Directional-gradient based radiomic descriptors from pre-treatment perfusion DSC-MRI to differentiate long-term from short-term survivors in Glioblastoma: Preliminary findings
Bolin Song1, Ramon correa1, Prateek Prasanna1, Niha Beig1, Anant Madabhushi1, and Pallavi Tiwari1

1Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States


We explored the utility of radiomic analysis to identify radiomic features (computer extracted features from MRI) that distinguish long-term survival patients from their short-term survival counterparts based on the pre-treatment perfusion DSC-MRI. Initial results indicate that dynamically extracted radiomic features from enhancing tumor and infiltrative edges on perfusion scans can segregate the 2 survival groups. A non-invasive means of predicting survival based on perfusion imaging may help clinicians to determine prognosis, and inform treatment strategy.


Glioblastoma (GBM) is a highly aggressive malignant primary brain tumor with an average survival of 14 months, with less than 5% patients surviving for over 3-years. With new experimental therapies in GBM under investigation, there is a pressing need for accurate risk stratification towards developing a more personalized approach in GBM management. We have previously shown success in developing a new radiomic feature COLLAGE that captures tumor heterogeneity by computing the entropy (measure of disorder) in pixel level edge directions1 from structural MRI scans. COLLAGE has previously been shown to distinguish different complex pathologies in brain tumors2. In this work, we present an extension of our COLLAGE feature called dynamic COLLAGE, to capture the subtle texture variations in perfusion traits of the tumors from pre-treatment DSC perfusion MRI scans, over multiple phases of contrast administration. We further investigate if the dynamic COLLAGE feature could discriminate long-term survivors (LTS) from short-term survivors (STS) in GBM. We hypothesize that the dynamic variations during the contrast uptake may perhaps be too subtle to be appreciable on the DSC MRI signal intensities alone and may be better captured via the changes in gradient entropy appearances over multiple phases of contrast administration.


Our dataset consisted of 64 perfusion DSC-MRI Glioblastoma cases from TCIA3, 35 of which were identified as STS (Survival<13 months) and 29 were LTS patients (Survival>13 months). Perfusion scans across all time phases were first registered to the corresponding T2w, FLAIR, and T1 Gadolinium contrast enhanced MRI scans using 3D Slicer (4.8.0). Annotations for enhancing tumor, and necrosis region were performed by an expert on T1contrast MRI scans, while the annotations for edema were performed using the FLAIR scans. These annotations were then transferred to the spatially registered DSC perfusion scans across all time-phases. We ensured that the time-phases for all patients were appropriately sampled at consistent time phases relative to the contrast injection time. 13 COLLAGE features were then extracted from each of the normalized time phases, followed by computing 4 statistics, mean, standard deviation, skewness, and kurtosis, for every COLLAGE feature, resulting in a total of 52 COLLAGE features for every tumor region, for every patient. These COLLAGE values for every time-phase on DSC perfusion MRI were then plotted as a temporal profile to reflect the kinetic changes in COLLAGE expression over time (Figures 1(a)). For classification, we split our data set into training (39 cases, 20 STS and 19 LTS) and hold-out set (25 cases, 15 STS and 10 LTS). In the training set, for each of the 52 COLLAGE features, we fed the corresponding dynamic COLLAGE feature vector obtained from across the time phases, into a linear discriminant analysis (LDA) classifier within a 3-fold cross validation setting. The area under ROC curve (AUC) and accuracy measures were recorded for each of the 52 COLLAGE features, and the top features were similarly evaluated in terms of their classification performance on the hold-out set.

Results and Discussion

The most significant radiomic feature from the edema region across STS and LTS subgroups was found to be kurtosis measurements of COLLAGE correlation (window size = 5) that captured correlations of co-occurrence of gradient orientations on perfusion DSC-MRI sequences. Using this feature, we obtained an AUC of 0.7537 on the training set and 0.9236 on an hold-out testing set. The best feature in the enhancing tumor region was found to be variance of COLLAGE sum of variance (window size = 5), which when using within the LDA classifier had an AUC of 0.8387 on the training set and 0.90 on the hold-out set. It is likely that GBMs for patients with short term survival are likely more aggressive and heterogeneous, compared to long-term survival patients. We observed a consistent trend reflecting higher values of dynamic COLLAGE from the enhancing lesion across all time-phases for STS patients, compared to LTS patients. (Figure 1(a)) This suggests that COLLAGE is likely picking up more heterogeneity on DSC MRI in poor survivors, as compared to patients with improved outcome.


Our results demonstrate feasibility of using dynamic COLLAGE radiomic analysis to discriminate the two distinct survival groups (LTS versus STS) from pre-operative perfusion DSC-MRI scans. Prospective validation of dynamic COLLAGE features in larger cohort may provide prognostic implications in designing treatment decisions for Glioblastoma patients.


Research reported in this publication was supported by Coulter Translational Award, National Cancer Institute of the National Institutes of Health underaward numbers R01CA136535-01, R01CA140772-01 & R21CA167811-01; the National Institute of Biomedical Imaging and Bioengineering of the NationalInstitutes of Health under award number R43EB015199-01.


1. Textural Kinetics: A Novel Dynamic Contrast-Enhanced (DCE)-MRI Feature for Breast Lesion Classification. Shannon C. Agner et al. Journal of Digital Imaging, Vol 24, No 3 (June), 2011.

2. Prasanna, P., Tiwari, P., & Madabhushi, A. (2014). Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe): Distinguishing Tumor Confounders and Molecular Subtypes on MRI. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2014 (pp. 73-80). Springer International Publishing.

3. Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. Journal of Digital Imaging. 2013; 26(6): 1045-1057. doi: 10.1007/s10278-013-9622-7.


Figure 1: (a) shows values of the best feature at five time phases in Edema for one STS patient and one LTS patient. (b) shows the annotations overlay with tumor (red refers to STS the patient and blue for LTS) and their corresponding feature heatmap for Edema

Figure 2: (a) shows the best feature in Edema at five time phases for STS and LTS patients in testing set. (b) shows the best feature in Enhancing tumor at five time phases for two groups of patients in testing set.

Figure 3: (a) classification accuracy using the best feature in Edema (CoLIAGe Kurtosis of correlation) for both LDA and QDA classifier. (b) shows the classification accuracy using the best feature in Enhancing lesion (CoLIAGe Sum of variance) for both LDA and QDA classifier.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)