Style-Invariant Cardiac Image Segmentation with Test-Time Augmentation

Deep models often suffer from severe performance drop due to the appearance shift in real clinical setting. Most of existing learning-based methods rely on images multiple sites/vendors or even corresponding labels. However, collecting enough unknown data robustly model segmentation cannot always hold since complex caused by imaging factors daily application. In this paper, we propose a novel style-invariant method for cardiac image segmentation. Based zero-shot style transfer remove and test-time augmentation explore diverse underlying anatomy, our proposed is effective combating shift. Our contribution three-fold. First, inspired spirit universal transfer, develop stylization content generate stylized that similarity images. Second, build up robust based U-Net structure. framework mainly consists two networks during testing: ST network removing network. Third, investigate transformed versions prediction results are merged. Notably, fully adaptation. Experiment demonstrate promising generic generalizing deep models.