[Paper Review] Phase Aberration Robust Beamformer for Planewave US Using Self-Supervised Learning

Shujaat Khan, Jaeyoung Huh+1 more|arXiv (Cornell University)|Feb 16, 2022

Ultrasound Imaging and Elastography4 citations

TL;DR

This paper proposes a self-supervised 3D CNN beamformer that mitigates phase aberration artifacts in planar wave ultrasound imaging by learning to reconstruct high-quality images from phase-aberrated inputs without requiring ground-truth speed-of-sound (SoS) maps. Trained to model SoS variations as stochastic distortions, the method achieves superior image quality and robustness to measurement loss and deep scan artifacts, outperforming conventional and SVD-based beamforming with comparable computational cost.

ABSTRACT

Ultrasound (US) is widely used for clinical imaging applications thanks to its real-time and non-invasive nature. However, its lesion detectability is often limited in many applications due to the phase aberration artefact caused by variations in the speed of sound (SoS) within body parts. To address this, here we propose a novel self-supervised 3D CNN that enables phase aberration robust plane-wave imaging. Instead of aiming at estimating the SoS distribution as in conventional methods, our approach is unique in that the network is trained in a self-supervised manner to robustly generate a high-quality image from various phase aberrated images by modeling the variation in the speed of sound as stochastic. Experimental results using real measurements from tissue-mimicking phantom and extit{in vivo} scans confirmed that the proposed method can significantly reduce the phase aberration artifacts and improve the visual quality of deep scans.

Motivation & Objective

• To address phase aberration in planar wave ultrasound imaging caused by tissue-specific speed-of-sound (SoS) variations.
• To develop a real-time, data-driven beamforming method that does not require explicit SoS estimation or iterative optimization.
• To improve image quality in deep scans where phase aberration severely degrades contrast and resolution.
• To enable robust performance under reduced planewave acquisition (e.g., 15 PWIs) and real-world measurement loss.
• To achieve computational efficiency comparable to standard delay-and-sum beamforming while surpassing SVD-based methods.

Proposed method

• A 3D convolutional neural network (CNN) is trained in a self-supervised manner using simulated phase-aberrated planewave images as input.
• The network learns to reconstruct high-quality B-mode images by modeling SoS variations as stochastic distortions, without requiring ground-truth SoS maps.
• Training data is generated by introducing random SoS deviations (±5% around 1540 m/s) in simulated planewave acquisitions.
• The network is optimized end-to-end using a perceptual loss function to preserve structural and contrast details.
• The beamformer operates directly on raw RF data from multiple planewave transmissions, performing coherent compounding in a learned, differentiable manner.
• The architecture is designed for low computational overhead, enabling real-time inference with latency comparable to standard CPC.

Experimental results

Research questions

• RQ1 Can a self-supervised deep learning approach effectively reduce phase aberration artifacts in planar wave ultrasound imaging without requiring ground-truth SoS maps?
• RQ2 How does the proposed method perform under reduced planewave acquisition (e.g., 15 PWIs) compared to conventional and SVD-based beamformers?
• RQ3 To what extent does the method improve image quality in deep scans where SoS variations are most pronounced?
• RQ4 Can the method generalize to real in vivo and phantom measurements with unknown SoS distributions?
• RQ5 Is the computational cost of the proposed method compatible with real-time ultrafast ultrasound imaging?

Key findings

• The proposed method achieved a 0.85 dB higher contrast ratio (CR) and 0.04 higher contrast-to-noise ratio (CNR) compared to standard coherent planewave compounding (CPC) at 15 planewaves.
• With 15 planewaves, the method improved GCNR by 0.05 units over SVD-based beamforming, with p-value < 10⁻³, indicating statistical significance.
• The method maintained high image quality even with only 15 planewaves, achieving performance comparable to SVD-BF with 31 planewaves.
• Computational latency was 47.74 ms for 15 planewaves, significantly lower than SVD-BF (22.56 seconds), making it suitable for real-time use.
• In pork phantom scans with 20 mm tissue layers, the proposed method restored visibility of hypo- and hyper-echoic structures lost in CPC images.
• In vivo scans confirmed that the method significantly reduced phase aberration artifacts and improved structural detail, especially in deep tissue regions.