3D Cross-Pseudo Supervision (3D-CPS): A Semi-supervised nnU-Net Architecture for Abdominal Organ Segmentation

Large curated datasets are necessary, but annotating medical images is a time-consuming, laborious, and expensive process. Therefore, recent supervised methods focusing on utilizing large amount of unlabeled data. However, to do so, challenging task. To address this problem, we propose new 3D Cross-Pseudo Supervision (3D-CPS) method, semi-supervised network architecture based nnU-Net with the method. We design preprocessing. In addition, set loss weights expand linearity each epoch prevent model from low-quality pseudo-labels in early training Our proposed method achieves an average dice similarity coefficient (DSC) 0.881 normalized surface distance (NSD) 0.913 2022-MICCAI-FLARE validation (20 cases).