高效多注意力融合的U-Net结直肠息肉图像分割算法

Efficient U-Net Colorectal Polyp Image Segmentation Algorithm Based on Multi-attention Fusion

针对结直肠内部环境复杂以及结直肠息肉边界模糊且颜色、形状、大小不一等问题,提出用于结直肠息肉图像分割的基于U-Net高效多注意力融合(efficient multi-attention fusion based on U-Net,EMAU-Net)算法。EMAU-Net算法利用快速行进方法(fast marching method,FMM)抑制结直肠镜检查中的高光噪声;在编码器部分采用轻量级网络SegFormer中的Transformer模块对U-Net做出改进,保留息肉纹理和位置信息,增强编码阶段网络对全局信息的捕捉能力;解码器部分应用改进的空间和通道挤压与激励块(spatial and channel squeeze&excitation block,scSE),降低参数量的同时又增强其特征,可以进一步提高对息肉边缘特征的捕捉能力。实验结果表明,EMAU-Net算法平均交并比(mean intersection over union,MIoU)达到92.97%,与其他算法对比,该算法表现更优,能够对结直肠癌的诊断和治疗提供有力帮助。

In view of the complex internal environment of colorectal and the blurred boundary and different color,shape and size of colorectal polyps,this paper proposes an efficient multi-attention fusion algorithm based on U-Net(EMAU-Net)for colorectal polyp image segmentation.EMAU-Net uses the fast marching method(FMM)to suppress the highlight noise in colonoscopy,and the Transformer module in the lightweight network SegFormer is used in the encoder to improve U-Net to retain polyp texture and position information and enhance the ability of the network to capture global information in the coding stage.In the decoder part,the improved space and channel extrusion and excitation blocks are applied to reduce the number of parameters and enhance their features at the same time,which can further improve the ability to capture the edge features of polyps.The experimental results show that the average cross-union ratio of EMAU-Net algorithm reaches 92.97%.Compared with the other algorithms,EMAU-Net algorithm performs better,which can provide powerful help for the diagnosis and treatment of colorectal cancer.