivalent to that of T1CE.

The outstanding aspect of this study was the evaluation of the enhanced performance of radiomic features when combined with dosiomic features. This integration was achieved by incorporating the most relevant dosiomic features into the leading prognostic CET features, resulting in a combined mode denoted as M_R-CET+D. This integrated model exhibited significantly improved performance, achieving a mean AUC of 0.96 0.07, surpassing both the M_R-CET and M_D models in effectively classifying GBM patients into long-term and short-term survival categories (p-value < 0.05), thereby validating our hypothesis. The radiomic and dosiomic features demonstrated superior predictive value in the integrated model compared to radiomic-based and dosiomic-based models alone. This observation suggests that these different feature sets may be more potent and relevant when utilized in combination. There has not yet been a study evaluating the potential of dosiomic features to elevate the survival prediction of radiomic features in patients with GBM. However, this potential has been explored in relation to other diseases.

The inclusion of the age variable as the sole prognostic patient characteristic for GBM survival did not boost the performance of any of the models. Although both the radiomic/dosiomic models and the combined radiomic/dosiomic + age models could effectively stratify patients into long-term and short-term classes, the models incorporating the age variable did not exceed the performance of the other models, except for the ED/nCET subregion-based model. In this case, the M_{R- ED/nCET + Age} model achieved a mean validation AUC of 0.81 0.14, surpassing the mean validation AUC of 0.76 0.10 obtained by the M_{R- ED/nCET} model (p-value < 0.05). These findings suggest that radiomic and dosiomic features may contribute importantly to survival discrimination in most models. However, in the M_{R- ED/nCET + Age} model, the age variable makes an important contribution alongside the radiomic features.

To enhance model interpretability, SHAP analysis was conducted to identify and explain the most influential features contributing to survival stratification. In the dosiomic-based model, the most impactful feature identified was D-wavelet-LLL-glszm- GrayLevelNonUniformityNormalized, a metric that quantifies the variation in gray-level intensities across the volume of interest (VOI), which in this study corresponds to the GTV. Elevated values of this feature, indicating a more heterogeneous dose distribution, led to negative SHAP values. This correlation suggests that greater dose heterogeneity is linked to a higher likelihood of shorter survival rather than longer survival. Across the dataset, this feature showed clear separation in SHAP value distributions between the two survival groups, indicating a meaningful contribution to class discrimination. This finding suggests that patients receiving a more heterogeneous tumor dose within the tumor may experience poorer outcomes, possibly due to under-coverage or suboptimal treatment planning.

The radiomic features that significantly influenced survival classification in the CET model included ET-T1-original-shape-MinorAxisLength and ET-T2-original-firstorder-90%ile features. The Minor Axis Length feature measures the minor axis of the shape of a defined region. Higher values of this feature were associated with negative SHAP values, explaining an increased likelihood of predicting shorter survival. This finding indicates that larger contrast-enhancing lesions may be indicative of a poorer prognosis, aligning with the clinical understanding of tumor aggressiveness. In contrast, the 90th percentile feature, reflecting the upper-end intensity distribution within the tumor, revealed positive SHAP values at elevated values, contributing to the prediction of prolonged survival. This suggests that tumors with higher voxel intensities may correspond to less aggressive tumor phenotypes and greater responsiveness to treatment.

Ultimately, the most predictive features of the individual radiomic and dosiomic models, including ET-T1-original-shape-MinorAxisLength, ET-T2-original-firstorder-90%ile, and D-wavelet-LLL-glszm-GrayLevelNonUniformityNormalized, emerged as the most prognostic indicators within the integrated radiomic and dosiomic model. This indicates that the two feature sets are complementary rather than redundant, providing non-overlapping information that enhances model discrimination.

The findings of the study should be interpreted with caution due to several limitations that need to be addressed in future research. A primary limitation of this study was the relatively small sample size, which restricted the application of a separate test set. The greater differences observed between the training and validation results in some models may, in part, reflect data heterogeneity. Therefore, it is crucial to validate the performance and generalizability of the proposed strategy in larger, multi-institutional datasets. Additionally, the study utilized a retrospective dataset, resulting in missing treatment information. The MR images were collected from different centers, which can offer advantages such as enhanced generalizability and adaptability of the model while also reducing center-specific biases. However, the use of different MRI machines and protocols across these centers introduces variability in the radiomic features. Although implementing image standardization techniques to normalize resolution and intensity may help mitigate such variations, it is unlikely to eliminate all inherent differences associated with multicenter data collection. To resolve this issue, harmonization methods such as ComBat have been suggested. However, given the diversity of vendors and scanner models, the dataset included 17 batches, several contained only one or two samples, which posed significant challenges for applying ComBat effectively. Future studies with larger batch sizes and more comprehensive harmonization strategies are warranted to further mitigate residual batch effects. Another limitation is that the radiotherapy data were sourced from a single center, which may lead to potential bias. Moreover, the only machine learning model suitable for this dataset was the LR model. Exploring other machine learning models on diverse datasets would be beneficial for future work. In this study, two subregions, CET and nCET/ED, were considered for radiomic analysis and model development. Necrosis was not analyzed as a separate subregion because post-surgical changes introduce substantial anatomical and contrast alterations, which make reliable delineation of necrosis on pre-radiotherapy MRI challenging. Moreover, not all patients exhibited necrotic regions on pre-radiotherapy imaging. Given the relatively limited sample size, excluding patients without necrosis would not have been methodologically practical.

Conclusion

In conclusion, the present study demonstrated the applicability of dosiomic features as prognostic indicators for stratifying a cohort of GBM patients into long-term and short-term survival groups using a LR machine learning model. Furthermore, pre-radiotherapy multimodal MRI scans were employed to evaluate the predictive power of each subregion and sequence individually. The CET subregion model utilizing multimodal MRI scans demonstrated significantly superior performance compared to the ED/nCET subregion model. However, the models based on various MRI sequences exhibited a consistent trend in their stratification capabilities. The noteworthy observation was a substantial enhancement in survival prediction when integrating dosiomic and radiomic features. These findings are promising, as they highlight tumor-specific markers that reflect the biological behaviors of GBM and the nuances of intra-tumor heterogeneity, thereby paving the way for personalized medicine approaches based on radiomics and dosiomics.

Author contributions

1 guarantor of the integrity of the entire study: SR. Mahdavi, H. Iraji, A. Mahmoudi. 2 study concepts and design: SR. Mahdavi, A. Mahmoudi. 3 literature research: A. Mahmoudi. 4 clinical studies: A. Mahmoudi, H. Iraji, M. Barahman. 5 experimental studies/data analysis: SR. Mahdavi, A. Mahmoudi, A. Zare Sadeghi, P. Saadatmand, E. Yazdani. 6 statistical analysis: A. Mahmoudi. 7 manuscript preparation: A. Mahmoudi. 8 manuscript editing: SR. Mahdavi, A. Zare Sadeghi, E. Yazdani

Funding

This research was supported by Iran University of Medical Sciences as a Ph.D. thesis number of 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/).

Declarations

Ethics approval and consent to participate

This research is a retrospective study that does not involve any experimental studies with human subjects. The dataset used in this research is publicly available. Therefore, an ethics statement is not required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.