Fig. 1.

Pipeline for the proposed radiomic feature robustness assessment. A set of single-center MRI images is selected. After pre-processing and automated tumor segmentation, the images are artificially perturbed. For each perturbation type, the robustness is metered by the intraclass correlation coefficient (ICC(2,1)). Measuring agreement and not only consistency of underlying features is key for transferring trained machine learning (ML) models to a different dataset. Redundant features are removed from the robust features. Subsequently, combinations of feature selectors and ML models are tested on different survival class boundaries. The best performing model is tested on a multi-center dataset (TCIA subset of BraTS)