基于深度学习建立睑结膜图像贫血筛查算法模型的研究  

An anemia screening tool based on deep learning with conjunctiva images

作　　者：胡小艳 李昊洋 刘祥 李雨捷 谭丽芳 李勇帅 陈芋文 易斌 HU Xiaoyan;LI Haoyang;LIU Xiang;LI Yujie;TAN Lifang;LI Yongshuai;CHEN Yuwen;YI Bin(Department of Anesthesiology,First Affiliated Hospital,Army Medical University,Army Medical University(Third Military Medical University),Chongqing,400038;Regiment Five,Basical Medicine College,Army Medical University,Army Medical University(Third Military Medical University),Chongqing,400038;Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences,Chongqing,400714,China)

机构地区：[1]陆军军医大学(第三军医大学)第一附属医院麻醉科,重庆400038 [2]陆军军医大学(第三军医大学)基础医学院学员五大队,重庆400038 [3]中国科学院重庆绿色与智能技术研究所,重庆400714

出　　处：《陆军军医大学学报》2023年第8期746-752,共7页Journal of Army Medical University

基　　金：国家重点研发计划(2018YFC0116700)。

摘　　要：目的探索深度学习技术在利用睑结膜图像作为输入的贫血自动筛查方面的应用。方法前瞻性收集和分析2021年3月18日到4月26日陆军军医大学第一附属医院麻醉科284例择期手术患者的睑结膜图像,依据对应的血红蛋白浓度,将图像标记为正常和贫血两个类别。基于InceptionV3、ResNet50V2、EfficientNetV2B0、DenseNet121等4种深度学习算法构建贫血预测模型。采用受试者特征工作曲线(receiver operating characteristic curve,ROC曲线)、准确性、敏感度、特异性、阳性预测值、阴性预测值评估模型性能。结果基于4种深度学习算法,贫血预测风险模型的ROC曲线下面积(area under curve,AUC)分别为:0.709(95%CI:0.643~0.769)、0.661(95%CI:0.594~0.725)、0.670(95%CI:0.603~0.733)、0.695(95%CI:0.628~0.756)。InceptionV3在测试集上的AUC(95%CI)、准确度、敏感度、特异性、阳性预测值、阴性预测值分别为:0.709(95%CI:0.643~0.769)、0.695、0.750、0.412、0.707、0.629,其模型性能最优;基于最优算法,开发了一个网络服务应用程序(http://150.158.58.4)用于对贫血的在线预测。结论本研究基于深度学习算法利用睑结膜图像建立的贫血筛查模型快速高效,InceptionV3模型的综合预测性能更优。ObjectiveTo explore the application of deep learning in automatic classification of anemia with conjunctival images as input.MethodsThe conjunctival images of 284 patients undergoing elective surgery in the Department of Anesthesiology of the First Affiliated Hospital of Army Medical University from March 18 to April 26,2021 were collected and analyzed prospectively.The images divided into 2 types:normal and anemia according to the corresponding hemoglobin concentration.Four deep learning algorithms,including InceptionV3,ResNet50V2,EfficientNetV2B0 and DenseNet121,were used to construct a prediction model for anemia.The performance of the model was evaluated by receiver operating characteristic(ROC)curve with accuracy,sensitivity,specificity,positive predictive value and negative predictive value.ResultsThe area under ROC curve(AUC)was 0.709(95%CI:0.643~0.769),0.661(95%CI:0.594~0.725),0.670(95%CI:0.603~0.733),and 0.695(95%CI:0.628~0.756),respectively for the 4 deep learning algorithms.The InceptionV3 model showed superior predictive performance on the test set,with an AUC value of 0.709(95%CI:0.643~0.769),an accuracy of 0.695,a sensitivity of 0.750,a specificity of 0.412,a positive predictive value of 0.707 and a negative predictive value of 0.629.Based on the optimal algorithm,a network service application which can be used for online prediction of anemia was developed(http://150.158.58.4).ConclusionOur model,which is established based on deep learning algorithm with conjunctiva image as input,has a good performance on fast and automatic prediction for anemia.The InceptionV3model has better comprehensive prediction performance.

关 键 词：深度学习 贫血 睑结膜 无创 

分 类 号：R319[医药卫生—基础医学] R322.91R556.0