Brain Tumor Image Segmentation Based On Improving U-net

Brain tumor is an abnormal group of cells that grow in the brain tissue of the skull,and is one of the more common tumor diseases that endanger human health.Brain tumors are not obvious in the early stages of symptoms and are not easy to attract people’s attention,but once symptoms appear,they can lead to serious consequences.Its clinical manifestations are that it will destroy and compress normal brain tissue,cause cerebral edema,and increase intracranial pressure,which can cause sensory and motor nerve disorders,and even affect the respiratory central nervous system.The mortality rate of patients with brain tumors is extremely high and the treatment is tricky,and "early prevention,early diagnosis,and early treatment" is an important and effective means to prevent and treat tumors.At present,the diagnosis of clinical brain tumors mainly relies on MRI detection technology,and the accurate segmentation of brain tumor sites based on MRI images can help doctors find lesions in time and accurately,and achieve more accurate treatment in surgery and radiotherapy.Therefore,in the treatment of brain tumors,it is important to improve the accuracy of brain tumor image segmentation.In the current segmentation methods,manual segmentation has the disadvantages of long time,high professional requirements and poor repeatability,but the deep learning method overcomes these problems effectively.In recent years,deep learning method has been widely used in the field of medical imaging,especially the U-net network has a good performance in image segmentation.However,U-net still exists in MRI brain tumor images and does not make good use of the contextual information in the images.It is necessary to learn the characteristic information of all pixels generated by the device,and contains irrelevant areas,time consuming and wasteful.To solve these problems,this paper proposes an improved U-net.The convolution module of U-net is replaced with residual structure to prevent the gradient from disappearing and achieve full feature extraction.Add the RFB cavity module to the network to increase the sensory field and reduce the loss of detail.Adding attention gates between jump connections makes the model more focused on the areas to be segmented and improves segmentation accuracy.Experiments on the open dataset Brats2018 and Brats2019 demonstrated the effectiveness of the approach.