Deep Learning-based Universal Beamformer for Ultrasound Imaging

Shujaat Khan, Jaeyoung Huh, Jong Chul Ye

Abstract

In ultrasound (US) imaging, individual channel RF measurements are back-propagated and accumulated to form an image after applying specific delays. While this time reversal is usually implemented using a hardware- or software-based delay-and-sum (DAS) beamformer, the performance of DAS decreases rapidly in situations where data acquisition is not ideal. Herein, for the first time, we demonstrate that a single data-driven adaptive beamformer designed as a deep neural network can generate high quality images robustly for various detector channel configurations and subsampling rates. The proposed deep beamformer is evaluated for two distinct acquisition schemes: focused ultrasound imaging and planewave imaging. Experimental results showed that the proposed deep beamformer exhibit significant performance gain for both focused and planar imaging schemes, in terms of contrast-to-noise ratio and structural similarity.

Introduction

Standard ultrasound image reconstruction relies on the Delay-and-Sum (DAS) algorithm. DAS applies time delays to signals from transducer elements and sums them. While efficient, DAS assumes a constant speed of sound and requires high data volumes. When data is subsampled to increase imaging speed or reduce power, DAS performance degrades, leading to low-contrast, blurry images. Previous deep learning solutions attempted to interpolate missing data before DAS. This paper introduces DeepBF, an end-to-end deep learning beamformer that replaces DAS entirely, directly transforming raw signals into high-quality images.

Methods

DeepBF is a deep convolutional neural network (CNN) designed as a universal beamformer. Unlike modular approaches, it takes raw or delayed RF channel data as input and directly outputs beamformed IQ (In-phase/Quadrature) data.

Architecture of the DeepBF network for focused (a) and planewave (b) imaging modes.

The architecture consists of 27 convolutional layers featuring a contracting path with concatenation, batch normalization, and ReLU activations. To capture spatial context, the model processes a 3D data cube of three adjacent depth planes to reconstruct the middle plane. This architecture proved versatile, performing effectively across both traditional focused B-mode imaging and ultrafast planewave imaging.

The model was trained and validated using in-vivo human carotid artery data, tissue-mimicking phantom data, and simulated subsampling (from 1x to 16x).

Results and Discussion

DeepBF demonstrated superior performance over DAS across all metrics. In focused mode, it improved the Contrast-to-Noise Ratio (CNR) from 1.69 to 2.16 and the Generalized CNR (GCNR) from 0.74 to 0.83 for fully sampled data.

Comparison of DAS and DeepBF performance across various subsampling factors.

A key finding is the "Super-Resolution" effect: DeepBF improved image quality even with full data, likely by learning to eliminate artifacts and optimize dynamic range better than DAS's rigid mathematical assumptions. Furthermore, its robustness is remarkable; at 16x subsampling (using only 4 channels), DeepBF achieved a CNR of 1.17, nearly matching DAS's performance with a full 64-channel set (1.38).

With an average reconstruction time of 4.8 milliseconds per depth plane, DeepBF is a strong candidate for real-time clinical applications. Its successful generalization from human data to phantom data indicates it learns fundamental beamforming principles rather than just memorizing anatomy.

Conclusion

DeepBF represents a significant shift toward data-driven ultrasound reconstruction. While currently tested on linear arrays, future work will explore its application to curved or phased arrays. Achieving the throughput required for live clinical scanning (thousands of planes per second) will likely necessitate optimized hardware implementations on GPUs or FPGAs.

Summary updated

Abstract

In ultrasound (US) imaging, individual channel RF measurements are back-propagated and accumulated to form an image after applying specific delays. While this time reversal is usually implemented using a hardware- or software-based delay-and-sum (DAS) beamformer, the performance of DAS decreases rapidly in situations where data acquisition is not ideal. Herein, for the first time, we demonstrate that a single data-driven adaptive beamformer designed as a deep neural network can generate high quality images robustly for various detector channel configurations and subsampling rates. The proposed deep beamformer is evaluated for two distinct acquisition schemes: focused ultrasound imaging and planewave imaging. Experimental results showed that the proposed deep beamformer exhibit significant performance gain for both focused and planar imaging schemes, in terms of contrast-to-noise ratio and structural similarity.
BibTeX
@inproceedings{khan2019deep,
  author = {Shujaat Khan and Jaeyoung Huh and Jong Chul Ye},
  title = {{Deep Learning-based Universal Beamformer for Ultrasound Imaging}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  url = {https://openreview.net/forum?id=lrgeRY9R3H},
}
Figure