空间双线性注意力网络识别溃疡性结肠炎与克罗恩病  

Identification of ulcerative colitis and Crohn’s disease based on spatial and bilinear attention network

作　　者：戚婧 阮广聪 杨毅 吴毅[1] 曹倩[3] 魏艳玲 粘永健[1] QI Jing;RUAN Guangcong;YANG Yi;WU Yi;CAO Qian;WEI Yanling;NIAN Yongjian(Department of Digital Medicine,School of Biomedical Engineering and Imaging Medicine,Army Medical University(Third Military Medical University),Chongqing,400038;Department of Gastroenterology,Army Medical Center of PLA,Chongqing,400042;Department of Gastroenterology,Sir Run Run Shaw Hospital,Zhejiang University School of Medicine,Hangzhou,Zhejiang Province,310016,China)

机构地区：[1]陆军军医大学(第三军医大学)生物医学工程与影像医学系数字医学教研室,重庆400038 [2]陆军特色医学中心消化内科,重庆400042 [3]浙江大学医学院附属邵逸夫医院消化内科,杭州310016

出　　处：《陆军军医大学学报》2023年第3期227-234,242,共9页Journal of Army Medical University

基　　金：重庆市研究生科研创新项目(CYS22746)。

摘　　要：目的 利用深度学习技术辅助内镜医师识别溃疡性结肠炎(ulcerative colitis, UC)与克罗恩病(Crohn’s disease, CD)。方法 收集2018年1月至2020年11月陆军特色医学中心消化内科与邵逸夫医院消化内科共1 576例受试者的内镜图像,包括CD、UC与正常三类共计34 300幅,并按照9∶1的比例随机划分训练集与测试集,用于对网络进行训练与测试。在ResNet50基础上构建新颖的空间双线性深度网络(SABA-ResNet),引入空间注意机制,通过膨胀卷积扩大感受野以联系上下文信息,并与普通卷积局部归纳特性相配合,自适应聚焦病变区域。利用双线性注意提高网络的特征表示能力,以二阶信息加权特征映射的通道信息,从而提高模型的分类性能。结果 SABA-ResNet在测试集上对CD、UC和正常识别的总体准确率为92.67%(95%CI:91.91~93.37),AUC分别为0.978(95%CI:0.972~0.983)、0.977(95%CI:0.971~0.982)和0.999(95%CI:0.998~1.000),灵敏度分别为88.40%、89.07%、98.89%,特异性分别为95.49%、94.88%、98.93%,F1值分别为88.80%、89.01%和98.60%。消融实验与类激活映射图表明空间注意与双线性注意可帮助模型捕获更多病变区域的特征。结论 所构建的网络将空间注意与双线性注意相结合,在CD、UC与正常的识别中取得了良好的性能,可以有效辅助内镜医师对UC与CD进行识别。Objective To identify ulcerative colitis(UC) and Crohn’s disease(CD) with aid of deep learning technology for endoscopists. Methods From January 2018 to November 2020, the endoscopic images of 1 576 subjects(including 34 300 CD, UC and normal images) were collected from the Department of Gastroenterology of Army Medical Center of PLA and Sir Run Run Shaw Hospital. The training set and test set were randomly divided according to the ratio of 9∶1 to train and test the neural network. A novel spatial and bilinear deep network(SABA-ResNet) was constructed on the basis of ResNet50. The spatial attention mechanism was introduced, and the receptive field was expanded by dilated convolution to leverage contextual information, which was combined with the local induction of standard convolution to adaptively focus the lesion region. Bilinear attention was applied to improve the feature representation ability of the network, and the second-order information was used to weight the channel information of the feature map, so as to improve the classification performance of the model. Results The overall accuracy of SABA-ResNet for the recognition of CD, UC and normal tissues on the test set was 92.67%(95%CI: 91.91 ~ 93.37), the AUC value was 0.978(95%CI: 0.972~0.983), 0.977(95%CI: 0.971~0.982) and 0.999(95%CI: 0.998~1.000), the sensitivity was 88.40%, 89.07% and 98.89%, the specificity was 95.49%, 94.88% and 98.93%, and the F1 value was 88.80%, 89.01% and 98.60%, respectively. The ablation experiment and the class activation map suggested that spatial attention and bilinear attention could help the model capture more features of the lesion region. Conclusion Our constructed network combines spatial attention and bilinear attention, achieves excellent performance in the recognition of CD, UC and normal tissue, and effectively assist endoscopists in the diagnosis of UC and CD.

关 键 词：炎症性肠病 深度学习 溃疡性结肠炎 克罗恩病 

分 类 号：R319[医药卫生—基础医学] R447R574.62