Datasetpublic

nvidia/NV-Raw2Insights-US

NV-Raw2Insights-US Simulations Dataset Description NV-Raw2Insights-US Simulations is a simulated full synthetic aperture (FSA) ultrasound dataset for training and evaluating neural networks on sound speed estimation, phase aberration correction, and tissue segmentation. Each sample is a single-frame FSA acquisition from a 180-element linear array simulated over a heterogeneous tissue phantom containing cysts. The dataset provides raw baseband IQ channel data… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/NV-Raw2Insights-US.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face

3likes336downloads

Dataset Card

NV-Raw2Insights-US Simulations

Dataset Description

NV-Raw2Insights-US Simulations is a simulated full synthetic aperture (FSA) ultrasound dataset for training and evaluating neural networks on sound speed estimation, phase aberration correction, and tissue segmentation.

Each sample is a single-frame FSA acquisition from a 180-element linear array simulated over a heterogeneous tissue phantom containing cysts. The dataset provides raw baseband IQ channel data alongside ground truth sound speed maps, binary cyst segmentation masks, and phase aberration values.

This dataset is ready for commercial use.

Dataset Owner(s)

NVIDIA Corporation

Dataset Creation Date

12/01/2025

License/Terms of Use

Governing Terms: This dataset is licensed under a Creative Commons Attribution 4.0 International License.

Intended Usage

NV-Raw2Insights-US Simulations is intended to be used by the research community to develop and benchmark methods for ultrasound image reconstruction, sound speed estimation, phase aberration correction, and tissue segmentation from raw channel data.

Dataset Characterization

Data Collection Method<br> Synthetic (k-Wave acoustic simulation)

Labeling Method<br> Synthetic (ground truth from simulation parameters)

Dataset Format

Raw Ultrasound Channel Capture Data (Apache Arrow)

Dataset Quantification

923 samples (830 train / 93 validation)

17 features per sample:

Feature	Shape	Dtype	Units	Description
`iq_real`	`[180, 180, 1024]`	float32	--	Real part of baseband IQ. Axes: `[n_tx, n_el, n_ax]`
`iq_imag`	`[180, 180, 1024]`	float32	--	Imaginary part of baseband IQ
`bmode`	`[507, 456]`	float32	linear	Synthetic aperture DAS b-mode (envelope, not log-compressed)
`bmode_focused`	`[507, 456]`	float32	linear	Focused transmit b-mode
`bmode_extent`	`[4]`	float32	mm	Spatial extent `[x_min, x_max, z_max, z_min]`
`sound_speed_map`	`[32, 32]`	float32	m/s	Ground truth speed of sound map
`sound_speed_extent`	`[4]`	float32	m	Spatial extent `[x_min, x_max, z_max, z_min]`
`segmentation_map`	`[507, 456]`	uint8	--	Binary segmentation: `0` = background tissue, `1` = cyst
`phase_error`	scalar	float32	radians	Per-sample phase aberration error
`fs`	scalar	float64	Hz	Sampling frequency (13.3 MHz)
`fc`	scalar	float64	Hz	Center frequency (6.5 MHz)
`fd`	scalar	float64	Hz	Demodulation frequency (6.5 MHz)
`t0`	scalar	float64	s	Time of transmit peak pressure
`c0`	scalar	float64	m/s	Background sound speed (1540 m/s)
`elpos`	`[3, 180]`	float32	m	Element positions `[x, y, z]` for each of 180 elements
`frame_idx`	scalar	int32	--	Frame index within source simulation (always `0` — each sample is a single-frame acquisition)

Pre-computed IQ normalization statistics:

Statistic	Value
`iq_rms_global`	0.6616
`iq_mean_real_global`	0.0021
`iq_mean_imag_global`	-0.0016
`iq_max_mean_global`	223.67

Supported Tasks

—Sound speed estimation: Predict sound_speed_map from iq_real / iq_imag
—Phase aberration estimation: Predict phase_error from IQ data
—Tissue segmentation: Predict segmentation_map from bmode or IQ data
—Image reconstruction: Reconstruct bmode from raw IQ data (learned beamforming)

Usage

python

from datasets import load_dataset

ds = load_dataset("nvidia/nv-raw2insights-us")

sample = ds["train"][0]
iq = sample["iq_real"] + 1j * sample["iq_imag"]  # [180, 180, 1024] complex64
sos = sample["sound_speed_map"]                    # [32, 32] float32
seg = sample["segmentation_map"]                   # [507, 456] uint8

Known Issues

—bmode_extent is in millimeters while sound_speed_extent and elpos are in meters.
—bmode_extent z-axis ordering is [x_min, x_max, z_max, z_min] (zmax before zmin).
—Cysts that touch the edge of the b-mode frame are not included in segmentation_map.
—Some regions of gross reverberation artifact are incorrectly segmented as cysts in segmentation_map.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Citation

bibtex

@misc{nv_raw2insights_us_simulations_2026,
  title={NV-Raw2Insights-US Simulations},
  author={Simson, Walter and Huver, Sean},
  year={2026},
  publisher={NVIDIA Corporation},
  howpublished={\url{https://huggingface.co/datasets/nvidia/nv-raw2insights-us-simulations}},
  license={CC BY 4.0}
}