基于双分支特征提取的息肉图像分割网络

Polyp Image Segmentation Network Based on Double Branch Feature Extraction

摘要: 结肠息肉分割是从结肠息肉图像中提取病理信息的关键步骤，对结直肠癌的诊断和治疗具有重要意义。针对结肠息肉分割中形状大小不一、病灶组织与背景差异性难以区分等问题，本文提出了一种基于CNN和Swin Transformer的双分支特征提取的医学息肉分割网络(DST-Net)，其充分考虑了卷积神经网络和Transformer在提取局部特征和全局特征各自的优势。DST-Net是一种编码器-解码器架构，首先设计了基于VGG和空洞卷积(AC)的两种编码器分别提取局部边界特征和多尺度特征；接下来在编解码器中间的底部模块使用两个连续的Swin Transformer模块，充分利用Transformer的远程依赖关系进一步加强网络的全局特征提取能力；最后，在编解码器之间的跳过连接中使用了通道注意力模块(CAB)，以更加关注可疑和复杂的区域。在CVC-Clinic DB和Kvasir两个公开的息肉数据集上验证了所提出方法，结果表明该模型优于现有的其他方法，可以准确有效地实现结肠息肉的分割任务。

Abstract: Colon polyp segmentation is a key step to extract pathological information from colon polyp images, which is of great significance for the diagnosis and treatment of colorectal cancer. Aiming at the problems such as different shapes and sizes in colon polyp segmentation and difficult to distinguish lesion tissue and background differences, this paper proposes a Dual-branch feature extraction medical polyp segmentation network (DST-Net), which fully considers the advantages of convolutional neural network and Transformer in extracting local features and global features. DST-Net is an encoder-decoder architecture. Firstly, two encoders based on VGG and void convolution (AC) are designed to extract local boundary features and multi-scale features respectively. Next, the bottom module in the middle of the codec uses two continuous Swin Transformer modules to make full use of the remote dependency of Transformer to further strengthen the global feature extraction capability of the network. Finally, the Channel Attention Module (CAB) is used in skip connections between codecs to pay more attention to suspicious and complex areas. The proposed method was validated on two publicly available polyp datasets, CVC-Clinic DB and Kvasir. The results show that the model is superior to other existing methods and can accurately and effectively achieve the task of colon polyp segmentation.