Haruka Itakura MD, Presenter: Nothing to Disclose

Achal Achrol, Abstract Co-Author: Nothing to Disclose

Tiffany Ting Liu BS, Abstract Co-Author: Nothing to Disclose

Sebastian Echegaray MS, Abstract Co-Author: Nothing to Disclose

Joshua Joseph Loya BA, MS, Abstract Co-Author: Nothing to Disclose

Abdullah H. Feroze BS, Abstract Co-Author: Nothing to Disclose

Lex Allen Mitchell MD, Abstract Co-Author: Nothing to Disclose

Scott Rodriguez, Abstract Co-Author: Nothing to Disclose

Erick Michael Westbroek, Abstract Co-Author: Nothing to Disclose

Samuel H. Cheshier MD, Abstract Co-Author: Nothing to Disclose

Gary K. Steinberg MD, PhD, Abstract Co-Author: Nothing to Disclose

Daniel L. Rubin MD, MS, Abstract Co-Author: Nothing to Disclose

Kristen W. Yeom MD, Abstract Co-Author: Nothing to Disclose

Sandy Napel PhD, Abstract Co-Author: Medical Advisory Board, Fovia, Inc Consultant, Carestream Health, Inc Scientific Advisor, EchoPixel, Inc

Griffith Harsh, Abstract Co-Author: Nothing to Disclose

Olivier Gevaert PhD, Abstract Co-Author: Nothing to Disclose

To develop and validate a univariate and multivariate model-based quantitative image signature to prognosticate survival in glioblastoma multiforme (GBM)

Preoperative MR imaging and survival data from 553 patients from two distinct cohorts with de novo GBM were analyzed. First, we analyzed single-institution data on 360 subjects with GBM at our medical center. A board-certified neuroradiologist delineated Regions-Of-Interest (ROIs) around areas of enhancement in each T1 post-contrast MR slice to define a 3D tumor volume. We computed quantitative image features (morphological characteristics and pixel density statistics) from these 3D ROIs and compared them to 2D features derived from the largest slice of the tumor volume. We applied Cox proportional hazards modeling to individual image features with correction for multiple hypothesis testing to identify markers significantly correlated with survival. We then performed multivariate Cox proportional hazards regression with L1-norm regularization to build a parsimonious model that best approximated the survival outcome. Finally, we validated this multivariate model on an independent, validation cohort, consisting of 193 subjects whose MR imaging and survival data were obtained from The Cancer Imaging Archive and The Cancer Genome Atlas, respectively, and processed in the same manner as above.

From the training and validation sets, we extracted 138 quantitative image features in 2D and 125 in 3D for each patient. In the univariate Cox proportional hazards model, 38 2D and 42 3D image features were significantly associated with survival after correcting for multiple hypothesis testing (P-value <0.05, FDR <0.05). In the multivariate Cox model, combinations of six 2D features (p=0.009), and two 3D features (p=0.0132), respectively, were significantly associated with survival. These particular features capture the variability of the boundary shape, with smooth shapes correlated to good prognosis and irregular shapes correlated with bad prognosis.

Univariate and multivariate combinations of quantitative image features from both 2D and 3D MR robustly predicted survival in GBM. The predictive strength of these features was further confirmed using an independent validation cohort.

A robust quantitative image signature may constitute the basis of a clinical tool for noninvasively prognosticating survival in patients with GBM.

Itakura, H, Achrol, A, Liu, T, Echegaray, S, Loya, J, Feroze, A, Mitchell, L, Rodriguez, S, Westbroek, E, Cheshier, S, Steinberg, G, Rubin, D, Yeom, K, Napel, S, Harsh, G, Gevaert, O, Development and Validation of a Quantitative Image Signature that Predicts Clinical Survival in Glioblastoma.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14009797.html