Multi-organ Segmentation Based on 2.5D Semi-supervised Learning

Automatic segmentation of multiple organs is a challenging topic. Most existing approaches are based on 2D network or 3D network, which leads to insufficient contextual exploration in organ segmentation. In recent years, many methods for automatic segmentation based on fully supervised deep learning have been proposed. However, it is very expensive and time-consuming for experienced medical practitioners to annotate a large number of pixels. In this paper, we propose a new twodimensional multi slices semi-supervised method to perform the task of abdominal organ segmentation. The network adopts the information along the z-axis direction in CT images, preserves and exploits the useful temporal information in adjacent slices. Besides, we combine CrossEntropy Loss and Dice Loss as loss functions to improve the performance of our method. We apply a teacher-student model with Exponential Moving Average (EMA) strategy to leverage the unlabeled data. The student model is trained with labeled data, and the teacher model is obtained by smoothing the student model weights via EMA. The pseudo-labels of unlabeled images predicted by the teacher model are used to train the student model as the final model. The mean DSC for all cases we obtained on the validation set was 0.5684, the mean NSD was 0.5971, and the total run time was 783.14 s.