Datasets:

nvidia
/

Open-H-RF

ArXiv:

License:

Dataset card Files Files and versions

xet

Community

Open-H-RF / wpi /README.md

wsimson

tristan-deep

Rename OpenH-RF to Open-H-RF (#93)

e51c7a4 15 days ago

preview code

Raw History Blame Contribute Delete

15.7 kB

---
name: wpi
pretty_name: "eSAF Rotational 3D US Raw Channel Data (Medical FUSION Lab, WPI)"
license: cc-by-4.0
task_categories:
  - image-to-image
tags:
  - ultrasound
  - rf
  - open-h-rf
  - 3d
  - beamforming
  - elevational-saf
language:
  - en
size_categories:
  - n<1K
---
# Rotational 3D Ultrasound Raw Channel Data for Elevational SAF

<p align="center"><img src="assets/esaf_topview_sweep.gif" width="55%" alt="Top-view MIP: the naive reconstruction (magenta) accumulating into a smear arc as the probe sweeps 180 degrees about the rotation axis, against the eSAF label (cyan)"></p>

*Top view down the rotation axis of a point target at 20 mm depth, 6 mm off-axis, [`data/tall_elev_H12__point_z020_r6.hdf5`](https://huggingface.co/datasets/nvidia/Open-H-RF/blob/main/wpi/data/tall_elev_H12__point_z020_r6.hdf5). Magenta is the naive per-angle reconstruction, filling in as the array (white line) sweeps 180° about the rotation axis (small circle); cyan is the paired eSAF label shipped in the file; white is where both agree. eSAF collapses the rotational smear arc back onto the target (arc-FWHM 1.26 → 0.39 mm).*

## Dataset Description

Rotational 3D ultrasound acquisitions of point, pair and off-axis targets, captured with an **elevation-focused 1D linear array rotated 180° about its axial axis** (1° steps). Each acquisition stores the **raw per-element channel RF** of a single normal plane-wave transmit at every rotation angle, i.e. the data before in-plane beamforming, which enables flexible offline beamforming and **elevational Synthetic Aperture Focusing (eSAF)**. Every file also carries the eSAF-reconstructed 3D volume as a paired label.

The release is mostly **simulated (Field II)**: a grid of 10 probe types × 19 targets spanning depths where the elevational beam thickness (the artifact eSAF corrects) varies. It is complemented by 5 **measured phantom** rotational scans with the physical Japan Probe 68-element array (same geometry as the simulation) at 10–45 mm depth. No human or animal subjects.

## Dataset Contributor(s)

- Ryo Murakami (contact)
- Medical FUSION Laboratory, Worcester Polytechnic Institute

## Dataset Creation Date

06/15/2026

## License / Terms of Use

[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/legalcode.en). Retain attribution and identify modifications when reusing the data.

## Intended Usage

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

## Dataset Characterization

- **Data Collection Method:** synthetic and phantom.
  - *Simulated (190):* Field II (MATLAB, `xdc_focused_array` + `calc_scat_multi`), scatterers rotated about the axial axis with the transducer fixed. **10 probe types** vary lateral aperture (`n_el` 32/68/128, pitch 0.1/0.2/0.3 mm), elevation height `H` (4/8/12 mm) and elevation focus `R` (25/45/90 mm or unfocused), encoded in the probe name (`tall_elev_H12` = `H` 12 mm; `baseline_R45_H8` = `R` 45 mm, `H` 8 mm) and in `probe/*`.
  - *Measured (5):* rotational scans of the Japan Probe 68-element array on a wire/point phantom with a Galil-controlled 180° rotation, acquired as 7-angle CPWC, time-tag-synced to the motor angles and reduced to the centre (0°) plane wave per angle to match the simulation schema.
- **Labeling Method:** synthetic ground truth. Target positions are exact, and each file's `description` attribute records probe and target, e.g. `probe=tall_elev_H12 target=point_z020_r6 r0=6.0mm z=20.0mm`. The paired eSAF volume (`custom/saf_bmode`) is the label.
- **Acquisition system:** Japan Probe JP_Linear_68 (baseline): 68 elements, pitch 0.2 mm, element width 0.15 mm, element height 8 mm, elevational lens focus 45 mm; center frequency 10 MHz; sampling 40 MHz; sound speed 1490 m/s; one normal plane-wave transmit per rotation angle; 180° rotation in 1° steps. Parameters match the paper simulation.

