What is OpenH-RF, a large-scale open ultrasound raw channel dataset?

Why raw channel data, not images, is the bottleneck

Most public ultrasound datasets release reconstructed images. The paper argues that this limits a specific class of problems: aberration correction and reverberation reduction, among other core technical challenges, depend on the signals recorded before beamforming, so an image-only dataset cannot support work on them. The authors also note that the smaller set of open raw channel datasets tends to be built for one reconstruction or imaging task, which makes it hard to develop general-purpose methods or foundation models that operate directly on ultrasound signals.

Raw channel data is also hard to obtain. It requires research scanners that expose the pre-beamforming signals, which the paper describes as costly and not available in every lab, and it is often treated as proprietary because it reveals details of the transmit sequences and system design. The stated consequence is that raw-channel research has largely stayed within groups that own such a setup, with methods developed and evaluated on in-house data rather than across institutions.

What the corpus actually contains

The headline scale is approximately 40 TB of raw channel data in 19,332 acquisitions, contributed by 38 institutions across academia and industry and organized into 33 contributions. Each contribution is annotated with the tasks it supports, drawn from categories including generalized reconstruction and beamforming, speed-of-sound estimation and aberration correction, image interpretation, motion tracking, image quality assessment, 3D and tracked reconstruction, blood flow and ultrasound localization microscopy, quantitative ultrasound, and clutter or reverberation suppression.

The composition is uneven, and the paper is explicit that the count depends on what unit you count. By acquisitions, simulated data dominates at 84% of the total (16,248 of 19,332), which the authors suggest may reflect how much easier it is to generate new simulated subjects. By frames, the picture inverts: 98% of all frames are in-vivo, and 85% of those in-vivo frames come from a single contribution, ULMShare, which holds long high-frame-rate acquisitions. Because consecutive frames in such acquisitions are highly correlated, the authors warn that frame counts overstate how much independent information a contribution adds, and that the number of distinct subjects, animals, or phantoms is a better proxy for dataset size.

The counting unit changes which data dominates

This is the paper's most consequential internal comparison, because it determines how a reader should sample the corpus. A benchmark built by drawing acquisitions at random would be dominated by simulation; one built by drawing frames would be dominated by a single in-vivo high-frame-rate source. Neither matches the anatomical and physiological diversity a reader might assume from the headline totals.

The per-medium breakdown makes the same point at a finer level. In-vivo acquisitions total 2,715 and are led by rectus femoris and biceps at 468 each, then breast at 354, calf at 342, heart at 203, carotid artery at 194, thyroid at 147, and brain at 139, with 331 labeled other and 69 unspecified. Phantom acquisitions total 369, led by liver fibrosis phantom at 4,000 frames-scale entries reported in the figure, alongside fetal phantom, breast, liver, and echogenicity or lesion phantoms. Simulation reaches 16,248 acquisitions, dominated by liver fibrosis and fetal phantom categories. Probe coverage is likewise skewed: linear arrays account for 10,353 acquisitions (53.6%), curved for 5,974 (30.9%), ring for 2,643 (13.7%), phased for 310 (1.6%), matrix for 43 (0.2%), and intravascular arrays for 9.

Task coverage is broad except for therapy

Measured by the number of contributions supporting each task, generalized reconstruction and beamforming is the best covered, supported by 27 contributions, and every other task category is covered by at least 5 contributions. The exception is the therapy category, which contains only one task, microbubble cavitation detection, and is covered only by the twente-cavitation dataset. A reader planning work in therapeutic ultrasound therefore has essentially a single source inside this corpus, while a reader working on reconstruction has many.

The paper also reports transmit-event counts by scheme, which show how lopsided the acquisition effort is: plane wave accounts for 225,720,883 transmit events, idealized simulated point source for 5,654,400, focused for 3,912,785, synthetic aperture for 2,612,088, mixed for 1,788,000, diverging wave for 517,214, single element for 480,600, coded or Hadamard for 384,660, and external source for 7,600. These are transmit events, not acquisitions or frames, and should not be read as a measure of how many independent datasets use each scheme.

A self-contained file format makes reuse possible

The barrier the paper targets is not only data scarcity but the domain knowledge needed to use raw channel data. Reconstructing an image requires a chain of steps — filtering, demodulation, beamforming, envelope detection — whose order and settings differ between contributions. OpenH-RF therefore asked each contributor to supply a reproducible reconstruction pipeline together with a reference image, and the curators checked that the pipeline produced the intended result before accepting the data. Each file records the version of the reconstruction library used to store it, so a recipient can reproduce what the original authors intended.

