PaperlayerView exact revision on arXiv

Source-linked AI summary

OpenH-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative

T. S. W. Stevens, S. W. Penninga, W. L. van Nierop, O. Nolan, F. Duelmer, B. N. Frey, B. Federici, V. van Hal, L. D. van Harten, H. de Hoop, R. Lopata, B. Luijten, M. van Sambeek, V. van der Schaft, H. -M. Schwab, M. F. Azampour, N. Navab, R. Ali, A. Andrzejek, J. Baek, B. Boitnott, T. Brevett, H. Hashemi, D. Hyun, J. H. Park, A. V. Telichko, C. Thomas, L. Zhuang, Y. Aklilu, H. Arora, J. Barker, A. K. Mackanic, P. McDowall, L. McHugh, I. Pirozzi, S. Simon-Trench, A. M. Bronstein, O. Senouf, S. Vedula, D. Zadok, H. Rivaz, M. Sharifzadeh, S. Chen, G. P. Luke, S. Wang, Yujia Wu, J. Verde, R. Göbl, O. Zettinig, E. Boctor, Yixuan Wu, H. -M. Bae, M. -G. Kim, Y. -M. Kim, H. Lee, S. -H. Oh, P. Pinto, M. Berrada-Sounni, M. Carley, R. J. Fedewa, S. Gavin, N. Goldsmith, R. Kassayan, C. Low, R. Mehranfar, G. Bardoscia, M. Boccardo, A. Botter, M. Carbonaro, G. L. Cerone, E. Cesti, K. Meiburger, S. Seoni, F. Truscello, B. J. Raiteri, L. Caron, M. Gasse, N. Ghigo, O. Gulenko, A. Leconte, S. A. Lee, A. Malescot, E. Martineau, J. Porée, J. Provost, G. Ramos-Palacios, B. Rauby, R. L. Rungta, A. F. Sadikot, A. Wu, P. Xing, A. Chavignon, C. Ferlay, V. Hingot, M. Reberol, A. Waelbroeck, H. -S. Kwon, J. Bucholz, I. Guracar, R. Loftman, C. Williams, T. Grutman, T. Ilovitsh, T. A. Aleef, P. B. Chen, W. Chen, Y. Chen, M. Frew, M. Honarvar, Y. C. Lee, Z. Li, H. Moradi, R. Rohling, S. E. Salcudean, Z. Wu, Q. Zeng, P. C. Cattin, S. El Hadramy, N. Bottenus, A. S. Perez, S. Feng, Y. Feng, P. Jin, Y. Lin, O. Ostras, G. Pinton, F. Santibanez, M. Sode, B. Turkbey, H. Wang, B. J. Wood, L. Zhang, S. Luo, J. Wiskin, A. Austeng, Y. Karabiyik, S. P. Näsholm, O. M. H. Rindal, D. Bustamante, M. Mehrmohammadi, H. Raeis, A. Hameddine, G. Lajoinie, H. de Roo, M. Versluis, R. Zorgdrager, J. S. Au, A. J. Y. Chee, J. Cohen, T. Gu, J. Y. -H. Hsu, S. Ling, M. Mourtzakis, H. Nahas, P. de la Torre, D. Xiao, A. C. H. Yu, E. Zafiris, P. Jarosik, P. Karwat, Z. Klimonda, M. Lewandowski, B. Byram, C. Khan, Z. Marshall, Y. -C. Pan, S. Liu, T. Sultan, C. Hoerig, J. Mamou, D. Auerbach, Y. C. Eldar, A. Wegerhoff, G. S. Fischer, S. Katayose, X. Ma, R. Murakami, Y. Tang, R. Tsumura, Y. Wang, H. K. Zhang, M. Azizian, S. Huver, J. J. Dahl, R. J. G. van Sloun, W. Simson

arXiv:2610.13616v1eess.SP

TL;DR

Open ultrasound research has few large, open raw-channel collections, despite the need for those signals in many imaging challenges. OpenH-RF assembles a diverse corpus under a unified, permissively licensed format. It comprises 40 TB across 19,332 acquisitions, with simulation accounting for 84 % of acquisitions.

  • Problem

    Open raw-channel ultrasound data are scarce and fragmented, limiting shared work on imaging challenges that require signal-level information.

    Source: §I · p. 3, §II · p. 4

  • Method

    OpenH-RF pools heterogeneous ultrasound acquisitions in a unified corpus with stored reconstruction pipelines and programmatic streaming access.

