Research On Segmentation Algorithm Of Brain Tumor MRI Image Based On U-Net

Glioma is a kind of primary brain tumor with very strong infiltration and high mortality rate,which is seriously threatening people’s life and health.Therefore,the early diagnosis of brain tumor is of great significance to improve the possibility of treatment,formulate surgical plan and improve the survival rate of patients.Brain tumor segmentation is a key task in the early diagnosis of tumors.It usually requires doctors to manually segment them based on experience from a large number of magnetic resonance imaging(MRI)images of the patient.However,the contrast between the soft tissues of the brain MRI image is low,and it is difficult to distinguish the tumor from the surrounding healthy tissues.Therefore,the manual segmentation method is time-consuming and laborious,and the segmentation results are also highly subjective.With the development of deep learning technology,convolutional neural network(CNN)algorithms are widely used in image processing.Among them,among which U-Net algorithm performs best in medical image segmentation task,but classical U-Net algorithm still has problems in brain tumor segmentation,such as undersegmentation and oversegmentation of tumor regions.Combined with the characteristics of brain tumor image,this paper studies the segmentation algorithm of brain tumor MRI image based on U-Net,and verifies it on public data sets.The researches work of this paper are as follows:(1)Aiming at the problem of low precision in tumor boundary segmentation of classical U-Net segmentation networks,based on U-Net an improved multi-scale feature extraction model called DCU-Net(Multi-scale U-Net with Dilated Convolution)is proposed.Firstly,MRI brain tumor images are clipped and preprocessed to reduce the input of background pixels to alleviate the class imbalance problem.Then,design the multi-scale spatial pyramid pool module to optimize part of the max-pooling layer in the classic U-Net model,so that the network expands the feature receptive field while maintaining the image resolution.The feature receptive field is expanded while the image resolution is maintained.Finally,the dilated convolution residual block is introduced through the Add operation to optimize the skip connection in the training network,and the features from the low level of the down sampling path are fused to improve the ability of the network to identify the details of the tumor.The experimental results show that the DCU-Net model can achieve the fusion and extraction of multi-scale brain tumor features,and effectively improve the segmentation accuracy of brain tumor region boundaries.(2)Aiming at the problem of fuzzy segmentation of brain tumor subregions in the classical U-Net model,based on U-Net an improved integrating feature attention mechanism model called DA-UNet(Dense Connected U-Net with Attention Gate)is proposed.Firstly,several dense layers are designed in the down sampling path of DA-UNet model.While improving the learning of detailed features,the model can avoid overfitting and gradient disappearance problems in the model training process to some extent.Secondly,the attention mechanism is introduced into the down sampling path of the model,which focused on improving the learning weight of the model to the brain tumor regions and inhibiting the learning weight of the model to the background features.Experiments show that,on the one hand,the DA-UNet model improves the segmentation accuracy of brain tumor internal sub-regions on multiple indicators,and on the other hand reduces the model complexity.