融合多种注意力机制的结直肠息肉分割神经网络  

Polyp Segmentation Network Combined With Multi-Attention Mechanism

作　　者：贾立新[1] 胡奕标 金燕[1] 薛智中 姜智伟 郑秋富 Jia Lixin;Hu Yibiao;Jin Yan;Xue Zhizhong;Jiang Zhiwei;Zheng Qiufu(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023)

机构地区：[1]浙江工业大学信息工程学院,杭州310023

出　　处：《计算机辅助设计与图形学学报》2023年第3期463-473,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(U22A2040);浙江省自然科学基金(LY17F010015)。

摘　　要：精确的息肉分割能够有效地帮助医生发现并切除异常组织,降低息肉转化为结直肠癌的风险.针对息肉具有不同大小、形态、颜色、纹理,类间相似度高、类内差异度大的特点,为了实现对息肉图像的精确分割,提出一种融合多种注意力机制的卷积神经网络.首先,以Res2Net为骨干网络提取图像特征,通过通道分组空间增强注意力增强特征;其次,使用轴向自注意力结合感受野策略获得同时具有边缘细节的低级特征和全局语义高级特征;最后,利用逆向注意力挖掘边界信息,增强网络对于较小且组织边界模糊的息肉的分割性能.在5个息肉分割数据集上,与其他5种基准方法比较的实验结果表明,平均Dice系数、平均交并比和平均绝对误差指标均有所提高,其分割精度和泛化性能均优于现有主流息肉分割网络.Accurate polyp segmentation can help doctors find and resect abnormal tissue,and decrease chances of polyps changing into colorectal cancer.Considering the characteristics of polyps with different sizes,shapes,colors,textures,high inter-class similarity,and intra-class variation,we proposed a polyp segmentation neural network combined with multi-attention to achieve accurate segmentation of polyp images.Firstly,Res2Net is used as a backbone network to extract image features,and features extracted by the backbone network are enhanced by channel group spatial enhanced module.Secondly,we use axial self-attention combined with receptive field block strategy to extract fine feature maps which both have low-level detail information and high-level global semantic information.Finally,we use reverse attention to mine boundary information to enhance the network’s segmentation performance for small polyps and its surrounding mucosa is not sharp.The experiments were conducted on five challenging polyp segmentation datasets and compared with five other benchmark methods.The results of experiments show that our method outperforms the compared methods in the mean Dice,mean IoU,MAE,segmentation accuracy,and performance of generalizability.

关 键 词：医学图像分割 息肉分割 结直肠癌 卷积神经网络 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]