    Source: §I · p. 3, 2) Data access · p. 7

  • Results

    OpenH-RF comprises 40 TB across 19,332 acquisitions, and simulation accounts for 84 % of all acquisitions.

    Source: §I · p. 3, §V · p. 11

  • Takeaways & Limitations

    The self-contained, streamable files lower the barrier to working with raw channel data and provide a starting point for community benchmarks and contributions.

    Source: §V · p. 12

Abstract

from arXiv · show

Medical ultrasound is a widely accessible, real-time imaging modality, yet progress in data-driven ultrasound research is constrained by the scarcity of large, open collections of raw channel data. Existing public datasets are predominantly based on reconstructed images, while the smaller number of open raw channel datasets is generally specialized to individual reconstruction or imaging tasks, limiting the development of general-purpose methods and foundation models that operate directly on ultrasound signals. While interpretive tasks can be performed on reconstructed images, many of the key technical challenges in ultrasound imaging, such as aberration correction and reverberation reduction, require the raw channel data itself. We introduce OpenH-RF, a community-driven open dataset initiative comprising approximately 40 TB of medical ultrasound raw channel data in 19,332 acquisitions, contributed by 38 institutions across academia and industry, spanning in-vivo, phantom, and simulated data. Each dataset is annotated with a number of supported tasks, ranging from generalized reconstruction, speed-of-sound estimation, and aberration correction to blood flow imaging, motion estimation, quantitative ultrasound, and image interpretation. OpenH-RF provides a unified, self-contained data format that captures raw channel data together with the acquisition, probe, metadata, and reconstruction information required for reproducible processing across scanners and acquisition schemes. The corpus is publicly hosted under a permissive CC BY 4.0 license and can be streamed programmatically and selectively at scale. By bringing heterogeneous ultrasound data into a common representation, backed by modern data infrastructure, OpenH-RF lowers the barrier to working with raw channel data and aims to open medical ultrasound research to new communities. See https://open-h.github.io/OpenH-RF.

I. INTRODUCTION

OpenH-RF addresses the scarcity and fragmentation of open raw ultrasound channel data, which limits shared evaluation and work on signal-level imaging challenges. It brings together a large, diverse corpus under a unified format, with broad task coverage and permissive access.

I. INTRODUCTION

  • Raw channel data preserves per-element amplitude, phase, and frequency information discarded by B-mode images, supporting challenges such as aberration correction and reverberation reduction.Interpretive tasks such as segmentation, quantification, and diagnosis can be performed on B-mode images, but many technical imaging challenges require raw channel data.Source: §I · p. 3
  • Raw-channel research faces limited scanner access, fragmented vendor and lab formats, and data rates that can reach gigabytes per second and terabytes per session.These factors have traditionally constrained the size of private and open datasets.Source: §I · p. 3
  • Earlier open raw-channel datasets target individual tasks, cover limited anatomies, probes, and transmit schemes, and use separate formats; none approaches the scale and variety intended for task-agnostic foundation-model training.Examples include PICMUS, CUBDL,, PALA, and ULMShare [33].Source: §I · p. 3
  • OpenH-RF comprises 40 TB of raw channel data from 19,332 acquisitions spanning in-vivo, phantom, and simulated sources.The collection covers a wide range of anatomies, phantoms, probe types, and transmit schemes.Source: §I · p. 3
  • OpenH-RF uses a unified format to make acquisitions from different scanners, probes, and transmit sequences interoperable, and releases them under CC BY 4.0 for academic and commercial use.The initiative organizes intended tasks into five categories and includes supporting metadata such as reference probe positions, flow maps, and SoS maps.Source: §I · pp. 3–4

II. ULTRASOUND IMAGING

Ultrasound frames are formed from digitized echoes recorded across transducer elements and transmit events. Converting raw channel data into B-mode images entails assumptions and information loss, making image formation an approximate inverse problem.

II. ULTRASOUND IMAGING

  • A frame contains C × M × N raw-channel samples: C receive elements, M transmit events, and N fast-time samples per element.Samples may be stored as real-valued RF or demodulated complex-valued IQ signals.Source: §II · p. 4
  • Delay-and-sum beamforming aligns and sums channel signals using time-of-flight delays that usually assume a constant sound speed of 1540 m/s.The beamformed signal is envelope-detected and log-compressed to form a B-mode image; motion across frames supports Doppler and vector-flow imaging.Source: §II · p. 4
  • Image formation from raw channel data is an inverse problem because processing discards information and relies on assumptions violated by heterogeneous sound speed, reverberation, and frequency-dependent attenuation.The acquisition scheme itself is also a design choice that can be optimized.Source: §II · p. 4

III. METHODS

OpenH-RF gathered contributions through a call for proposals that required shareable channel data and a common submission format. The core team reviewed metadata and reconstruction reproducibility across contributions from 38 institutions.

