Source-linked AI summary
Adaptive Ultrasound Beamforming using Deep Learning
Ben Luijten, Regev Cohen, Frederik J. de Bruijn, Harold A. W. Schmeitz, Massimo Mischi, Yonina C. Eldar, Ruud J. G. van Sloun
TL;DR
The paper addresses the computational burden of adaptive ultrasound beamforming while seeking high image quality with limited data and low-rate measurements. It introduces ABLE, a compact neural network structured around adaptive beamforming constraints to predict apodization weights. Across plane-wave and synthetic-aperture imaging, ABLE produces high-quality images with lower computational burden and maintains high contrast under undersampling, though artifacts appear at 25% channel subsampling.
Problem
Adaptive beamforming can improve ultrasound image quality, but its computational burden conflicts with low-cost, real-time imaging and low-rate acquisition.
Method
ABLE uses a compact neural network that adopts adaptive beamforming structure and constraints to learn apodization weights from received RF data.
Results
ABLE provides high-contrast, high-resolution imaging comparable to MV with substantially lower reconstruction time and computational burden across multiple ultrasound systems.
Takeaways & Limitations
ABLE supports fast adaptive ultrasound beamforming and reconstructs high-contrast images from subsampled data, while strong point scatterers remain visible under reduced-channel acquisition.
Takeaways & Limitations
At 25% channel subsampling, artifacts become noticeable, predominantly in speckle regions that are challenging to capture with limited receiving channels.
Abstract
from arXiv · showhide
Biomedical imaging is unequivocally dependent on the ability to reconstruct interpretable and high-quality images from acquired sensor data. This reconstruction process is pivotal across many applications, spanning from magnetic resonance imaging to ultrasound imaging. While advanced data-adaptive reconstruction methods can recover much higher image quality than traditional approaches, their implementation often poses a high computational burden. In ultrasound imaging, this burden is significant, especially when striving for low-cost systems, and has motivated the development of high-resolution and high-contrast adaptive beamforming methods. Here we show that deep neural networks that adopt the algorithmic structure and constraints of adaptive signal processing techniques can efficiently learn to perform fast high-quality ultrasound beamforming using very little training data. We apply our technique to two distinct ultrasound acquisition strategies (plane wave, and synthetic aperture), and demonstrate that high image quality can be maintained when measuring at low data-rates, using undersampled array designs. Beyond biomedical imaging, we expect that the proposed deep~learning based adaptive processing framework can benefit a variety of array and signal processing applications, in particular when data-efficiency and robustness are of importance.
I. INTRODUCTION
Ultrasound offers low-cost, portable, real-time imaging but conventional beamforming limits image quality or computational efficiency. ABLE applies a compact, task-specific neural network to adaptive beamforming across plane-wave and synthetic-aperture imaging.
- Ultrasound is low-cost, portable, minimally invasive, and free of ionizing radiation, but generally provides lower image quality than other modalities.
- Focused sonification increases scan time with scan lines and imaging depth, creating a trade-off between frame rate and lateral resolution.
- Static DAS enables real-time reconstruction but compromises resolution and contrast, especially under unfocused transmissions with wide-area clutter.
- MV improves resolution without compromising contrast by adaptively placing sidelobes, but its computational cost reduces ultrasound interactability.
- ABLE restricts network degrees of freedom by embedding adaptive beamforming structure, enabling data-efficient prediction of apodization weights from received RF data.
- ABLE is evaluated on plane-wave and synthetic-aperture imaging, including subsampled synthetic-aperture transmit designs, against DAS, iMAP, and MV.
- The framework is presented as applicable beyond ultrasound to content-adaptive prediction of signal-processing parameters.
II. EXISTING BEAMFORMERS
Ultrasound beamforming first corrects received channel signals for time of flight, then combines focused channel responses using spatially varying weights. DAS commonly uses fixed window weights such as boxcar, Hanning, or tapered cosine windows.
- Time-of-flight correction: Time-of-flight correction applies geometry-dependent delays to focus each receiving channel on an imaging point.The focused response is generated pixel-wise for every receiving array element.
- Time-of-flight correction: Each receiving channel contributes a pixel-wise response after dynamic focusing to the corresponding imaging point.The channel signal for element n is denoted by xn[t].
- Delay-and-sum beamforming: DAS forms an image by multiplying focused channel contributions by predetermined weights and summing them.The weights may vary spatially across image coordinates.
- Delay-and-sum beamforming: Common DAS weighting choices include boxcar, Hanning, and tapered cosine windows.The boxcar setting is written as w = 1N.
