Deep Learning-Based Universal Beamformer for Ultrasound Imaging

Shujaat Khan, Jaeyoung Huh, Jong Chul Ye

TL;DR

Researchers have developed "DeepBF," an end-to-end deep learning beamformer that replaces the traditional "Delay-and-Sum" (DAS) algorithm in ultrasound imaging. Unlike previous methods that merely use AI to fill in missing data, this single neural network directly transforms raw signals into high-quality images across different acquisition schemes and significant data subsampling rates.

What Problem Does This Paper Address?

In ultrasound imaging, the standard way to create an image is through beamforming. The most common method, Delay-and-Sum (DAS), works by applying time delays to signals received by different transducer elements and then summing them up. While DAS is simple and fast, it has a major weakness: it assumes a constant speed of sound and requires a high volume of data to produce clear images.

When researchers try to speed up imaging or reduce power consumption by "subsampling" (collecting less data), DAS performance plummets, resulting in blurry images with low contrast. Previous deep learning attempts tried to fix this by using neural networks to interpolate (guess) the missing data before handing it off to a standard DAS beamformer. This paper argues that we should skip the "band-aid" interpolation and let a neural network handle the entire beamforming process from start to finish.

Methods

The authors propose DeepBF, a deep convolutional neural network (CNN) designed to be a "universal" beamformer.

  • End-to-End Architecture: Instead of just pre-processing data, the network takes raw (or delayed) RF channel data as input and directly outputs beamformed IQ (In-phase/Quadrature) data.
  • The Network: It consists of 27 convolutional layers featuring a "contracting path" with concatenation, batch normalization, and ReLU activation functions.
  • Contextual Input: To improve accuracy, the model doesn't just look at one "slice" of data; it takes a 3D data cube consisting of three adjacent depth planes to generate the target image for the middle plane.
  • Versatility: The same architecture was tested on two very different acquisition methods: Focused B-mode imaging (traditional) and Planewave imaging (ultrafast).

Datasets

The study utilized three primary data sources:

  1. In-vivo Human Data: RF data from the carotid artery area of 10 volunteers, acquired using an Alpinion E-CUBE 12R system with a linear array transducer.
  2. Phantom Data: Data from an ATS-539 multipurpose tissue-mimicking phantom, used specifically to test how well the model generalizes to new, unseen objects.
  3. Simulated Subsampling: The researchers took full-rate data and artificially removed channels to create six different subsampling levels (from 1x down to 16x).

Key Insights and Ideas

  • Universal Expressiveness: The most striking finding is that a single network architecture can handle various detector configurations and subsampling rates. This suggests that the "physics" of beamforming can be generalized within the weights of a deep network.
  • The "Super-Resolution" Effect: Surprisingly, DeepBF improved image quality even when provided with full data (no subsampling). The authors attribute this to a "performance boosting" behavior similar to what is seen in super-resolution AI, where the network learns to eliminate artifacts and optimize dynamic range better than the mathematical assumptions of DAS.
  • Simplicity Wins: By removing the intermediate interpolation step used in previous state-of-the-art methods, the pipeline becomes simpler and more robust.

Key Results

  • Contrast and Clarity: In focused mode, DeepBF 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.
  • Robustness to Data Loss: Even at 16x subsampling (using only 4 channels instead of 64), DeepBF produced images with a CNR of 1.17, which is remarkably close to what DAS achieves with the full 64-channel dataset (1.38).
  • Speed: The average reconstruction time was approximately 4.8 milliseconds per depth plane, making it a viable candidate for real-time clinical use if optimized further.
  • Generalization: The model trained on human carotid data performed excellently on the tissue phantom, proving it wasn't just "memorizing" human anatomy but actually learning how to beamform.

Limitations and Future Work

While the results are promising, the authors note that the current implementation was tested on specific linear array transducers. Future work would likely involve testing the "universality" across different probe types (like curved or phased arrays).

Additionally, while 4.8ms is fast, clinical ultrasound requires processing thousands of planes per second for high-frame-rate imaging. The authors suggest that optimized hardware implementation (like specialized GPUs or FPGAs) will be necessary to bring this "Deep Beamformer" into live clinical scanners for applications like cardiac or fetal imaging.

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},
}

Loading related content...

Figure