基于自分块轻量化Transformer的医学图像分割网络

Medical image segmentation network based onself-partitioning lightweight Transformer

传统的医学图像分割网络参数量大、运算速度缓慢,不能有效应用于即时检测技术。为解决该问题,提出了一种轻量化的医学图像分割网络SPTFormer。该网络构建了自分块Transformer模块,其通过自适应的分块策略重塑特征图,利用并行化计算在提高Attention运算速度的同时关注局部细节特征;还构建了SR-CNN模块,使用平移加复位操作提升对局部空间信息的捕获能力。在ISIC 2018、BUSI、CVC-ClinicDB和2018 data science bowl四个模态数据集上进行了实验,与基于Transformer的TransUNet网络相比,所提网络SPTFormer精度分别提高了4.28%、3.74%、6.50%和1.16%,GPU计算耗时降低58%。该网络在医学图像分割应用中具有更优的性能,可以良好地兼顾网络精度和复杂度,为计算机即时辅助诊断提供了新方案。

The traditional medical image segmentation network has a large number of parameters and slow computing speed,and cannot applies effectively to the real-time detection technology.To address this issue,this paper proposed a lightweight medical image segmentation network called SPTFormer.Firstly,this network constructed a self-blocking Transformer module,which reshaped the feature map through an adaptive blocking strategy and utilized parallel computing to improve the attention operation speed while paying attention to local detail features.Secondly,this network constructed an SR-CNN module,which used the shift-restored operation to improve the ability to capture local spatial information.By experimenting on ISIC 2018,BUSI,CVC-ClinicDB and 2018 data science bowl,compared with the TransUNet model based on Transformer,the accuracy of the proposed network improves by 4.28%,3.74%,6.50%,and 1.16%,respectively,the GPU computation time reduces by 58%.The proposed network has better performance in medical image segmentation applications,which can well balance the network accuracy and complexity,and provides a new solution for real-time computer-aided diagnosis.