基于改进三基编解码U-Net的脑肿瘤分割方法OA

The study of brain tumor segmentation based on deep learning has important clinical and scientific significance.To ad-dress the shortcomings of the current brain tumor segmentation network based on deep learning in processing multimodal informa-tion,voxel imbalance,and key information loss,the paper proposes a new network model based on the Three-Base Codec U-Net(TBCU).To effectively extract detailed,global,and multi-scale features while reducing information redundancy and loss,the TBCU model is optimized from the original E1D3 U-Net.It adopts a three-encoder-path structure and optimized basic convolu-tion blocks,which include expanded convolution blocks and multi-scale residual blocks.In the decoder part,the mixed atten-tion block is introduced to integrate global and detailed information,improve attention to the lesion area,and enhance the de-coder's complexity to cope with complex information.The experimental results showed that on the BraTS 2021 dataset,the Dice scores of the TBCU model in the whole tumor,tumor core,and enhanced tumor area are 0.9132,0.9013,and 0.8913,respec-tively,which are 0.8,3.5,and 2.3 percentage points higher than the original model.The comprehensive segmentation perfor-mance is superior to that of U-Net and other algorithms,and the model also achieves favorable results in transfer experiments.The TBCU model demonstrates high segmentation accuracy and stability,providing a clearer and more accurate basis as well as stronger support for the clinical diagnosis and decision-making of brain tumors.