Multi-Scale Segmentation Network Of Brain Tumor Based On U-net Block

In recent years,the segmentation of brain tumor has become a hot issue in the field of medical image research.Brain tumor is formed by the accumulation of a large number of abnormal cells in the brain,which is very harmful to human health.Gliomas are brain tumors originating from the glial cells in the central nervous system.They are the most common tumors in brain cancer with a high mortality rate.Gliomas can be divided into low grade gliomas(LGG)and high grade gliomas(HGG).Although the therapeutic technique has improved,the prognosis is so poor that the survival duration of patients with HGG is less than two years.Compared with HGG,the therapeutic effects and prognosis of LGG are more effective.Therefore,in order to improve patient survival duration,it is important to segment tumors accurately and develop appropriate therapeutic plans.With the development of diagnostic imaging technology,there are many imaging methods of brain tumors.Magnetic resonance imaging(MRI)has high soft tissue resolution,and no radiation damage,provides a wide range of physiological significance comparison,and the characteristics from MRI can greatly improve the accuracy of evaluating the malignant degree of brain tumors,so it is extensively employed in diagnosing brain and nervous system abnormalities.However,brain tumor segmentation remains a challenging task,because differentiating brain tumors from normal tissues is difficult,tumor boundaries are often ambiguous and there is a high degree of variability in the shape,location and extent of the patient.It is therefore desired to devise effective image segmentation architectures.In the past few decades,many algorithms for automatic segmentation of brain tumors have been proposed.Methods based on deep learning have achieved favorable performance for brain tumor segmentation.In this paper,we propose a multi-scale brain tumor(MRI)segmentation network based on u-net block,which mainly includes two stages:encoding and decoding.In the encoding stage,convolution layer,u-net block and multi-pooling block are used to capture the long-distance spatial information under different resolutions.In the decoding stage,the feature maps of different resolutions are respectively upsampled to extract and utilize sufficient features.In the u-net block,we use the Xception block and 3D depthwise separable convolution to protect the integrity of information and reduce the computation of the network.The shallow convolution layer can learn the local and low-dimensional features of the image,while the deep convolution layer can learn the global and high-dimensional features.These features can be directly transferred to the decoding layer by skip connection,which can provide much detailed and sufficient information for the network.In the multi-pooling block,we use different pooling kernels to obtain the feature information of different scales.In order to verify the performance of the network,we performed a large number of contrast and elimination experiments on the BraTS 2015 dataset,and we obtained dice scores of 0.85,0.72 and 0.61 for the whole tumor,tumor core,and enhancing tumor,respectively.The segmentation performance of our network was competitive compared to other state-of-the-art methods.