Coronary Artery Identification on Echocardiograms for Kawasaki Disease Diagnosis

Kawasaki disease’s consequences may result in vasculitis, myocarditis and coronary dilatation causing long term heart complications by damaging blood vessels all over the body. It is the most common acquired heart condition affecting young children in developed countries. Follow up of Kawasaki disease patients is done by 2D echocardiograms in order to detect coronary artery abnormalities. Such inspection is very difficult to automatize and has to be done manually since the size and shape of the heart in young children varies significantly. We present a solution to ease and speed-up the diagnosis of Kawasaki disease based on Convolutional Neural Networks. More specifically, our work can automatically detect which frames of a 2D echocardiogram contain coronary arteries. A Convolutional Neural Network has been designed with this purpose, and its performance has been compared to those of VGG16 and Resnet50 networks. To evaluate our approach a specific echocardiogram dataset for Kawasaki disease has been created in collaboration with 12 de Octubre Hospital in Madrid. This solution can be considered as a first step in the development of a fully automated solution for its diagnosis.