Ultrasound Engineering
Ultrasound Fractional Delays: A Finer Grid Does Not Guarantee Flat Gain
Separate fractional-delay grid resolution from interpolation gain and phase error, with a reproducible RF example and an OEM acceptance workflow.
By ALEVA | Orange County, California, USA
Does a finer delay grid guarantee a faithful ultrasound beamformer? No. The requested time offset and the frequency response of the interpolator are separate specifications. An implementation can place its nominal delay accurately while changing the amplitude of the signal it is meant to align.
Separate delay selection from waveform reconstruction
For sampled RF data, a focusing delay usually falls between available samples. The integer part selects a sample neighborhood; the fractional part determines how that neighborhood is combined. A 2025 Forum Acusticum proceedings paper by Cruza and colleagues illustrates this distinction in an FPGA beamformer: the integer delay addresses the sample buffer, while the fractional delay selects a coefficient set. Its four fractional positions are a property of that reported architecture, not a universal ultrasound requirement. [1]
Our engineering question is narrower than choosing the propagation model or calculating delays efficiently: once a delay has been requested, how faithfully does the implementation reconstruct the waveform at that offset?
A reproducible half-sample example
Consider a two-tap linear fractional-delay filter, with a fixed fractional delay a between zero and one sample:
y[n] = (1-a) x[n] + a x[n-1]
H(w) = (1-a) + a exp(-j w)
w = 2 pi f / FsHere Fs is the sample rate and f is the signal frequency. An ideal delay of a samples has response exp(-j w a): unit magnitude and the desired phase. At a = 0.5, the implemented response becomes exp(-j w/2) cos(w/2). Below Nyquist its phase is the desired half-sample delay, but its magnitude is cos(w/2), not one.
For a synthetic Fs = 40 MHz example, a 5 MHz sinusoid has gain cos(pi/8), approximately 0.9239, or -0.69 dB. A 10 MHz sinusoid has gain cos(pi/4), approximately 0.7071, or -3.01 dB. These values follow directly from the filter equation using 20 log10 of the magnitude. They are not measurements of an ALEVA product, a probe, tissue or the 2025 paper's implementation.
The example deliberately assumes an infinite, noiseless sinusoid, fixed delay and exact arithmetic. It isolates interpolation amplitude error. It does not predict image contrast, spatial resolution or diagnostic performance. At a = 0 the same filter passes the signal unchanged; consequently, testing only integer delays would miss the demonstrated loss.
Why the fractional position belongs in the test
In a receive beamformer, different channels and image locations can request different fractional offsets. The example shows why a frequency-dependent interpolation response can also create delay-dependent weighting. A single gain correction measured at one frequency and one offset cannot establish flat response across the complete operating band and all offsets.
This is also why a finely represented delay number is not sufficient evidence of fine waveform accuracy. Distinguish delay-grid quantization, interpolation error and coefficient/arithmetic quantization. Improving one does not automatically bound the others.
The resource trade-off is established rather than new. Tomov, Nikolov and Jensen's 2011 work investigates combining filter-based interpolation with focusing and summation to reduce implementation demands. The accessible DTU file is explicitly an early/preprint version associated with the published IUS proceedings paper. It is useful architectural context, not a benchmark of a current ALEVA system. [2] TI's older SPRABO0 white paper likewise explains polyphase decomposition: selecting the required fractional phase avoids explicitly computing all zero-inserted samples. Its 2011 copyright date must not be mistaken for a new processor announcement. [3]
A proposed OEM acceptance test
First define whether interpolation acts on RF or complex baseband data, the sample rate, occupied frequency band, delay sign convention, supported fractional positions and behavior at record boundaries. An RF-band result cannot simply be reused for a different IQ representation.
Then sweep deterministic tones across that declared band at every supported fractional position, comparing complex output with an analytic ideal delay after excluding boundary transients. Record amplitude error and phase error separately. For a continuously variable implementation, use a documented dense sweep and identify it as sampled evidence rather than a proof over all possible delays.
Repeat with finite-bandwidth pulses and realistic changing delay trajectories. Include transitions where the fractional part wraps and the integer sample index changes. These tests assess indexing and transient behavior that a fixed-delay frequency response cannot establish. Retain boundary rules and output-validity flags so zero padding is not silently interpreted as acquired data.
Finally compare the quantized implementation with its floating-point reference using the same input and delay sequence. Evaluate point-target location, sidelobes and image uniformity across the intended field, alongside sustained throughput, memory use and processing latency. Define project-specific acceptance limits before selecting a preferred result; the algebra above provides no universal pass threshold.
The design decision
Longer or differently designed fractional-delay filters may improve a specified band's approximation, but their arithmetic, storage and latency must be measured in the intended architecture. The deliverable should therefore include the operating band, delay grid, coefficient version, error measurements and resource cost together. For an ultrasound beamformer OEM evaluation, that evidence is more informative than a delay-resolution number alone.
This article presents an analytical example and a proposed engineering workflow. It makes no claim that ALEVA commercially implements the cited architectures and reports no clinical validation.
Sources and publication status
Cruza J.F., Mateos R., Bueno R., Montaldo G., Camacho J. Modular and flexible hardware architecture for real-time 3D ultrasound imaging. Forum Acusticum / Euronoise 2025, Malaga, June 23-26, 2025. Published conference proceedings, Section 2.3. DOI: 10.61782/fa.2025.0583. Official proceedings PDF. Not a journal article.
Tomov B.G., Nikolov S.I., Jensen J.A. Scalable Intersample Interpolation Architecture for High-channel-count Beamformers. IEEE IUS 2011, pp. 381-384. DOI: 10.1109/ULTSYM.2011.0091. DTU-hosted early/preprint version; version status explicitly stated by the repository.
Karadayi K., Lee C., Kim Y. Software-based Ultrasound Beamforming on Multi-core DSPs. TI-hosted technical white paper SPRABO0, copyright 2011, Sections 2 and 4. Manufacturer-hosted PDF. Not a peer-reviewed journal article; no precise issue date inferred from the copyright year.