The pipelines are implemented with GPU-accelerated, differentiable operations in a machine-learning framework, which the authors present as a way to integrate ultrasound reconstruction directly into model training. The corpus itself is versioned, and the paper describes version 1.0, so follow-up work can state exactly which release it used. Access is designed around the fact that a few frames can already reach multiple gigabytes: users can filter by anatomy, probe type, or intended task, and within each file the raw channel data is split into compressed chunks along the transmit axis so that selected frames or transmits can be read locally or streamed remotely without loading a full acquisition.

What the dataset does not settle

OpenH-RF is a data resource, not a benchmark result. The paper reports no model trained on the corpus and no comparison of methods across it; the evidence is the composition, annotation, and infrastructure of the collection. Claims about what foundation models could learn from it are framed as opportunities, including self-supervised pretraining conditioned on the probe geometry and transmit parameters stored in every file, generative models for dehazing or channel compression, and patient-adaptive acquisition. Those are proposed directions, not demonstrated outcomes.

The composition limits also travel with the data. Because simulated acquisitions dominate the count while in-vivo frames dominate the frame total and cluster in one contribution, any sampling scheme inherits a specific bias, and the authors state that the unevenness depends heavily on whether one counts acquisitions, frames, or transmit events. Therapy is represented by a single dataset. And the corpus is presented as a starting point rather than a finished resource, with the format and submission process designed to accept new contributions.

Who should start here, and with which slice

The practical priority that follows from the paper's own numbers is to choose the counting unit before choosing the data. A team building a reconstruction or beamforming benchmark has the widest choice, since 27 contributions support that task, and should sample by distinct subjects rather than by frames to avoid letting ULMShare's long high-frame-rate acquisitions dominate. A team working on speed-of-sound estimation, aberration correction, or clutter suppression has multiple contributions to draw on but should check per-contribution acquisition counts, since these range from a handful to thousands. A team working on microbubble cavitation detection has one dataset, and a team needing matrix or intravascular arrays has 43 and 9 acquisitions respectively.

The paper's own framing of value is that raw channel data has been confined to groups with their own scanners and in-house data, and that a shared permissively licensed corpus changes what can be compared across institutions. The most defensible reading of this release is therefore as infrastructure for cross-institution comparison and for methods that operate on signals before beamforming — with the caveat that its balance of media, anatomy, and task coverage is a property of what contributors chose to share, not a designed sampling of clinical ultrasound.

The open question this release poses

The paper closes on a direction rather than a result: moving ultrasound research beyond the beamformed image toward methods that use the full information content of the raw signal. Whether that happens depends less on the 40 TB total than on whether the corpus's metadata — probe geometry, transmit parameters, and per-file reconstruction pipelines — is rich enough for one model to handle different probes and transmit schemes. That is the specific bet OpenH-RF makes, and it is the claim a follow-up study would need to test.

Where the evidence stops

Everything above describes the dataset as documented in this paper: its size, composition, annotations, format, and licensing under CC BY 4.0. The paper does not report downstream model performance, cross-institution benchmark results, or validation that the corpus improves any specific task. Readers should treat the stated opportunities — foundation models on raw channel data, generative priors for cognitive ultrasound, benchmarks for the machine-learning and computational-imaging communities — as the authors' proposals, and evaluate the corpus on the composition numbers reported here rather than on anticipated outcomes.

The decision this paper supports

For a reader deciding whether OpenH-RF fits a project, the decisive question is not total size but task coverage and counting unit together. Reconstruction and beamforming work has 27 supporting contributions and can be sampled broadly; therapy work has one dataset; and any evaluation that counts frames rather than subjects will be dominated by a single in-vivo high-frame-rate contribution. The paper's contribution is to make that trade-off visible and to supply the metadata and per-file pipelines needed to act on it.

Scope and limitations of the release

The corpus spans in-vivo, phantom, and simulated data, and the paper documents ethical approval and de-identification for in-vivo contributions in each data card. It is hosted publicly under CC BY 4.0 and can be streamed selectively. The limitations that matter for interpretation are compositional rather than technical: simulated acquisitions dominate by count, in-vivo frames concentrate in one contribution, therapy rests on a single dataset, and matrix and intravascular arrays are represented by 43 and 9 acquisitions. The paper presents the corpus as version 1.0 and as a starting point open to new contributions.

What to watch in follow-up work

The paper's stated next step is a foundation model for raw channel data, pretrained with self-supervised objectives on the full corpus and conditioned on the probe geometry and transmit parameters stored in every file. If such work appears, the comparison that matters is whether one model conditioned on that metadata can handle the probe and transmit-scheme diversity documented here — linear arrays at 53.6% of acquisitions against matrix arrays at 0.2%, and plane-wave transmits at 225,720,883 events against 7,600 from an external source. That spread, not the 40 TB headline, is the condition any general-purpose raw-signal model would have to survive.

