OpenBreastUS: Ultrasound CT Benchmark

• OpenBreastUS is a large-scale breast ultrasound CT benchmark featuring 8,000 phantoms and over 16 million wave simulations to bridge theoretical models with practical imaging.
• It employs advanced neural operators and simulation techniques to solve high-frequency Helmholtz equations for both forward modeling and inverse reconstruction.
• The term broadly encompasses open breast-ultrasound datasets, codebases, and preprocessing pipelines, supporting clinical risk assessment and image-based CAD.

Searching arXiv for the named resources and closely related breast-ultrasound works. OpenBreastUS most specifically denotes a large-scale wave-imaging dataset and benchmarking suite for breast ultrasound computed tomography (USCT), introduced as “OpenBreastUS: Benchmarking Neural Operators for Wave Imaging Using Breast Ultrasound Computed Tomography” (Zeng et al., 20 Jul 2025). In a broader breast-ultrasound research context, the same label is also used informally for adjacent open resources: the combined release of the Breast Lesion Dataset and the PersonalizedUS system for malignancy risk assessment (Fröhlich et al., 2024), and, in spirit, the breast subset of the OpenUS-46 corpus used for self-supervised ultrasound representation learning (Megahed et al., 27 Oct 2025). The term is therefore not fully uniform across the literature. Its most explicit and official use refers to the 2025 USCT benchmark; elsewhere it functions as a convenient umbrella label for open breast-ultrasound datasets, codebases, and preprocessing or modeling pipelines.

1. Definition and scope

In its official 2025 sense, OpenBreastUS is “a large-scale wave equation dataset designed to bridge the gap between theoretical equations and practical imaging applications,” containing “8,000 anatomically realistic human breast phantoms and over 16 million frequency-domain wave simulations using real USCT configurations” (Zeng et al., 20 Jul 2025). The dataset is intended for both forward simulation and inverse imaging, with particular emphasis on benchmarking neural operators under clinically realistic wave-physics conditions.

A distinct but related usage arises from the paper introducing PersonalizedUS, where “OpenBreastUS” does not appear in the paper itself but “naturally refers to the combined resource the authors commit to releasing after peer review: an open, curated breast ultrasound dataset plus open-source code for the PersonalizedUS system” (Fröhlich et al., 2024). In that usage, the term maps to two components: the Breast Lesion Dataset, a public CSV of biopsy-verified lesions, and the PersonalizedUS codebase and web application.

A third, “in spirit” usage appears in the context of USF-MAE, whose OpenUS-46 corpus aggregates multiple public ultrasound datasets. The breast-related subset includes BrEaST, BUS (Dataset B), BUS_UC, BUS_UCML, and S1, and can be assembled directly from the repository link list (Megahed et al., 27 Oct 2025). This suggests that “OpenBreastUS” may function as an editorial shorthand for an openly assembled breast-ultrasound subset, even when that term is not the official project title.

A recurring misconception is that any open breast-ultrasound or breast-USCT resource is formally named OpenBreastUS. That is not generally the case. For example, the stochastic numerical breast-phantom resource for USCT explicitly states that “OpenBreastUS” is “a misnomer for this work” and should not be used in place of its official dataset and code DOIs (Li et al., 2021).

2. OpenBreastUS as a USCT wave-imaging benchmark

The official OpenBreastUS benchmark targets the computational bottleneck of USCT: repeated accurate solution of high-frequency wave equations during forward modeling and inverse reconstruction. The underlying forward model is the heterogeneous Helmholtz equation,

$$\left[ \nabla^{2}+\left( \frac{\omega}{c(x)} \right)^2 \right] u(x)=-s(x),$$

with a Sommerfeld-type radiation condition outside the region of interest (Zeng et al., 20 Jul 2025). The inverse problem is formulated as PDE-constrained minimization of data misfit over the sound-speed field $c(x)$, with adjoint-state gradients derived from forward and adjoint Helmholtz solves.

The dataset scale is unusually large for realistic biomedical wave imaging. It contains 8,000 anatomically realistic human breast phantoms, each paired with 256 sources and 8 frequencies, giving $8{,}000 \times 256 \times 8 = 16{,}384{,}000$ complex wavefield simulations (Zeng et al., 20 Jul 2025). The phantoms are 2D slices at $480 \times 480$ resolution on a $240 \text{ mm} \times 240 \text{ mm}$ domain with 0.5 mm grid spacing, surrounded by water and interrogated by a 220 mm diameter annular array of 256 transducers.

