Differentiable Beamforming for Distributed Attenuation Estimation and Spatial Gain Compensation (SGC)

Benjamin N. Frey1, Dongwoon Hyun2,3, Walter Simson3, Thurston Brevett4,
Louise Zhuang4, Jihye Baek3, Sergio J. Sanabria3, Jeremy J. Dahl3
Departments of 1Applied Physics, 3Radiology, and 4Electrical Engineering, Stanford University, Stanford, CA 94305
2Siemens Healthineers, Palo Alto, CA 94305
Paper Video

Overview

Ultrasound imaging is a non-invasive and cost-effective diagnostic tool widely used in clinical practice. However, ultrasound image quality can be significantly degraded by spatially-varying acoustic attenuation in tissue, leading to artifacts such as shadowing and reduced contrast. To address this challenge, we propose a novel framework called Spatial Gain Compensation (SGC) that leverages differentiable beamforming and automatic differentiation to estimate distributed attenuation maps and normalize ultrasound images. The goal of SGC is to provide normalized ultrasound images with improved visibility and contrast, thereby enhancing diagnostic accuracy.

Spatial Gain Compensation Framework

Spatial Gain Compensation (SGC) utilizes iterative optimization with automatic differentiation to provide ultrasound image normalization via distributed attenuation estimation.

Technical approach diagram

(1) First, an initial estimate of the attenuation map is made which is used to develop a spatially-varying over-unity apodization pattern. (2) A synthetic-aperture ultrasound image is beamformed using these apodization profiles. (3) A novel log-difference subaperture image quality metric is computed. Here, with proper spatial gain compensation, all sub-aperture reconstructions should be similar (despite differences in "views from different sub-apertures"). (4) Sub-aperture image differences are effectively loss values that can be backpropagated through the differentiable beamformer to update the attenuation map.

Simulation Results

Model performance on heterogeneous attenuation k-Wave simulations was evaluated. Five simulation layouts were created: (1) a positive-contrast attenuating inclusion (attenuation greater than the background), (2) a negative-contrast attenuating inclusion (attenuation less than the background), (3) another positive-contrast attenuating inclusion with a smaller diameter, (4) hyperechoic inclusion in a uniform attenuation background, and (5) a hypoechoic inclusion in a uniform attenuation background.

Simulation sample results 1-3 show that the attenuation inclusions are accurately estimated, and image normalization is visually apparent, especially in the reduction of shadowing beneath inclusion. Simulation sample results 4-5 show that the attenuation estimates are uniform, and the model is not misled by echogenicity changes.

The blue squares in the B-mode images indicate the regions where the image quality metric was computed. For more information on this metric, please see our proceedings paper.

Experimental Results

Model performance on experimental phantom datasets were investigated using a Verasonics L12-3v linear array and CIRS 040 GSE bi-partition (one-half 0.5 dB/cm/MHz, one-half 0.7 dB/cm/MHz) attenuation phantom. In total, three data samples were obtained: (1) one acquisition where the transducer straddled the attenuation boundary, (2) one acquisition where the transducer was positioned over the high-attenuation region with inclusions lower than the background, and (3) one acquisition where the transducer was titled so as to layer the attenuation regions.

In the bi-partition sample result (1), significant image normalization occurs and spatial gain compensation is visually apparent. In the inclusion sample result (2), the inclusions are not visible in the attenuation estimate, but vertical bands form near the lateral extents of the inclusion. In the bilayer sample result (3), the boundary is not detected.

Related Projects

Check out some related projects on adaptive ultrasound imaging with automatic differentiation!

Differentiable Beamforming for Ultrasound Autofocusing

Walter Simson, Louise Zhuang, Sergio Senabria, Neha Antil, Jeremy Dahl, Dongwoon Hyun

MICCAI, 2023

project page / video / pdf

Applying differentiable beamforming to auto-focus ultrasound and high-resolution estimate sound speed in tissue.

BibTeX


      @inproceedings{frey2024differentiable,
        title={Differentiable Beamforming for Distributed Attenuation Estimation and Spatial Gain Compensation (SGC)},
        author={Frey, Benjamin N and Hyun, Dongwoon and Simson, Walter and Brevett, Thurston and Zhuang, Louise and Baek, Jihye and Sanabria, Sergio J and Dahl, Jeremy J},
        booktitle={2024 IEEE Ultrasonics, Ferroelectrics, and Frequency Control Joint Symposium (UFFC-JS)},
        pages={1--5},
        year={2024},
        organization={IEEE}
      }