RyoMurakami/OpenH-RF-eSAF

OpenH-RF Sub-Dataset — Rotational 3D US Raw Channel Data for Elevational SAF (Simulated + Measured Phantom)

One .hdf5 per acquisition, zea file format. Raw per-element channel data (pre-beamforming). See data/manifest.json for the full list and reconstruct.py

• pipeline.yaml for the reference reconstruction.

Paths in this card are relative to the package root (this file ships as the package README.md; the .hdf5 live in data/). Files marked (repo) belong to the full project repository at https://github.com/RyoMurakami-MedRobo/OpenH-RF-eSAF and are not bundled in this portable package — they require MATLAB / Field II / the acquisition drives.

Dataset Description

Synthetic rotational 3D ultrasound acquisitions of point, pair, and off-axis targets, captured with an elevation-focused 1D linear array that is rotated 180° about its axial axis (1° steps, 180 frames). Each acquisition stores the raw per-element channel RF for a single normal plane-wave transmit at every rotation angle — i.e. the data before in-plane beamforming — which is what enables flexible offline beamforming and the elevational Synthetic Aperture Focusing (eSAF) method. The targets span a wide depth range to capture the depth-dependent elevational beam thickness (the artifact eSAF corrects). The release is mostly simulated (Field II), complemented by a small set of real measured phantom rotational scans acquired with the physical Japan Probe 68-element array (same geometry as the simulation) over a shallow-to-focal depth series (10–45 mm). Simulated + measured phantom data (no in-vivo / subjects).

Dataset Contributor(s)

Medical FUSION Laboratory, Worcester Polytechnic Institute. Contact: Ryo Murakami.

Dataset Creation Date

06/15/2026.

License / Terms of Use

CC BY 4.0 (full text in LICENSE). The release contains simulated (Field II) and real measured phantom rotational scans — inanimate phantom only, so there are no IP, subject-consent, or IRB constraints.

Citation. When using this dataset, please cite:

R. Murakami et al., "Elevational Synthetic Aperture Focusing for Rotated Array-Based Three-Dimensional Ultrasound Imaging," IEEE Access, 2025.

Intended Usage

Advanced beamforming and elevational resolution recovery for rotational 3D US (eSAF), elevation-PSF / aperture-growth studies, and as a reproducible raw-channel-data benchmark for rotational synthetic-aperture reconstruction.

Dataset Characterization

• Data Collection Method: synthetic, generated with Field II (Jensen) run in MATLAB. The main release is a probe × target grid produced by sim/batch_sim_probe_target.m (repo) (with sim/sim_probe_catalog.m / sim/sim_target_catalog.m, repo): for each (probe type, target) it uses xdc_focused_array
  • calc_scat_multi to produce the raw per-element channel RF at every rotation angle (scatterer rotated about the axial axis, transducer fixed), then sim/sim_dataset_to_zea.py (repo) repackages every case into the zea format here. 10 probe types span lateral aperture (n_el 32/68/128, pitch 0.1/0.2/0.3 mm), elevation height H (4/8/12 mm), and elevation focal depth R (25/45/90 mm + unfocused) — see the probe table in data/manifest.json. (The earlier 18-acquisition set generated by sim/batch_generate_fieldii.m + sim/mat_to_zea.py (repo), and a pure-Python analytic forward model sim/forward_sim.py (bundled), remain available as compatible alternatives with the identical schema.) Measured phantom acquisitions (5): real rotational scans of the physical Japan Probe 68-element array on a wire/point phantom, acquired with CPWC channel-RF capture (experiment/Ryo_SetUp_JP68_PWCompound_3D_ChannelRF.m, repo) and a Galil-controlled 180° rotation. Each scan is time-tag-synced (frames → motor angles) and reduced to the single centre (normal) plane wave per angle by experiment/sync_channel_rf.m (repo) (so the schema matches the simulation, n_tx = 1), then converted with the same sim/sim_dataset_to_zea.py (repo). They span a shallow-to-focal depth series (10, 20, 30, 40, 45 mm nominal target depth).
• Labeling Method: synthetic ground truth (exact target positions known; in data/manifest.json).
• Acquisition system (simulated): Japan Probe JP_Linear_68 — 68-element linear array, pitch 0.2 mm, element width 0.15 mm, element height 8 mm, elevational lens focus 45 mm (Field II xdc_focused_array with 500 elevation math sub-elements); center frequency 10 MHz; sampling 40 MHz (NS200BW, 4 samples/wavelength); speed of sound 1490 m/s; single normal plane-wave transmit per rotation angle; 180° rotation, 1° step (180 frames). Parameters match the paper simulation (makeslices_field2.m).

