Research On Brain Tumor Image Segmentation Based On Dense Residual Block And Dual Attention Mechanism

Gliomas are the malignant tumors with the highest mortality rate.Because brain tumors are infiltrative and quickly spread to various areas of the brain and can be lifethreatening in severe cases,early diagnosis and accurate segmentation of brain tumors are the basis for subsequent treatment decisions.Magnetic resonance imaging(MRI)is a commonly used imaging technology in clinical examination,which has the advantages of non-invasive,high soft tissue resolution,no radiation to the human body,etc.At present,brain tumor images generally are segmented by physicians,but manual segmentation is laborious and time-consuming,which depends more on the professional level of physicians.Besides,it is affected by subjective factors.In recent years,brain tumor image segmentation based on deep learning methods has become a hotspot in the field of medical images.However,the location and morphology of brain tumors are variable,and there is a class imbalance between tumor sub-regions,so most methods are not ideal for segmentation in small brain tumor sub-regions.To solve the above problems,this paper studies in-depth on the deep learning technology in brain tumor segmentation to achieve higher segmentation accuracy.The main works of this paper are as follows:(1)Based on the problems of low contrast of brain tumor images and insufficient extraction of deep features,the U-Net structure is improved by adding multi-scale feature extraction and dense residual block.This paper designs a brain tumor segmentation model based on multi-scale dense residual block.The multi-scale feature extraction block can fuse feature maps of different receptive fields to capture more contextual information.The dense residual block transmits the features in cascading,and the output features of all the previous layers are fused with the current layers,which improves the transmission efficiency of the features between the layers and reduces the amount of network parameters.The experimental comparative analysis shows that the multi-scale dense residual block can improve the accuracy of brain tumor segmentation.(2)To solve the problems of imbalance of brain tumor sub-regions and low accuracy of tumor sub-region segmentation,attention mechanism is introduced based on the above model.This paper designs a brain tumor segmentation model that combines dense residual block with dual attention mechanism.The dual attention module exerts attention on the feature map based on the two dimensions of channel and spatial to enhance the expression ability of tumor features.The dual attention module is embedded in the jump connection,which can avoid the semantic gap caused by the direct fusion of low-level feature maps and high-level feature maps.The experimental comparative analysis verifies that the dual attention mechanism can improve the performance of small brain tumor sub-regions and enhance the detailed characteristics of tumor contours.(3)To reduce the impact of data imbalance on the accuracy of the model,this paper uses a mixed loss function composed of cross-entropy loss and dice loss.The cross-entropy loss function can converge quickly,but it is difficult to process brain tumors images with complex structures.The dice loss function can effectively deal with the imbalance of data in medical images and improve the training performance.(4)There is a magnetic field shift phenomenon during the brain MRI imaging,which can lead to uneven brightness of the image.This paper uses the NI4ITK method to correct bias field of the image,and expands the sample size by data augmentation to improve the generalization ability of the model.(5)Based on the BraTS2018 dataset,we perform ablation experiments on the proposed model.The model is evaluated by multiple evaluation indicators.The model in this paper is compared with other state-of-the-art methods to verify the performance of the model.