基于眼心关联的先天性心脏病围手术期转归多实例预测  

Multi-case learning model for predicting the perioperative outcome of CHD based on eye-heart connection

作　　者：孔令聪 杨欢 胡联亭 李聪 陈炫卉 刘华章 况宇 杨小红 梁会营 KONG Lingcong;YANG Huan;HU Lianting;LI Cong;CHEN Xuanhui;LIU Huazhang;KUANG Yu;YANG Xiaohong;LIANG Huiying(School of Medicine,South China University of Technology,Guangzhou 510006,Guangdong Province,China;Medical Big Data Center,Guangdong Provincial People's Hospital,Guangdong Academy of Medical Sciences;Department of Ophthalmology,Guangdong Eye Institute)

机构地区：[1]华南理工大学医学院,广州510006 [2]广东省人民医院,广东省医学科学院医学大数据中心 [3]广东省眼病防治研究所眼科

出　　处：《中国数字医学》2023年第1期33-37,94,共6页China Digital Medicine

基　　金：国家自然科学基金面上项目(62076076);国家自然科学基金青年项目(82122036)。

摘　　要：目的:构建先天性心脏病(先心病)围手术期转归多实例预测模型,提高先心病预后评估的精确性。方法:基于先心病患者眼底彩照图像,引入多实例学习算法,结合长短注意力Transformer结构处理多图像实例特征序列,构建基于眼心关联的先心病围手术期转归多实例预测模型。结果:该模型有效规避了深度学习模型限制输入图像尺寸所导致的细节丢失问题,在测试集中对围手术期复合不良结局预测的准确率为0.78,AUC为0.82,优于CANet、DLI、ResNet50、InceptionV3等目前常用模型;处理每张图片耗时0.104 s,与上述模型相当。结论:基于眼心关联的先心病围手术期转归预测模型可以有效提升模型的图像细节捕捉能力,提高先心病预后评估的效能,同时拓展了人工智能模型的临床适用范围。Objective To construct a multi-case learning model for predicting the perioperative outcome of congenital heart disease(CHD)and improve the accuracy of prognostic assessment.Methods A multi-case learning algorithm was introduced based on fundus images of patients with CHD,and a multi-case learning model for predicting the perioperative outcome of CHD based on eye-heart connection was constructed by adopting the Long and short attention Transformer structure to process multiple-image case feature sequences.Results The model effectively avoided the problem of detail loss caused by deep learning models which limit the input image size.In the test set,the accuracy of predicting perioperative complex adverse outcomes was 0.78,and the AUC was 0.82,which were better than CANet,DLI,ResNet50,InceptionV3 and other commonly used models.The processing time of each image was 0.104 s,which is similar to those of the above methods.Conclusion The learning model for predicting the perioperative outcome of CHD based on eye-heart connection can effectively improve the detail capturing ability of the model,improve the efficacy of prognostic assessment of CHD,and expand the clinical application scope of artificial intelligence models.

关 键 词：先天性心脏病 眼底彩照 多实例学习 转归 

分 类 号：R319[医药卫生—基础医学] R197.323.2