Medical Image Segmetation Algorithm Based On Attention Mechanism And Multi-modality U-Net Neural Network

Medical image segmentation is an important research direction in the field of computer vision,as it can help doctors diagnose diseases more accurately and timely,improve the effectiveness of diagnosis and treatment,and increase patient survival rates.Brain tumors are a common neurological disorder with a high incidence and diagnostic difficulty.Accurately and quickly locating the location and type of brain tumors is crucial for the formulation of diagnostic and treatment plans and prognosis evaluation.In recent years,with the development of deep learning technology,multi-modal brain tumor segmentation based on convolutional neural networks has become a hot research topic.However,due to the significant heterogeneity in the shape,location,and size of brain tumors in different individuals,and limitations in network complexity,the accuracy and speed of existing brain tumor segmentation algorithms still need improvement.This paper proposes three brain tumor segmentation algorithms based on deep convolutional neural networks and multi-modal magnetic resonance imaging(MRI),and evaluates their performance on public datasets.The main research contents of this article are as follows:(1)The PSA-UNet algorithm based on pyramid split attention and 2D U-Net for MRI brain tumor segmentation is proposed.PSA-UNet proposes an improved pyramid split attention module,which is concatenated with the convolutional blocks in the 2D U-Net network to obtain multi-scale information of brain images.Instance normalization is used to process the data,reducing interference between different brain tumor patient data and improving the stability of the model.Focal loss and multi-class Dice loss functions are used to reduce the influence of pixel class imbalance in brain tumor patient data on the segmentation results.The PSA-UNet algorithm is trained and tested on the Bra TS2020 dataset,which is evaluated mainly through the scores of evaluation metrics and visualization results.The results show that the Dice coefficients of the whole tumor and core area in PSA-UNet are 0.901 and 0.804,respectively,and the Hausdorff95 distance of the core area is 10.518.Moreover,the algorithm can also segment brain tumors with diverse morphologies and small lesion pixel proportions effectively.Therefore,the PSA-UNet algorithm proposed in this paper has certain advantages in the performance of brain tumor segmentation.(2)The 3D-CAS-UNet algorithm based on hybrid attention and 3D UNet for MRI brain tumor segmentation is proposed.3D-CAS-UNet introduces improved 3D Self Attention and 3D Coordinate Attention modules at the skip connections of the 3D UNet network,constructing 3D-SA-UNet and 3D-CA-UNet subnetworks to extract tumor contours and detail information more accurately.Furthermore,3D-CAS-UNet uses residual blocks and group normalization to improve model stability and reduce bias between different sample distributions,effectively preserving the original features.The proposed 3D-CAS-UNet algorithm is trained and tested on the Bra TS2020 and Bra TS2018 datasets.Results show that the Dice coefficient of the whole tumor segmentation for 3D-SA-UNet and 3D-CA-UNet are 0.900 and 0.898,0.899 and 0.903 respectively on the two datasets.The two models still achieve good segmentation results in cases of disconnected brain tumor areas and small lesions.Therefore,the proposed algorithm in this chapter demonstrates significant advantages in segmenting the entire brain tumor and exhibits a certain degree of generalization ability.(3)The lightweight DAS-CA-UNet algorithm based on atrous spatial pyramid and3D-CA-UNet for MRI brain tumor segmentation is proposed.Firstly,different dilated rates of the spatial pyramid pooling module are introduced at the bottom layer of the network to extract features at different scales.Secondly,3D depthwise separable convolutions are used to replace standard convolutions in all residual modules to reduce model parameters and accelerate inference speed.Finally,the proposed DAS-CA-UNet algorithm is evaluated on the Bra TS2020 dataset.The experimental results demonstrate that the use of spatial pyramid pooling and 3D depthwise separable convolution can achieve lightweight performance while maintaining model accuracy.Specifically,the DAS-CA-UNet achieves a Dice coefficient of 0.896 and 0.727 for the entire tumor and enhanced region,respectively,and a Hausdorff95 distance of 6.538 for the core region.Notably,the DAS-CA-UNet still achieves good segmentation performance in the presence of blurred boundaries and irregular shapes and sizes in the enhanced region of brain tumors.These results suggest that the proposed DAS-CA-UNet algorithm provides good brain tumor segmentation performance while achieving lightweight performance.In summary,this article proposes three brain tumor segmentation algorithms,namely PSA-UNet,3D-CAS-UNet,and DAS-CA-UNet,based on deep convolutional neural networks and multi-modality MRI.The performance of these algorithms is evaluated on the Bra TS2018 and Bra TS2020 datasets.The experimental results demonstrate that the proposed algorithms have good brain tumor segmentation performance and lightweight advantages,and can to some extent address typical segmentation problems such as small lesions,non-connected regions,blurred boundaries,and irregular shapes.Therefore,the proposed algorithms have good research significance and application value in the fields of machine learning theory and clinical medicine.