Research On MRI Brain Tumor Segmentation Algorithm Based On Attention Mechanism And ResU-Net

Brain tumors are defined as abnormal or uncontrolled cells.According to cancer reports published by the World Health Organization(WHO),brain cancer accounts for less than 2%of human cancer cases.However,once detected,brain tumors can be a serious threat to a patient’s life.Brain tumors can be classified as cancerous(malignant)or non-cancerous(benign).The difference between benign and malignant tumors is that benign tumors do not spread to other tissues and can be surgically removed.Radiologists use Magnetic Resonance Imaging(MRI)and Computed Tomography(CT)scan images to classify diseases such as tumors.However,manual segmentation of brain tumors by specialists is time-consuming and prone to small errors.And segmentation techniques also exhibit inter-patient variability,making it difficult to reproduce results.Therefore,the use of computer methods to assist in brain tumor segmentation is playing an increasingly important role in the modern medical field.In recent years,Deep Convolutional Neural Network has been widely applied to medical images and achieved good segmentation results.In the brain tumor segmentation task,the residual U-Net network cannot extract global features and skip the connection to directly splice the low-level features in the encoder with the corresponding decoder features for feature fusion,although it uses the residual module to alleviate the problem of network gradient disappearance.But these low-level feature information contains a lot of redundant information.Secondly,convolution has localized features,so the convolution-based methods are not able to model long-distance dependencies well.Therefore,for the above problems,we use the expanded feature pyramid model,CBAM attention mechanism and Transformer self-attention mechanism to solve them respectively.The expanded feature pyramid model can extract feature maps of different sizes;the CBAM dual-attention mechanism effectively improves the utilization of useful information for segmentation tasks and suppresses useless information;the Transformer mechanism is proposed based on the self-attention mechanism,which completely eliminates convolution and can handle global information well.The works done in this thesis are.(1)In this thesis,the brain tumor dataset of Medical Image Computing and Computer Aided Intervention(MICCIA)multimodal brain tumor segmentation challenge(BraTS)is used as the dataset for experiments,and a novel three-dimensional MRI brain tumor image segmentation.The method inspired by the attention mechanism is proposed,which is based on the dual attention mechanism CBAM module and the expanded feature pyramid module modified by the residual U-Net,i.e.RDAU-Net network model.Due to the importance of multi-scale features,this thesis added DA blocks to expand the perceptual field and obtain image information of different sizes.In this thesis,We insert a CBAM block after each skip connection layer to improve the extraction of channel information and spatial information and reduce the redundant information of low-level features.For the experimental results,this thesis uses DSC(Dice Similarity Coefficient),Hausdorff distance,sensitivity,and specificity metrics to evaluate and achieve superior metric scores.And by comparing with other state-of-the-art methods,the method proposed by this thesis achieved the best results.This proves that our method can effectively improve the segmentation of brain tumors.(2)This thesis also proposes a new method for MRI brain tumor segmentation that uses the Transformer structure,a dual-attention mechanism CBAM to improve on the residual U-Net network.The brain tumor datasets from the BraTS 2018 and 2019 were used as the datasets for the experiments.In this thesis,the experimental results were evaluated using Dice similarity coefficient,Hausdorff distance,sensitivity and specificity metrics for the whole tumor region(WT),enhanced tumor core(ET)and tumor core(TC).The proposed method was validated using the 2019 Multimodal Brain Tumor Segmentation Challenge dataset.The evalution results for different segmented regions on the validation set showed that Dice scores were 0.9101 on WT,0.9052 on TC and 0.8282 on ET,Hasdorff distance scores were 5.6428 on WT,6.0206 on TC and 2.9892 on ET.The specificity scores were 0.99 on WT,0.9972 on TC and 0.9948 on ET,and sensitivity scores were 0.9252 on WT,0.9004 on TC and 0.8686 on ET.The specificity scores were 0.99 on WT,0.9972 on TC and 0.9948 on ET,and the sensitivity scores were 0.9252 on WT,0.9004 on TC and 0.8686 on ET,respectively.In summary,the method proposed by this thesis achieved superior index scores.And by comparing with the RDAU-Net method,the method achieved good results for segmentation on WT,proving that the method in this thesis can effectively improve the segmentation of brain tumors.