B. Iterative Maximum-A-Posteriori
Adaptive beamformers estimate signal structure from delayed channel data to improve weighting, while regularization and subspace methods address stability and image quality. Regularized MV remains computationally demanding because covariance-matrix inversion dominates its computation, whereas EBMV adds further complexity.
- Iterative Maximum-A-Posteriori: iMAP estimates signal and noise variances, computes a MAP beamformed-signal estimate, and repeats the updates until a stopping criterion.The paper uses iMAP2, defined as two iterations of the algorithm.
- Minimum-variance beamforming: MV beamforming selects pixel-specific apodization weights that minimize output variance while preserving a distortionless response.The optimization uses the sample covariance matrix of delayed receiving-array signals and a steering vector set to a = 1N.
- Minimum-variance beamforming: Regularized MV uses spatial smoothing and diagonal loading to stabilize covariance estimation before solving for the adaptive weights.Diagonal loading adds a trace-determined factor to the covariance-matrix diagonal.
- Minimum-variance beamforming: Covariance-matrix inversion dominates regularized MV computation because the matrix has size N × N.This computational burden applies to the resulting solution for the adaptive weights.
- Eigen-based minimum variance beamforming: EBMV projects the signal subspace formed by dominant desired eigenvectors onto the weight vector to update the beamforming weights.The procedure is repeated for every pixel to produce an EBMV-beamformed image.
III. ABLE: ADAPTIVE BEAMFORMING BY DEEP LEARNING
ABLE replaces a computationally expensive adaptive processor with a neural network that predicts apodization weights within a structured beamforming pipeline. Its compact fully connected architecture uses an inner latent space and dropout to support compact representations and reduce overfitting.
- ABLE architecture: ABLE adaptively calculates apodization weights from input channel data using a neural network inspired by MV beamforming.The network tunes reconstruction parameters within a predictable beamforming structure.
- ABLE architecture: The network produces a beamformed pixel by multiplying time-of-flight-corrected channel data by neural-network-generated apodization weights.The weight-generating network is denoted fθ : RN → RN.
- Network design: ABLE uses four fully connected layers, with N nodes in the outer layers and N/4 nodes in the inner layers.The inner layers create a compact latent space tied to the receiving aperture size N.
- Network design: The N/4-node inner layers force a compact data representation that aids noise suppression.This bottleneck is part of the stated architectural rationale.
- Network design: Dropout with probability 0.2 is applied between every fully connected layer to prevent overfitting.Nodes are randomly dropped during the network operation described in the architecture.
B. Activation functions
The proposed antirectifier layer is designed for RF data, preserving both positive and negative signal components while introducing nonlinearity. This addresses limitations of ReLU and bounded tanh activations in ultrasound beamforming networks.
- B. Activation functions: ReLU may impair RF-data training because its nonnegative output produces many ‘dying’ nodes for signals containing positive and negative values.ReLUs are computationally efficient and provide sparse representations, but their suitability is limited for signed RF data.
- B. Activation functions: Unlike ReLU, tanh preserves negative values but is bounded between -1 and 1 and therefore tends to saturate.
- B. Activation functions: The antirectifier combines sample-wise ℓ2 normalization with concatenated positive and negative input components.It introduces nonlinearity while preserving negative signal contributions and the input’s dynamic range.
1) Image Loss:
The image loss is designed to make ABLE’s output resemble a high-quality beamformed target, while the training setup uses acquired and simulated plane-wave and synthetic-aperture data. The target beamformer is empirically fine-tuned EBMV.
- 1) Image Loss:: SMSLE is introduced because logarithmic compression and ultrasound’s large backscattering dynamic range make MSE and MAE poorly balanced image-deviation measures.The loss splits beamformed data into positive and negative parts and sums their mean-squared-logarithmic errors.
- 1) Image Loss:: The datasets comprise 1000 acquired frames for each of plane-wave and synthetic-aperture imaging, plus 100 simulated point-scatterer frames for each method.The acquisitions include in-vivo carotid, coronary-artery, phantom, and stent data.
- 1) Image Loss:: ABLE maps time-of-flight-corrected RF inputs to high-quality beamformed targets generated with an empirically fine-tuned EBMV beamformer.
B. Validation
Validation measures resolution, contrast, and computational complexity for the beamforming methods. Resolution uses point scatterers, contrast uses low- and high-intensity regions, and complexity counts only adaptive operations.
- B. Validation: Resolution is assessed using lateral and axial FWHM measurements of simulated point scatterers, with polar-domain lateral evaluation for circular-array synthetic-aperture scans.
- B. Validation: CNR measures contrast between low- and high-intensity regions using simulated anechoic cysts for plane-wave imaging and an in-vivo coronary artery for synthetic-aperture imaging.The high- and low-intensity regions are represented by their means and variances.
