BiUNet:基于双层路由注意力的轻量化医学分割网络

BiUNet:Lightweight Medical Segmentation Network Based on Bi⁃Level Routing Attention

针对视觉Transformer骨干提取网络计算开销大,模型训练缓慢的问题,同时为了进一步提升Transformer结构在医学图像领域的分割性能,提出一种名为BiUNet的新型轻量级U型架构的医学图像分割网络。将输入医学图像整切成若干图像块后,送入一种基于双层路由动态稀疏注意力机制的BiFormer转换器中,通过组合下采样和特定块数的BiFormer模块,构建多级金字塔结构实现特征提取。随后通过组合上采样和卷积模块,相应构建多级金字塔结构进行特征解码,进而实现像素级语义分割。该模型在3个医学数据集上依次取得了90.2%, 93.7%和85.6%的mIoU值以及5.55 G的Flops和28.10 M的参数量。结果表明,BiUNet能够以轻量化的效果有效提升医学图像分割的精度。

To solve the problems of high computing cost and slow model training in vision Transformer backbone extraction network,and to further improve the performance of Transformer structure in the field of medical image segmentation,a new lightweight U-architecture medical image segmentation network named BiUNet was proposed.The input medical image was cut into several blocks,and then the blocks were fed into the BiFormer based on the dynamic sparse attention mechanism of Bi-level routing.By com⁃bining downsampling and BiFormer modules with a specific number of blocks,a multi-level pyramid struc⁃ture was constructed to achieve feature extraction.Subsequently,the feature map output from the encoder was decoded by a multi-level pyramid structure which was constructed by combining the upsampling and convolution modules, and pixel-level semantic segmentation was realized. The model achieved 90. 2%,93. 7% and 85. 6% mIoU values as well as 5. 55 G Flops and 28. 10 M parameters on the three medicaldatasets sequentially. The results show that BiUNet can effectively improve the accuracy of medicalimage segmentation with a lightweight effect.