Deep learning (DL) algorithms are increasingly used in cardiac imaging. We aimed to investigate the utility of DL algorithms in de-noising transthoracic echocardiographic images and removing acoustic shadowing artefacts specifically in patients with congenital heart disease (CHD). In addition, the performance of DL algorithms trained on CHD samples was compared to models trained entirely on structurally normal hearts. Deep neural network based autoencoders were built for denoising and removal of acoustic shadowing artefacts based on routine echocardiographic apical 4-chamber views and performance was assessed by visual assessment and quantifying cross entropy. 267 subjects (94 TGA and atrial switch and 39 with ccTGA, 10 Ebstein anomaly, 9 with uncorrected AVSD and 115 normal controls; 56.9% male, age 38.9 ± 15.6 years) with routine transthoracic examinations were included. The autoencoders significantly enhanced image quality across diagnostic subgroups (p < 0.005 for all). Models trained on congenital heart samples performed significantly better when exposed to examples from congenital heart disease patients. Our study demonstrates the potential of autoencoders for denoising and artefact removal in patients with congenital heart disease and structurally normal hearts. While models trained entirely on samples from structurally normal hearts perform reasonably in CHD, our data illustrates the value of dedicated image augmentation systems trained specifically on CHD samples.
Keywords: Adult congenital heart disease; Autoencoder; Congenital heart disease; De-noising; Deep-learning; Echocardiography.