- B. Validation: Computational complexity is evaluated in FLOPs while excluding time-of-flight correction, array-signal multiplication, and summation shared by all beamformers.The comparison therefore focuses on adaptive complexity.
1) iMAP:
The computational analysis accounts for the operations required by iMAP, regularized MV, and EBMV beamforming. It also describes beam-profile and subsampling evaluations involving plane-wave and synthetic-aperture imaging.
- 1) iMAP:: Each iMAP iteration estimates signal and noise variances and updates the MAP estimate, with total complexity depending on the number of iterations I.
- 1) iMAP:: Regularized MV complexity is dominated by covariance-matrix inversion with complexity O(N^3), plus operations for calculating the weight vector.
- 1) iMAP:: EBMV adds an eigen-decomposition of complexity N^3 and a dominant-eigenvector projection to the regularized MV computation.
- 1) iMAP:: The implementation projects approximately 50% of the eigenvectors, using k = 0.5.
- 1) iMAP:: Normalized beam profiles are evaluated along simulated point scatterers for plane-wave and synthetic-aperture imaging.
- 1) iMAP:: The proposed subsampling schemes compare randomly selected active array elements with a deterministic selection pattern.Black boxes indicate active elements in the visualization.
3) ABLE:
ABLE’s inference complexity is computed from its fully connected layers and activation functions, while its subsampling variant trains on reduced array signals with fully sampled EBMV targets.
- Inference complexity: A fully connected layer with Ni inputs and Ni+1 outputs requires 2NiNi+1 + Ni+1 FLOPs during inference.The calculation models the layer as Ax + b.
- Inference complexity: ReLU activations require 2 FLOPs each, while antirectifier activations require 4 FLOPs per node; dropout adds no inference cost.The resulting network FLOP expression includes the fully connected layers and activation costs.
- Subsampling: ABLE’s subsampling evaluation uses random or transmit-proximal receiving-element subsets at 25% and 50% input rates.The transmit-proximal scheme favors specular reflections.
- Subsampling: The network is trained on subsampled array signals using fully sampled EBMV reconstructions as targets.Two receiving-element selection schemes are evaluated.
V. RESULTS
ABLE produces high-contrast, high-resolution ultrasound images across plane-wave and synthetic-aperture settings while substantially reducing computational cost. It also reconstructs undersampled data, though artifacts and strategy-dependent image differences remain.
- Image quality: ABLE outperforms DAS and iMAP2 on axial resolution, lateral resolution, and CNR for both plane-wave and synthetic-aperture imaging.It even achieves better lateral resolution than its EBMV training target.
- Plane-wave results: ABLE and EBMV produce narrower sidelobes and stronger clutter suppression than DAS in plane-wave images.The comparison includes simulated point scatterers and carotid artery images.
- Synthetic-aperture results: In synthetic-aperture images, ABLE and EBMV reduce sidelobe intensity, clarify stent struts, and reveal coronary-plaque substructures.The improvements are associated with reduced clutter and improved resolution.
- Computational complexity: ABLE requires 71,232 FLOPs for the 128-channel linear array and 47,930 FLOPs for the 105-channel IVUS array, versus 6,186,599 and 414,994 for EBMV.In practice, this produced an approximately 400-fold speedup.
- Subsampling: ABLE reconstructs high-contrast images from both random and deterministic subsampling, but artifacts become especially visible at 25% channel subsampling.Strong point scatterers remain clearly visible, whereas speckle regions are more difficult to reconstruct.
- Subsampling: Random subsampling yields wider sidelobes but less pronounced grating lobes and better overall contrast than deterministic subsampling.The contrast difference is attributed to stronger suppression of noise and clutter in mostly empty regions.
- Computational complexity: ABLE has lower computational cost than matrix-inversion-based MV methods across array sizes, with the advantage becoming especially important for large arrays.The comparison is presented in FLOPs as a function of input array size.
- Limitations and future work: The method’s data efficiency comes partly from pixel-wise operation, which prevents it from using neighboring-pixel context to suppress some artifacts such as grating lobes.The paper identifies spatially aware beamformers as a direction for future development.
VII. CONCLUSION
ABLE uses a compact, model-based deep-learning architecture to reconstruct high-quality ultrasound images across multiple imaging systems. It achieves resolution and contrast comparable to a complex adaptive beamformer with substantially lower reconstruction time and computational burden.
- ABLE is a compact, model-based deep-learning architecture for improving conventional ultrasound beamforming.
- Its resolution and contrast are up to par with a complex state-of-the-art adaptive beamformer.
- ABLE substantially reduces reconstruction time and computational burden compared with complex adaptive beamforming.
- The approach paves the way toward real-time adaptive beamforming in ultrasound systems.