arxiv: 2604.06257 · v1 · submitted 2026-04-06 · ⚛️ physics.med-ph · eess.IV· eess.SP

Recognition: no theorem link

mach: ultrafast ultrasound beamforming

Alexander P. Rockhill, Charles Guan, Gianmarco Pinton, Masashi Sode

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:40 UTC · model grok-4.3

classification ⚛️ physics.med-ph eess.IVeess.SP

keywords ultrafast ultrasoundbeamformingGPU accelerationCUDA kernelreal-time reconstructionvolumetric imagingdelay-and-sumPyMUST benchmark

0 comments

The pith

Mach enables real-time 3D ultrafast ultrasound by beamforming at 1.1 trillion points per second on consumer GPUs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents mach, an open-source GPU-accelerated beamformer designed for volumetric ultrafast ultrasound imaging. It uses a hybrid delay computation strategy in an optimized delay-and-sum CUDA kernel to cut memory requirements while reusing calculations across frames through shared memory. This design delivers processing rates of 1.1 trillion points per second, allowing full 3D reconstructions in just 0.23 milliseconds on a consumer GPU. The speed surpasses the acoustic travel time to the imaging depth and outperforms prior open-source tools by over ten times. By removing the main computational limit, mach opens the door to real-time use of advanced ultrasound techniques in research and clinical settings.

Core claim

mach achieves 1.1 trillion points per second throughput, enabling real-time 3D ultrafast ultrasound reconstruction for the first time on consumer-grade hardware. The system relies on a hybrid delay computation strategy that substantially reduces memory overhead compared to fully precomputed approaches, paired with a CUDA implementation that optimizes memory layout for coalesced access and reuses delay computations across frames via shared memory. Validation on the PyMUST rotating disk dataset shows reconstruction in 0.23 ms with numerical errors below -60 dB for Power Doppler and -120 dB for B-mode, confirming accuracy against other beamformers.

What carries the argument

The hybrid delay computation strategy within the optimized delay-and-sum CUDA kernel, which reduces memory overhead and enables reuse of delay computations across frames via shared memory.

If this is right

Real-time 3D volumetric ultrasound reconstruction is now achievable on consumer-grade GPUs.
Beamforming computational demands no longer limit research throughput in ultrafast ultrasound modalities.
Applications such as 3D functional neuroimaging, intraoperative guidance, and ultrasound localization microscopy can operate in real time.
The open Python interface facilitates easy integration and further development by the research community.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This performance level could support even denser reconstruction grids or higher frame rates in future studies.
The memory-efficient design may extend to portable or embedded ultrasound systems.
Optimization patterns from the CUDA kernel could transfer to accelerating similar summation-based computations in other imaging domains.

Load-bearing premise

The hybrid delay computation strategy reduces memory overhead without introducing numerical errors that affect image quality in downstream applications.

What would settle it

If the reconstructed images from mach on the PyMUST benchmark show errors greater than -60 dB in Power Doppler or -120 dB in B-mode when compared to reference beamformers, the claim of maintained numerical accuracy would be falsified.

read the original abstract

Purpose: Volumetric ultrafast ultrasound produces massive datasets with high frame rates, dense reconstruction grids, and large channel counts. Beamforming computational demands limit research throughput and prevent real-time applications in emerging modalities such as elastography, functional neuroimaging, and microscopy. Approach: We developed mach, an open-source, GPU-accelerated beamformer with a highly optimized delay-and-sum CUDA kernel and an accessible Python interface. mach uses a hybrid delay computation strategy that substantially reduces memory overhead compared to fully precomputed approaches. The CUDA implementation optimizes memory layout for coalesced access and reuses delay computations across frames via shared memory. We benchmarked mach on the PyMUST rotating disk dataset and validated numerical accuracy against existing open-source beamformers. Results: mach processes 1.1 trillion points per second on a consumer-grade GPU, achieving $>$10$\times$ faster performance than existing open-source GPU beamformers. On the PyMUST rotating disk benchmark, mach completes reconstruction in 0.23~ms, 6$\times$ faster than the acoustic round-trip time to the imaging depth. Validation against other beamformers confirms numerical accuracy with errors below $-60$~dB for Power Doppler and $-120$~dB for B-mode. Conclusions: mach achieves 1.1 trillion points per second throughput, enabling real-time 3D ultrafast ultrasound reconstruction for the first time on consumer-grade hardware. By eliminating the beamforming bottleneck, mach enables real-time applications such as 3D functional neuroimaging, intraoperative guidance, and ultrasound localization microscopy. mach is freely available at https://github.com/Forest-Neurotech/mach

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

mach gives a practical, open-source speed jump to 3D ultrafast ultrasound beamforming with measured 1.1T points/sec throughput and direct validation on public data.