The anatomy is generated with the FDA’s VICTRE 3D simulation tool and then sliced into 2D tissue maps. Four density categories are represented with explicit train/test splits and storage footprints: HET $1{,}800/200$, FIB $2{,}700/300$, FAT $1{,}800/200$, and EXD $900/100$, totaling approximately 28.8 TB (Zeng et al., 20 Jul 2025). The simulations use 8 operating frequencies from 300 to 650 kHz in 50 kHz steps. The dataset primarily varies sound speed; density and attenuation are not modeled, which the paper states as a limitation.

Ground-truth fields are generated with the Convergent Born Series algorithm, selected for convergence and numerical stability in strongly scattering heterogeneous Helmholtz problems (Zeng et al., 20 Jul 2025). This makes OpenBreastUS a benchmark not merely for surrogate learning on simplified PDEs, but for testing whether learned operators can cope with high-frequency, clinically motivated acoustic heterogeneity.

3. Neural-operator benchmarking and in vivo reconstruction

OpenBreastUS was designed as a benchmarking platform for both forward operators and inverse solvers. The forward task learns the map

$$\mathcal{G}_{\omega_j}:\left(c(x),s(x)\right)\mapsto u(x),$$

while inverse learning targets either a direct mapping from multi-frequency measurements to sound-speed images or a model-based iterative reconstruction in which learned forward surrogates replace repeated numerical solves (Zeng et al., 20 Jul 2025).

The forward benchmarks compare UNet, FNO, AFNO, BFNO, and MgNO. Across test frequencies from 300 to 600 kHz, MgNO achieves the lowest reported errors. At 300 kHz, for example, the reported relative RMSEs are UNet 0.1236, FNO 0.0269, AFNO 0.0165, BFNO 0.0113, and MgNO 0.0028; at 600 kHz they are 0.1969, 0.0644, 0.0383, 0.0285, and 0.0092, respectively (Zeng et al., 20 Jul 2025). The inverse comparisons show a similar ranking shift when reconstruction is embedded in a model-based framework: direct inversion with DeepONet, InversionNet, and NIO yields PSNR/SSIM of 16.65/0.8572, 20.26/0.8640, and 18.06/0.8692, whereas neural-operator-accelerated FWI yields 20.02/0.8674 for UNet-FWI, 25.84/0.9104 for FNO-FWI, 24.85/0.8506 for AFNO-FWI, 27.91/0.9193 for BFNO-FWI, and 30.48/0.9381 for MgNO-FWI (Zeng et al., 20 Jul 2025).

The paper also studies frequency scaling, source-position generalization, breast-type generalization, and media transfer. A consistent pattern is that simplified models generalize poorly outside their training distribution, whereas global or multiscale neural operators degrade more gracefully. This supports the paper’s central claim that realistic wave-imaging datasets are necessary to avoid overly optimistic conclusions drawn from oversimplified PDE benchmarks.

The most consequential result is a reported in vivo demonstration. Clinical data from Karmanos Cancer Institute, acquired with a 22 cm ring transducer array of 1024 elements and pulse center frequency 2.5 MHz, were reconstructed using forward surrogates trained on OpenBreastUS combined with gradient-based inversion over 8 frequencies between 300 and 650 kHz (Zeng et al., 20 Jul 2025). In the malignant case, BFNO-based reconstruction reportedly delineated the malignant tumor and tissue layers, while NIO produced incorrect structures. In a benign cyst case, BFNO and FNO recovered breast anatomy and the cyst’s smooth boundary. This is presented as the first demonstration of efficient in vivo imaging of the human breast using neural operator solvers.

4. OpenBreastUS as an open clinical risk-assessment resource

In the PersonalizedUS literature, “OpenBreastUS” designates a different but complementary resource: the release of an open multicenter breast-ultrasound lesion table together with interpretable malignancy-risk software (Fröhlich et al., 2024). The Breast Lesion Dataset contains $c(x)$0 biopsied breast lesions with clinical, BI-RADS, and Doppler features derived from a multicenter observational study with retrospective and prospective cohorts. Inclusion criteria were “breast lesions classified as BI-RADS at least 3 identifed by ultrasound and submitted to percutaneous core needle biopsy and/or excisional biopsy,” while exclusion criteria included BI-RADS 2 lesions, lesion size greater than 30 mm, age less than 18 years, non-primary breast pathologies, and lesions submitted exclusively to fine-needle aspiration cytology (Fröhlich et al., 2024).

