Prediction of Glioblastoma Multiforme Patient Survival Using MR Image Features and Gene Expression Control/Tracking Number: 11-O-1519-ASNR Nicolasjilwan, M.1·Clifford, R.2·Raghavan, P.1·Wintermark, M.1·Hammoud, D.3·Huang, E.4·Jaffe, C.5·Freymann, J.2·Kirby, J.2·Buetow, K.4·Huang, S.6·Holder, C.6·Gutman, D.6·Flanders, A. E.7 1University of Virginia, Charlottesville, VA, 2SAIC-Frederick, Inc., Frederick, MD, 3National Institute of Health, Bethesda, MD, 4National Cancer Institute, Bethesda, MD, 5Boston University School of Medicine, Boston, MA, 6Emory University Hospital, Atlanta, GA, 7Thomas Jefferson University Hospital, Philadelphia, PA.

DISCLOSURE Nothing to disclose

PURPOSE  Utilize conventional MRI imaging features to predict survival of patients with glioblastoma multiforme (GBM) after initial diagnosis.  Linear regression models incorporating MR imaging features and tumor gene expression to predict patient survival.  Important role in selecting treatment options.

Materials & Methods  The study is part of The Cancer Genome Atlas (TCGA) MR imaging (MRI) characterization project of the National Cancer Institute.  MR images for 70 GBM patients made available through the National Biomedical Imaging Archive were reviewed independently by six neuroradiologists.

 The VASARI feature scoring system for human gliomas, developed at Thomas Jefferson University Hospital, was employed.  30 features clustered by categories. –Lesion Location –Morphology of Lesion Substance –Morphology of Lesion Margin –Alterations in Vicinity of Lesion –Extent of resection   620 genes associated with angiogenesis used in this investigation.

 Survival was recoded as a binary categorical variable: survival less than or greater than 1 year.  Associations between imaging features and survival were assessed using linear regression models. Survival was the outcome; imaging features were the predictors.

Well marginated Non-enhancing F4 Enhancement Quality: 1=None 2=Mild/Minimal 3=Marked/Avid F13 Definition of the non-enhancing margin 1= n/a 2= Smooth 3= Irregular Courtesy Dr Adam Flanders

Predominantly Non-enhancing F5 Proportion Enhancing: 1= n/a 2=None (0%) 3= 95% 8=All (100%) Courtesy Dr Adam Flanders

RESULTS

Univariate Analysis of Association between VASARI features and survival

 Individually, 6 MRI features show association to survival with an unadjusted p-value <  Negative correlation with survival:  Ependymal extension (F19), (P = )  Longest dimension of lesion size (F29)  Deep white matter invasion (F21)  The presence of satellites (F24).  Positive correlation with survival:  Location of the tumor in the right (usually non-dominant) hemisphere (F2).  Frontal lobe location (feature F1a), (P = ).

Linear regression models incorporating the most significant VASARI feature, F19 (ependymal extension), and expression of angiogenesis-related genes.  4 genes individually improve the predictive power of F19 (ependymal extension).  Expression of ANG (angiogenin) and TGFB2 (TGF- beta 2) genes negatively correlates with survival.  CCL5 (chemokine (C-C motif) ligand 5) and TNF (tumor necrosis factor) positively correlate with survival.  Feature F19 correctly predicts survival for 72% of the cases. A model based on ependymal extension, CCL5, ANG, TGFB2 and TNF correctly predicts survival for 82% of patients.

CONCLUSION  A subset of VASARI imaging features correlate well with patient survival.  Linear regression models incorporating multiple imaging features or a single VASARI feature (ependymal extension) and tumor gene expression can be used to predict patient survival.  We are refining these models and are investigating whether including patient clinical characteristics into linear models can improve their predictive power.

AKNOWLEDGMENT University of Virginia SAIC-Frederick National Institute of Health National Cancer Institute Thomas Jefferson University Hospital Emory University Hospital Boston University School of Medicine