Comparison of Multiview Echocardiography Fusion Using Traditional and Deep Learning Models

Sharanya Balachandran, Michelle Noga, Harald Becher, Jonathan Windram, Pierre Boulanger, Kumaradevan Punithakumar

Keywords: Cardiac Image Fusion, Robotic arm, Deep Learning, Medical Image Analysis

TL;DR: Deep learning network for cardiac image fusion with the data acquired using robotic arm and benchmarking with the existing methods.

Abstract: Cardiovascular disease is the second leading cause of death in Canada, and echocardiography is considered one of the preferred ways of examining cardiovascular diseases because it is non-invasive, cost-effective, portable, and easy to use. However, echocardiography scans are limited by a low signal-to-noise ratio and occasional signal dropout. These limitations can be overcome by developing an image fusion process that combines images from multiple views of the heart to enhance information and quality, thereby advancing the diagnostic value of cardiac ultrasound images. In this study, a unified unsupervised image fusion network, U2Fusion, is applied to fuse echocardiographic images, eliminating the need to extract handcrafted features used in traditional fusion. It enables precise visualization of cardiac structures in the fused images. The datasets are acquired using an ultrasound transducer mounted on Universal Robots UR10e arm. The study demonstrated the benefit of using an unsupervised deep-learning approach to fuse 2D images from human volunteers, which outperforms other traditional methods, such as average, maximum, as well as wavelet fusion and deep learning models such as U-Net, DenseFuse, and ACU$^2$E-Net, for fusion in terms of metrics including PSNR, SSIM, entropy, and CNR. The clinical significance of this method lies in the ability to provide accurate visualization of the heart structures, including the edges of chambers and other anatomical details. The clinical application would provide a better assessment of cardiac structures and could improve the diagnosis of various heart diseases. The clinical significance of the method would be an accurate visualization of the focused region of heart structures, including vital information such as the edges of chambers and other details.

Submission Number: 27