A. Call for Contributions

  • Contributors submitted data under CC BY 4.0, with in-vivo ethics approval documented and protected health information removed.Accepted proposals received a submission guide requiring zea format, a reconstruction script, and a license file.Source: A. Call for Contributions · p. 5
  • The core team checked submissions for missing or spurious metadata and verified that reconstruction scripts produced the intended results against contributor-provided reference images.The submission process required a reference image to support reproducibility checks.Source: A. Call for Contributions · p. 5
  • The initiative collected contributions from 38 institutions across academia and industry, producing 33 contributions.Table I is captioned as organizing acquisitions by medium and intended tasks.Source: A. Call for Contributions · p. 5, Table I · p. 5

B. Task Categories

OpenH-RF groups intended applications into probe motion, tissue motion, and inaccurate image-formation assumptions. Access to channel data supports tasks that depend on pose, phase, arrival-time, or aperture-domain information.

1) Probe motion:

  • Tracked or mechanically moved probes and matrix arrays support 3D volume reconstruction, while phase-preserving measurements across probe positions can synthesize a larger aperture.Probe pose allows measurements from different positions to be combined.Source: 1) Probe motion · p. 5

2) Tissue motion:

  • Channel data supports tissue displacement and blood-flow estimation by providing control over focusing, phase-sensitive estimation, beamforming, angle compounding, and tissue suppression.Applications include shear-wave elastography, structure tracking, Doppler flow estimation, and microbubble localization for ultrasound localization microscopy.Source: 2) Tissue motion · p. 5

3) Inaccurate assumptions:

  • Raw channel access allows reconstruction methods to be compared on identical acquisitions and preserves arrival times and phase needed for aberration correction.Channel-domain lag-one coherence can also support evaluation; aperture-domain methods exploit coherence and arrival structure lost after summation for clutter suppression and image-quality assessment.Source: 3) Inaccurate assumptions · pp. 5–6

4) Tissue characterization:

The section describes how raw channel data supports tissue characterization and how acoustic emissions can be monitored during ultrasound therapy. It also illustrates that acquisitions can contain interleaved transmit tracks.

4) Tissue characterization:

  • Channel data preserves arrival times and phase for phase-based measures used in speed-of-sound estimation, and preserves backscatter frequency content for quantitative ultrasound.The cited applications include estimating tissue properties such as attenuation and backscatter coefficients.Source: 4) Tissue characterization · p. 6
  • Raw-to-insight models can identify or delineate structures and classify findings, and can study how reconstruction choices affect downstream performance.Applications include breast lesion diagnosis, tumor volumetry, vessel-wall delineation, and standard-view recognition.Source: 4) Tissue characterization · p. 6

5) Therapy:

  • Microbubble cavitation detection characterizes bubble emissions during ultrasound exposure; sufficiently wideband waveforms and array data support monitoring and passive location mapping.The passage connects this monitoring to the safety and efficacy of ultrasound drug delivery and blood-brain-barrier opening.Source: 5) Therapy · p. 6

C. Data Organization

OpenH-RF uses a unified, self-contained file format and access tooling to organize acquisitions and support reproducible processing. The format keeps raw measurements with acquisition and probe information, references, and reconstruction support.

C. Data Organization

  • Selected frames or transmits can be streamed and cached without loading the full acquisition.The data group is chunked along the transmit axis to support local or remote access to selected portions.Source: Figure 6 · p. 7

1) File format:

  • Each file stores one acquisition in a fixed hierarchy: raw channel data, scan parameters needed for reconstruction, and probe geometry and properties.Optional groups hold provenance and annotations, quality metrics, and contributor-specific arrays.Source: C. Data Organization · p. 6, 1) File format · p. 7
  • Task-specific references, including spatial maps, probe positions, microbubble positions, and segmentations, are stored alongside raw channel data.Derived maps include beamformed data, attenuation maps, images, and speed-of-sound maps used to evaluate aberration correction.Source: 1) File format · p. 7
  • A track is a labeled pair of data and scan groups for one transmit sequence; tracks share acquisition-level groups and a global schedule preserves transmit timing.Each track can be processed by its own pipeline while the acquisition remains in one file.Source: 1) File format · p. 7

2) Data access:

OpenH-RF pairs contributor-provided reconstruction pipelines with selective access to its large corpus. Its examples include tracked swept synthetic aperture and varied 3D acquisitions.

