Switchable Deep Beamformer

Shujaat Khan, Jaeyoung Huh, Jong Chul YearXiv 2020 · arXiv · OpenReview

Abstract

Recent proposals of deep beamformers using deep neural networks have attracted significant attention as computational efficient alternatives to adaptive and compressive beamformers. Moreover, deep beamformers are versatile in that image post-processing algorithms can be combined with the beamforming. Unfortunately, in the current technology, a separate beamformer should be trained and stored for each application, demanding significant scanner resources. To address this problem, here we propose a switchable deep beamformer that can produce various types of output such as DAS, speckle removal, deconvolution, etc., using a single network with a simple switch. In particular, the switch is implemented through Adaptive Instance Normalization (AdaIN) layers, so that various output can be generated by merely changing the AdaIN code. Experimental results using B-mode focused ultrasound confirm the flexibility and efficacy of the proposed methods for various applications.

Introduction

In ultrasound (US) imaging, beamforming is the critical process of converting raw signals from probe elements into a coherent image. While the traditional Delay-and-Sum (DAS) method remains the industry standard, it often produces blurry images characterized by significant speckle noise. Recent deep learning approaches have significantly improved image quality by integrating beamforming with post-processing tasks like deconvolution or noise reduction.

However, these advancements face a major practical hurdle: model bloat. Current technology typically requires a separate, memory-heavy neural network for each specific output type (e.g., one for despeckling, another for resolution enhancement). Storing and switching between multiple massive models is computationally expensive and memory-intensive, making it impractical for resource-constrained hardware in portable medical ultrasound scanners. This paper introduces a "Switchable Deep Beamformer" that performs multiple reconstruction tasks within a single neural network by toggling a simple "style code."

Methods

The proposed architecture combines a "Universal Deep Beamformer" (DeepBF) backbone with Adaptive Instance Normalization (AdaIN) layers. This design allows the network to learn a shared representation of ultrasound physics while specializing the final output via a switch.

Architecture and the AdaIN Switch

The backbone is a deep convolutional neural network that processes raw radio frequency (RF) data. The "switch" is implemented using AdaIN, a technique originally designed for artistic style transfer. AdaIN works by adjusting the mean and variance of feature maps in the network's bottleneck. Instead of altering the entire network's weights, the user inputs a small "style code" vector. This code passes through a generator that instructs the AdaIN layers on how to transform internal data to produce a specific output, such as DAS, Despeckle, or Deconvolution.

Architecture of the Switchable Deep Beamformer featuring AdaIN layers for style-based output selection.

Mathematically, this process is grounded in Optimal Transport theory. The authors suggest that AdaIN provides a minimal-cost path (Wasserstein-2 distance) to transform features from one signal domain to another, proving that adjusting mean and variance is a theoretically sound way to map between Gaussian feature distributions.

Experimental Setup

The model was evaluated using three distinct datasets collected with an E-CUBE 12R US system:

  1. In-vivo Carotid/Thyroid: 400 images from 10 volunteers for training and testing.
  2. In-vivo Calf/Forearm: 100 images used as an independent test set to evaluate generalization across different anatomies.
  3. Tissue-Mimicking Phantom: 218 frames of standardized phantoms (anechoic and hyper-echoic regions) to provide objective measurements of contrast and resolution.

Results and Discussion

The switchable beamformer successfully produced four distinct output types—DAS, Despeckle, Deconvolution, and a combined Deconv+Despeckle—using a single set of weights.

Quantitative and Qualitative Performance

The model demonstrated superior performance compared to standard methods. Quantitative gains were observed in the Generalized Contrast-to-Noise Ratio (GCNR). For instance, in DAS mode, the switchable beamformer's GCNR was consistently higher than conventional DAS by 0.005 to 0.02 units.

Comparison of various beamforming outputs (DAS, Despeckle, Deconvolution) generated by the single switchable network on phantom data.

Efficiency and Generalization

A key advantage is speed; the average reconstruction time per depth plane was approximately 0.165 milliseconds. This is significantly faster than previous universal deep beamformers (~4.8 ms), enabling real-time clinical application. Furthermore, the model generalized well to the independent calf/forearm dataset, indicating that it learns fundamental ultrasound physics rather than just memorizing specific anatomical structures.

Conclusion

The Switchable Deep Beamformer addresses the challenge of model bloat in deep learning-based ultrasound imaging. By treating different signal processing algorithms as "styles" controlled via AdaIN, the system achieves high versatility with minimal memory overhead. While the current model is trained to mimic existing math-based filters, future work could explore unsupervised learning to discover novel image enhancement techniques. This approach paves the way for high-performance, multi-functional ultrasound imaging on portable and resource-constrained devices.

Summary updated

Abstract

Recent proposals of deep beamformers using deep neural networks have attracted significant attention as computational efficient alternatives to adaptive and compressive beamformers. Moreover, deep beamformers are versatile in that image post-processing algorithms can be combined with the beamforming. Unfortunately, in the current technology, a separate beamformer should be trained and stored for each application, demanding significant scanner resources. To address this problem, here we propose a {\em switchable} deep beamformer that can produce various types of output such as DAS, speckle removal, deconvolution, etc., using a single network with a simple switch. In particular, the switch is implemented through Adaptive Instance Normalization (AdaIN) layers, so that various output can be generated by merely changing the AdaIN code. Experimental results using B-mode focused ultrasound confirm the flexibility and efficacy of the proposed methods for various applications.
BibTeX
@misc{khan2020switchable,
  author = {Shujaat Khan and Jaeyoung Huh and Jong Chul Ye},
  title = {{Switchable Deep Beamformer}},
  year = {2020},
  eprint = {2008.13646},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2008.13646},
}
Figure