Brain Tumor MRI Image Segmentation Based On Improved U-Net

Brain tumor is a common tumor disease,which shows an aggressive growth pattern and seriously endangers the physical and mental health of patients.Medical images can clearly present the detailed situation of the lesion,providing a basis for doctors to understand the condition,diagnose and treat it.As a part of computer aided diagnosis system,medical image segmentation of brain tumor is to separate brain tumor from normal brain tissue in the image,and then obtain the segmentation effect image of brain tumor.With the rise of deep learning,a large number of scholars have invested in the field of brain tumor medical image segmentation,proposed many network frameworks,and achieved good segmentation results.Among them,the U-Net segmentation network framework is the most widely used.However,there are also some problems in the process of brain tumor segmentation,such as noise,brain tumor MRI image sample imbalance affecting segmentation accuracy,and the current depth learning network is complex and difficult to meet real-time requirements.Therefore,this paper proposes two improved brain tumor segmentation algorithms based on the U-Net network framework for the problems faced in the process of brain tumor MRI image segmentation.(1)In order to improve the segmentation accuracy of brain tumor regions,a brain tumor segmentation method based on improved U-Net was proposed.An improved residual block is introduced into the network to expand the Receptive field and solve the problem that the gradient may disappear,so as to speed up the convergence of the network;introducing attention mechanism in the long connection of encoding and decoding paths allows the network to autonomously learn to distinguish important and secondary information,enhance meaningful information,suppress noise and irrelevant information,avoid information redundancy,and improve the network’s expressive ability;because there is sample imbalance in brain tumor MRI images,which seriously affects the segmentation accuracy of brain tumor,a mixed loss function is designed to solve this problem,which solves the sample imbalance problem and improves the segmentation accuracy of brain tumor;in the evaluation index of Dice coefficient,the segmentation accuracy of the whole tumor region,the core region of brain tumor and the enhanced region of brain tumor reached91.02%,82.27% and 78.73%,respectively,2.88%,6.46% and 5.58% higher than that of UNet.(2)Aiming at the problem that U-Net and U-Net++lack the ability to explore information from the full scale,and cannot determine the organ to be segmented,the location and boundary of the focus,a brain tumor segmentation method based on improved U-Net3+is proposed.The proposed lightweight feature extraction module has been added to the network to extract features from multi-scale directions;The residual CBAM attention mechanism is used to solve the redundancy problem caused by the network in the process of multi-level feature fusion,so that the model focuses on tumor regions and further improves the segmentation accuracy of tumors.The Dice values in the whole tumor region,brain tumor core region and brain tumor enhancement region reach 91.85%,83.20% and 80.72%.At the same time,the network is lightweight in terms of reducing network hierarchy and reducing the number of channels to reduce network complexity.The Params and FLOPs are6.74 M and 43.79 GMAc,respectively.Compared to U-Net3+networks,the Params and FLOPs are reduced by 20.29 M and 33.84 GMAc,respectively.