Generative models for ultrasound image reconstruction from single plane-wave simulated data

Abstract

Ultrasound image reconstruction from a single plane-wave transmission is required for many applications, However, imaging quality can be degraded when using conventional delay-and-sum (DAS) beamforming. This paper evaluates the performance of diffusion models (Diff) and conditional Generative Adversarial Networks (cGAN) for ultrasound image reconstruction when using the same base architecture, a UNet. Models were trained using a simulated dataset of 12500 acquisitions. Each sample featured a randomly positioned anechoic cyst in a medium with uniform sound speed, with downsampled IQ channel data serving as input. Results demonstrated that diffusion models could generate B-mode images of similar or improved contrast than the cGANs. On average, they exhibited a higher contrast-to-noise ratio (1.32 for Diff vs 1.11 for cGAN) and gCNR (0.83 for Diff vs 0.76 for cGAN).