## Processing the Dataset

The acquisitions can be processed with the `reconstruct.py` [script](https://github.com/open-h/Open-H-RF/blob/main/datasets/wpi/reconstruct.py) as provided in the [Open-H-RF GitHub repository](https://github.com/open-h/Open-H-RF), together with the `pipeline.yaml` definition in this folder and the [zea library](https://github.com/tue-bmd/zea). The script streams the data from the Hugging Face Hub, beamforms the rotation frame closest to ±90° (where an off-axis target lies in-plane) and plots it beside the per-frame rotation angle from `metadata/probe_pose`. Set `ZEA_FILE` and `FRAME` at the top of the script to pick another acquisition or frame. The default, frame 90 (~−90°) of [`data/experiment__acq_exp_30mm.hdf5`](https://huggingface.co/datasets/nvidia/Open-H-RF/blob/main/wpi/data/experiment__acq_exp_30mm.hdf5), gives:

<p align="center"><img src="assets/bmode.png" width="70%" alt="B-mode frame of the measured phantom at -90 deg probe rotation, with the per-frame rotation-angle trajectory alongside it"></p>

The rotational eSAF itself (in-plane DAS → `recon_3d` → `safrot_backproj`) is the published MATLAB implementation; its output is shipped in every file as `custom/saf_bmode`:

```python
with zea.File("data/baseline_R45_H8__point_z080_r4.hdf5") as f:
    saf = {e.name: e for e in f.custom}  # custom/saf_bmode elements
    volume_db = saf["values"].data  # (1, z, x, y) float32 dB
    coordinates = saf["coordinates"].data  # (z, x, y, 3) float32 m
    print(saf["values"].description)  # axes + eSAF parameters + arc-FWHM
```

## Dataset Format

[zea v0.1.6](https://github.com/tue-bmd/zea)

One zea HDF5 file per acquisition, single track. The raw channel RF and scan parameters are in the standard data/scan groups (`tracks/track_0`); the probe rotation per frame is the **`metadata/probe_pose`** trajectory (`euler_xyz`, radians, rotation in the z component, zero translation; `sampling_frequency = 1.0 Hz` is nominal, one pose per frame).
The paired eSAF label is the zea custom field **`custom/saf_bmode`**: `values` is `(1, z, x, y)` float32 in dB (0 dB = volume max, empty pixels −inf) over a ±2 mm depth window about the target, and `coordinates` holds the per-pixel `[x, y, z]` in metres. It lives in `custom` rather than the data group because it is a single frame, while `raw_data` has ~180. The per-case arc-FWHM before/after is in the `description` of `custom/saf_bmode/values`.

Pre-processing: none for the simulations beyond the forward model; for the measured scans, frame-to-angle synchronisation, per-angle dwell averaging and centre-plane-wave selection. Both remain raw per-element RF (not demodulated or decimated).

## Dataset Quantification

**Current Open-H-RF release:** 195 HDF5 files; 1.87 GB (1,870,462,976 bytes) stored; root `zea_version` **0.1.6**. Sizes include all HDF5 contents and use decimal units (MB = 10^6 bytes, GB = 10^9 bytes, TB = 10^12 bytes), not decoded-array memory or original-source download sizes.

- **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 axis is listed above under Data Collection Method; each file's own `description` attribute gives its exact probe/target/r0/depth. (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 `metadata/probe_pose` — z Euler component, radians — not necessarily uniform).
- **Stored HDF5 size:** 1.87 GB (1,870,462,976 bytes), including the paired `saf_bmode` label volumes.
- **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).

