Research On 2D/3D MRI Image Segmentation Of Brain Tumor Based On Fully Convolution Neural Network

Biomedical images,like computed tomography(CT),magnetic resonance imaging(MRI),ultrasound imaging(US),nuclear medicine imaging,and other medical imaging technologies,are used a lot in clinical diagnosis and treatment today.Among them,magnetic resonance imaging,as a non-invasive soft tissue contrast imaging method,can provide valuable information about the shape,size and location of brain tumor.It is the main method used to examine patients with brain tumor,and plays an important role in the segmentation task of brain tumor.However,due to the complex and changeable shape of brain tumor itself,fuzzy boundary,low contrast and complex sample gradient,the high-precision segmentation of brain tumor MRI images is very challenging.At present,the manual segmentation mainly relies on professional doctors,which is time-consuming and poor repeatability.In recent years,the U-Net model based on fully convolutional neural network architecture has been widely applied in the field of medical image segmentation and has made significant progress.This thesis aims to investigate 2D medical image segmentation and 3D medical image segmentation networks from the perspective of enhancing the U-Net model and enhancing segmentation accuracy.Experiments are conducted on 2D datasets of skin lesion segmentation and 3D datasets of brain tumor MRI images in an effort to improve segmentation accuracy and the U-Net model.The main research contents are:(1)For 2D medical images segmentation model,we propose an improved model based on 2D U-Net,i.e.,CSPU-Net,by combining low-resolution information and high-resolution information to achieve high-precision 2D medical image segmentation.CSPU-Net incorporates two types of cross-stage local network structures(Cross Stage Partial Module,CSP)to extract image features,and segment the gradient stream by split-merge strategy to alleviate the drawback of using explicit feature mapping replication for connectivity,reduce information loss,and enhance the model feature learning capability.In the brain tumor segmentation task,two loss functions,GDL(General Dice Loss)and WCE(Weighted Cross Entropy),are combined to solve the training sample category imbalance problem and help the model achieved better segmentation results.Finally,experiments are conducted using brain tumor segmentation dataset and skin lesion segmentation dataset respectively,and the experimental results demonstrate that the CSPU-Net proposed in this thesis can effectively improve the 2D medical image segmentation accuracy.(2)In order to compensate the information loss of 3D brain tumor MRI images processed into 2D slices and retain more brain tumor information,a channel pyramidal segmentation attention brain tumor segmentation model is proposed.Improving 3D U-Net using cross-stage local network structure will cause great machine burden due to the large 3D model parameters and computation.In contrast,the attention mechanism is more lightweight,which can make the model more focused on segmentation content and capture a large amount of contextual information and multi-scale features with fewer parameters and computation,effectively improving the performance of the model.In this thesis,we adopt a new channel pyramid segmentation attention module and add it to the lateral connection with 3D U-Net model to help the model capture important information to improve the segmentation accuracy through multi-scale channel attention extraction.Extensive experiments show that the new 3D PSAUNet model embedded with the channel pyramidal segmentation attention module can achieve higher segmentation accuracy.