Impact of image preprocessing on the scanner dependence of multi-parametric MRI radiomic features and covariate shift in multi-institutional glioblastoma datasets.

Abstract

Recent advances in radiomics have enhanced the value of medical imaging in various aspects of clinical practice, but a crucial component that remains to be investigated further is the robustness of quantitative features to imaging variations and across multiple institutions. In the case of MRI, signal intensity values vary according to the acquisition parameters used, yet no consensus exists on which preprocessing techniques are favorable in reducing scanner-dependent variability of image-based features. Hence, the purpose of this study was to assess the impact of common image preprocessing methods on the scanner dependence of MRI radiomic features in multi-institutional glioblastoma multiforme (GBM) datasets. Two independent GBM cohorts were analyzed: 50 cases from the TCGA-GBM dataset and 111 cases acquired in our institution, and each case consisted of 3 MRI sequences viz. FLAIR, T1-weighted, and T1-weighted post-contrast. Five image preprocessing techniques were examined: 8-bit global rescaling, 8-bit local rescaling, bias field correction, histogram standardization, and isotropic resampling. A total of 420 features divided into 8 categories representing texture, shape, edge, and intensity histogram were extracted. Two distinct imaging parameters were considered: scanner manufacturer and scanner magnetic field strength. Wilcoxon tests identified features robust to the considered acquisition parameters under the selected image preprocessing techniques. A machine learning-based strategy was implemented to measure the covariate shift between the analyzed datasets using features computed using the aforementioned preprocessing methods. Finally, radiomic scores (rad-scores) were constructed by identifying features relevant to patients overall survival after eliminating those impacted by scanner variability. These were then evaluated for their prognostic significance through Kaplan-Meier and Cox hazards regression analyses. Our results demonstrate that overall, histogram standardization contributes the most in reducing radiomic feature variability as it is the technique to reduce the covariate shift for 3 feature categories and successfully discriminate patients into groups of different survival risks.