Radiomic features from the peritumoral brain parenchyma on treatment-naïve multi-parametric MR imaging predict long versus short-term survival in Glioblastoma Multiforme: Preliminary Findings

Abstract

Objective

Despite 90% Glioblastoma (GBM) recurrences occurring in the peritumoral brain zone (PBZ), its contribution in patient survival is poorly understood. The current study leverages computerized texture (i.e. radiomic) analysis to evaluate the efficacy of PBZ features from preoperative MRI in predicting long (>18-months) versus short-term (<7-months) survival in GBM.

Methods

65 patient exams (29 short-term, 36 long-term) with Gadolinium-contrast T_1w, FLAIR, T_2w sequences from the Cancer Imaging Archive were employed. An expert manually segmented each study as: enhancing lesion, PBZ, and tumour necrosis. 402 radiomic features (capturing co-occurrence, gray-level dependence, directional gradients) was obtained for each region. Evaluation was performed using 3-fold cross validation, such that a subset of studies was used to select the most predictive features, and the remaining subset was used to evaluate their efficacy in predicting survival.

Results

A subset of 10 radiomic “peritumoral” MRI features, suggestive of intensity heterogeneity and textural patterns, was found to be predictive of survival (p = 1.47 × 10⁻⁵), as compared to features from enhancing tumour, necrotic regions, and known clinical factors.

Conclusion

Our preliminary analysis suggests that radiomic features from the PBZ on routine pre-operative MRI may be predictive of long-, versus short-term survival in GBM.

Introduction

Glioblastoma Multiforme (GBM) is a highly aggressive brain tumour with dismal prognosis. Despite aggressive treatment with surgical resection, chemo-radiation-therapy, and anti-angiogenesis therapy with bevacizumab, the median survival time after diagnosis for GBM is only 12 months. A minority of all GBM patients, roughly 5–10%, survive between 3 to 12 years.¹ With new monoclonal antibodies, vaccines, and gene therapies currently under investigation,² there is a pressing need for accurate risk stratification towards a more personalized approach in GBM management.

Tumour heterogeneity both within and around the tumour margin is a well-known contributor of poor-survival in GBM³. Multiple studies^4,5,6 have demonstrated that GBM heterogeneity is not limited to the tumour margins but also involves the peritumoral brain parenchymal zone (PBZ), with roughly 90% of GBM recurrences occurring in the PBZ.⁷ Few studies⁸ have investigated the PBZ and its microenvironment suggesting that interaction of specific cells (i.e. glioma cells, vascular endothelial, neuroglial, and microglial cells^9,10) and molecular events in PBZ contribute to tumour infiltration, blood brain barrier compromise, and micro vascularity, ultimately leading to poor survival in GBM. Therefore, studying the role of PBZ may have significant therapeutic and patient outcome implications¹¹.

Magnetic resonance imaging (MRI) is the modality of choice in neuro-oncology for diagnosis and treatment response evaluation of GBM. The PBZ on MRI is defined as the brain area surrounding the tumour without contrast enhancement in T1-gadolinium-enhanced MRI (Gd-T_1w). This region is often hyper-intense in T2-weighted (T_2w), and FLAIR acquisition, which reflects vasogenic oedema in the vicinity of the tumour and suggests tumour infiltration.⁴ While several studies have shown the association of pre-operative MRI features from within and around the tumor with patient survival in GBM,^12,13 the contribution of imaging features from the PBZ in patient survival is poorly understood. This is largely due to the challenges in capturing tumour infiltration and the subtle differences in intra-tumoral heterogeneity in PBZ on MRI.

Radiomics¹⁴ is the study of capturing subtle quantitative measurements on radiographic imaging (i.e. MRI) by computing local macro and micro-scale morphologic changes in texture patterns (e.g. roughness, image homogeneity, regularity, edges) within the lesion. These features have the potential to reflect the underlying pathophysiology of the disease by capturing statistical inter-relationships between voxels with similar (or dissimilar) contrast values. For example, the Haralick entropy feature¹⁵ measures the randomness of the gray-level distribution of intensities and has been identified as a surrogate measure of intra-tumoral heterogeneity on imaging.^16,17,18. Similarly, Laws features¹⁹ capture the responses to pre-defined filters that identify textural patterns corresponding to spots, level, waves, ripples, or edges in an image. Changes in these texture patterns may be reflective of the breakdown in micro-architecture in a lesion. Hence, our current work is based on the hypothesis that (a) radiomic features, combined across multi-parametric MRI sequences (Gd-T_1w, T_2w, FLAIR) from peritumoral regions, which we termed in this paper as “peritumoral radiomic features” can capture subtle quantitative attributes associated with tumour aggressiveness that are otherwise visually not appreciable, and (b) these attributes are different between long-term versus short GBM survivors.