The associated system, PersonalizedUS, combines regularized logistic regression with Locart-based local conformal prediction. Residuals on the calibration set are defined as

$c(x)$1

and subgroup-specific prediction sets are constructed through a decision-tree partition of feature space and local quantile calibration (Fröhlich et al., 2024). The resulting prediction set is

$c(x)$2

This framework is intended to deliver interpretable risk scores together with local coverage guarantees. The main reported logistic-regression performance is Sensitivity 0.900, Specificity 0.888, PPV 0.897, NPV 0.892, with AUROC 0.958, AUPRC 0.960, and log-loss 0.268 (Fröhlich et al., 2024). For local conformal prediction at $c(x)$3, most leaves reportedly had average set size close to 1 and “Truth Only (%)” above 90%, whereas two “hard” leaves had larger set sizes and lower truth-only percentages, together covering about 20% of the test set (Fröhlich et al., 2024).

A clinically salient case study concerned BI-RADS 4a/4b lesions in the prospective test set. At threshold $c(x)$4, only 64 of 196 biopsies would be requested, corresponding to a greater than 65% reduction, and the model “wouldn't miss any cancer cases among BI-RADS 4a and 4b lesions” (Fröhlich et al., 2024). The web application was deployed at four Brazilian reference centers: HCRP-USP, MATER, Hospital de Amor de Barretos, and Hospital de Amor de Campo Grande.

5. Data curation, preprocessing, and image-based CAD components

Any broader OpenBreastUS ecosystem requires preprocessing and artifact control because real-world breast ultrasound contains irregularities not reliably captured in metadata. BUSClean was developed precisely for this setting as open-source software for BUS scan filtering, cleaning, and knowledge extraction (Bunnell et al., 2024). Its pipeline detects invalid scans, non-B-mode images, dual-view frames, scan-area boundaries, lesion calipers, and burnt-in annotations; it also extracts BI-RADS labeling and measurement fields from text.

On an internal testing set of 430 unseen clinical BUS images, BUSClean reported 100% sensitivity and 99.5% specificity for enhanced/non-B-mode detection, 96.7% sensitivity and 93.3% specificity for lesion caliper detection, 100% sensitivity and 98.6% specificity for dual-view scan detection, and 100% sensitivity and 100% specificity for invalid scan detection (Bunnell et al., 2024). On the external BUSI dataset, text-presence detection achieved 88.6% sensitivity and 98.3% specificity, and enhanced/non-B-mode detection achieved 90.9% sensitivity and 99.9% specificity. Caliper detection improved from 43.3% sensitivity and 93.3% specificity out of the box to 92.1% sensitivity and 92.3% specificity after a Hough-based adaptation for the external caliper style (Bunnell et al., 2024). These results indicate that preprocessing itself is a domain-adaptation problem rather than a one-time data-cleaning step.

For image-based CAD under limited data, the BIRADS-SDL framework offers a different strand of methodology. It transforms ultrasound images into BIRADS-oriented feature maps,

$c(x)$5

where $c(x)$6 is distance to the lesion boundary (Zhang et al., 2019). The method feeds these maps into a semi-supervised stacked convolutional autoencoder classifier. On UDIAT, BIRADS-SDL reported ACC $c(x)$7, AUC $c(x)$8, MCC $c(x)$9, SEN $8{,}000 \times 256 \times 8 = 16{,}384{,}000$0, and SPE $8{,}000 \times 256 \times 8 = 16{,}384{,}000$1; on UTSW, the corresponding values were $8{,}000 \times 256 \times 8 = 16{,}384{,}000$2, $8{,}000 \times 256 \times 8 = 16{,}384{,}000$3, $8{,}000 \times 256 \times 8 = 16{,}384{,}000$4, $8{,}000 \times 256 \times 8 = 16{,}384{,}000$5, and $8{,}000 \times 256 \times 8 = 16{,}384{,}000$6 (Zhang et al., 2019). The best-performing Gaussian width was reported around $8{,}000 \times 256 \times 8 = 16{,}384{,}000$7 pixels.

For semantic segmentation, a fuzzy FCN with anatomy-constrained CRF provides a five-class breast-ultrasound labeling scheme over fat, mammary layer, muscle layer, background, and tumor (Huang et al., 2019). The reported performance of the full method is tumor TPR 90.33%, FPR 9.00%, tumor IoU 81.29%, and overall mIoU 80.47% across five categories (Huang et al., 2019). A plausible implication is that an “OpenBreastUS” platform oriented toward image understanding would require not only open lesion labels but also robust segmentation and artifact-management primitives.

