Abstract
The use of machine learning (ML) in the medical field is hindered by the scarcity of high-quality data. This work tackles the deficiency of echocardiogram pictures (echoCG) by using advanced generative models for synthetic creation. We provide a comprehensive practical assessment of these approaches and ascertain their limits. Our study examines the efficacy of a cycle-consistent generative adversarial network (CycleGAN), a contrastive unpaired translation (CUT) approach, and a latent diffusion model (Stable Diffusion 1.5). We outline our data generating approach, exhibit picture examples, describe our assessment strategy, and offer findings from a user study measuring the perceived quality and medical validity of the produced samples, done with licensed cardiologists and surgeons.