EPFL Ultrafast Ultrasound Datasets

Deep Learning-based Inpainting for Sparse Arrays in Ultrafast Ultrasound Imaging: Dataset

Overview

This website provides access to the dataset associated with the publication:

R. Viñals, J.-P. Thiran — Deep Learning-based Inpainting for Sparse Arrays in Ultrafast Ultrasound Imaging, IEEE Transactions on Computational Imaging, 2025.

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/. Please cite this paper when using the dataset. 

The EPFL Ultrafast Ultrasound Dataset is a large-scale collection of in vivo and in vitro ultrafast ultrasound acquisitions designed for research on deep learning, image reconstruction, beamforming, and sparse-array imaging. It includes over 20,000 in vivo acquisitions, detailed metadata, and complete acquisition settings.

Download the Dataset

For stable and reproducible access, we recommend downloading via the direct links below or using the S3 interface (instructions provided further down).

Core Files

• README: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/README.md
• In vivo metadata (CSV): https://datasets.epfl.ch/epfl_ultrafast_ultrasound/invivo_dataset.csv
  • Contains one row per acquisition, with: id,volunteer_id,body_region
• Acquisition & beamforming settings (ZIP): https://datasets.epfl.ch/epfl_ultrafast_ultrasound/settings.zip
• In vitro phantom dataset: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/invitro_CIRS054GS.zip

In Vivo Datasets

• Volunteer 1:
  • Abdomen: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_1_abdomen.zip
  • Breast, carotid & shoulders: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_1_breast_carotid_shoulders.zip
  • Legs: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_1_legs.zip
  • Arm + Back: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_1_arm_back.zip
• Volunteers 2–9:
  • Volunteer 002: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_002.zip
  • Volunteer 003: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_003.zip
  • Volunteer 004: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_004.zip
  • Volunteer 005: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_005.zip
  • Volunteer 006: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_006.zip
  • Volunteer 007: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_007.zip
  • Volunteer 008: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_008.zip
  • Volunteer 009: https://datasets.epfl.ch/epfl_ultrafast_ultrasound/volunteer_009.zip

Access via S3 (recommended for large transfers)

The dataset is hosted on an S3-compatible endpoint and can be accessed using rclone.

1. Create an rclone remote
   • Add to ~/.config/rclone/rclone.conf:
     • [datasets_epfl]
     • type = s3
     • provider = Other
     • endpoint = https://datasets.epfl.ch
     • region = us-east-1
     • env_auth = false
     • access_key_id =
     • secret_access_key =
   • Or create it automatically: rclone config create datasets_epfl s3 provider Other endpoint https://datasets.epfl.ch region us-east-1
2. List all files: rclone ls datasets_epfl:epfl_ultrafast_ultrasound
3. Download a file:rclone copy datasets_epfl:epfl_ultrafast_ultrasound/volunteer_002.zip . -P
4. Download the full dataset: rclone sync datasets_epfl:epfl_ultrafast_ultrasound ./epfl_ultrafast_ultrasound -P

Dataset Description

• In Vivo Data

  • Total number of acquisitions: 20,000
  • Participants: 9 volunteers
  • Format: One .zip archive per volunteer (multiple archives for Volunteer 1)
  • Acquisition Distribution by Body Region
    • Abdomen: 6,599
    • Neck: 3,294
    • Breast: 3,291
    • Lower limbs: 2,616
    • Upper limbs: 2,110
    • Back: 2,090
  • Additional (unused in publications):
    • Shoulders: 123 (volunteer_001)
    • Arms: 19 (invivo_04451.npz → invivo_0469.npz)
  • File Naming: invivo_00000.npz → invivo_20141.npz

• In Vitro Data

  • Total number of acquisitions: 2,179
  • Phantom: CIRS Model 054G
  • File Naming: invitro_00000.npz → invitro_02178.npy

Acquisition & Beamforming Parameters

Provided in the settings folder:

File Format

All acquisitions are stored as NumPy arrays: import numpy as np data = np.load("file.npz")

Publications Using This Dataset

Selected In Vivo Acquisitions by Paper

In Vitro Acquisitions Used in Publications

Training / Validation / Test Splits

• Default split (used in most papers):
  • Training: 1, 2, 3, 6, 7, 9
  • Validation: 4
  • Test: 5, 8
• Split used in IUS 2023 (Quality Enhancement of Ultrafast Ultrasound Images with Deep Networks and Transfer Learning)
  • Training: 1, 2, 3, 4, 6, 7, 9
  • Validation: 8
  • Test: 5

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/

Please cite the corresponding publication when using this dataset.