2) Data access:

  • Contributors provide a reproducible reconstruction pipeline and reference image, which are used to verify the intended output.Stored pipeline steps and the zea library version allow recipients to reproduce processing; GPU-accelerated differentiable operations also support integration into model training.Source: 2) Data access · p. 7

IV. RESULTS

  • Tracked swept synthetic aperture examples combine multiple individual acquisitions into an SSA reconstruction.The figure shows colorado-boulder and tumunich contributions, with individual acquisitions on the left and tracked SSA on the right.Source: Figure 7 · p. 7

2) Data access:

  • Compressed transmit-axis chunks let users read or remotely stream selected frames or transmits instead of loading an entire acquisition.Users can filter the corpus by anatomy, probe type, or intended task; downloading the full 40 TB corpus is impractical for most users.Source: 2) Data access · p. 7

IV. RESULTS

  • The 3D examples span catheter pullback, robot-tracked sweeps, probe rotation, and subject movement through a ring array.The figure also shows tracked probe paths and the distinct acquisition geometries.Source: Figure 8 · p. 8

1) 3D / tracked reconstruction (3D):

Tracked motion supports swept synthetic aperture reconstruction and lets acquisitions represent larger 2D or 3D regions. Examples span multiple tracking approaches and acquisition geometries.

1) 3D / tracked reconstruction (3D):

  • Six of 33 contributions record frame positions using tracking or positioning systems, and tracking can expand a small acquisition into a larger 2D or 3D one.The sources include robots, optical or electromagnetic tracking, a rotation stage, a catheter encoder, and a positioning system.Source: 1) 3D / tracked reconstruction (3D) · p. 8

a) Swept Synthetic Aperture

  • 3.4× and 1.8× narrower −6 dB lateral widths were measured after coherent summation on phantom wire targets for colorado-boulder and tumunich, respectively.The respective sweeps cover 160 mm with a 20 mm array and 70 mm with a 38 mm array.Source: 1) 3D / tracked reconstruction (3D) · p. 8, Table II · p. 8

1) 3D / tracked reconstruction (3D):

  • The 3D examples include an IVUS catheter pullback, robot-translated or curved-path sweeps, probe rotation, and a forearm stepped through a ring array.Frame spacing ranges from 0.04 mm for tumunich to 9 mm between us4us slices.Source: 1) 3D / tracked reconstruction (3D) · p. 8

2) Blood flow / ultrasound localization microscopy:

OpenH-RF includes flow and localization-microscopy datasets spanning simulations, phantoms, and in-vivo acquisitions, alongside contributions for sound-speed estimation and aberration correction. Stored sound-speed references show how reconstruction changes with the assumed medium.

2) Blood flow / ultrasound localization microscopy:

  • Six of 33 contributions target flow estimation across simulation, phantom, and in-vivo data, in both 2D and 3D.These include 99 mouse-brain acquisitions, human transcranial perfusion and carotid vector-flow data, a flow phantom, and simulations with known microbubble positions.Source: 2) Blood flow / ultrasound… · pp. 8–9

3) Speed-of-sound estimation / aberration correction:

  • In the nv-raw2insights-us phantom, a 100–200 m/s sound-speed mismatch smears a cyst into an ellipse, while the true-map reconstruction corrects the assumption.In dartmouth-uct, the water path blurs the skin and merges inclusions under the uniform 1540 m/s assumption; unc-liver changes little because its slow and fast layers average near 1540 m/s. Fig. 10 compares uniform-medium, stored-map, and true-map reconstructions.Source: Figure 10 · p. 9, 3) Speed-of-sound estimation /… · p. 9

2) Blood flow / ultrasound localization microscopy:

  • The mouse-brain collection contains 99 contrast-enhanced acquisitions from 61 animals; 12.8 seconds of one acquisition yield 3,016 microbubble tracks with ULM.Power Doppler resolves the vasculature from the same acquisition.Source: 2) Blood flow / ultrasound… · p. 9

3) Speed-of-sound estimation / aberration correction:

  • Nine of 33 contributions target sound-speed estimation or aberration correction, with references ranging from full maps and measured speeds to biopsy labels or no numerical reference.Four simulations store the full sound-speed map; two datasets provide measured sound speeds, while three in-vivo or phantom datasets lack a numerical reference.Source: 3) Speed-of-sound estimation /… · p. 9

B. Dataset Composition

OpenH-RF v1.0 combines 19,332 acquisitions across in-vivo, phantom, and simulated targets with diverse probes and transmit schemes. Its in-vivo data also spans a wide range of lag-one coherence values.

