Imaging descriptors improve the predictive power of survival models for glioblastoma patients

Neuro-Oncology 15(10):1389 – 1394, 2013. doi:10.1093/neuonc/nos335 Advance Access publication February 7, 2013 N E U RO - O N CO LO GY Imaging descriptors improve the predictive power of survival models for glioblastoma patients Department of Radiology, Duke University Medical Center, Durham, North Carolina, (M.A.M.); The Preston Robert Tisch Brain Tumor Center, Duke University, Durham, North Carolina, (A.D.); Department of Electrical & Computer Engineering, Duke University, Durham, North Carolina (J.M.M.) Background. Because effective prediction of survival time can be highly beneficial for the treatment of glioblastoma patients, the relationship between survival time and multiple patient characteristics has been investigated. In this paper, we investigate whether the predictive power of a survival model based on clinical patient features improves when MRI features are also included in the model. Methods. The subjects in this study were 82 glioblastoma patients for whom clinical features as well as MR imaging exams were made available by The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA). Twenty-six imaging features in the available MR scans were assessed by radiologists from the TCGA Glioma Phenotype Research Group. We used multivariate Cox proportional hazards regression to construct 2 survival models: one that used 3 clinical features (age, gender, and KPS) as the covariates and 1 that used both the imaging features and the clinical features as the covariates. Then, we used 2 measures to compare the predictive performance of these 2 models: area under the receiver operating characteristic curve for the 1-year survival threshold and overall concordance index. To eliminate any positive performance estimation bias, we used leave-one-out cross-validation. Results. The performance of the model based on both clinical and imaging features was higher than the performance of the model based on only the clinical features, in terms of both area under the receiver operating characteristic curve (P , .01) and the overall concordance index (P , .01). Conclusions. Imaging features assessed using a controlled lexicon have additional predictive value compared with clinical features when predicting survival time in glioblastoma patients. Received August 22, 2012; accepted December 10, 2012. Corresponding Author: Jordan Malof, PhD, Department of Electrical & Computer Engineering, Duke University, 130 Hudson Hall, Durham, NC 27708 (). Keywords: glioblastoma, MRI, proportional hazards, survival analysis, VASARI. G lioblastoma (GBM) is the most commonly occurring type of malignant primary brain tumor and the second most common type of primary brain tumor in general.1 It is characterized by very poor survival rates: a 1-year survival rate of 35.2% and a 5-year survival rate of only 4.7%.1 Accurate prognosis for individual patients could be of high benefit to them, and thus multiple studies have been published examining the impact of various factors on time to death. Lacroix et al2 have shown that a high (≥98%) extent of tumor resection gives a significant survival advantage compared with a low (,98%) extent of resection. The dependence of survival on complete resection of the enhancing tumor was further confirmed by Stummer et al.3 Regarding clinical features, it has been demonstrated that age2,4 and Karnofsky Performance Status (KPS)2,4,5,6 are significant predictors of survival. Multiple recent studies focus on genomic predictors of survival in GBM patients. One among the most prominent studies is that of Verhaak et al,7 who found a gene expression– based classification for GBM patients that relates well to their clinical outcomes. Although notably less attention has been given to the predictive value of pre- and postoperative medical imaging scans, some studies on the topic are available. Lacroix et al2 examined 7 different features based on pre- and postoperative MRI scans and showed that 4 of them were significant predictors of survival: tumor functional grade (proximity to eloquent brain), necrosis grade, edema grade, and enhancement grade. Pope et al8 evaluated the impact of 15 MRI variables on survival in GBM patients and found that noncontrast-enhancing tumor (nCET), edema, satellites, and multifocality were significant predictors of survival. # The Author(s) 2013. Published by Oxford University Press on behalf of the Society for Neuro-Oncology. All rights reserved. For permissions, please e-mail: . Maciej Andrzej Mazurowski, Annick Desjardins, and Jordan Milton Malof Mazurowski et al.: Imaging descriptors improve the power of GBM survival models Materials and Methods Patient Population In this study, we used data provided by The Cancer Genome Atlas (TCGA) that contained clinical as well as genomic information for patients. The data in the TCGA set were collected according to appropriate institutional review board approval (TCGA Research Network 2008).11 For the subset of the GBM patients from TCGA, the MRI exams were made available by The Cancer Imaging Archive (TCIA) through a collaborative effort between the National Cancer Institute (NCI) and multiple clinical institutions in the United States. For this study, out of these we identified 82 GBM subjects for whom both the clinical information of interest (age, gender, and KPS) and imaging features extracted from the MRI exams by radiologists were available. Each of these cases was scrutinized and assigned MRI features by a panel of radiologists using the standardized VASARI lexicon (http://cabig.cancer. gov/action/collaborations/vasari/). For each case, a consensus rating was established by summarizing the radiologists’ ratings. Each case was assigned to at least 3 radiologists for rating, and 76/82 cases were in fact rated by at least 3 radiologists. For the majority (68/82), the consensus was based on ratings by exactly 3 radiologists per case. Of the remaining 14 cases, 5 were rated by 6 radiologists, 3 were rated by 4 radiologists, 3 were rated by 2 radiologists, and 3 were rated by 1 radiologist. The reason for a case being rated by fewer than 3 radiologists (6 total cases) is that, 1 or 2 radiologists were not able to identify all necessary exams in the database (eg, 1 of the 3 MRI modalities) or deemed the case exams unsuitable for rating. In such situations, an additional arbiter investigated the case to resolve the conflict. 1390 NEURO-ONCOLOGY † OCTOBER 2013 The image annotations were collected through an NCIcoordinated multi-institutional effort by members of the TCGA Glioma Phenotype Research Group and were made available to us by the group. Patient Features Each patient was characterized by a set of clinical and imaging (ie, VASARI) features. Specifically, the clinical features were age (in days), gender, and KPS. The VASARI lexicon for MRI annotation contains 26 imaging descriptors based on different MRI modalities, including T1 and T2/fluid attenuated inversion recovery (FLA (...truncated)