Room: Room 205
Purpose: To evaluate the effects of several image preprocessing techniques on the robustness of Random Forest (RF) classifiers for the prediction of MGMT methylation status in patients with glioblastoma multiforme (GBM).
Methods: Two independent GBM cohorts were retrospectively analyzed. Fifty cases from the TCGA-GBM dataset were used as the training set. A second cohort acquired at our institution (n = 111) was used for validation. A total of 147 texture features were extracted from the segmented tumor volumes using three sequences viz. FLAIR (Fluid-Attenuated Inversion Recovery), T1-weighted, and T1-weighted post-contrast images. Features were computed from original images and those preprocessed with (a) bias field correction or (b) histogram standardization. Haralick features were computed using different bin numbers (16, 32, 64, 128), resulting in 12 different settings for generating features. Features were selected according to their relevance to MGMT methylation status using the Wilcoxon test (p value < 0.1), and those found to be relevant were used in RF classifiers using hold-out validation, where one-third of the discovery samples were randomly selected for testing. The best-performing RF classifier was validated on the second cohort. Kaplan-Meier analysis was performed to assess the feasibility of MGMT methylation status determined using the RF model for predicting overall survival (OS).
Results: Evaluation of 12 RF models indicates that histogram standardization resulted in the best classification scheme for predicting MGMT methylation status, achieving an area under the receiver-operating characteristic (ROC) curve of 0.75 (95% CI: 0.72-0.77) and 0.60 (95% CI: 0.44-0.65), a sensitivity of 0.84 and 0.76, and a specificity of 0.65 and 0.44, in the discovery and validation cohorts, respectively. MGMT methylation status predictions made by this classifier predicted survival (p=0.048) in the validation set.
Conclusion: The results demonstrate image preprocessing can improve the prognostic utility of machine-learning classifiers constructed with MRI features.
Funding Support, Disclosures, and Conflict of Interest: This work was in part through the NIH/NCI Cancer Center Core Support Grant P30 CA008748, and the MSK Department of Radiology, Brain Tumor Center and Neuro-Oncology Research Translation in Humans.