read the letter

The key takeaway is that this work ships a CUDA beamformer that hits 1.1 trillion points per second on consumer GPUs and finishes the PyMUST rotating-disk reconstruction in 0.23 ms, which is fast enough to remove the main compute barrier for real-time volumetric imaging. The hybrid delay approach plus coalesced memory layout and shared-memory reuse across frames is what drives the gain over prior open-source GPU codes, and they back it with side-by-side timing and error numbers against existing implementations. Numerical checks show errors below -60 dB on Power Doppler and -120 dB on B-mode, which is solid for downstream use. The code is on GitHub with a Python front end, so labs can actually run it without reimplementing everything. The main soft spot is that the exact hybrid delay equations are not laid out in full detail, which could slow down someone trying to port the trick to a different architecture, though the measured accuracy on the public dataset still holds. This is aimed at ultrasound groups doing 3D functional imaging, elastography, or localization microscopy who need faster reconstruction on standard hardware. Readers who care about reproducible benchmarks and open tools will find it directly useful. The work is grounded enough in external data and comparisons that it deserves a serious referee rather than a desk reject.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces mach, an open-source GPU-accelerated beamformer for volumetric ultrafast ultrasound imaging. It employs a highly optimized CUDA kernel using a hybrid delay computation strategy to reduce memory overhead, along with coalesced memory access and shared-memory reuse of delays across frames. Benchmarks on the public PyMUST rotating-disk dataset report 0.23 ms reconstruction time, numerical accuracy with errors below -60 dB (Power Doppler) and -120 dB (B-mode) relative to other open-source beamformers, and a peak throughput of 1.1 trillion points per second on consumer-grade hardware, enabling real-time 3D reconstruction.

Significance. If the measured performance holds, the work is significant for medical physics and ultrasound imaging. It directly addresses the computational bottleneck that has limited real-time volumetric ultrafast applications in elastography, functional neuroimaging, and ultrasound localization microscopy. Strengths include the open-source release, direct validation against external open-source implementations on a public dataset, and concrete quantitative error bounds rather than qualitative claims. These elements support reproducibility and practical adoption.

minor comments (3)

[Approach] Approach section: the hybrid delay computation strategy is described at a high level as reducing memory overhead without numerical errors, but the exact formula or pseudocode for combining precomputed and on-the-fly delays is not provided. Including this would strengthen reproducibility without altering the central performance claims.
[Results] Results section: the 1.1 trillion points per second throughput figure should explicitly state the GPU model, whether the metric includes host-device transfers, and whether it represents sustained or peak performance to allow direct comparison with other implementations.
[Conclusions] The manuscript would benefit from a short limitations paragraph addressing edge cases (e.g., very high channel counts or non-linear imaging modes) where the hybrid delay approximation might require additional validation.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their supportive review and recommendation of minor revision. We are pleased that the significance of mach for enabling real-time volumetric ultrafast ultrasound in applications like elastography and ultrasound localization microscopy is recognized, along with the value of our open-source implementation and quantitative validations. As the report does not include any specific major comments requiring response, we have no point-by-point rebuttals. We remain available to address any minor issues or suggestions from the editor.

Circularity Check

0 steps flagged

No significant circularity; results are empirically measured against external benchmarks

full rationale

The paper describes a GPU-accelerated beamforming implementation and reports directly measured throughput (1.1 trillion points per second) and timing (0.23 ms on PyMUST rotating-disk dataset) on consumer hardware. Numerical fidelity is validated by explicit comparison to independent open-source beamformers (errors < -60 dB Power Doppler, < -120 dB B-mode). No equations, predictions, or first-principles derivations are presented; performance claims rest on code execution and external dataset comparisons rather than any self-referential fitting, renaming, or self-citation chain. The central claims are therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The work is an engineering implementation that rests on the standard delay-and-sum beamforming algorithm and CUDA programming model; no new physical parameters, axioms, or entities are introduced.

axioms (2)

domain assumption Delay-and-sum beamforming correctly reconstructs ultrasound images when delays are computed from geometry and speed of sound
Invoked implicitly as the foundation for the optimized kernel.
standard math CUDA memory coalescing and shared-memory reuse improve throughput without altering numerical results
Core optimization premise of the implementation.

pith-pipeline@v0.9.0 · 5611 in / 1434 out tokens · 35776 ms · 2026-05-10T18:40:40.629727+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

1 extracted references

[1]

Ultrafast imaging in biomedical ultrasound,

1 M. Tanter and M. Fink, “Ultrafast imaging in biomedical ultrasound,”IEEE Trans. Ultrason. Ferroelectr. Freq. Control61, 102–119 (2014). 2 J. Provost, C. Papadacci, J. E. Arango,et al., “3d ultrafast ultrasound imaging in vivo,” Physics in Medicine & Biology59(19), L1 (2014). 3 E. Mac ´e, G. Montaldo, I. Cohen,et al., “Functional ultrasound imaging of th...

2014