Segmentation Of Multi Sequence Brain Tumors Based On Residual And Attention Mechanism

The brain is the nerve center of human body,and various brain diseases threaten human health seriously.Brain tumors continue to plague patients and doctors due to uncontrolled spontaneous reproduction.With the rapid development of medical imaging technology,MRI has become the main imaging method in medical practice due to its small damage,high resolution and three-dimensional imaging.Surgical resection is the most effective treatment in many tumor treatment schemes.To know the shape and size of tumor and segment tumor is the first step of surgical resection.Artificial cutting takes a long time,requires high professional knowledge and has poor reproducibility.Therefore,it is an urgent problem to realize automatic and accurate segmentation of brain tumors.Because brain tumor cells can reproduce autonomously and affect the surrounding normal tissue cells,the boundary of brain tumor MRI image is blurred,and the individual differences of brain tumor are large.The existing segmentation algorithms achieve better segmentation in single sequence or few sequence brain tumor segmentation,but there is still space for improvement in the segmentation of some tumor substructures.In this paper,the segmentation of brain tumor based on deep convolution network is studied.(1)To solve the problem of fuzzy tumor boundary and difficult segmentation in MRI tumor image,this paper proposes a method of tumor segmentation based on encoder and decoder with residual network.In the classical method of segmentation,the encoder structure and decoder structure are symmetrically distributed,and the same level of encoder and decoder are connected by skipping.On this basis,this paper introduces the residual network,and uses multiple residual units in series to replace the original network,to achieve the function of encoding and decoding.More local features are extracted from jump links in residual cells,and more global features are extracted from jump links between residual cells at the same level.In addition,to solve the problem of segmentation category imbalance caused by the large proportion difference between the target region and the background region in the brain tumor image,this paper combines the cross entropy loss and dice loss linearly and weights dice loss to solve the problem.This method can get more context information and improve the ability of feature extraction.Experiments on different tasks of brain tumor image segmentation in data set show the effectiveness and adaptability of the proposed algorithm.(2)This paper proposes a segmentation method of brain tumor based on attentionmechanism.In the slice of brain tumor image,the target tumor is small,and the tumor is heteromorphic,so there is more redundant information in MRI brain tumor image.In order to reduce the interference of redundant information and achieve task focus,this paper introduces attention mechanism on the basis of classical segmentation network.In this paper,in addition to the calculation of self attention,low-level semantic information and border information are embedded to achieve adaptive weighting of the input image and feature map,and guide the model to focus on the target tumor area.Experiments on different tasks of brain tumor segmentation verify the effectiveness and adaptability of the proposed algorithm.