Association of geometric features with genetic markers in glioblastoma multiforme
Adam Hasse1, Mark Dapash2, Yong Jeong3, Star Su4, Abrianna Cummings4, Sameer Ansari5, Daniel Ginat6, and Timothy J Carroll6

1Graduate Program in Medical Physics, University of Chicago, Chicago, IL, United States, 2Pritzker School of Medicine, University of Chicago, Chicago, IL, United States, 3Department of Biomedical Engineering, Northwestern University, Evanston, IL, United States, 4MedIx REU Program, Depaul University, Chicago, IL, United States, 5Department of Radiology, Northwestern University, Chicago, IL, United States, 6Department of Radiology, University of Chicago, Chicago, IL, United States


A data analysis pipeline was developed to study relationships between geometric imaging features and the underlying tumor phenotype. This pipeline was run on the tumors of 203 patients clinically diagnosed with glioblastoma multiforme. For each tumor, the tumor bulk and percent non-enhancing volume was calculated, along with the surface regularity of any tumor with a 3D T1-weighted post-contrast MR scan. These features were compared to the expression of P53 and Ki67 sampled from these tumors. Although there were significant differences between multiple features for both genes, only the surface regularity was a significant predictor of Ki-67 proliferative index.

Target Audience

Researchers and clinicians interested in radiomics, specifically radiomics of glioblastoma multiforme (GBM).


GBM is the most common type of primary brain tumor, accounting for about 50% of gliomas. Even with an aggressive treatment regiment of surgical resection followed by radiation and chemotherapy, the median survival time is under 15 months, with a 5-year survival rate of about 10%.1,2 This poor prognosis can be attributed in part to the large phenotypic heterogeneities present in GBM.3 Because of this, much work has been done to further understanding of the various tumor phenotypes of GBM to advance treatment options. However, these phenotypes can only be determined by biopsy or surgical resection, both of which present significant risk to the patient. By correlating imaging features from standard MRI images with these phenotypes, decisions about treatment options can be more informed and more selective.


A cohort of 203 patients (142 males, 111 females, average age of 57.5 years) with confirmed diagnoses of GBM were selected from a larger dataset. These were specifically chosen due to the presence of some sort of genetic data and T1-weighted, post-contrast MR images, both of which were required for this analysis. Software was developed in MATLAB so a trained operator could outline the contrast-enhanced tumor on axial T1-weighted post-contrast images, along with the necrotic core if it was present. The contrast-enhanced outlines were then slightly modified by a dilation, followed by a gray level thresholding method to more accurately conform to the outline of the contrast-enhanced region.4 This “shrink-wrapping” method was validated using two separate users on twenty cases, resulting in a dice similarity coefficient of overlap outlines of 0.92. Further, the non-enhancing region of the tumor was also outlined, if present. These outlines were confirmed by a board-certified radiologist with 7 years of experience.

To obtain the tumor bulk, defined as the contrast-enhanced region and the associated necrosis, the outlines were smoothed in the axial direction to reduce the sharp steps from slice-to-slice. Tumor bulk (TB) then was calculated for each imaging set by summing the volumes of truncated pyramids (see Equation 1), where h is the slice thickness and An is the area of the tumor bulk on the nth slice.

$$$TB = \sum_{n} \frac{h}{3} (A_{n} + A_{n+1} + \sqrt{A_{n}A_{n+1}})$$$ Equation 1

Similarly, the volume of the non-enhancing area was calculated.

Additionally, the surface regularity was calculated for each outline.5 However, due to the dependence of this value on the surface area (TS), only 3D T1-weighted post-contrast images were used for this analysis. The surface regularity (SR) was calculated using Equation 2.

$$$SR = 6\sqrt{\pi} \frac{TB}{\sqrt{TS^{3}}}$$$ Equation 2

After these features were calculated, two genetic markers were highlighted for analysis due to the large number of cases present. P53, a tumor suppression gene, and Ki67, a proliferative marker, were analyzed in this study. Because these genes are measured by the fraction of cells stained via immunohistochemistry, the median value was chosen for each (P53 labeling index (P53-LI) = 0.1, Ki67 proliferative index (ki67-PI) = 0.2) to separate each data set into two classes (see Figure 1). T-tests and ROC analysis were performed to determine separation of the classes and the predicting power of each feature.


Figure 2 shows a synopsis of the results of the unpaired t-test between the two classes and the corresponding geometric features. If the separation between the two classes was significant, ROC analysis was performed, and the AUC and its corresponding significance above the guessing line was calculated. Because the only significant AUC occurred with Ki67-PI and surface regularity, the corresponding ROC curve is shown in Figure 3.


Although there was significant separation between tumor bulks of the two classes of P53-LI (P53-LI overexpression has a significantly higher tumor bulk than P53-LI nominal expression), there was no predictive power in using tumor bulk. This is similar for the percent non-enhancing volume of P53-LI and Ki67-PI as well. Since symptoms of GBM can present with or without symptoms, it is difficult to determine the initiation of the tumor, which would need to be normalized for a proper analysis. However, surface regularity, which just states how spherical a tumor is, does not have the same issues. Overexpression of Ki67-PI has a significantly higher surface regularity than nominal expression of Ki67-PI, and it has significant predicting power. Thus, it is possible to conclude that surface regularity may be a geometric biomarker of Ki67-PI in GBM.


This work is supported, in part, by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under grant number T32 EB002103, in part, by the Division of Information and Intelligent Systems at the National Science Foundation under grant number 1659836, and in part by the Pritzker School of Medicine Summer Research Program.


1. Hanif F, Muzaffar K, Perveen K, Malhi SM, Simjee SU. Glioblastoma Multiforme: A Review of its Epidemiology and Pathogenesis through Clinical Presentation and Treatment. Asian Pac J of Cancer Prev. 18(1), 3-9 (2017).

2. Stupp R et al. Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial. Lancet Oncol. 10(5), 459-66 (2009).

3. Verhaak, RGW et al. An integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDFGRA, IDH1, EGFR, and NF1. Cancer Cell 17(1), 98 (2010).

4. Otsu, N. A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 9(1), 62-6 (1979).

5. Perez-Beteta, J. et al. Tumor Surface Regularity at MR Imaging Predicts Survival and Response to Surgery in Patients with Glioblastoma. Radiology 00, 1-8 (2018).


Separation of classes based on P53-LI and Ki67-PI. Number of 3D cases for surface regularity calculations are shown in parentheses.

Results showing significant separation (column 3) and significant AUC (column 4).

ROC curve for Ki67-PI and surface regularity

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