Coronary Artery Segmentation on Echocardiograms for Kawasaki Disease Diagnosis

Symptoms such as vasculitis, myocarditis and coronary dilatation are possible consequences of Kawasaki disease. Damage to the blood vessels caused by these symptoms can lead to long-term heart problems. It is the most common acquired heart disease affecting young children in developed countries. One way to detect abnormalities in the coronary arteries caused by this disease is by monitoring patients with 2D echocardiograms. This monitoring must be performed manually due to the great difficulty of automation, as the spectrum of heart size and shape in infants can be very diverse. This paper presents a solution to help diagnose Kawasaki disease in a faster and simpler way by using convolutional neural networks. More specifically, this work can automatically segment coronary arteries from 2D echocardiography images. These frames are relevant as Kawasaki disease can be detected by measuring the opening of the coronary artery. Different U-net based models have been developed and evaluated. In addition, an echocardiogram dataset specific for Kawasaki disease has been created in collaboration with the Hospital 12 de Octubre in Madrid. This solution can be considered as another step in the development of a fully automated solution for diagnosis [1].