Studying Robustness of Semantic Segmentation Under Domain Shift in Cardiac MRI

Cardiac magnetic resonance imaging (cMRI) is an integral part of diagnosis in many heart related diseases. Recently, deep neural networks have demonstrated successful automatic segmentation, thus alleviating the burden time-consuming manual contouring cardiac structures. Moreover, frameworks such as nnU-Net provide entirely auto- matic model configuration to unseen datasets enabling out-of-the-box application even by non-experts. However, current studies commonly neglect clinically realistic scenario, which a trained network applied data from different domain deviating scanners or protocols. This potentially leads unexpected performance drops learning models real life applications. In this work, we systematically study challenges and opportunities transfer across images multiple clinical centres scanner vendors. order maintain usability, build upon fixed U-Net architecture configured nnU-net framework investigate various augmentation techniques batch normalization layers easy-to-customize pipeline component general guidelines on how improve generalizability abilities existing methods. Our proposed method ranked first at Multi-Centre, Multi-Vendor & Multi-Disease Image Segmentation Challenge (M&Ms).