基于Swin-Unet改进的医学图像分割算法
Improved Medical Image Segmentation Algorithm Based on Swin-Unet

在过去的十几年时间里，基于CNN的神经网络在医学图像分割领域取得了突破性进展。尤其是以unet为代表的u形网络架构和跳跃连接被广泛应用于一系列的医学图像任务。由于CNN的内在局限性，不能够很好的获取到全局和远程语义信息交互。由于腹部器官复杂，容易发生形变、边缘模糊、体积小等原因导致分割比较困难。因此在Swin-Unet的基础上改进，首先末端编码器与首个解码器之间引入多尺度模型提取模块，增强不同形状大小信息特征提取。其次将最后两个编码器和解码器的swin transformer Block引入残差机制来缓解模型深度带来的梯度弥散现象。并且在最后的编码器末端引入ASSP模块获取多尺度细节信息。最后，在跳跃连接中引入通道注意力机制(CAM)，可以让模型强化重要信息特征通道，弱化特征不相关通道，最后达到有效提高模型分割精度和准确度的效果。
In the past ten years, CNN-based neural networks have made breakthrough progress in the field of medical image segmentation. In particular, the U-shaped network architecture and skip connections represented by unet are widely used in a series of medical image tasks. Due to the inherent limitations of CNN, global and remote semantic information interaction cannot be well obtained. Due to the complexity of abdominal organs, which are prone to deformation, blurred edges, and small size, segmentation is difficult. Therefore, based on Swin-Unet, a multi-scale model extraction module is introduced between the terminal encoder and the first decoder to enhance the extraction of information features of different shapes and sizes. Secondly, the residual mechanism is introduced into the swin transformer Block of the last two encoders and decoders to alleviate the gradient dispersion phenomenon caused by the depth of the model. And the ASSP module is introduced at the end of the final encoder to obtain multi-scale detailed information. Finally, introducing the channel attention mechanism (CAM) in the skip connection allows the model to strengthen important information feature channels and weaken feature irrelevant channels, and finally achieve the effect of effectively improving the model segmentation precision and accuracy.