边缘先验信息下的多类型肠道息肉图像分类网络  

Multi-Category Intestinal Polyp Image Classification Network Based on Edge Prior Information

作　　者：李胜[1] 曹婧 叶淑芳[2] 代飞 何熊熊[1] Li Sheng;Cao Jing;Ye Shufang;Dai Fei;He Xiongxiong(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310012,China;Center of Digestive Endoscopy,People′s Hospital of Lishui,Lishui 323020,Zhejiang,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310012 [2]丽水市人民医院消化内镜中心,浙江丽水323020

出　　处：《中国生物医学工程学报》2022年第4期431-442,共12页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(61873239);浙江省重点研发计划项目(2020C03074)。

摘　　要：肠道息肉的分级能够为内窥镜医生提供辅助诊断,对需要及时处理的高风险息肉和可以暂缓处理的低风险息肉进行区分。现有的基于深度学习息肉分类算法不能很好地区分类间相似性高的图像,针对息肉分级任务有待改进。提出一个包含边缘检测阶段、边缘特征描述提取阶段以及息肉分类阶段的边缘先验信息下的多类型肠道息肉图像分类网络。首先,在边缘检测阶段的跳跃连接层处,设计并嵌入反向注意力边缘监督模块以更好地捕获息肉边缘细节信息;其次,在内窥镜医生先验知识的指导下分别通过统计息肉边缘像素点个数和凹凸性来表示息肉边缘周长大小和光滑性特征,以此来补充神经网络特征提取的不足;最后,在分类网络的DenseBlock4后加入通道注意力自适应地捕获判别性特征。所构建的数据集来自丽水市人民医院消化内镜中心2018年至2019年的脱敏数据,样本量含1 050幅原始图像。在构建的四分类数据集上进行五折交叉验证,达到了77.29%的总体准确率,相比于已有算法的最好结果提高了6.46%。融合边缘先验信息的分类网络能够有效地对非腺瘤性息肉与低级别腺瘤性息肉、高级别腺瘤性息肉与腺癌这两组类间高相似度的息肉图像进行区分,增加网络的鲁棒性并提高网络的分类性能,在有限的训练数据集下为医生诊断提供辅助意见。The classification of intestinal polyps can help endoscopists to assist in diagnosis and distinguish between high-risk polyps requiring immediate treatment and low-risk polyps that can be deferred. Existing polyp classification algorithms based on deep learning can′t distinguish the high degree of inter-class similarities images, and need to be improved for multi-category polyp classification task. In this paper, a multi-category polyp image classification network based on edge prior information was proposed, including edge detection stage, edge feature descriptor extraction stage and polyp classification stage. Firstly, at the skip connection layer in the edge detection stage, a reverse attention edge monitoring module was designed and embedded to better capture the details of polyp edge. Secondly, under the guidance of the prior knowledge of the endoscopist, the perimeter size was represented by counting the number of pixels on the edge of the polyp, and the concavity and convexity were used to represent the smoothness feature, so as to supplement the insufficiency of neural network feature extraction. Finally, the channel attention was inserted after DenseBlock4 of the classification network to adaptively capture discriminative features. The private dataset was consisted of 1 050 desensitized original images that are collected from the Digestive Endoscopy Center of Lishui People′s Hospital within the year 2018 to 2019. Five-fold cross-validation was conducted in the polyp four-category dataset constructed in this paper, and the overall accuracy reached 77.29%, which was 6.46% higher than the best results of existing algorithms. The classification network fused with edge prior information can effectively discriminate two groups of polyp images with high degree of inter-class similarities, namely non-adenomatous polyps and low-grade adenomatous polyps, high-grade adenomatous polyps and adenocarcinoma. The established network in this paper increased the robustness and improved classification performance, prov

关 键 词：肠道息肉分类 边缘特征提取 注意力机制 临床先验知识 

分 类 号：R318[医药卫生—生物医学工程]