In this study, we will extract a set of radiomic features that capture these subtle changes from regions of (a) enhancing tumour, (b) the PBZ, and (c) tumour necrosis, across all 3 routine multi-parametric MRI scans (T_1w, T_2w, FLAIR); and compare their combined ability to distinguish short- from long-term GBM survivors, with that of the known clinical parameters (e.g. age, gender, Karnofsky performance score (KPS)). Our ultimate goal is to develop robust, non-invasive techniques to predict overall GBM survival using routine imaging data that can be easily translated in a clinical setting for improving GBM patient management.

Materials and methods

Study population

Our cohort consisted of a total of 65 treatment-naïve multi-parametric MRI scans from The Cancer Imaging Archive (TCIA). TCIA is an open archive of cancer-specific medical images and associated clinical metadata established by the collaboration between the National Cancer Institute (NCI) and participating institutions in the United States.²⁰ The Health Insurance Portability and Accountability Act (HIPPA) compliant project in the cancer genome atlas (TCGA) was conducted in compliance with regulations and policies for the protection of human subjects, and approvals by institutional review boards were appropriately obtained. The preoperative MRIs of the corresponding patients of the project were made available for public download from TCIA. Our inclusion criteria consisted of the following: 1) availability of all 3 routine MRI sequences (Gd-T_1w, T_2w, FLAIR) for treatment-naïve patients, 2) MRI scans with diagnostic image quality, 3) availability of individual overall survival information. A total of 65 studies were identified that fulfilled these inclusion criteria and were then divided into two survival categories as short-term survival (STS) and long-term survival (LTS). The STS group consisted of patients with overall survival time (OS) between 30 and 213 days post initial imaging, while the LTS group included patients with OS between 541 and 2094 days. All MR images were acquired in axial sections with a 1.5 Tesla (T) or 3.0 T scanner. Table 1 shows the baseline characteristics of the study population.

Table 1.

Patient demographics for the two subset of studies.

Cohort	Mean Overall Survival (days)	Mean Age (years)	Mean KPS	Sex
Short Term Survivors (29 patients)	117 (Range: 30 – 213)	65.69 (Range: 38 – 84)	74.14 (Range: 40 – 100)	15 Females 14 Males
Long Term Survivors (36 patients)	948 (Range: 541 – 2129)	53.47 (Range: 14 – 75)	82.64 (Range: 60 – 100)	13 Females 23 Males

Pre-processing and Registration

For every patient study, the 2 MRI sequences, T_2w, and FLAIR were co-registered with reference to Gd-T_1w MRI using 3D affine registration. The registration was performed using 3D Slicer.²¹ To account for resolution variability, during registration, every MRI slice was resampled to a uniform pixel spacing of 0.5 × 0.5 mm². Similarly, every MRI volume was interpolated to 3 mm slice thickness. We then corrected every study for intensity non-standardness. Intensity non-standardness refers to the inherent drift between different MRI acquisitions, due to which image intensity values lack tissue specific meaning between studies and across groups. Correction for intensity non-standardness was implemented using the approach as described in²² and implemented in MATLAB R2014b (Mathworks, Natick, MA). Figure I in the supplementary document demonstrates the impact of intensity standardization on multi-institutional FLAIR scans. It is clear in Figure I(a) that different studies have different intensity ranges and are not in alignment. As a result of intensity standardization (Figure I(b)), the distributions across studies from different institutions are no longer misaligned, suggesting successful correction of the drift artefact. Additional pre-processing involving skull stripping and bias field correction was conducted. Skull stripping was performed via the skull-stripping module in 3D Slicer.²³

Segmentation of tumour and peritumoral regions on MRI

Every 2-D MRI slice with visible tumour was annotated by an expert reader into 3 regions (1) PBZ, (2) tumour necrosis, and (3) enhancing tumor.²⁴ Tumour necrosis on Gd-T_1w was identified as areas of relatively hypo-intense regions (occasionally with ring-enhancement) frequently centrally located in the tumoral region. T_2w and FLAIR scans were used to identify oedematous regions (to define PBZ), while necrosis and enhancing tumour were delineated based on post gadolinium T_1w MRI. Figure 1 shows a representative Gd-T1w and FLAIR image with expert annotations for necrosis, enhancing tumour, and oedema, outlined in green, red, and yellow respectively.

