Pancreas Segmentation Inspired By Curriculum Learning

Semantic segmentation of medical images is an important part of computer aided diagnosis.Pancreas,as one of the important endocrine organs in abdominal organs,needs to improve its segmentation accuracy compared with other abdominal organs.Pancreas segmentation is a challenge for deep convolutional neural networks due to its very small proportion in abdominal computed tomography images,no more than 0.5%,and its anatomically complex morphology.The existing coarse-to-fine segmentation method provides an efficient framework for solving the problem of pancreas segmentation.However,how to improve the segmentation accuracy of pancreas to the level of other abdominal organs segmentation is still a very challenging problem.The work of this paper focuses on the design of pancreas segmentation method based on convolutional neural network to improve the accuracy of pancreas segmentation,and its contributions can be summarized as follows:1.At present,pancreas segmentation is usually performed by coarse-to-fine method,the two-stage segmentation method believes that although the coarse model can not segment the pancreas well,it can detect the approximate location of the pancreas.However,the pancreas occupies a very small proportion and is too similar to the surrounding tissues.In many cases,the location of the pancreas detected by the coarse model is incomplete or completely absent.However superior the performance of the fine model is,there is an upper limit of the overall segmentation accuracy.In this paper,a pancreas segmentation method inspired by curriculum learning is proposed to help the coarse model gradually acquire the ability to identify pancreatic pixels by setting the segmentation tasks from simple to difficult, and improve the accuracy of the coarse model to detect pancreas.2.In this paper,the momentum update strategy of parameters is adopted to assist model training.The model continuously inherits the knowledge learned in the previous training stage for subsequent training tasks in the process of switching various segmentation tasks,ensuring the continuity and efficiency of knowledge transformation,and eliminating redundant training steps in the two-stage segmentation framework,greatly shortening the training time.In the national institutes of health pancreas segmentation data set,the pancreas segmentation method inspired by curriculum learning achieved high segmentation accuracy in different backbones according to the accuracy of Dice-S?rensen Coefficient.In comparison with the methods of U-shaped network as the main backbone,the method in this paper achieves the highest segmentation accuracy of 85.42%.