scidata-benchmark/BreastUSCT

This dataset contains 1,089 two-dimensional slices extracted from distinct three-dimensional numerical breast phantoms (NBPs) for ultrasound computed tomography (USCT) research, covering the four ACR BI-RADS breast composition categories (Types A–D). Derived from anatomical models developed by the VICTRE project and adapted for USCT studies, each sample includes spatial maps of speed of sound, density, acoustic attenuation, and tissue labels at a resolution of 2560 × 2560 pixels.

Breast types:

• A – almost entirely fatty
• B – scattered areas of fibroglandular density
• C – heterogeneously dense
• D – extremely dense

Each slice comes from a different 3D phantom, so no phantom appears in more than one split.

Usage

One HDF5 file per split: train.h5, validation.h5, test.h5. Download only the split you need and index it like a numpy array; only the samples you access are read from disk.

from huggingface_hub import hf_hub_download
import h5py

f = h5py.File(hf_hub_download("scidata-benchmark/BreastUSCT", "test.h5", repo_type="dataset"))

len(f["sos"])            # 110
f["sos"][0]              # (2560, 2560) float32
f["label"][:4]           # (4, 2560, 2560) uint8
f["type"].asstr()[:]     # ['A', 'A', ..., 'D']
f["subject_id"][:], f["y"][:]

With 🤗 datasets: load_dataset("scidata-benchmark/BreastUSCT", data_files={"test": "test.h5"}, split="test", streaming=True).with_format("numpy").

Splits

80/10/10, stratified by breast type:

split	A	B	C	D	total
train	80	322	387	82	871
validation	10	40	48	10	108
test	10	41	49	10	110

Fields

Each split file contains, for its N samples:

field	description	dtype / shape
sos	speed of sound [mm/μs]	float32, N × 2560 × 2560
den	ambient density [kg/mm³]	float32, N × 2560 × 2560
att	acoustic attenuation, power-law prefactor [dB / (MHz^y · mm)]	float32, N × 2560 × 2560
label	tissue labels: water 0, fat 1, skin 2, glandular 29, ligament 88, lesion 200	uint8, N × 2560 × 2560
y	attenuation power-law exponent	float64, N
type	breast type (A–D)	bytes, N
subject_id	phantom identifier (original file {type}{subject_id}.mat)	int64, N

Pixel size is 0.1 mm × 0.1 mm. Arrays use the same (row, column) orientation as the original MATLAB files. Samples are sorted by original file name.

Attribution

Source: Illinois Data Bank

Original dataset: https://databank.illinois.edu/datasets/IDB-5648161

Original authors: Li, Fu and Villa, Umberto and Park, Seonyeong and Anastasio, Mark

This repository redistributes the original dataset, repackaged as one HDF5 file per split. The train/validation/test split is an addition and is not part of the original release.

If you use this dataset in your research, please cite the original paper:

@data{illinoisdatabankIDB-5648161,
doi = {10.13012/B2IDB-5648161_V1},
url = {https://doi.org/10.13012/B2IDB-5648161_V1},
author = {Li, Fu and Villa, Umberto and Park, Seonyeong and Jeong, Gangwon and Anastasio, Mark A.},
publisher = {University of Illinois Urbana-Champaign},
title = {2D Acoustic Numerical Breast Phantoms for Ultrasound Computed Tomography},
year = {2025}
}