Research On Brain Tumor Segmentation Based On Improved U-Net

With the change in people’s lifestyles in recent years,the incidence rate of brain tumors is increasing,and the treatment of brain tumors requires neurosurgery to remove the focus area,so the division of the focus boundary is the key to surgery.Usually,brain tumors are divided manually by several senior brain experts or by computer-assisted semi-automatic methods and automatic brain tumor image segmentation methods.However,the semi-automatic segmentation methods based on traditional image segmentation have low segmentation accuracy and cannot identify different types of tumors,while the automatic segmentation method U-Net is not sensitive to edge details and cannot effectively handle samples that are difficult to classify and segment,Therefore,this paper proposes an improved U-Net network,which can segment brain tumor images with higher accuracy.To design a brain tumor segmentation network with higher segmentation accuracy,this paper proposes an improved U-Net network,which is studied from four aspects:convolution module,overall architecture,loss function,and optimization mode.Aiming at the deficiency of the U-Net backbone network in extracting tumor location information,a mixed convolution group is used instead of the convolution layer to increase the network’s extraction of location information and global information by increasing the backbone network receptive field.Given the shortcomings of the U-Net network in terms of low segmentation accuracy of brain tumor images and insufficient ability of edge segmentation,the tandem architecture is used to overlay two improved U-Net,and each layer of the feature layer and the decoded image layer of the two networks are connected in series through multiple connection operations to increase the fusion and secondary extraction of features of the network and enhance the ability of segmentation of edge details.As ordinary loss functions can only focus on segmentation performance or resolution performance alone and can only assign static weights to training targets,the Focal-Dice Loss Function is proposed.By dynamically changing the Dice loss weight and cross-entropy loss weight of each target,the problem of difficulty to segment and classifying targets can be handled more effectively.Aiming at the problem that the tandem architecture increases the depth of the network and the training is easy to fall into the saddle point,a nested optimization function is proposed.The hot restart is used to make the network jump out of the local minimum value during the iteration and speed up the model convergence.In the experimental part,the segmentation effect of the improved model on the brain tumor dataset is compared with depth learning algorithms and traditional segmentation algorithms.The results show that the improved algorithm has higher segmentation accuracy than other algorithms.This paper uses a new loss function,an embedded optimization function,a new network architecture,and a hybrid expansion convolution to effectively solve the problem that the model is not sensitive to the difficult segmentation and classification samples,and is not sensitive to the edge details and global location information,thus improving the segmentation accuracy of brain tumor images.This method can speed up doctors’ diagnosis and treatment of brain tumor patients and provides the possibility for intelligent medical treatment and intelligent surgery.Figure [89] Table [12] Reference [84]...