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

Brain tumor is one of the most aggressive diseases.It not only shorten life of patients,but also is often accompanied by pain.Accurate and timely diagnosis of brain tumor type and its removal is the key to improve the survival rate of patients.One of the most important tools for discovering brain tumors is magnetic resonance imaging(MRI).MRI images of brain tumors are irregularly shaped and unevenly distributed,so it require expert to identify and segment them.In recent years,with the rapid development of deep learning,especially the advancement of computer vision technology,deep learning has become the main way to solve image tasks.Therefore,this article mainly studies the classification and segmentation of brain tumor images based on convolutional neural networks.This paper proposes a model called Global Average Pooled Residual Network(GResNet)for brain tumor classification.The model has the following characteristics: 1)Applying the well-established CNN architecture in the field of deep learning named ResNet34 for the classification task.2)To reduce the number of parameters and avoid overfitting,we use the global average pooling layer instead of the flattened layer for classification.3)In order to be able to fuse the low-level and high-level features of the network to improve the classification accuracy,we concatenate the feature vectors of different layers.4)We define a loss function,which is sum of the interval loss and the cross entropy loss.The total loss increases the penalty for misclassification.Experiment show that our model achieves a classification accuracy of 95.00%.This paper proposes a model called self-calibrated four-path 3D U-Net(3D SCFU-Net)for the segmentation of complete brain tumor regions in 3D brain tumor MRI images.The segmentation work has the following aspects.1)3D self-calibrated convolution is proposed and applied in 3D U-NET model.In order to improve the fields-of-view of convolution and the learning ability of the convolution layer,this paper modified the convolution layer of the encoder part into self-calibrated convolution based on the use of 3D U-NET model.2)A four-path attention mechanism is proposed for 3D U-NET model.Traditional triplet Attention mechanism takes advantage of channel and spatial context information,this article carries on the modification,joined a slice path to capture 3D image information.According to the structural characteristics,it is called the four-path attention mechanism.It can capture the width,height,channel,section four dimensions of mutual information to calculate the weight.The experiment show that the 3D SCF-U-NET segmentation model achieves the comparable results with a DICE coefficient of 85.1% and a sensitivity of90.7% on the Bra TS2019 data set.