Research On Brain Tumor Segmentation Algorithm Based On Convolutional Neural Network

Brain tumor is a kind of tumor disease with extremely high fatality rate,which seriously threatens the life and health of patients.During the diagnosis and treatment of brain tumors,doctors usually use three-dimensional MRI images as the basis for judgment.At present,the diagnosis of brain tumors mostly relies on the manual segmentation of brain tumors by doctors.The workload is huge,and it relies heavily on the doctor’s personal experience.The subjectivity is strong and it is difficult to achieve rapid and unified precise segmentation of brain tumor tissues.Research on a fully automatic brain tumor segmentation model based on convolutional neural networks will help reduce the workload of doctors and relieve strained medical resources.Compared with manual segmentation,the fully automatic segmentation model can effectively increase the segmentation speed of brain tumor lesions and effectively improve the efficiency of disease screening,thereby saving more patients’lives.In view of the large differences in clinical images in the study of brain tumor segmentation,the diverse tumor morphology and location,and the difficulty of distinguishing the boundaries of the lesion area,in order to achieve fine segmentation of different areas of brain tumors and improve the overall segmentation accuracy of brain tumors,this paper fully investigates the convolution The application of neural network in image segmentation,especially medical image segmentation,and on this basis,researches on brain tumor image segmentation algorithms based on 2D convolutional neural network and 3D convolutional neural network have been carried out:(1)A brain tumor image segmentation model based on 2D convolutional neural network is proposed.The main design idea of the model refers to the U-Net network structure.In order to improve U-Net’s ability to segment image details,this research proposes a cascaded U-Net network model based on multi-scale dilated convolution.The model uses a multi-scale dilated convolution path as the basic feature extraction unit.A short connection is added between the two-level codec structure before and after the network.The attention module is added to the jump connection part,and the supervising of the shallow model is strengthened by introducing deep supervision,which has improved the model’s ability to segment the details of different areas of brain tumors.(2)A brain tumor image segmentation model based on 3D convolutional neural network is proposed,which strengthens the use of the three-dimensional spatial information of brain tumor MRI.The model is based on 3D U-Net,introduces a multi-depth residual structure in the coding stage of the model,adds an attention threshold module to the jump connection,and makes full use of the multi-level and multi-scale features of the network decoding stage through a multi-scale deep supervision mechanism,which improves the model’s fine segmentation ability of brain tumors.Finally,through comparative experiments,it is proved that the multi-scale dilated convolutional cascade U-Net model based on 2D convolutional neural network and the improved residual structure 3D U-Net model based on 3D convolutional neural network are compared with similar models.The segmentation model has a significant performance improvement in the segmentation of the whole tumor area,core tumor area and enhanced tumor area of the brain tumor lesion tissue,and the overall segmentation accuracy of the model has been improved to a certain extent.