Research On Brain Tumor MRI Image Segmentation Based On Attention Res-Unet

With the development of society and technology,medical imaging technology has been widely used in various medical fields and is an important auxiliary tool for doctors to diagnose patients’ conditions.Nowadays,brain tumor disease has become a modern disease that seriously affects human health due to its high morbidity and mortality rate.Therefore,accurate medical imaging aids diagnosis can help doctors provide timely treatment to patients and thus save their lives.Magnetic Resonance Imaging(MRI),as a non-invasive brain tumor imaging technique,is the main imaging method for diagnosing and treating brain tumors.The tumor image will be better assisted by further segmentation operation to complete the final diagnosis,which is currently a time-consuming,laborious and high error rate manual segmentation.With the application of deep learning technology with its unique characteristics in the field of medical image segmentation,the brain tumor segmentation method combined with convolutional neural network(CNN)can effectively improve the quality of tumor image segmentation.However,brain tumors have large shape differences,uneven distribution of location and size,complex boundaries and other factors,and the sample size of medical images is small,so it is very difficult to perform brain tumor semantic segmentation.To improve the accuracy of brain tumor MRI image segmentation,this paper further investigates the MRI image segmentation technique based on deep learning and proposes two improved Res-Unet brain tumor MRI image segmentation methods.(1)To the problems of missing feature information and low segmentation accuracy of full convolutional neural network in medical image segmentation,a brain tumor segmentation method integrating attention mechanism and Res-Unet network model is proposed.In the neural network design,the deep Res-Unet network is constructed by adding the depth residual module to make the number of convolutional layers of the network reach 104 layers to compensate for the defects that the U-Net network is not deep enough and the representation of features is not accurate enough,and to improve the segmentation accuracy and efficiency of 2D tumor MRI images.At the same time,the dropout layer is added after each convolutional block stack to reduce the overfitting of training.Finally,the Squeeze and Excitation(SE)attention mechanism is introduced to make full use of the contextual information of brain tumor MRI images to improve the segmentation accuracy of the model.(2)When segmenting brain tumor images,there are often problems such as inconspicuous boundary detail features,loss of internal texture features,poor segmentation results when the texture shapes of different case data in the images differ greatly,and the network cannot establish remote dependencies and global contextual connections due to the limitation of perceptual fields in convolutional operations.In this study,we use 3D data,add the Transformer module to the original Res-Unet network structure,combine the Self-Attention mechanism and convolution,and use the convolutional layer to extract enough local features while capturing long-range and global features with the help of Self-Attention.In which,the tokenized image blocks of the convolutional neural network feature maps are encoded by the Transformer as input sequences for extracting global context,and then the decoder upsamples the encoded features and finally combines them with the high-resolution CNN feature maps to achieve accurate information localization.Compared with the network without adding the self-attention mechanism,its segmentation results are significantly improved in terms of detail performance,and its segmentation yields a finer internal texture of the brain tumor.