Dual Task Semi-supervised Pancreatic Segmentation Based on Prior Information and Multiple Regularization

Abstract

Due to the relatively small size and complex internal structure of the pancreas, the segmentation is often inaccurate during image processing. A more effective automatic segmentation method is proposed to solve this problem. A multi-task deep neural network architecture based on V-Net architecture is proposed. By capturing the relationship between the prior positions of the pancreas, the target of the pancreas can be constrained at the regional level. In addition, this study uses dual task training methods to simultaneously perform segmentation tasks and regression tasks, and generate high-quality pseudo-label graphs to better utilize the valid information in a large number of unlabeled data. At the same time, this study also introduces the idea of consistency regularization, which uses the consistency regularization of prior information and the consistency regularization of noise disturbance and network disturbance to optimize the segmentation network between double decoders and double tasks, so as to further improve the segmentation effect and generalization ability of the model. Experiments show that compared with the benchmark method, the Dice coefficient of the method in this study is improved by 5.40% (for 10% labeled data) and 3.46% (for 20% labeled data) respectively, which proves the efficiency of the method in processing unlabeled medical images described in this chapter.