Research And Application Of Convolutional Neural Network Based Segmentation And Classification Of Brain Tumors

Brain tumors pose a significant threat to the body and are one of the leading causes of cancer-related deaths.Magnetic resonance imaging(MRI)is a commonly used clinical method for detecting brain tumors,which provides multimodal images through various imaging techniques and plays a great role in the medical field.In recent years,with the rapid development of computer technology in image processing,deep learning is often combined with medical research to use image processing technology for segmentation and classification of brain tumor images to assist medical staff in tumor diagnosis.Thesis mainly studies the automatic segmentation and classification of brain tumors based on Convolutional Neural Networks(CNN)technology.The specific situation is as follows:1.Regarding brain tumor segmentation,a U-shaped network design with a fused dual-attention mechanism is proposed to address the issue of insufficient feature extraction of the convolutional neural network.First,residual modules are introduced in the U-Net network to transmit shallow information to the later network layers to solve the problem of shallow features being lost during training.Then,the Dynamic Rectified Linear Unit(Dy-Re LU)activation function is used to replace the Rectified Linear Unit(Re LU)activation function in the residual blocks to dynamically adjust the activation function with input information and improve the model feature representation ability.Next,the CBAM mechanism is introduced to make the network pay more attention to certain feature layers and spatial regions,suppress non-lesion areas features,and improve tumor segmentation accuracy.At the same time,the crossentropy loss and the sum of generalized dice loss are used as a hybrid loss function during training to alleviate the problem of class imbalance to some extent.2.Regarding brain tumor classification,a brain tumor classification method based on SE-Net is proposed to address the problem of information loss in deep networks.First,after batch normalization and feature fusion,the Swish activation function is used instead of the Re LU activation function to improve the model accuracy in deep networks.Secondly,the ECA and improved BAM attention modules are added after the first and second convolutional layers,respectively,to extract features from both spatial and channel directions to fully focus on the target features.Finally,the global maximum pooling is added to the SE attention module to use the dual-channel pooling layer to extract effective features,suppress invalid features,and improve the model expression ability.In addition,to verify the effectiveness of the improved modules on the network model,ablation experiments and comparative experiments are conducted,and the heat maps generated by the gradient-weighted class activation mapping method are used to explain the attention modules,fully demonstrating the feasibility of the improved model.3.To help medical staff better diagnose patients,system based on the brain tumor classification and segmentation model are developed to facilitate medical staff in providing better treatment for brain tumor patients.