Multi-modal MRI Brain Tumor Segmentation Based On The Combination Of U-Net And Transformer

Brain tumors are essentially the growth of irregular cells in the body’s central nervous system.Among the brain tumors,the diagnosis and treatment of primary brain tumors and central nervous system malignancies are the most difficult.Glioma,which accounts for approximately forty-five percent of brain tumors,is one of the most commonly seen primary brain tumors and poses a serious threat to patients’ lives.Magnetic resonance imaging(MRI)with multimodal protocols has become the most mainstream method for acquiring brain tumors today,and therefore segmentation of brain tumor MRI medical images has become a necessary prerequisite for brain tumor prediction assessment and treatment planning.However,manual medical image segmentation of brain tumors is a highly technical and time-consuming task,and it is difficult to obtain commonalities among different brain tumors by comparison because of their differences in size,shape and location.Coupled with the ambiguity of lesion boundaries in medical images,this remains a huge obstacle for accurate brain tumor segmentation.In recent years,rapidly developing deep learning(DL)algorithms have achieved great success in image classification and have also become the most popular method for medical image segmentation of brain tumors today.The U-Net architecture has been widely applied to medical image segmentation,and has become the most popular model for brain tumor segmentation.This model takes medical images as input directly and uses an encoder-decoder structure with jump connections to enhance detail retention for more accurate and efficient feature extraction.In this thesis,two improved two-dimensional U-Net variant networks are proposed to solve the problems of brain tumor segmentation.1.Considering the importance of local and global features in medical image segmentation,a 2D deep residual U-Net(2DRMUnet)with Transformer architecture is proposed in this thesis.In the encoder this thesis uses stacked bottleneck residual blocks and adds dropout after each convolutional block stack in the encoder and decoder.In the bridge the Transformer module is used for global feature modeling.A cascade connection from the encoder output after the Transformer block is added to the decoder.This thesis conducts extensive experiments and evaluates the proposed network on the Bra TS2020 and Bra TS2021 datasets.The comparative ablation experiments show that this thesis achieves a high-performance,low-computation brain tumor segmentation architecture based on a two-dimensional convolutional network compared to a three-dimensional network.2.In the brain tumor segmentation task,all three tumor regions are irregular and variable,with large differences in their areas and shapes,which are not conducive to multi-scale feature extraction.Moreover,the accurate segmentation of the boundaries of the three tumor regions is also crucial,which often requires high-resolution feature extraction,yet the input feature maps of the Transformer module as a bridge have been downsampled to very low resolution.To remedy the above problem,a deformable residual module(DRM)and multi-scale Transformer aggregation U-Net variant network(2DRMUnet)are proposed in this thesis.DRM is added to the jump connection,which enables the convolution operation to automatically change the size of the perceptual field for different tumor regions as a way to obtain shape perceptual feature information.Multi-scale Transformer aggregation is used to model global top-down information for feature maps of different resolutions.The performance of 2DRMUnet in this experiment has an all-around improvement over 2DRTUnet,which proves the effectiveness of DRM and multi-scale Transformer aggregation.