Adaptive and Compressive Beamforming using Deep Learning for Medical Ultrasound

Shujaat Khan, Jaeyoung Huh, Jong Chul YearXiv 2019 · arXiv · OpenReview

Abstract

In ultrasound (US) imaging, adaptive beamforming techniques aim to improve the resolution and contrast-to-noise ratio (CNR) compared to standard delay-and-sum (DAS) beamformers. However, these methods often degrade under inaccurate modeling or reduced channel counts. This paper introduces DeepBF, a deep learning-based beamformer designed to process full or sub-sampled radio-frequency (RF) data. By leveraging a convolutional neural network (CNN), DeepBF generates high-quality ultrasound images across varying measurement conditions and subsampling rates. Experimental results demonstrate significant improvements in image quality and robustness to data compression.

Introduction

The industry standard for ultrasound image reconstruction, Delay-and-Sum (DAS), is computationally efficient but prone to low resolution and side-lobe artifacts. While adaptive beamformers like the Minimum Variance (MV) method offer superior resolution, they are computationally intensive and sensitive to noise. Furthermore, the advent of 3D ultrasound necessitates compressive beamforming to manage massive data rates. Existing compressive techniques often rely on slow iterative optimization. DeepBF addresses these challenges by providing a single, fast, and adaptive solution for both high-quality beamforming and data compression.

Methodology

DeepBF utilizes an encoder-decoder CNN architecture to map raw, time-delayed RF data directly to In-phase/Quadrature (IQ) data. To manage GPU memory constraints while exploiting spatio-temporal redundancies, the network processes "sub-cubes" of three adjacent depth planes.

The DeepBF pipeline: from raw RF data to log-compressed B-mode images.

A key theoretical insight is that the CNN, through its ReLU activations, acts as a collection of distinct linear mappings. This allows the network to exhibit "input-dependent adaptivity," effectively mimicking the behavior of expensive adaptive beamformers by adjusting its weights based on the incoming signal. The model is trained on diverse targets, including DAS, MV, and deconvolution-based targets, allowing it to learn to reverse physical blurring artifacts.

Results and Discussion

The efficacy of DeepBF was evaluated using in vivo human scans (carotid and thyroid) and tissue-mimicking phantoms.

Visual comparison of DAS, MV, and DeepBF on phantom data, showing superior contrast and resolution.

Key findings include:

  • Enhanced Contrast: DeepBF achieved a gain of 1.23 dB in Contrast Recovery (CR) and 0.07 units in CNR over DAS in phantom tests.
  • Robust Sub-sampling: The model maintained high image quality even with 75% to 93% channel reduction (4x to 16x sub-sampling), where traditional DAS and MV methods failed.
  • Generalization: A model trained on neck scans generalized effectively to leg muscle scans at different transducer frequencies.
  • Efficiency: With a reconstruction time of approximately 4.8 ms per depth plane, DeepBF is significantly faster than iterative optimization methods, though further optimization is required for full-frame real-time clinical use.

Conclusion

DeepBF represents a significant step toward universal, adaptive ultrasound beamforming. By replacing traditional pipelines with a single CNN, researchers have demonstrated a robust method for high-resolution imaging that is resilient to extreme data compression. Future work will focus on end-to-end learning of time-delays and further hardware acceleration for real-time deployment.

Summary updated

Abstract

In ultrasound (US) imaging, various types of adaptive beamforming techniques have been investigated to improve the resolution and contrast-to-noise ratio of the delay and sum (DAS) beamformers. Unfortunately, the performance of these adaptive beamforming approaches degrade when the underlying model is not sufficiently accurate and the number of channels decreases. To address this problem, here we propose a deep learning-based beamformer to generate significantly improved images over widely varying measurement conditions and channel subsampling patterns. In particular, our deep neural network is designed to directly process full or sub-sampled radio-frequency (RF) data acquired at various subsampling rates and detector configurations so that it can generate high quality ultrasound images using a single beamformer. The origin of such input-dependent adaptivity is also theoretically analyzed. Experimental results using B-mode focused ultrasound confirm the efficacy of the proposed methods.
BibTeX
@misc{khan2019adaptive,
  author = {Shujaat Khan and Jaeyoung Huh and Jong Chul Ye},
  title = {{Adaptive and Compressive Beamforming using Deep Learning for Medical Ultrasound}},
  year = {2019},
  eprint = {1907.10257},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.10257},
}
Figure