Feng Xiaomeng, Wang Taiping, Yang Xiaohang, Zhang Minfei, Guo Wanpeng, Wang Weina
Department of Mathematics, School of Sciences, Hangzhou Dianzi University, Hangzhou 310018, China.
Hangzhou Medipath Intelligent Technology Co., Ltd, Hangzhou, China.
Math Biosci Eng. 2023 Jan;20(1):128-144. doi: 10.3934/mbe.2023007. Epub 2022 Sep 29.
Convolutional Neural Network (CNN) plays a vital role in the development of computer vision applications. The depth neural network composed of U-shaped structures and jump connections is widely used in various medical image tasks. Recently, based on the self-attention mechanism, the Transformer structure has made great progress and tends to replace CNN, and it has great advantages in understanding global information. In this paper, the ConvWin Transformer structure is proposed, which refers to the W-MSA structure in Swin and combines with the convolution. It can not only accelerate the convergence speed, but also enrich the information exchange between patches and improve the understanding of local information. Then, it is integrated with UNet, a U-shaped architecture commonly used in medical image segmentation, to form a structure called ConvWin-UNet. Meanwhile, this paper improves the patch expanding layer to perform the upsampling operation. The experimental results on the Hubmap datasets and synapse multi-organ segmentation dataset indicate that the proposed ConvWin-UNet structure achieves excellent results. Partial code and models of this work are available at https://github.com/xmFeng-hdu/ConvWin-UNet.
卷积神经网络（CNN）在计算机视觉应用的发展中起着至关重要的作用。由U形结构和跳跃连接组成的深度神经网络在各种医学图像任务中被广泛使用。最近，基于自注意力机制，Transformer结构取得了很大进展并倾向于取代CNN，并且它在理解全局信息方面具有很大优势。本文提出了ConvWin Transformer结构，它参考了Swin中的W-MSA结构并与卷积相结合。它不仅可以加快收敛速度，还能丰富补丁之间的信息交换并提高对局部信息的理解。然后，将其与医学图像分割中常用的U形架构UNet集成，形成一种名为ConvWin-UNet的结构。同时，本文改进了补丁扩展层以执行上采样操作。在Hubmap数据集和突触多器官分割数据集上的实验结果表明，所提出的ConvWin-UNet结构取得了优异的结果。这项工作的部分代码和模型可在https://github.com/xmFeng-hdu/ConvWin-UNet获取。
