Abstract:

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 ResNet18. 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.

Key words: Kawasaki disease, echocardiography, deep learning, classification diagnosis