Figure 1.

Annotations of necrotic core and enhancing tumour as delineated by an expert for a representative Gd-T_1w MRI slice, are outlined in green and red respectively, while the annotations for PBZ as delineated on FLAIR are shown in yellow.

Radiomic MRI features

For every region (enhancing tumour (ET), PBZ, and necrosis (N)), and for every MRI protocol, a set of 134 radiomic features was obtained, which resulted in a total of 9 feature sets denoted as $F_a^i$, such that a ∈ {ET, PBZ, N}, and i ∈ {T_1w, T_2w, FLAIR}. Every feature set $F_a^i$ included 13 Haralick (capture spatial gray-level characteristics),¹⁵ 25 Laws¹⁹ (capture presence of spots, edges, waves, and ripples in an image), 38 Laplacian pyramid (capture edge characteristics at different resolution scales),²⁵ 48 Gabor (capture structural detail at different orientations and scales)²⁶ and 20 Histogram of Gradient orientations (HoG) (captures frequency of occurrence of spatial gradient orientation characteristics)²⁷ features. Additionally, for every region, we concatenated the features from across the 3 protocols to obtain three additional sets of multi-parametric (MP) radiomic features, denoted by $F_a^{MP}=[F_a^{T_{1w}};F_a^{T_{2w}};F_a^{\mathit{FLAIR}}]$, where a ∈ {ET, PBZ, N}. The radiomic features were first computed on a per-voxel basis. A median feature value was then calculated from the feature responses of all voxels within the region of interest. All feature calculations were performed using in-house software implemented in MATLAB R2014b platform. Detailed description of the set of features employed in this work and its possible relationship to the pathophysiology of GBM is provided in Table 2.

Table 2.

Pathophysiological significance of radiomic features reflecting the possible biological traits being captured on imaging that may be relevant in predicting overall survival in GBM patients.

Feature category	Descriptor	Intuitive Description	Relevance to GBM pathophysiology
Haralick features (Repeated occurrence of grey level configuration in the texture represented via the grey-level co-occurrence matrix (GLCM), which varies rapidly with distance in fine textures and slowly in large textures)	Inverse Difference Moment (IDM)	IDM is a reflection of the presence or absence of uniformity, and hence is a measure of local regions of homogeneity High IDM: Higher presence of locally uniform windows in GLCM Low IDM: Higher presence of locally heterogeneous windows in GLCM	Captures underlying lesion heterogeneity.
	Correlation	Quantifies the linear patterns in an image based on the distance parameter.	Increased presence of linear patterns yield higher correlation values, lack of image linearity yield lower correlation values
	Sum Entropy	Measure of GLCM relationship to distribution of intensity with respect to entropy (measure of disorder).	Higher entropy is indicative of more chaotic arrangement in areas of high viable cell population
	Sum Variance	Measure of GLCM relationship to distribution of intensity with respect to variance. High sum variance: greater standard deviation of sum average. Low sum variance: low standard deviation of sum average	Possibly accounting for greater variation of scattered atypia and local accumulation of mitotic processes as observed on histopathology.
Laws features	E5, L5, S5, R5 (combination in both X and Y directions)	E- Edges L- Level S- Spots R- Ripples	Accounting for characteristic qualitative appearance of wave, ripple, edge, and spots within an ROI
Histogram of Gradient orientations (HoG)	Per-pixel intensity orientations	Intensity orientation captures prominent direction of intensity change in X and Y direction for every pixel.	Possibly accounting for high cellular activity in enhancing lesion area.
Laplacian pyramids		Multi-resolution filters capture edges at different levels	Enhances prominent edges not discernible on the original scale

Feature selection

We used the minimum Redundancy Maximum Relevance (mRMR) feature selection scheme²⁸ to independently identify an ensemble of the most predictive features for each of the 12 feature sets, $F_a^i$, where a ∈ {ET, PBZ, N}, and i ∈ {T_1w, T_2w, FLAIR, MP}. The mRMR scheme attempts to simultaneously optimize two distinct criteria. The first is “maximum relevance” which selects features that have the maximal mutual information (MI) with respect to the corresponding label. The second is “minimum redundancy” which ensures that selected features are those, which have the minimum MI with respect to each other. For each of $F_a^i$, we identified a total of 10 most predictive features to be employed in the prediction model, such that that they were maximally dissimilar with respect to each other, while maximally similar with respect to the class labels (LTS versus STS in our case).

