ALEVA

Insights

Ultrasound Frequency Compounding: Measure Independent Information, Not Band Count

Evaluate ultrasound frequency compounding through inter-band correlation, spatial response and implementation cost, with a reproducible covariance model and source-led OEM acceptance guidance.

By ALEVA | Orange County, California, USA

When an ultrasound system has a fixed usable bandwidth, how many frequency bands should its image-processing pipeline combine? The engineering decision should account for correlation between the images, the spatial response of each band, and the cost of processing them. A smoother image is useful evidence, but it cannot establish that small targets remain distinguishable.

Separate a filter-bank choice from an acquisition change

Frequency compounding combines images with different spectral content. In a receive-side implementation, a retained RF acquisition can feed several filters, followed by envelope or intensity detection and combination. The release specification should identify exactly where filtering occurs, how bands are normalized, and whether combination acts on complex samples, envelopes, intensities, or displayed pixels. Those operations define different outputs.

Li and colleagues examined Gaussian-pulse frequency compounding through analysis, simulation and phantom experiments. For a fixed band width and fixed total spectral span, their results show that speckle reduction approaches a limit as more bands are added. Their work makes the axial-resolution trade-off explicit. It was published online on July 17, 2018, and in the January 2019 issue of IEEE Transactions on Medical Imaging; it is foundational research, not a new 2026 product announcement. Li et al., 2019.

Recent work also changes the acquisition itself. Kim and colleagues' peer-reviewed July 2026 study combines fundamental-frequency and second-harmonic images using a phase-inverted transducer and phase-dependent processing. The university's public abstract describes wire-phantom, tissue-mimicking-phantom and ex vivo experiments. That architecture is not equivalent to splitting one receive spectrum into more software bands. Its reported performance should not be transferred to another probe or conventional filter bank. Here it provides research context only; the full publisher text was not accessible during this review. Kim et al., 2026.

A calculation that exposes correlated bands

Let I1 through IN be linear-intensity images of the same stationary, homogeneous region. For this illustrative model, assume each has mean μ, variance σ², and the same pairwise correlation ρ. Use an equally weighted arithmetic mean. Expanding the covariance of that mean gives:

Var(Imean) = σ² [1 + (N − 1)ρ] / N

Define intensity speckle contrast as standard deviation divided by mean. Under these assumptions, the compounded-to-single-band contrast ratio is sqrt([1 + (N − 1)ρ] / N). An equivalent independent-image count is Neff = N / [1 + (N − 1)ρ]; this is a variance-equivalent count for this model, not a count of acquired frames.

Four independent images, with ρ = 0, give a contrast ratio of 0.50 and Neff = 4. If all pairwise correlations are 0.50, the ratio is approximately 0.79 and Neff = 1.6. Increasing that correlated set to eight images gives a ratio of 0.75 and Neff approximately 1.78. Doubling the number of processed bands therefore adds little variance reduction in this example. These values are direct algebra, not measured ultrasound data or predicted ALEVA performance.

Real filter banks rarely have equal statistics or one common correlation. For weights w summing to one and measured intensity covariance matrix Σ, the more general variance is wᵀΣw, while the output mean is wᵀμ. Estimate both from defined regions and independent acquisitions. If normalization changes the individual means or noise levels, recompute the covariance after normalization. Do not infer independence from the number of different center frequencies.

Measure what a quieter texture may conceal

For a family of Gaussian pulses with the same shape, reducing spectral width broadens the temporal envelope. In a simple pulse-echo model, temporal response maps to axial distance through z = ct/2. Thus, halving the spectral-width parameter doubles the corresponding temporal and axial width in that idealized family. This scaling is not a universal millimeter-resolution specification: the measured system response also depends on the probe, propagation, focusing and filtering. The frequency-compounding trade-off has been studied for decades, including the experimental work of Trahey and colleagues in 1986. Trahey et al., 1986.

A proposed acceptance experiment should retain the full-band reference and evaluate each candidate bank on the same input acquisition. Use homogeneous regions at specified depths for intensity mean, variance and inter-band correlation. Use point targets to measure axial and lateral response widths and sidelobes. Use separate contrast targets to inspect boundary preservation. Freeze the display mapping for visual comparison, and calculate linear-domain statistics before log compression. A low speckle-contrast number does not establish better task performance.

Include a weak-signal region. Equalizing a band with low useful signal can increase its noise contribution, so a normalized brightness match is not sufficient. Record band spectra and signal/noise conditions with depth. Keep evaluation regions away from boundaries when estimating homogeneous speckle statistics, and repeat independent acquisitions to assess variability. Many adjacent pixels share spatial correlation; treating them as independent repeats can overstate confidence.

Make the OEM decision reproducible

Before optimizing an FPGA or GPU implementation, define the filter coefficients, sample rate, band centers and widths, transition regions, normalization, detection domain, combination weights and output precision. Store test inputs and numerical reference outputs with that configuration. If filters are applied to RF in one implementation and IQ in another, document the frequency translation and equivalent passbands before comparing their results.

Measure end-to-end latency and peak working memory on the intended platform, including filtering, detection, combination and transfers. A streaming implementation can accumulate bands without retaining every intermediate image, but the implementation still needs an explicit numerical and scheduling specification. Select a setting only after its accepted spatial response, depth coverage, texture statistics and resource cost have been recorded together.

For an OEM/ODM team, the useful deliverable is a reproducible operating point and its limitations. More bands are justified when the measured information gain supports the implementation cost and the target task remains acceptable. Neither band count nor a smooth display is a substitute for that evidence. This article proposes an engineering evaluation method; it makes no clinical-performance, certification or commercial-use claim for ALEVA hardware.

References

  1. Li Y., Winetraub Y., Liba O., de la Zerda A., Chu S. Optimization of the Trade-Off Between Speckle Reduction and Axial Resolution in Frequency Compounding. IEEE Transactions on Medical Imaging, 38(1):107–112, January 2019; online July 17, 2018. Peer-reviewed journal article. DOI: 10.1109/TMI.2018.2856857.

  2. Kim G.S., Moon H.H., Lee G.Y., Ra G.L., Jeong J.S. Phase-frequency compound imaging with extended depth-of-field using a phase-inverted ultrasound transducer. Sensors and Actuators A: Physical, 404:117743, July 1, 2026. Peer-reviewed journal article; public university abstract and metadata reviewed. DOI: 10.1016/j.sna.2026.117743.

  3. Trahey G.E., Allison J.W., Smith S.W., von Ramm O.T. A quantitative approach to speckle reduction via frequency compounding. Ultrasonic Imaging, 8(3):151–164, July 1986. Peer-reviewed journal article. DOI: 10.1177/016173468600800301.