federated medical image segmentation with concept shift

LAScarQS dataset was obtained from MICCAI 2022 Left Atrial and Scar Quantification & Segmentation Challenge. It collected late gadolinium-enhanced (LGE) magnetic resonance imaging (MRI) scans of left atrial and their segmentation from three centers, i.e., 20 cases from Beth Israel Deaconess Medical Center (Center A), 20 cases from Imaging Sciences at King’s College London (Center B) and 100 cases from Utah School of Medicine. We randomly selected 30 cases from the dataset in Utah as generic test data. To simulate the label bias scenario, we equally split the rest of the data from Utah into two centers (Center C and D) and performed random erosion and dilation to their annotations, respectively. The morphology transformations of the annotations are implemented based on the Markov process similar to the previous work [62] (Please refer to supplemental materials for details). The information about data partition and transformation is summarized in Table. 2. All images were resampled to the resolution of 0.6 × 0.6 × 1.25mm, cropped into the size of 256 × 256, and normalized with Z-score normalization. Data augmentations such as random rotation, flip, elastic deformation, and Gaussian noise were applied during training.