Classification

After feature selection, we employed a random forest (RF) classifier²⁹ to determine the ability of each of $F_a^i$ in reliably distinguishing LTS from STS studies. RF is a commonly used ensemble classifier that combines predictions from several weak decision tree classifiers to generate a more accurate and stable classifier. Treebagger implementation of the RF classifier in MATLAB was employed, with a total of 50 trees used for training the classifier. Gini impurity was used as the criterion to measure the quality of split. The RF classifier has previously been successfully employed for various biomedical classification applications³⁰. Advantages of RF include, (1) ability to integrate a large number of input variables, (2) robustness to noise in the data, and (3) relatively few tuning-parameters.

Evaluation

To ensure robustness of the classifier to training and testing data, a randomized three-fold cross-validation procedure was implemented. In a single cross-validation run, the 65 studies being considered were divided into three randomized subsets of 22, 22, and 21 studies each. Two subsets were considered as training subsets and the remaining as testing subset, following which classification was performed. This was repeated until all three subsets were classified. The entire cross-validation procedure was iterated 50 times to avoid bias. Feature selection and classifier construction were done separately for each set of training data (for all three folds over all 50 runs), with corresponding testing data only used for evaluation of classifier performance.

Kaplan Meier (KM) survival analysis was used to compare survival times between the LTS versus STS patients. The horizontal axis on the survival curve shows the time and the vertical axis shows the probability of survival. Any point on the survival curve reflects the probability that a patient in each group would remain alive at that time. Optimal classifier predictions would show maximum separation between the survival curves (Figure 3(e)). The difference between the 2 groups for survival analysis was assessed by the prediction of random forest classifier, aggregated over the 50 runs within the cross-validation.

Figure 3.

Kaplan Meier Survival curves for classification to predict short-term (shown in red curve) from long-term survivors (shown in blue curve) using clinical features like (a) age, (b) gender, (c) Karnofsky Performance Score (KPS), (d) combination of clinical features (age, gender, KPS) and top 10 peritumoral radiomic features across multi-parametric MRI sequences, as compared to the Kaplan Meier survival curve obtained from the “ground truth” labels.

Statistical Analysis

Survival curves were compared statistically by a Cox proportional hazards model.^31,32,33,34 All statistical analyses were performed using the statistical package in R³⁵. Hazard ratios (HR) were used as a means to quantify the direction of individual feature effects on survival. Features yielding negative regression coefficients (i.e. low feature values correlated with long term survival) in our Cox Model produce a HR between 0 and 1; features yielding positive regression coefficients (i.e. low feature values correlated with short term survival) produce a HR between 1 and infinity.

We also computed Concordance indices (CI)^36,37 for each of our univariate and multivariate analysis experiments using the Survival Analysis package in R. CI or C-statistic is the fraction of all pairs of subjects whose predicted survival times are correctly ordered (i.e. concordant with actual survival times). CI = 1 indicates that the model has perfect predictive accuracy, and CI = .5 indicates that the model is not better than random chance.

Results

Experiment 1: Identify the relative role of every region within and around the tumour in predicting long versus short-term GBM survival

Figure 2 shows the KM curves obtained from tumour necrosis (top row), enhancing tumour (middle row) and the PBZ (bottom row) for every MRI sequence as well as for multi-parametric features for every region. The corresponding concordance values and hazard ratio ranges are listed in Table 3.

Figure 2.

Kaplan Meier Survival curves obtained using radiomic features to predict short-term (shown in red curve) from long-term survivors (shown in blue curve). Top row shows the survival curves obtained by using the top 10 features from the tumour necrosis region alone across T1w, T2w, FLAIR, and multi-parametric MRI. Middle row shows the corresponding survival curves obtained by using the top 10 features from the enhancing lesion alone. Similarly, the bottom row shows the survival curves obtained by using the top 10 peritumoral radiomic features, for T1w, T2w, FLAIR, and multi-parametric MRI respectively.

Table 3.

Table listing the hazard ratios, concordance and statistical significance using clinical and top 10 radiomic features obtained from different compartments (edema, necrosis, enhancing lesion) as well using combined multi-parametric MRI features.

