UCGNet: Capsule-Guided GAN for Ultrasound Image Reconstruction from Single-Plane Wave RF Data

Maryam Asad  Samani; Ali  Gharekhani; Parastoo  Farnia; Bahador  Makki Abadi

doi:10.18502/fbt.v13i2.21930

Maryam Asad Samani Department of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran
Ali Gharekhani Department of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran
Parastoo Farnia Department of Medical Physics and Biomedical Engineering, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran
Bahador Makki Abadi Department of Medical Physics and Biomedical Engineering, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran

DOI: https://doi.org/10.18502/fbt.v13i2.21930

Keywords: Capsule Network; Generative Adversarial Network; Ultrasound Image Reconstruction; Lightweight Deep Learning; Single-Plane Wave.

Abstract

Purpose: This study aims to improve ultrasound image reconstruction from single-plane wave RF data using Capsule Neural Networks, which can produce comparable image quality to Convolutional Neural Networks while requiring significantly fewer parameters. In addition to reducing model size, the proposed approach preserves clinically important image features and is better suited for real-time implementation in embedded systems with constrained computational resources.

Materials and Methods: We propose a novel ultrasound image reconstruction architecture, UCGNet (U-Caps-GAN Network), which combines Capsule Networks with a Generative Adversarial Network framework. UCGNet reconstructs high-quality B-mode ultrasound images directly from single-plane wave RF data and is evaluated on the Plane-wave Imaging Challenge in Medical Ultrasound (PICMUS) dataset. Capsule Networks play a key role in achieving parameter efficiency by encoding spatial hierarchies through vectorized feature representations. Their dynamic routing mechanism captures part–whole relationships and pose variations, enabling the network to preserve fine structural details essential for diagnostic imaging, without relying on deep, redundant convolutional layers. This makes the proposed architecture particularly well-suited for real-time applications in embedded systems with limited computational resources.

Results: The reconstructed images achieved a mean Signal-to-Noise Ratio (SNR) of 18.45 and a Peak Signal-to-Noise Ratio (PSNR) of 40.92, outperforming the baseline UNet model in terms of image quality. Additionally, UCGNet required about 23% of the training parameters compared to UNet, demonstrating its suitability for real-time applications on resource-constrained devices.

Conclusion: UCGNet provides an efficient and accurate solution for ultrasound image reconstruction from raw RF data. Its improved image fidelity and reduced computational complexity make it a strong candidate for practical use in portable and embedded medical imaging systems.