Ultrasound Sound-Speed Calibration: Keep Tuning Separate from Validation

By ALEVA | California, USA

When comparing two ultrasound beamformers, are we measuring a reconstruction improvement—or a better choice of propagation speed? An OEM acceptance test should make that distinction explicit.

Why the baseline matters

The CUBDL organizers reported conflicting sound-speed metadata between contributed datasets and phantom manufacturer information. They optimized sound speed within a common delay-and-sum reference to support fair comparisons. This is a lesson from the 2020 challenge, documented on a results page updated in 2021, not a new 2026 product announcement.

Recent research makes the calibration question more practical. De la Torre, Xiao and Yu published a single-shot global sound-speed estimator in Ultrasonics in 2025. Their public repository provides an example and a sound-speed-aware delay-and-sum function. This article checks the public publication record and repository; it does not independently reproduce their performance.

A reproducible depth check

Consider a hypothetical homogeneous medium with true sound speed 1500 m/s and a reflector 30 mm from the acoustic origin. Assume straight-path pulse-echo propagation, zero electronic timing offset and no refraction. The round-trip time is t = 2d/c = 40 microseconds. Reconstructing that same time with 1540 m/s gives d_display = c_assumed × t / 2 = 30.8 mm: a +0.8 mm depth error.

This calculation is a scale-error illustration, not a measured ALEVA result or a prediction of lateral resolution. It does not model heterogeneous tissue, lens delays, clutter or the finite bandwidth of a pulse.

A proposed OEM test contract

First, record the phantom identity, temperature, nominal acoustic properties, probe geometry, sampling clock and timing origin. Preserve both the supplied sound-speed value and any estimated replacement, with its estimation method and software version.

Second, separate calibration data from held-out evaluation captures. If a candidate adapts sound speed on every frame, freeze its algorithm and permitted inputs before evaluation; do not silently tune its objective using known test-target positions. Report a fixed-speed reference and an adaptive configuration separately.

Third, evaluate geometry as well as appearance. Compare target location, point-spread width and contrast across depths and lateral positions, while recording processing latency. A brighter or narrower target alone is not an acceptance decision.

The trade-off and limit

A global estimate offers a compact calibration parameter, but one scalar does not describe a spatially varying medium. A May 2026 preprint by Zhuang and colleagues investigates wavefield-correlation reconstruction for local sound-speed optimization and identifies straight-ray assumptions and reverberation clutter as limitations of earlier approaches. That source is a preprint, not treated here as independently verified clinical evidence.

For OEM/ODM transfer, the deliverable should therefore include the calibration rule, its allowed inputs, failure handling and an untouched evaluation dataset—not only a selected image. These are proposed engineering acceptance practices, not an ALEVA implementation, certification or medical-performance claim.

Sources

1. CUBDL, Results and Sound Speed Optimization. IEEE IUS 2020 challenge; page updated July 5, 2021.

2. De la Torre P, Xiao D, Yu ACH. One shot, one SoS: A real-time, single-shot global speed of sound estimator. Ultrasonics 153 (2025), 107612. DOI; authors' public code and data repository. Publisher full text was not accessible in this review.

3. Zhuang L et al. A Wavefield Correlation Approach to Improve Sound Speed Estimation in Ultrasound Autofocusing. arXiv:2602.12805v2, revised May 7, 2026; preprint. DOI.