[Paper Review] Efficient B-mode Ultrasound Image Reconstruction from Sub-sampled RF Data using Deep Learning

Yeo Hun Yoon, Shujaat Khan+2 more|arXiv (Cornell University)|Dec 17, 2017

Advanced MRI Techniques and Applications33 references4 citations

TL;DR

This paper proposes a deep learning method for efficient B-mode ultrasound image reconstruction from sub-sampled RF data by exploiting redundancy in the Rx-Xmit domain using a convolutional neural network (CNN) inspired by low-rank Hankel matrix decomposition. The approach achieves superior image quality with significantly lower computational cost than compressed sensing methods like ALOHA, enabling high-fidelity reconstruction without hardware changes or iterative optimization.

ABSTRACT

In portable, three dimensional, and ultra-fast ultrasound imaging systems, there is an increasing demand for the reconstruction of high quality images from a limited number of radio-frequency (RF) measurements due to receiver (Rx) or transmit (Xmit) event sub-sampling. However, due to the presence of side lobe artifacts from RF sub-sampling, the standard beamformer often produces blurry images with less contrast, which are unsuitable for diagnostic purposes. Existing compressed sensing approaches often require either hardware changes or computationally expensive algorithms, but their quality improvements are limited. To address this problem, here we propose a novel deep learning approach that directly interpolates the missing RF data by utilizing redundancy in the Rx-Xmit plane. Our extensive experimental results using sub-sampled RF data from a multi-line acquisition B-mode system confirm that the proposed method can effectively reduce the data rate without sacrificing image quality.

Motivation & Objective

• To address image degradation from RF data sub-sampling in portable, 3D, and ultra-fast ultrasound systems.
• To overcome limitations of compressed sensing methods that require hardware modifications or suffer from high computational complexity.
• To develop a data-driven deep learning approach that leverages signal redundancy in the Rx-Xmit domain for efficient RF interpolation.
• To achieve high image quality with minimal runtime overhead, suitable for real-time clinical applications.
• To provide a theoretically grounded deep learning framework by linking it to annihilating filter-based low-rank Hankel matrix methods (ALOHA).

Proposed method

• The method uses a deep CNN to directly interpolate missing RF data by exploiting spatial redundancy in the Rx-Xmit and Rx-SL domains.
• The network is designed based on theoretical insights from low-rank Hankel matrix decomposition, specifically the ALOHA framework, to ensure interpretability and performance.
• The architecture uses 28 residual blocks with 3×3 convolutions and 64 filters per layer, enabling efficient feature learning without matrix inversions.
• The model is trained end-to-end on sub-sampled RF data from both linear and convex array transducers to ensure generalization.
• The method operates directly on RF data in the Rx-Xmit domain, avoiding the need for beamforming before reconstruction.
• The approach avoids iterative optimization, resulting in run-time complexity several orders of magnitude lower than ALOHA.

Experimental results

Research questions

• RQ1 Can deep learning effectively interpolate sub-sampled RF data in ultrasound imaging while preserving diagnostic image quality?
• RQ2 How does the performance of the proposed deep learning method compare to traditional compressed sensing and interpolation techniques like ALOHA and linear interpolation?
• RQ3 To what extent does the network generalize across different transducer types (linear vs. convex array) and sub-sampling schemes?
• RQ4 Can a deep learning model be designed with theoretical justification based on signal structure, such as low-rank Hankel matrices?
• RQ5 What is the computational efficiency gain of the proposed method compared to iterative compressed sensing algorithms?

Key findings

• The proposed method achieved 24.11% higher PSNR than sub-sampled input and 19.05% higher than linear interpolation in ×4 Rx sub-sampling.
• In ×8 Rx sub-sampling, the method improved PSNR by 23.19% over sub-sampled input and 25.93% over linear interpolation.
• For 4×2 Rx-Xmit sub-sampling, the method achieved 4.20% higher PSNR than sub-sampled input and 7.19% higher than linear interpolation.
• The method improved CNR by 12.18%, 15.65%, and 14.62% over ALOHA in ×4 Rx, ×8 Rx, and 4×2 Rx-Xmit sub-sampling, respectively.
• The PSNR improvement over ALOHA was 0.98 dB, 1.09 dB, and 1.05 dB in ×4 Rx, ×8 Rx, and 4×2 Rx-Xmit sub-sampling schemes.
• The SSIM improvement over ALOHA was 4.65%, 4.94%, and 1.22% in ×4 Rx, ×8 Rx, and 4×2 Rx-Xmit sub-sampling, respectively.