Despeckling Echocardiograms Using Generative Adversarial Networks

Abstract

Previous research had shown that generative adversarial networks (GANs) are capable of despeckling echocardiograms (echos) through image-to-image translation in real-time once trained. However, only limited information regarding the quality of denoised echos and explainability of useful GAN components is provided. Therefore, this thesis conducted an ablation study and utilised patch-based no-reference (NR) metrics for the evaluation of the despeckling quality. The training data consists of 4000 low resolution apical 4-chamber (A4CH) echos of 2000 distinct patients. Moreover, it was investigated if despeckling is an advantageous preprocessing step to improve the segmentation of cardiac structures in the left side of the heart in the A4CH view.

The results of the NR metrics demonstrate that the implemented denoising GAN (DnGAN) outperforms the used non-local low-rank framework filter which produced the denoised echos for the translation task. Especially the introduction of a perceptual loss component is attributed with training stability and noise reduction. In terms of segmentation, both filtered echos and the DnGAN showed no significant improvement over the raw echos. The residuals, calculated as the difference between the raw and despeckled echos, were lowest (best) for the DnGAN indicating potential, in particular for echos of poor quality.