A Domain Translation Framework with an Adversarial Denoising Diffusion Model to Generate Synthetic Datasets of Echocardiography Images

Abstract

Currently, medical image domain translation operations show a high demand from researchers and clinicians. Amongst other capabilities, this task allows the generation of new medical images with sufficiently high image quality, making them clinically relevant. Deep Learning (DL) architectures, most specifically deep generative models, are widely used to generate and translate images from one domain to another. The proposed framework relies on an adversarial Denoising Diffusion Model (DDM) to synthesize echocardiography images and perform domain translation. Contrary to Generative Adversarial Networks (GANs), DDMs are able to generate high quality image samples with a large diversity. If a DDM is combined with a GAN, this ability to generate new data is completed at an even faster sampling time. In this work we trained an adversarial DDM combined with a GAN to learn the reverse denoising process, relying on a guide image, making sure relevant anatomical structures of each echocardiography image were kept and represented on the generated image samples. For several domain translation operations, the results verified that such generative model was able to synthesize high quality image samples: MSE: 11.50 +/- 3.69, PSNR (dB): 30.48 +/- 0.09, SSIM: 0.47 +/- 0.03. The proposed method showed high generalization ability, introducing a framework to create echocardiography images suitable to be used for clinical research purposes.

References in corpus (8)

• Diffusion Models Beat GANs on Image Synthesis
• Deep Unsupervised Learning using Nonequilibrium Thermodynamics
• Score-Based Generative Modeling through Stochastic Differential Equations
• Deep Learning for Segmentation using an Open Large-Scale Dataset in 2D Echocardiography
• Tackling the Generative Learning Trilemma with Denoising Diffusion GANs
• Memory-efficient GAN-based Domain Translation of High Resolution 3D Medical Images
• A Data Augmentation Pipeline to Generate Synthetic Labeled Datasets of 3D Echocardiography Images using a GAN
• Unsupervised Medical Image Translation with Adversarial Diffusion Models

Cited by in corpus (2)

• Cardiac valve event timing in echocardiography using deep learning and triplane recordings
• Graph Convolutional Neural Networks for Automated Echocardiography View Recognition: A Holistic Approach