基于深度学习的川崎病合并冠状动脉瘤超声分类诊断

Ultrasonic classification diagnosis of Kawasaki disease complicated with coronary aneurysms based on deep learning

目的 探讨基于深度学习的川崎病合并冠状动脉瘤超声分类诊断的可行性。方法 收集复旦大学附属儿科医院诊断为川崎病患者的心超图像,选取胸骨旁大动脉短轴切面,非彩色多普勒成像的二维图像,剔除冠状动脉图像缺失、图像质量差,图像不完善的病例,共纳入研究图片1 328张,其中包括冠状动脉瘤664张图片,冠状动脉正常664张图片。利用所采集图片基于典型的深度神经网络AlexNet、LeNet、VggNet19、ResNet18进行分类诊断,其中1 000张作为训练集,164张作为验证集,164张作为测试集,约各占总图像数量的80%、10%、10%。结果 基于AlexNet的深度神经网路对川崎病合并冠状动脉瘤的超声图像分类结果最好,可达98%准确率,且该深度分类模型在参数数量及训练效率上均有明显优势。结论 基于深度卷积神经网络的川崎病合并冠状动脉瘤的超声分类诊断具有可行性。

Objective To investigate the feasibility of ultrasonic classification diagnosis based on deep learning for Kawasaki disease complicated with coronary aneurysms.Methods Echocardiography images of Kawasaki patients diagnosed in the Pediatric Hospital of Fudan University were collected.The parasternal short axis sections were selected and two-dimensional images without color Doppler imaging were included.In addition,the cases with missing images,poor image quality and imperfect images were eliminated.The collected dataset contains 664 images of coronary aneurysms and 664 images of normal coronary arteries.The collected images are used for classification diagnosis based on four classical deep neural networks including AlexNet,LeNet,VggNet19 and ResNet 18.The dataset was divided into a training set of 1000 images,a validation set of 164 images,and a test set of 164 images,accounting for approximately 80%,10%,and 10%of the total image count,respectively.Results The AlexNet based deep neural network has the best results for ultrasonic image classification of Kawasaki disease complicated with coronary aneurysms and the diagnostic accuracy can reach 98%.In addition,this depth model has obvious advantages in the parameter number and training efficiency.Conclusion The ultrasonic classification diagnosis of Kawasaki disease complicated with coronary aneurysms based on deep convolutional neural network is feasible.