B. Dataset Composition

  • 19,332 acquisitions, 25,092,732 frames, and 241,078,230 transmit events make up OpenH-RF v1.0’s 33 contributions.The collection includes in-vivo and phantom targets alongside many in-silico simulations, covering a wide range of imaging subtypes.Source: B. Dataset Composition · p. 9
  • Contributions span linear, curved, ring, phased-array, and IVUS probes, alongside plane-wave, synthetic-aperture, focused, diverging-wave, coded, and mixed transmit schemes.Probe element counts and center frequencies also vary; mixed schemes combine multiple transmit types within one track.Source: B. Dataset Composition · p. 9

C. Lag-One Coherence (LOC)

  • The supplied in-vivo acquisitions span lag-one coherence (LOC), a normalized cross-correlation of delayed signals from neighboring array elements that estimates achievable contrast and CNR.The plotted subset includes acquisitions with at least three frames using full-aperture focused imaging, single-element transmits, or line-scanning, across the listed anatomy contributions.Source: C. Lag-One Coherence (LOC) · pp. 9–10

V. DISCUSSION

OpenH-RF combines a broad, heterogeneous raw-channel corpus with task annotations and a common format. Its composition varies substantially by counting unit, and the discussion identifies opportunities for cross-institution evaluation and data-driven methods.

V. DISCUSSION

  • The collection's figures characterize acquisition composition by medium and subtype, probe type and frequency, transmit scheme, and lag-one coherence by anatomy.The coherence plot shows individual focused in-vivo acquisitions and anatomy-specific means, with transducer pitch separated vertically.Source: Figure 11 · p. 10, Figure 12 · p. 10, Figure 13 · p. 11
  • 84% of acquisitions are simulated, while in-vivo data make up 98% of frames, 85% of which come from ulmshare.Because consecutive frames in long, high-frame-rate acquisitions are highly correlated, distinct subjects, animals, or phantoms are a better proxy for diversity when sampling.Source: §V · p. 11
  • Generalized reconstruction / beamforming is supported by 27 contributions; every other task has at least 5, while the sole therapy task appears only in twente-cavitation.Aberration correction and speed-of-sound estimation are one annotation category counted under both labels.Source: §V · pp. 11–12, Table III · p. 12
  • OpenH-RF supports evaluation across scanners, probes, and institutions, providing a common basis for benchmarking and generalization studies.Parallel tracks image the same target using different transmit schemes, enabling comparisons of their information gain or quality improvement.Source: §V · p. 12
  • Raw-channel learning opportunities include supervised regression and classification, unsupervised generative modeling, and a prospective foundation model conditioned on probe geometry and transmit parameters.The discussion also identifies possible uses in image-quality enhancement, compression, and patient-adaptive acquisition.Source: §V · p. 12
  • Each acquisition is a self-contained file with its reconstruction pipeline and can be streamed from Hugging Face Hub, lowering the access barrier for researchers outside ultrasound.The format and submission process are designed to accept future contributions, benchmarks, and shared models.Source: §V · p. 12

VI. CONCLUSION

OpenH-RF is presented as a large, openly licensed collection of raw ultrasound data from 38 institutions, spanning diverse tasks. The authors position it as a shared starting point for raw-signal research and future foundation-model development.

VI. CONCLUSION

  • OpenH-RF comprises approximately 40 TB of compressed raw channel data from 19,332 scans contributed by 38 academic and industry institutions.It covers tasks from general reconstruction to motion and blood-flow estimation, as well as tracked ultrasound.Source: §VI · p. 12
  • The authors present OpenH-RF as the largest openly licensed collection of raw medical ultrasound data to date, emphasizing accessibility through its permissive license and open-source ecosystem.They hope it supports rigorous evaluation and discovery in a research space they describe as traditionally restricted.Source: §VI · p. 12
  • A proposed direction is an ultrasound foundation model trained on raw channel data and conditioned on the provided metadata.The authors frame the dataset as a shared starting point for methods that use raw-signal information beyond beamformed images.Source: §VI · p. 12
Loading 2610.13616v1…