Contact

For questions or issues:

• Roser Viñals: 📧 [email protected]
• Jean-Philippe Thiran: 📧 [email protected]

Multi-Anatomical Ultrafast Ultrasound RF Dataset: Raw RF Signals acquired with a GE 11L-D probe

This website provides access to the dataset associated with the Zenodo Repository: Multi-Anatomical Ultrafast Ultrasound RF Dataset: Raw RF Signals acquired with a GE 11L-D probe, accompanying our data paper published in Data in Brief (doi: 10.1016/j.dib.2026.113259).

Important Note: Due to the large volume of data (20,431 acquisitions), the actual raw RF files are hosted on the EPFL servers to ensure high-speed access and efficient data management. This Zenodo record contains the permanent DOI, dataset metadata, acquisition parameters, and documentation required for citation.

📄 Citation

If you use this dataset in your research, please cite both the Zenodo repository and the associated data paper:

1. Zenodo Repository:

Roser Viñals, Jean-Philippe Thiran. (2026). Multi-anatomical ultrafast ultrasound dataset: raw RF signals acquired with a GE 11L-D probe. Zenodo. https://doi.org/10.5281/zenodo.20054330

2. Associated Data Paper:

Roser Viñals, Jean-Philippe Thiran (2026). A multi-anatomical in vivo and in vitro ultrafast ultrasound dataset: 20,431 acquisitions with 105 plane waves using the GE 11L-D probe and a vantage system. Data in Brief, Volume 69, 113259. https://doi.org/10.1016/j.dib.2026.113259

📊 Dataset Overview

This dataset comprises 20,431 acquisitions (20,121 in vivo from 16 volunteers and 310 in vitro from a CIRS phantom). It is designed for benchmarking beamforming algorithms, validating image reconstruction techniques, and training deep learning models.

• System: Verasonics Vantage 256 / GE 11L-D Probe

• Format: NumPy arrays (.npz)

• Total Size: 4.0 TB

📥 Data Access

Documentation & Metadata

Available for direct download and via Zenodo

• 📄 README: Download

• 📊 Acquisition Metadata (CSV): Download

• ⚙️ Acquisition Settings (ZIP): Download

In Vitro Volunteer Datasets (ZIP): Download

In Vivo Volunteer Datasets (ZIP):

🔗 Bulk Download via S3

To download the full dataset, we recommend using the S3-compatible endpoint to sync data via rclone.

# 1. Create the remote
rclone config create datasets_epfl s3 provider Other endpoint https://datasets.epfl.ch region us-east-1

# 2. Sync the entire repository
rclone sync datasets_epfl:epfl_ge11ld_ultrafast_ultrasound ./local_folder -P

🔬 Research Context & Reproducibility

This repository contains the complete dataset described in the following paper:

R. Viñals and J.-P. Thiran, “A multi-anatomical in vivo and in vitro ultrafast ultrasound dataset: 20,431 acquisitions with 105 plane waves using the GE 11L-D probe and a vantage system,” Data in Brief, vol. 69, p. 113259, 2026, doi: 10.1016/j.dib.2026.113259.

A subset of this dataset was utilized in the following peer-reviewed study:

R. Viñals and J. -P. Thiran, “Deep Learning-Based Inpainting for Sparse Arrays in Ultrafast Ultrasound Imaging,” in IEEE Transactions on Computational Imaging, vol. 12, pp. 187-202, 2026, doi: 10.1109/TCI.2025.3648531.

Note for Reproducibility:

• Angle Selection: While the original acquisitions contain 105 plane waves, the cited paper utilized only the 103 plane waves with angles closest to 0°.

• Test Sets: Volunteers 005, 008, and 017 were used as the test set in the cited study.

📜 License

This dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Contact

For questions regarding the data or issues with access, please contact:

Roser Viñals ([email protected])

Jean-Philippe Thiran ([email protected])