ROOD-MRI: Benchmarking the robustness of deep learning segmentation models to out-of-distribution and corrupted data in MRI

Lyndon Boone, Mahdi Biparva, Parisa Mojiri Forooshani, Joel Ramirez, Mario Masellis, Robert Bartha, Sean Symons, Stephen C. Strother, Sandra E. Black, Chris Heyn, Anne L. Martel, Richard H. Swartz, Maged Goubran

Abstract: Highlights•Developed open-source benchmarking platform and metrics for robustness of DNNs.•Quantified sensitivity of DNNs to OOD data on three neuroimaging segmentation tasks.•Modern CNNs are highly susceptible to distribution shift, corruptions and artifacts.•Simple augmentation strategies improve robustness for anatomical segmentation tasks.•Vision transformers exhibit improved robustness over FCNs.