EchoXFlow

A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function

 

_{A small subset of an echocardiographic examination. Top row, left to right: parasternal long axis, tissue Doppler, colour Doppler, strain imaging, parasternal short axis. Second row: pulsed-wave tissue velocity. Third row: transmitral and aortic flow. Last row: electrocardiogram. Reproduce this figure with scripts/render_banner.py from the GitHub repository.}

EchoXFlow is 666 routine-care echocardiographic examinations with 35,377 recordings, stored in the scanner's native beamspace (polar) grid before scan conversion. Every recording carries a co-registered ECG. Recordings include time-resolved 1D, 2D and 3D B-mode, colour, tissue, pulsed-wave and continuous-wave Doppler, together with dense annotations (2D speckle-tracking contours, 3D left-ventricular meshes) and sparse annotations (sample volumes, measurement points and spectral traces).

The dataset paper is published in the NeurIPS 2026 Evaluations and Datasets track.

This repository holds the Croissant metadata plus one uncompressed tar archive per examination. The data is read with the EchoXFlow code repository, which provides the readers, rendering and filtering.

Data types (see Table 3 of the paper for details)

Data type	Items	Unique exams	Sampling rate (Hz, median)
Electrocardiogram	35,377	666	600
Brightness mode
1D (M-mode)	2,527	643	496
2D (single plane)	26,356	666	35
2D (dual plane)	657	497	24
2D (triplane)	573	479	47
3D	716	487	28
Doppler
2D Tissue	1,194	469	30
2D Color	8,451	663	21
1D Pulsed Wave	5,452	661	534
1D Continuous Wave	2,807	657	501
Dense annotations
3D LV	201	199	27
2D LV A2C	443	334	49
2D LV A3C	277	273	49
2D LV A4C	443	334	49
2D LA A2C	9	9	56
2D LA A4C	11	11	79
2D RV A4C	25	25	49
Sparse annotations
Sample volumes, 1D Pulsed Wave	5,445	661
Sample volumes, 1D Continuous Wave	2,760	652
Points and markers	3,843	561

Contents

Path	What it is
croissant.json	Croissant 1.1 metadata: recordings, arrays, streams, recording_links and merged_recordings record sets, plus RAI fields
exams/<exam_id>.tar	One archive per examination containing exams/<exam_id>/<recording_id>.zarr/, the Zarr v2 recording stores
splits.csv	Patient-level five-fold split (exam_id,five_fold_split): no two examinations of the same patient are in different folds
view_classifier_probabilities.csv	Per-recording view probabilities from the scanner's own view detector, EchoPrime and EchoFine, for all single-plane B-mode recordings
LICENSE	CC BY-NC-SA 4.0

Download and extraction

Download the repository with the Hugging Face CLI:

pip install -U huggingface_hub
hf download Ahus-AIM/EchoXFlow --repo-type dataset --local-dir EchoXFlow

or from Python:

from huggingface_hub import snapshot_download
snapshot_download("Ahus-AIM/EchoXFlow", repo_type="dataset", local_dir="EchoXFlow")

Git with LFS works as well: git lfs install && git clone https://huggingface.co/datasets/Ahus-AIM/EchoXFlow.

Then extract every exam archive under one data root. The archives contain the exams/<exam_id>/... prefix, so all of them extract into the layout that croissant.json refers to:

cd EchoXFlow
mkdir -p data
cp croissant.json splits.csv view_classifier_probabilities.csv data/
for archive in exams/*.tar; do
  tar -xf "$archive" -C data
done
export ECHOXFLOW_DATA_ROOT="$PWD/data"

Code

Install the EchoXFlow package and use the GitHub repository for the readers, rendering, filtering and dataset building:

pip install git+https://github.com/Ahus-AIM/EchoXFlow

Recording stores

Each <recording_id>.zarr is a directory-backed Zarr v2 group with consolidated metadata and these namespaces:

• data/: raw samples as acquired (B-mode, Doppler IQ-derived colour samples, spectral strips, ECG), in beamspace.
• derived/: quantities computed from the raw samples.
• coordinates/: calibration and geometry (depth, angles, PRF, Nyquist, segment offsets).
• timestamps/: one time axis per stream, on the recording clock, aligned with the ECG.
• annotations/: contours, meshes, QRS triggers, measurements and spectral traces.

Root attributes list content_types, frame_counts_by_content_type, stream_summaries and the array inventory. The arrays record set in croissant.json carries the shape, dtype, role and SHA-256 of every array.

Views

Every single-plane B-mode recording carries labels from two published view classifiers, EchoPrime and EchoFine, run on eight evenly spaced scan-converted frames. The labels are predictions, not expert annotations. Each record exposes the top-1 label and probability of both classifiers and a coarse view that is set only when the two agree: A2C, A3C, A4C, PLAX, PSAX, Other, or none on disagreement. Multi-plane B-mode recordings are not classified. The native_* columns are the scanner's own view-detector scores, kept as exported.

Splits

splits.csv assigns each of the 666 examinations to one of five folds (133 or 134 per fold). The split is made at patient level: every examination of the same patient is in the same fold, so no patient appears in two folds.

Limitations

Single-centre, single-vendor data with a selected population. The dataset is not representative of all institutions, vendors, acquisition protocols or patient subgroups, and it is not validated for clinical diagnosis, triage, treatment or prognosis. See the rai: fields of croissant.json.

License

CC BY-NC-SA 4.0, see LICENSE. Direct and indirect patient identifiers are removed; released metadata uses anonymised public IDs.

Citation

@inproceedings{stenhede_2026_echoxflow,
  title = {EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function},
  author = {Stenhede, Elias and Sulkowska, Joanna and Orstad, Eivind Bj{\o}rkan and Schirmer, Henrik and Ranjbar, Arian},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year = {2026},
  eprint = {2605.05447},
  doi = {10.48550/arXiv.2605.05447},
  url = {https://arxiv.org/abs/2605.05447}
}

If you use the view labels, also cite EchoPrime (Vukadinovic et al., Nature 2026, doi:10.1038/s41586-025-09850-x) and EchoFine (Naidoo et al., Medical Image Analysis 2026, doi:10.1016/j.media.2026.104006).