The priority for readers

The most useful way to read this paper is as a map of what raw ultrasound data now exists and where it is thin. Reconstruction and beamforming is the well-covered task with 27 contributions; therapy, matrix arrays, and intravascular imaging are the thin edges. Between them sit speed-of-sound estimation, aberration correction, motion, flow, and quantitative ultrasound, each supported by several contributions whose acquisition counts vary widely. A reader should pick the task first, then check the per-contribution acquisition counts and the counting unit before assuming the corpus's overall scale carries over to their slice.

The judgment this paper invites

OpenH-RF's value is concentrated in the tasks where raw channel access is unavoidable and where multiple institutions have now contributed: reconstruction and beamforming above all, with speed-of-sound estimation, aberration correction, and clutter suppression close behind. Its weakest points are equally specific — a single therapy dataset, 43 matrix-array acquisitions, and an in-vivo frame total that one contribution largely supplies. The judgment the paper supports is that shared raw channel data is now available at a scale that permits cross-institution comparison, provided the comparison is defined at the level of subjects and tasks rather than at the level of terabytes or frames.

The comparison that matters

The paper's own framing sets raw channel data against reconstructed images, and its numbers show why the distinction is not cosmetic: 19,332 acquisitions across 33 contributions and 38 institutions, with per-file reconstruction pipelines and versioned releases, exist so that methods can be compared on the same signals rather than on each group's in-house scanner. Whether that comparison materializes is the open question, and the corpus's composition — simulation-heavy by count, in-vivo-heavy by frames, reconstruction-heavy by task — is the condition under which it will be attempted.

The bottom line for a new user

A new user should begin with the task annotation rather than the total size. If the task is reconstruction or beamforming, the corpus offers 27 supporting contributions and enough probe and transmit diversity to test generalization. If the task is therapy, matrix imaging, or intravascular imaging, the corpus offers one, 43, and 9 acquisitions respectively, and should be treated as a seed rather than a benchmark. In between, the per-contribution acquisition counts — from a handful to thousands — determine how much independent evidence any single task comparison actually rests on.

What this release changes

The change the paper claims is access: raw channel data that previously required a research scanner and in-house pipelines is now available under CC BY 4.0 with per-file reconstruction scripts, reference images, and versioned releases, streamable by anatomy, probe type, or task. The change it does not yet claim is performance: no model results are reported. The paper's contribution is the infrastructure and the composition map, and the judgment it supports is that cross-institution work on raw ultrasound signals is now feasible for the well-covered tasks while remaining thin for therapy and specialized array geometries.

The specific bet OpenH-RF makes

The bet is that a common representation plus stored acquisition, probe, and reconstruction metadata is enough for one model to handle different probes and transmit schemes. The evidence for the bet is the format and the metadata, not a trained model. The evidence against overstating it is the composition: 84% of acquisitions are simulated, 98% of frames are in-vivo but 85% of those come from one contribution, and therapy has a single dataset. A reader evaluating the bet should look at whether follow-up work conditions on the stored metadata and reports results per task and per probe type, rather than reporting a single aggregate number over the whole corpus.

How to read the numbers together

Three counts describe the same corpus and disagree by design. Acquisitions (19,332) favor simulation at 84%. Frames favor in-vivo at 98%, concentrated in ULMShare. Transmit events favor plane wave at 225,720,883 out of roughly 235 million reported across schemes. None of these is the dataset size in the sense a reader usually means; the paper's own suggestion is that distinct subjects, animals, or phantoms is the better proxy. Any claim about OpenH-RF's coverage should therefore name its counting unit, and any sampling plan should be built on subjects rather than on whichever total is largest.

The practical threshold for using this corpus

The threshold is task coverage, not terabytes. Tasks supported by many contributions — reconstruction and beamforming at 27 — can support cross-institution comparison today. Tasks supported by a few contributions can support method development but not broad benchmarking. Tasks supported by one contribution, namely microbubble cavitation detection, can support only work that does not require diversity. That ordering, derived from the paper's own task-coverage table and per-contribution acquisition counts, is the most useful thing a prospective user can take from this release.

The final priority

The priority this paper establishes is to treat raw channel data as the object of comparison and to define comparisons at the level of subjects and tasks. OpenH-RF supplies 19,332 acquisitions, 33 contributions, 38 institutions, per-file reconstruction pipelines, and a CC BY 4.0 license; what it does not supply is evidence that any model trained on it generalizes. The next meaningful result would be a model conditioned on the stored probe and transmit metadata and evaluated per task, because that is the claim the corpus's heterogeneity — linear arrays at 53.6% of acquisitions against matrix arrays at 0.2%, and one therapy dataset — is built to test.

Sources used in this answer