Ultrasound Signal Processing: From Models to Deep Learning

Ben Luijten, Nishith Chennakeshava, Yonina C. Eldar, Massimo Mischi, Ruud J. G. van Sloun

Abstract

Medical ultrasound imaging relies heavily on high-quality signal processing algorithms to provide reliable and interpretable image reconstructions. Hand-crafted reconstruction methods, often based on approximations of the underlying measurement model, are useful in practice, but notoriously fall behind in terms of image quality. More sophisticated solutions, based on statistical modelling, careful parameter tuning, or through increased model complexity, can be sensitive to different environments. Recently, deep learning based methods have gained popularity, which are optimized in a data-driven fashion. These model-agnostic methods often rely on generic model structures, and require vast training data to converge to a robust solution. A relatively new paradigm combines the power of the two: leveraging data-driven deep learning, as well as exploiting domain knowledge. These model-based solutions yield high robustness, and require less trainable parameters and training data than conventional neural networks. In this work we provide an overview of these methods from the recent literature, and discuss a wide variety of ultrasound applications. We aim to inspire the reader to further research in this area, and to address the opportunities within the field of ultrasound signal processing. We conclude with a future perspective on these model-based deep learning techniques for medical ultrasound applications.

Introduction

Medical ultrasound is a cornerstone of modern diagnostics due to its real-time capabilities, non-ionizing nature, and cost-effectiveness. The core of ultrasound imaging lies in the signal processing pipeline, particularly beamforming—the transformation of raw echoes into coherent images. Traditionally, this has relied on physics-based models that, while fast, often oversimplify the complexities of human tissue. Conversely, modern data-driven deep learning offers high performance but often lacks the interpretability and reliability required for clinical use. This paper explores the evolution of these techniques and highlights the emergence of Model-Based Deep Learning as a powerful middle ground that marries physical domain knowledge with neural network flexibility.

The Spectrum of Signal Processing Approaches

The field can be categorized into three distinct paradigms based on how they balance physical models and data-driven learning. These range from purely model-based algorithms to entirely data-driven deep learning architectures.

The spectrum of ultrasound signal processing approaches, from traditional model-based algorithms to data-driven deep learning.
  1. Model-Based Algorithms: Traditional methods derived from first principles, such as Delay-and-Sum (DAS) or Minimum Variance (MV) beamforming. While DAS is fast, it assumes a constant speed of sound, leading to artifacts. MV is more adaptive but computationally intensive.
  2. Data-Driven Deep Learning: These approaches use generic architectures (e.g., U-Nets) to replace the processing chain entirely. While flexible, they often ignore the underlying physics of wave propagation, requiring massive datasets to generalize.
  3. Model-Based Deep Learning: This "sweet spot" embeds physical domain knowledge directly into neural network architectures. By using techniques like Deep Unfolding, where iterative algorithms are mapped to network layers, or Hybrid Architectures, where specific model components are replaced by small networks, researchers achieve high robustness with fewer parameters.

Adaptive Beamforming by Deep Learning (ABLE)

A prime example of model-based deep learning is the ABLE method. Traditional adaptive beamforming (like MV) requires calculating complex weights to suppress noise, which is computationally slow. ABLE replaces this specific calculation with a small, specialized neural network while maintaining the overall physical structure of the beamformer.

The ABLE architecture, an example of model-based deep learning that adaptively calculates beamforming weights.

By forcing the network to operate within the framework of a beamforming operation, it achieves the quality of MV beamforming at a fraction of the computational cost, making high-resolution imaging viable for real-time clinical use.

Datasets and Benchmarking

Evaluating these diverse methods requires standardized data. The paper highlights several key resources:

  • CUBDL (Challenge on Ultrasound Beamforming with Deep Learning): A benchmark dataset providing a common ground for evaluating deep learning beamformers.
  • Simulated Data: Tools like Field II are used to generate ground-truth reflectivity maps, which are impossible to obtain from real human tissue.
  • In-vivo Data: Real-world testing on volunteers and clinical scans ensures that models generalize beyond simulation and handle the variability of human anatomy.

Key Insights and Performance Gains

The integration of physics and AI yields several critical advantages over "black box" approaches:

  • Inductive Bias: By constraining networks with physical models, they require significantly less training data and generalize better to new patients and transducer geometries.
  • Efficiency: Hybrid methods like ABLE achieve a 2 to 3 order of magnitude reduction in computational complexity compared to traditional adaptive beamforming.
  • Data Compression: Techniques like LISTA (Learned Iterative Shrinkage and Thresholding Algorithm) have demonstrated an 88% reduction in data rates for 3D ultrasound without sacrificing quality.
  • Contrast Enhancement: Methods like ADMIRE show a 10-20dB improvement in Contrast-to-Noise Ratio (CNR) by explicitly modeling and removing clutter and multi-path scattering.

Conclusion and Future Perspectives

While model-based deep learning shows great promise, challenges remain. Clinical adoption requires a shift toward software-based reconstruction on powerful GPUs. Furthermore, ensuring that these models do not "hallucinate" features is critical for clinical safety. Future research aims to develop "universal" beamformers that can adapt to varying transducer geometries and frequencies on the fly, further bridging the gap between acoustic physics and artificial intelligence.

Summary updated

Abstract

Medical ultrasound imaging relies heavily on high-quality signal processing algorithms to provide reliable and interpretable image reconstructions. Hand-crafted reconstruction methods, often based on approximations of the underlying measurement model, are useful in practice, but notoriously fall behind in terms of image quality. More sophisticated solutions, based on statistical modelling, careful parameter tuning, or through increased model complexity, can be sensitive to different environments. Recently, deep learning based methods have gained popularity, which are optimized in a data-driven fashion. These model-agnostic methods often rely on generic model structures, and require vast training data to converge to a robust solution. A relatively new paradigm combines the power of the two: leveraging data-driven deep learning, as well as exploiting domain knowledge. These model-based solutions yield high robustness, and require less trainable parameters and training data than conventional neural networks. In this work we provide an overview of these methods from the recent literature, and discuss a wide variety of ultrasound applications. We aim to inspire the reader to further research in this area, and to address the opportunities within the field of ultrasound signal processing. We conclude with a future perspective on these model-based deep learning techniques for medical ultrasound applications.
BibTeX
@misc{luijten2022ultrasound,
  author = {Ben Luijten and Nishith Chennakeshava and Yonina C. Eldar and Massimo Mischi and Ruud J. G. van Sloun},
  title = {{Ultrasound Signal Processing: From Models to Deep Learning}},
  year = {2022},
  eprint = {2204.04466},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2204.04466},
}
Figure