Swin-EchoNet: Deep Learning-based Two-Chamber Segmentation of 2D Echocardiography using Swin Transformer

Abstract

The initial crucial stage in recognizing heart-related problems involves making a precise diagnosis. Live heart images can be obtained through techniques such as MRI, CT scan, and Echocardiography. Determining significant cardiac parameters for disease diagnosis, such as systolic and diastolic volumes, ejection fraction, and left atrium (LA), requires accurate segmentation of the left ventricle in echocardiography images. However, automated segmentation of these images is a complex and challenging task. Therefore, there is a requirement for a fully automatic method that can accurately segment cardiac images and save time. We utilized our proposed model Swin-echonet architecture for accurate left ventricle segmentation in echocardio-graphy images. Our method has been successfully tested on two separate datasets, namely the CAMUS dataset with 1800 echocardiographic images and data from a hospital with 1550 echo images. On the CAMUS dataset, we achieved a mean dice coefficient of 0.951±0.2271. Additionally, our method produced a mean IoU of 0.7263 and a mean Dice coefficient of 0.9738 on the second dataset. The obtained results demonstrate the efficiency of our method across diverse datasets, indicating its potential to assist medical professionals in the detection and treatment of cardiac problems.