Abdomen Multi-organ Segmentation Using Pseudo Labels and Two-Stage

Recently, the nnU-Net network had achieved excellent performance in many medical image segmentation tasks. However, it also had some obvious problems, such as being able to only perform fully supervised tasks, excessive resource consumption in the predict. Therefore, in the abdominal multi-organ challenge of FLARE23, only incomplete labeled data was provided, and the size of them was too large, which made the original nnU-Net difficult to run. Based on this, we had designed a framework that utilized generated pseudo labels and two-stage segmentation for fast and effective prediction. Specifically, we designed three nnU-Net, one for generating high-quality pseudo labels for unlabeled data, the other for generating coarse segmentation to guide cropping, and the third for achieving effective segmentation. Our method achieved an average DSC score of 88.87% and 38.00% for the organs and lesions on the validation set and the average running time and area under GPU memory-time cure are 45 s and 3000 MB, respectively.