Dataset Format

zea file format (HDF5), one file per acquisition. Pre-processing — simulated: none beyond the forward model (raw RF, not demodulated/decimated); measured: time-tag frame→angle synchronisation, per-angle dwell averaging, and centre-plane-wave selection (still raw per-element RF, not demodulated/decimated). Rotation angle per frame is stored as a zea custom element at custom/scan/rotation_angles_deg. Every file is written with zea.File.create() (sim/sim_dataset_to_zea.py, repo) and carries a zea_version stamp, so zea loads it natively (not as a legacy file).

Paired pre-/post-SAF labels (the dataset's target output). Each file also carries the elevational-SAF reconstructed 3D B-mode volume at custom/labels/saf_bmode_volume (with custom/labels/saf_{x,y,z}_mm axes). This is the post-SAF output/label paired with the pre-beamformed input (data/raw_data): the raw channel RF is back-projected through the published eSAF algorithm (matlab/saf/safrot_backproj.m, repo: in-plane DAS → recon_3d → safrot_backproj, elevational focus 45 mm, f-number 45/8) into a 3D volume B_SAF(x,y,z), generated by sim/make_saf_all.m → experiment/run_esaf_synced.m (repo) and written into the zea file by sim/pack_saf_labels.py (repo). The stored volume is the normalized envelope over a thin depth window (±2 mm) about the target; per-case arc-FWHM before/after and gain are in data/manifest.json and on the dataset (arc_fwhm_before_after_mm attribute).

A MATLAB .mat version of the same raw channel data + metadata, plus a reference eSAF-beamformed result and a _ref.png figure, is provided per acquisition alongside the source grid as sim_dataset_out/<probe>/<target>.mat and ..._ref.png (each .mat holds the raw RF, the in-plane DAS, the metadata and the eSAF output produced with the published algorithm matlab/saf/safrot_backproj.m (repo): in-plane DAS → recon_3d → safrot_backproj, f-number 45/8). A FWHM-vs-depth overview across probes is sim_dataset_out/dataset_overview_r4.png (sim/dataset_overview.m, repo). The zea .hdf5 acquisitions are hosted on Hugging Face at https://huggingface.co/datasets/RyoMurakami/OpenH-RF-eSAF (git-LFS; not committed to the GitHub code repo). The MATLAB .mat/_ref.png intermediates are reproducible from source and kept on lab storage.

Dataset Quantification

• Acquisitions: 195 = 190 simulated + 5 measured phantom.
  • Simulated (190): 10 probe types × 19 targets (16 single points over depth {20,45,80,130} mm × radial offset from the rotation centre {0,2,4,6} mm, plus 3 pair/oblique cases). The probe and target axes are listed in data/manifest.json. (The earlier compatible set has 18 acquisitions.)
  • Measured (5): real rotational phantom scans at nominal depths {10,20,30,40,45} mm (experiment__acq_exp_*.hdf5), centre plane wave, ~182 measured rotation angles over ~180°.
• Frames per acquisition: simulated 180 (one per 1° step); measured ~182 (the actual encoder angles are stored in custom/scan/rotation_angles_deg, not necessarily uniform).
• Total size on disk: simulated 0.6–5 MB per case (zea gzip; point-target RF is sparse), measured ~80–92 MB per case (dense tissue RF); **1.5 GB** for the full set (including the paired saf_bmode_volume labels).
• Train/val/test split: N/A (benchmark / characterization set; the probe × depth × radius axes are the intended study dimensions).

Per-sample feature table

Shapes use placeholders because dimensions vary across the probe grid and between simulated and measured scans: n_frames = 180 (simulated, one per 1° step) or ~182 (measured encoder angles); n_el ∈ {32, 68, 128} (probe grid; 68 for the baseline and all measured scans); n_ax = axial sample count (per case); n_z = depth samples of the label volume (target ± ~2 mm window).

field (HDF5 path)	shape	dtype	units	description
data/raw_data	(n_frames, 1, n_ax, n_el, 1)	float32	a.u.	raw per-element channel RF; dims = (frame=rotation, tx, axial, element, ch)
scan/probe_geometry	(n_el, 3)	float32	m	element positions (lateral x, 0, 0)
scan/sampling_frequency	scalar	float32	Hz	4.0e7
scan/center_frequency	scalar	float32	Hz	1.0e7
scan/sound_speed	scalar	float32	m/s	1490
scan/initial_times	(1,)	float32	s	t0 (first-sample time)
scan/t0_delays	(1, n_el)	float32	s	transmit delays (0; normal plane wave)
scan/polar_angles	(1,)	float32	rad	transmit steering (0)
custom/scan/rotation_angles_deg	(n_frames,)	float32	deg	probe rotation angle per frame (zea custom element)
custom/labels/saf_bmode_volume	(n_el, n_el, n_z)	float32	a.u.	paired label: elevational-SAF reconstructed 3D B-mode volume B_SAF(x,y,z) (normalized envelope)
custom/labels/saf_x_mm / saf_y_mm	(n_el,)	float32	mm	lateral / elevation axes of saf_bmode_volume
custom/labels/saf_z_mm	(n_z,)	float32	mm	depth axis of saf_bmode_volume (target ± ~2 mm)

Subject Metadata

N/A (synthetic phantom; no subjects/PHI).