Feature	Hazard Ratio (Range reflects values across top 10 features)	p-value	CI
Age	1.025	.002	.625
Gender	1.609	.09	.563
KPS	.9735	.0015	.647
Gd-T1w enhancing lesion	[.5095 – 1.6849]	0.095	.666
Gd-T1w necrosis	[.4704 – 2.8855]	0.006	.691
Gd-T1w edema	[.1162 – 2.3760]	0.046	.67
T2w enhancing lesion	[.04750 – 1.16488]	0.003	.578
T2w necrosis	[.2117 – 1.9732]	0.004	.646
T2w edema	[.1465 – 7.8025]	0.0006	.637
FLAIR enhancing lesion	[.1302 – 1.755e6]	0.039	.629
FLAIR necrosis	[.2991 – 2.1138]	0.294	.663
FLAIR edema	[.3712 – 2.5728]	0.003	.694
Enhancing lesion – Multi-parametric MRI	[.02892 – 1.23991]	0.0002	.63
Necrosis – Multi-parametric MRI	[.7456 – 1.2107]	0.008	.656
Edema – Multi-parametric MRI	[.7149 – 1.5840]	0.000015	.702

Peritumoral radiomic features were found to be predictive across T_2w (p = 0.0006, CI = 0.637), and FLAIR sequences (p = 0.003, CI = 0.694), as compared to tumour enhancement or tumour necrosis features. On Gd-T_1w MRI, radiomic features from tumour necrosis (p = 0.006, CI = 0.69) were found to be more predictive of long-term versus short-term survivors than the PBZ features. Interestingly, peritumoral radiomic features (Figure 2(l)) when combined across multi-parametric sequences (p = 1.47 × 10⁻⁵, CI = 0.70), mostly constituting of Laws features (capturing ripple, edge, and wave like texture patterns), Haralick energy (capturing image heterogeneity), and gray-level correlation features (Table 4) outperformed radiomic features from the other regions (enhancing tumour, necrosis), and from individual sequences in their ability to distinguish STS and LTS studies.

Table 4.

List of the 5 most-predictive radiomic features from PBZ across the three MR sequences Gd- T_1w, T_2w, FLAIR, as well as across the multi-parametric set. Note that W5W5 represents a laws feature that captures wave patterns in an image using a 5×5 filter. Similarly, R5R5 represents a laws feature that captures ripple patterns, S5S5 captures spot patterns, while E5E5 captures the edge patterns in an image.

T1w	T2w	FLAIR	Multi-parametric MRI
W5W5	E5E5	W5W5	R5R5 (T1w)
R5R5	R5R5	R5R5	Sum Variance (T1w)
Sum Variance	W5W5	S5S5	R5R5 (T2w)
E5E5	S5S5	E5E5	E5E5 (T2w)
L5L5	Correlation	L5L5	R5R5 (FLAIR)

Three representative Haralick features (entropy, correlation and sum entropy features) for a long-term (bottom row) and short-term (top row) patient on Gd-contrast T_1w MRI are shown in Figure 4.

Figure 4.

A single 2-dimensional Gd-T_1w MRI slice for two different patients with STS (a) and LTS (g) respectively. Expert annotated region bounded in green is necrosis; region bounded in orange is enhancing tumour, while the region bounded in black is oedema. The corresponding per-voxel representations of 3 Haralick descriptors are shown for entropy (d, j), Correlation (e, k), and Sum Entropy features (f, l).

Experiment 2: Compare the value of the PBZ radiomic features in predicting long-term versus short-term survivors, with known clinical variables (age, KPS, gender)

Clinical variables including age, gender and KPS were evaluated in a univariate fashion to identify their ability in distinguishing STS and LTS using KM curves as shown in Figure 3. While age (p=0.002), and KPS (p=0.001) and were found to be significantly different between LTS and STS, gender was not found to be a predictive factor (p=0.09) of STS and LTS survival.

These clinical features, when combined with the multi-parametric peritumoral radiomic features (yielded highest statistical significance in Experiment 1) in a multi-variate analysis, were found to be more predictive of overall survival (CI = 0.735) as compared to using peritumoral radiomic features alone (CI = 0.702). These findings are consistent with the recently reported results in³⁸, where the combination of clinical and radiomic features were found to improve prediction of GBM survival, as compared to radiomic features alone.

Discussion

The variety in overall survival and response to treatment in GBM is largely due to the fact that, unlike other tumours, GBM is highly heterogeneous exhibiting aggressive biological traits across tumours, as well as within a single tumor.³ Interestingly, studies^4,6 have suggested that heterogeneity extends beyond the tumour margins into the peritumoral brain parenchyma, and ultimately affects overall survival in GBM. In this work, we investigated the role of “peritumoral” radiomic features computed from the PBZ region on routine multi-parametric MRI protocols (T_1w, T_2w, FLAIR), in predicting long-term versus short-term survivors of GBM.

