Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative

summary

The gist

Medical ultrasound is a widely accessible, real-time imaging modality, yet progress in data-driven ultrasound research is constrained by the scarcity of large…

In short

Open-H-RF is a massive, community-driven dataset of 40 TB of raw medical ultrasound channel data from 19,332 acquisitions across various sources. It provides a unified format including raw data and reconstruction information for reproducible processing. This initiative enables new AI research by allowing models to learn directly from the full signal content rather than just processed images.

Key concepts

Open-H-RF
A large, open dataset comprising approximately 40 TB of raw medical ultrasound channel data. It includes acquisition details, probe information, metadata, and reconstruction parameters in a unified format to ensure results are reproducible across different scanners and setups.
Raw Channel Data
The fundamental data collected directly from the ultrasound transducer elements. This includes arrival times and phase information across the acquisition. Using this raw data allows researchers to perform advanced tasks like adaptive beamforming, which goes beyond what standard displayed images can support.
Unified Format (zea file)
A standardized file structure used by Open-H-RF that stores a single acquisition along with all necessary parameters and metadata. This format is designed to capture the raw channel data alongside the specific reconstruction steps applied, making it self-contained for processing.

Terminology used across episodes

This episode discusses

The paper

Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative · Read on arXiv

NVIDIA Eindhoven University of Technology Department of Computer Science and Engineering, Stanford University

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 Open-H-RF, a community-driven open dataset initiative comprising approximately 40 TB of medical ultrasound raw channel data in 19,332 acquisitions, contributed by 39 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. Open-H-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, Open-H-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/Open-H-RF.

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative".

Paper summary: Rosa: So, to wrap up this discussion about Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative, it’s clear that the sheer volume and variety of raw channel data is what matters here.

Dev: The authors are providing a unified format for that forty terabytes of data, which is organized so you can process it reproducibly across different scanners and acquisition schemes <ref:2610.13616#pg3>.

Taro: It’s really about lowering the barrier for researchers who might not be deep in ultrasound to start building things on this foundation.

Rosa: They are opening up possibilities beyond just the images we see, focusing on tasks like tissue characterization and correcting those tricky inaccurate assumptions.

Dev: The next natural step they point towards is training an ultrasound foundation model that uses this raw channel data and conditions it with all that available metadata.

Taro: So it’s moving the field away from just the beamformed image towards methods that actually exploit the full information content of the raw signal.

Rosa: That’s what they are aiming for, using this dataset as a starting point for everyone to build upon and explore new data-driven solutions in ultrasound.

Conclusion: Rosa: So this paper is about Open-H-RF, which is this massive collection of raw ultrasound data they’ve put together from almost forty terabytes across all kinds of scans.

Dev: It’s essentially a shared data initiative, a way for everyone to get access to the actual raw channel information instead of just the final pictures.

Rosa: The authors are bringing together people from academia and industry to contribute this data, spanning everything from living tissue to simulations.

Dev: They’ve built this unified format so you can process it reproducibly no matter what scanner or acquisition setup you started with.

Rosa: What this means is that we’re moving past just looking at the final image and starting to work with the raw signal itself, which opens up ways to do much deeper analysis.

Dev: It’s about building things like foundation models that are trained directly on these raw channel data sets, conditioned by all that acquisition information they include.

Rosa: So it’s a big step toward getting datadriven solutions that really understand the underlying physics of how the ultrasound is made.

Dev: This moves us toward studying things like tissue characterization and motion estimation at a much more fundamental level than we’ve done before.

Rosa: The sheer volume and variety suggest this is going to be a major resource for researchers trying to push past just beamformed images.

More episodes

← Home