Objectives: The aim of this study was to investigate whether radiomic analysis with random survival forests (RSFs) can predict overall survival from T1-weighted contrast-enhanced baseline magnetic resonance imaging (MRI) scans in a cohort of glioblastoma multiforme (GBM) patients with uniform treatment.
Materials and methods: This retrospective study was approved by the institutional review board and informed consent was waived. The MRI scans from 66 patients with newly diagnosed GBM from a previous prospective study were analyzed. Tumors were segmented manually on contrast-enhanced 3-dimensional T1-weighted images. Using these segmentations, P = 208 quantitative image features characterizing tumor shape, signal intensity, and texture were calculated in an automated fashion. On this data set, an RSF was trained using 10-fold cross validation to establish a link between image features and overall survival, and the individual risk for each patient was predicted. The mean concordance index was assessed as a measure of prediction accuracy. Association of individual risk with overall survival was assessed using Kaplan-Meier analysis and a univariate proportional hazards model.
Results: Mean overall survival was 14 months (range, 0.8-85 months). Mean concordance index of the 10-fold cross-validated RSF was 0.67. Kaplan-Meier analysis clearly distinguished 2 patient groups with high and low predicted individual risk (P = 5.5 × 10). Low predicted individual mortality was found to be a favorable prognostic factor for overall survival in a univariate Cox proportional hazards model (hazards ratio, 1.038; 95% confidence interval, 1.015-1.062; P = 0.0059).
Conclusions: This study demonstrates that baseline MRI in GBM patients contains prognostic information, which can be accessed by radiomic analysis using RSFs.
Trial registration: ClinicalTrials.gov NCT01089868.