Data Validation

reconstruct.py (runnable, verified — official zea API, no fallback code) loads one zea acquisition, reads its acquisition parameters via zea.Config.from_path('pipeline.yaml') + File.load_parameters, beamforms the rotation frame closest to ±90° rotation magnitude (the frame where an off-axis target lies in-plane; this handles signed encoder angles too — measured scans run 0 → ~−180°) with the native zea.Pipeline op chain Cast → Demodulate → Beamform(delay_and_sum) → EnvelopeDetect → Normalize → LogCompress defined in pipeline.yaml, and writes a two-panel PNG: the B-mode image, and the per-frame probe rotation angle (custom/scan/rotation_angles_deg) so downstream users know how to interpret the frame axis — the special data this dataset adds, per reviewer feedback:

python reconstruct.py --input data/<probe>__<target>.hdf5 --output out.png

The rotational eSAF across frames — the contribution of this dataset — is implemented in matlab/saf (repo) (recon_3d → safrot_backproj); per-probe before/after eSAF reference images and a FWHM-vs-depth overview accompany the MATLAB .mat release (sim/dataset_overview.m, repo), and the resulting paired custom/labels/saf_bmode_volume is stored in every .hdf5 (see Dataset Format above).

Known Issues

• Paired SAF label — on-axis targets (r0 = 0) do not narrow, by design. eSAF refocuses the rotational elevation smear; a target sitting on the rotation axis has essentially no smear, so its saf_bmode_volume is not sharper than the input (arc-FWHM gain ≈ 1). This is expected physics, not a defect — the 40 on-axis cases (median gain 1.00×) are included so the pair covers the degenerate no-smear case. Off-axis targets (n=120, median gain 1.75×, up to 12×) and paired/oblique targets (n=30, median 3.71×) improve clearly; targets at the focal depth (45 mm) and weak-elevation-focus probes (efocus_deep_90, elev_unfocused) have less smear to recover. Across all 195 cases, median arc-FWHM gain is 1.36× (42 cases < 1×, mostly the on-axis/near-focus group above). Arc-FWHM is measured on a centred reconstruction: the smear circle passes through both the rotation axis and the target (not a circle centred on the rotation axis). The eSAF back-projection uses a fixed elevational focus of 45 mm; per-depth focus tuning (see docs/eSAF_focus_depth_study_JP.md, repo) can further sharpen deep off-axis cases but was not applied here (single as-designed focus).
• Measured phantom depth window. The real reflector bead sits ~4 mm off the rotation axis (not on-axis) and, for each scan, slightly deeper than the folder's nominal depth label; labels are reconstructed over the interactively-identified reflector depth window (not a naive nominal-depth ± 2 mm window), which matters because a mis-centred window can pick up near-axis clutter instead of the actual bead.
• Simulated data (Field II spatial-impulse-response model): realistic transducer field, but no tissue attenuation, aberration, multiple scattering, or electronic noise. Not a substitute for measured data.
• Speed of sound is 1490 m/s, matching the paper Table 1 and the experiment.
• A single normal plane-wave transmit per rotation angle is simulated (the dataset stores n_tx = 1); multi-angle compounding is left to downstream users.
• Measured scans: acquired as 7-angle CPWC; only the centre (0°) plane wave is kept here to match the n_tx = 1 schema. The dwell frames per angle are averaged before storage (noise reduction). Real reflectors are not ideal point scatterers — expect reverberation/clutter near the surface and specular layering; rotation angles are the measured encoder values (slightly non-uniform, full span ≈ 180°, sign per encoder direction). The elevational lens focus is the nominal 45 mm, but the effective back-projection focus for eSAF is depth-dependent on real data (see docs/eSAF_focus_depth_study_JP.md, repo).

Raw Source Data

The raw, pre-conversion acquisition/simulation outputs that were processed into the zea .hdf5 files above are archived (same CC BY 4.0 license) at https://huggingface.co/datasets/RyoMurakami/OpenH-RF-eSAF-raw: the raw Verasonics per-line channel-RF captures (RFdata_line*.mat + encoder logs) for the 5 measured acquisitions, and the per-case MATLAB intermediates (raw RF, in-plane DAS, eSAF output) for the 190 simulated cases. See that repository's README for how each maps to data/*.hdf5 here.

Ethical Considerations

None. The data is either fully synthetic (Field II) or measured on an inanimate phantom — no human or animal subjects, no PHI, no consent/IRB constraints.