Role of PBZ in predicting long-term versus short-term GBM survival in GBM

In our per-region analysis, peritumoral radiomic features captured from across multi-parametric MRI were found to be most predictive, as compared to the features from other regions (tumour enhancement, tumour necrosis) and across protocols. While several studies^39,40 have reported the severity of peritumoral brain oedema on diagnosis MRI as a negative prognostic factor, the exact role of PBZ in GBM prognosis remains controversial.⁴¹ For example, Schoenegger et al.,³⁹ in a multi-institutional study, identified peritumoral oedema as an independent prognostic factor of GBM survival; with patients exhibiting major oedema identified to have significant shorter overall survival compared to patients with minor oedema. However, in a large series of 416 glioblastoma MRI patient studies, Lacroix et al.⁴² did not find the extent of oedema to be an independent prognostic factor of GBM survival. Thus, the results concerning the prognostic impact of brain oedema in GBM patients have not been conclusive in the literature. We believe that one of the limitations of the existing studies in the literature has been that they have so far only investigated gross volumetric measurement of the PBZ as a prognostic measure.

Role of Radiomic features in predicting long-term versus short-term GBM survival on multi-parametric MRI sequences

Our peritumoral radiomic analysis on combined multi-parametric MRI identified Laws features (provide a quantitative measurement of presence of waves, spots, edges, and ripple textural patterns in an image), as one of the most predictive feature sets of long-term versus short-term GBM survivors. We believe that the changes due to proliferation of focal endothelial cells, and neovascularization (linked to poor outcome⁴³), are manifested as hypo-intense spot and ripple appearances on MRI, and may possibly be captured by the Laws features. Similarly, the other set of radiomic features that were found to be predictive were Haralick co-occurrence descriptors (inverse difference moment (IDM), sum variance and measure of correlation). All of these features are measures of local image homogeneity. For example, correlation is a measure of relative linear dependence in gray levels between the pixels at the specified positions. The more similar the pixels, the higher the correlation. Similarly, IDM is lower for images that have more local variation in image intensities (more heterogeneity), and relatively higher for images with more homogeneous local intensities.

Our work presented one of the first attempts at comprehensively evaluating radiomic features from different tumour-specific regions and across MRI sequences, subsequently evaluating the role of radiomic features derived from the PBZ in predicting long-term versus short-term survival patients. The criterion for overall short-term (< 7-months) and long-term (>18months) survival used in this study was relatively strict, as compared to prior studies.^1,44,45 The strict criterion allowed for identification of radiomic feature based MR imaging profiles that are distinct between patients with STS and LTS.

Limitations

This study did have some limitations. The reported results are preliminary, as our study was limited by a relatively small sample size. In light of a relatively small cohort, an independent validation of the radiomic features was not performed. Because of heterogeneity in imaging parameters within the TCIA cohort, we only included studies with routine sequences (T_1w, T_2w and FLAIR) and have not investigated radiomic features derived from advanced MR sequences, such as tumour perfusion or DWI. Segmentation variability has not been taken into consideration since this study was performed based on annotations from a single expert reader.

Concluding Remarks

In this study, we investigated whether computer-extracted radiomic features from peritumoral brain parenchyma on treatment-naïve routine MRIs can predict GBM patients with different overall survivals i.e less than 7 months (short-term) versus greater than 18 months (long-term). The results suggest that radiomic features from the peritumoral zones appear to be predictive of overall patient survival. The addition of complementary imaging parameters in future studies may further improve survival prediction using radiomic analysis. Future work will also focus on a larger GBM cohort with inclusion of intermediate survival patients as well as independent validation of the radiomic features identified in this preliminary analysis, while taking into account the extent of resection and subsequent treatment.

Key points.

1. Radiomic features from peritumoral regions can capture Glioblastoma heterogeneity to predict outcome.

2. Peritumoral radiomics along with clinical factors are highly predictive of Glioblastoma outcome.

3. Identifying prognostic markers can assist in making personalized therapy decisions in Glioblastoma.

List of Abbreviations

GBM

Glioblastoma Multiforme

PBZ

Peritumoral Brain Zone

KM

Kaplan-Meier

STS

Short-term Survival

LTS

Long-term Survival

OS

overall survival

T

Tesla

Gd

Gadolinium

KPS

Karnofsky performance score

RF

Random Forest

HIPAA

The Health Insurance Portability and Accountability Act

TCGA

The Cancer Genome Atlas