针对结直肠息肉图像分割时动态信息处理和边缘细节捕捉能力不足，导致边界信息损失和错误分割等问题，本文提出一种建立在Swin Transformer框架上的线索交叉聚合(cross-cue fu-sion，CCF)结肠息肉分割方法。该方法首先通过Transformer编码器对图像的病变特征进行逐级提取。其次利用改进的二阶通道注意力(second-order channel attention，SOCA)机制增强跨层级信息交互能力，有效提取丰富的多尺度上下文特征信息。再次采用反向通道频率注意力(reverse channel frequency attention，RCFA)机制中的离散余弦变换(discrete cosine transform，DCT)，突出多尺度上下文信息的通道特征。最后通过CCF模块从动态和静态深度两个层面增强图像特征，进而提升动态信息处理和细节捕捉能力。在数据集CVC-ClinicDB、Kvasir、CVC-ColonDB和ETIS-LaribPolypDB上进行测试，Dice指数分别为0。942、0。924、0。800和0。774。MIoU指数分别为0。896、0。878、0。726和0。697。实验数据表明，本文提出的方法能有效分割结直肠息肉图像，为结直肠息肉的诊断提供了新思路。
Colorectal polyp segmentation method fusing Transformer and cross-cue fusion
In order to solve the problems of insufficient dynamic information processing and edge detail capture in colorectal polyp image segmentation,such as boundary information loss and wrong segmentation,this paper proposes a colorectal polyp segmentation method based on Swin Transformer framework.Firstly,Transformer encoder is used to extract the pathological features of the image step by step.Secondly,the improved second-order channel attention(SOCA)mechanism is used to enhance cross-level information interaction ability and effectively extract rich multi-scale context feature information.Furthermore,the discrete cosine transform(DCT)in the attention mechanism of reverse frequency channel is used to highlight the channel characteristics of multi-scale context information.Finally,the image features are enhanced from both dynamic and static depth through the cross-cue fusion(CCF)module to improve the dynamic information processing and detail capture capabilities.When tested on the datasets CVC-ClinicDB,Kvasir,CVC-ColonDB,and ETIS-LaribPolypDB,Dice indices are 0.942,0.924,0.800 and 0.774,respectively.The MIoU indices are 0.896,0.878,0.726 and 0.697,respectively.The experimental data show that the proposed method can effectively segment colorectal polyp images and provide a new idea for the diagnosis of colorectal polyp.
image segmentationcolorectal polypsTransformercross-cue fusion(CCF)modulereverse channel frequency attention(RCFA)module
NETL
NSTL
万方数据