6. Relation to foundation models, generative models, and other “open” breast-ultrasound resources

The USF-MAE work situates breast ultrasound within a larger modality-wide self-supervised corpus. OpenUS-46 contains 374,192 ultrasound images from 46 public datasets spanning 23 clinical applications, and its breast subset aggregates BrEaST, BUS (Dataset B), BUS_UC, BUS_UCML, and S1 (Megahed et al., 27 Oct 2025). Using this corpus, USF-MAE pretrains a masked autoencoder on ultrasound alone and then fine-tunes on BUS-BRA, where it reports precision 81.3%, recall 82.0%, and F1 81.6% (Megahed et al., 27 Oct 2025). This does not officially define OpenBreastUS, but it demonstrates a concrete route for assembling an openly accessible breast-ultrasound subset from existing public sources.

BUSGen represents another adjacent development, but with a different emphasis: synthetic data generation rather than direct open release of clinical images (Yu et al., 12 Jan 2025). BUSGen is a conditional diffusion model pretrained on over 3.5 million breast ultrasound images and 3,749 lesions. It conditions on lesion bounding boxes, pathology labels, and device type, and uses the diffusion objective

$8{,}000 \times 256 \times 8 = 16{,}384{,}000$8

In diagnosis on the BUSI test set, a BUS-DM trained on 1 million generated images achieved AUC 0.929, compared with 0.927 for NYU-AI trained on 288,767 real samples and 0.876 for Baseline-CLIP (Yu et al., 12 Jan 2025). Because the real BUS-3.5M and downstream clinical datasets cannot be publicly released, BUSGen proposes synthetic, fully de-identified sharing as an alternative open-data pathway.

By contrast, the USCT stochastic phantom resource and the open-UST hardware platform are breast-ultrasound resources but are not OpenBreastUS in the formal naming sense. The phantom work provides 2D and 3D acoustic numerical breast phantoms and simulated USCT data, including 52 sets of simulated measurement data and open-source generation code, but explicitly rejects the OpenBreastUS label as a misnomer (Li et al., 2021). The open-UST project contributes a 256-element ring-array system manufactured for approximately £2k, with measured SNR 61.2 dB, $8{,}000 \times 256 \times 8 = 16{,}384{,}000$9 dB bandwidth 54%, lateral opening angle 55.4°, and elevational beamwidth 16.3 mm, making it relevant to breast UST research rather than to breast B-mode lesion CAD (Roberts et al., 2023).

Taken together, these works show that “OpenBreastUS” names a specific USCT wave-imaging benchmark in one paper, but also points more loosely to an ecosystem of open breast-ultrasound resources: biopsy-verified lesion tables and interpretable risk tools, preprocessing and OCR pipelines, public breast-image subsets within larger ultrasound corpora, synthetic-data generators, numerical phantoms, and open hardware.

7. Limitations, ambiguities, and future directions

A central ambiguity is terminological. The 2025 USCT paper uses OpenBreastUS as an official project name (Zeng et al., 20 Jul 2025), whereas the PersonalizedUS paper treats it only as a natural label for a future open release (Fröhlich et al., 2024), and the stochastic-phantom paper explicitly rejects it (Li et al., 2021). This means that any technical discussion should specify whether the subject is wave-imaging USCT, lesion-level clinical risk assessment, or an assembled open breast-ultrasound corpus.

Each instantiation also carries domain-specific limitations. The official OpenBreastUS benchmark is restricted to 2D simulations, models only sound speed, and omits density, attenuation, explicit transducer directivity, calibration effects, and measurement noise (Zeng et al., 20 Jul 2025). PersonalizedUS is trained on suspicious lesions selected for biopsy and therefore reflects a diagnostic rather than screening population; its conformal guarantees rely on exchangeability and adequate calibration-set size per subgroup (Fröhlich et al., 2024). BUSClean requires adaptation to local overlays, fonts, and caliper styles (Bunnell et al., 2024). USF-MAE’s breast transfer results depend on harmonized preprocessing and public dataset assembly choices (Megahed et al., 27 Oct 2025). BUSGen’s privacy analysis is empirical rather than formally guaranteed, and generated samples may require quality filtering (Yu et al., 12 Jan 2025).

Despite those differences, a coherent pattern emerges. OpenBreastUS, understood broadly, denotes the move from isolated breast-ultrasound datasets toward openly shareable, reproducible infrastructures that connect curation, preprocessing, uncertainty quantification, representation learning, synthetic generation, and physical simulation. In the narrowest and most precise sense, it is the 2025 breast-USCT neural-operator benchmark. In the broader research vernacular, it signals an evolving open ecosystem for breast-ultrasound science.