ALEVA

Insights

Ultrasound Apodization: Preserve Coherent Gain Without Hiding Noise

Distinguish coherent signal gain from noise normalization using an exact eight-channel example, with limits for correlated noise, adaptive weights and OEM verification.

By ALEVA | Orange County, California, USA

Engineering analysis with an exact synthetic calculation. No measured hardware or clinical performance is claimed.

The engineering question

When a receive aperture changes with depth, how should an ultrasound beamformer scale its weighted channel sum? Dividing by the number of active channels is correct for an equally weighted average. With unequal weights, that same divisor changes coherent signal gain. Dividing by the sum of the weights fixes that particular gain error, but it does not make output noise independent of the aperture.

This distinction matters when comparing FPGA and GPU implementations, changing a coefficient table, or disabling a receiver channel. A matched signal amplitude and a matched noise floor are separate acceptance criteria.

Two properties of the same weight vector

Consider already delayed real channel samples x_i = s + n_i. The desired component s is identical across channels. For this calculation, weights w_i are fixed, real and nonnegative; noise has zero mean, equal variance sigma², and is independent between channels. Define S = sum(w_i) and Q = sum(w_i²), with S > 0.

For the normalized output y = sum(w_i x_i) / S, the coherent signal remains s. The output noise variance is sigma² Q / S². Define an effective channel count N_eff = S² / Q: the noise variance is sigma² / N_eff. This is a noise-equivalent count under these assumptions, not a number of physical channels or an image-resolution measure.

The calculation is a direct specialization of weighted-sum beamforming and covariance propagation. Neves and colleagues present the weighted output, output-power quadratic form and distortionless constraint in their peer-reviewed 2025 study; those mathematical objects distinguish signal response from power. The numerical example here is our own analytical construction, not a result reported in that study. [1]

An eight-channel example

With eight unit weights, S = 8 and Q = 8. Sum normalization gives unit coherent gain and noise variance 0.125 sigma².

Now use weights [1, 2, 3, 4, 4, 3, 2, 1]. Here S = 20, Q = 60, and N_eff = 400/60, approximately 6.667. Sum normalization still gives unit coherent gain, while noise variance becomes 0.15 sigma². Relative to the uniform case, that is 1.2 times the variance, or approximately 0.792 dB, calculated as 10 log10(1.2).

Multiplying every weight by any positive constant leaves the normalized output unchanged. Consequently, the example could use weights scaled to a maximum of one without changing either result. It is an arithmetic example, not a claim that this taper is optimal for a particular probe.

If the divisor were sqrt(Q), output noise variance would instead equal sigma² under the same independent-noise model. Coherent gain would then be S/sqrt(Q), which varies with the weights. No single scalar can hold both properties fixed across arbitrary weight vectors with different N_eff.

Where the simple model stops

Actual channels need not contain identical desired amplitudes. Element response, focusing errors and off-axis signals change the steering vector. Cross-channel disturbances need not be independent. For a fixed real weight vector and noise covariance R_n, the sum-normalized noise variance is w^T R_n w / S². A disturbance identical on every channel passes through the unit-coherent-gain normalization without the independent-noise averaging benefit.

Nor does electronic noise describe all background structure in an ultrasound image. Speckle and coherent clutter require their own measurements. A taper that improves a spatial sidelobe criterion may still change main-lobe width, background statistics or weak-target visibility. The effective channel count does not evaluate those effects.

Learned and adaptive weights need additional care. The OATS journal paper uses simulation-trained, depth-dependent apodizations, illustrating why a receive weight table can encode more than a conventional fixed window. Its bibliographic date is online publication on September 24, 2025, with the March 2026 Ultrasonics issue. This reference establishes the design context; it does not validate the normalization example or an ALEVA implementation. [2] If weights depend on the noisy data being tested, treating them as independent fixed coefficients may give the wrong noise prediction. Complex or signed weights also require a steering-vector response and a defined safeguard against a near-zero normalization denominator.

A proposed OEM verification record

First, feed identical aligned samples into every active channel and verify coherent gain before envelope detection, log compression or automatic display scaling. Repeat across depth, steering and every supported channel-disable mask.

Second, use independent zero-mean synthetic noise with a declared variance and verify the predicted output variance, allowing explicitly stated finite-sample uncertainty. Third, inject a common disturbance to expose the independent-noise assumption. These tests address normalization; separate point-target and image tests must assess spatial performance.

Record the actual quantized weights, active-channel mask, S, Q, normalization divisor, arithmetic rounding and overflow behavior. Recompute S and Q from the coefficients actually used by the implementation. A divisor generated from an unquantized table can disagree with the implemented sum. Keep any deliberate depth gain separate in the record so it cannot silently conceal a normalization change.

The OEM/ODM decision

Specify which response the normalization preserves, which noise model supports the predicted floor, and where gain is measured. Accept a coefficient or aperture change against both signal and noise evidence, then evaluate its spatial trade-offs. This makes implementation comparisons interpretable without pretending that equal brightness means equal beamforming behavior.

References

  1. Neves L. C. et al. Improving Ultrasound B-Mode Image Quality with Coherent Plane-Wave Compounding Using Adaptive Beamformers Based on Minimum Variance. Sensors 25(5), 1306, February 21, 2025. Peer-reviewed journal article. DOI: 10.3390/s25051306. Mathematical context: equations 1, 3 and 4.

  2. Xiao D., Nahas H., Hiroshima M., Kishimoto A., Yu A. C. H. Optimized Apodizations with Training Simulations (OATS): Learned depth-dependent apodizations via differentiable beamforming for reduced operator tuning. Ultrasonics 159, 107827, March 2026; online September 24, 2025. Peer-reviewed journal article; abstract and bibliographic record reviewed. DOI: 10.1016/j.ultras.2025.107827.