UCGNet: GAN for Ultrasound Beamforming through Capsule Layers from Single-Plane Wave RF Data

Abstract

Purpose: This study aims to implement Capsule Networks for ultrasound beamforming and image reconstruction, addressing the limitations of conventional Convolutional Neural Networks (CNNs) in embedded systems. The goal is to enhance image quality from single-plane wave transmission using fewer parameters while maintaining diagnostic accuracy.

Materials and Methods: We propose a novel image reconstruction architecture, UCGNet (U-Caps-GAN Network), which integrates Capsule Networks (U-Caps) within a Generative Adversarial Network (GAN) framework. The method is applied to reconstruct high-quality ultrasound images from single-plane wave data and is evaluated using the Plane-wave Imaging Challenge in Medical Ultrasound (PICMUS) dataset.

Results: The reconstructed images achieved a mean signal-to-noise ratio (SNR) of 18.4383 and a peak signal-to-noise ratio (PSNR) of 41.0226, outperforming the baseline UNet model in terms of accuracy. Moreover, UCGNet used less than 25% of the training parameters compared to UNet.

Conclusion: UCGNet provides an effective and lightweight solution for ultrasound image reconstruction. Its improved accuracy and reduced parameter count make it well-suited for practical medical imaging applications, particularly in resource-constrained environments.