Paths below are inside each `.hdf5`; with `zea.File` use `f.data` / `f.scan` / `f.metadata.probe_pose`, and `f.custom` for the SAF label volume.

| field (HDF5 path)                 | shape                          | dtype   | units | description |
|-----------------------------------|--------------------------------|---------|-------|-------------|
| `tracks/track_0/data/raw_data` (`f.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) |
| `probe/probe_geometry`            | (n_el, 3)                      | float32 | m     | element positions (lateral x, 0, 0) |
| `tracks/track_0/scan/sampling_frequency` | scalar                  | float32 | Hz    | 4.0e7 |
| `tracks/track_0/scan/center_frequency`   | scalar                  | float32 | Hz    | 1.0e7 |
| `tracks/track_0/scan/demodulation_frequency` | scalar              | float32 | Hz    | 1.0e7 (= center frequency; used by the reference pipeline's demodulate op) |
| `tracks/track_0/scan/sound_speed` | scalar                         | float32 | m/s   | 1490 |
| `tracks/track_0/scan/initial_times` | (1,)                         | float32 | s     | t0 (first-sample time) |
| `tracks/track_0/scan/t0_delays`   | (1, n_el)                      | float32 | s     | transmit delays (0; normal plane wave) |
| `tracks/track_0/scan/polar_angles` | (1,)                          | float32 | rad   | transmit steering (0) |
| `metadata/probe_pose/rotation`    | (n_frames, 3)                  | float32 | rad   | probe pose per frame, `euler_xyz`; rotation about the axial (z) axis is the z component |
| `metadata/probe_pose/translation` | (n_frames, 3)                  | float32 | m     | probe tip translation (all zero — pure rotation) |
| `metadata/probe_pose/sampling_frequency` | scalar                  | float32 | Hz    | 1.0 — **nominal** one-pose-per-frame rate, not a physical value |
| `metadata/credit` (`f.metadata.credit`) | scalar                   | str     | –     | dataset credit / attribution (lab, contact, citation, license) |
| `metadata/subject/type` (`f.metadata.subject.type`) | scalar       | str     | –     | `simulated phantom` (Field II sims) or `phantom` (measured `experiment__*` scans) |
| `custom/saf_bmode/values` (custom field, via `f.custom`) | (1, n_z, n_el, n_el) | float32 | dB | **paired label**: elevational-SAF reconstructed 3D B-mode volume, log-compressed normalized envelope (0 dB = max, empty pixels −inf); dims = (frame, z=depth, x=lateral, y=elevation) |
| `custom/saf_bmode/coordinates` | (n_z, n_el, n_el, 3)             | float32 | m     | per-pixel `[x, y, z]` positions of the label volume (target ± ~2 mm depth window) |
## Subject Metadata
No human or animal subjects / no PHI. Each file stores `metadata/subject/type`: `simulated phantom` for the Field II simulations, `phantom` for the measured `experiment__*` scans. Creator attribution is stored per file in `metadata/credit`.
## Data Validation
`reconstruct.py` beamforms a single rotation frame with the `pipeline.yaml` chain (Cast → Demodulate → Beamform(delay_and_sum) → EnvelopeDetect → Normalize → LogCompress) and plots the stored rotation trajectory beside it, as a check on the acquisition parameters and the frame-to-angle mapping.
The eSAF labels were checked per case with the arc-FWHM of the rotational smear before and after eSAF, shown here annotated for the case in the hero image:
<p align="center"><img src="assets/esaf_topview_before_after.png" width="60%" alt="Annotated top-view MIP: mm axes, the rotation axis and its 180 degree sweep, a color legend, and the arc-FWHM numbers"></p>
## 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` label 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`, source) 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`, source).
## 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 or IRB constraints.
## Raw Source Data
The raw pre-conversion outputs are archived under the same CC BY 4.0 license at
<https://huggingface.co/datasets/RyoMurakami/OpenH-RF-eSAF-raw>: the raw Verasonics per-line
channel-RF captures (with 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.
## 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.

	---
	name: wpi
	pretty_name: "eSAF Rotational 3D US Raw Channel Data (Medical FUSION Lab, WPI)"
	license: cc-by-4.0
	task_categories:
	- image-to-image
	tags:
	- ultrasound
	- rf
	- open-h-rf
	- 3d
	- beamforming
	- elevational-saf
	language:
	- en
	size_categories:
	- n<1K
	---

	# Rotational 3D Ultrasound Raw Channel Data for Elevational SAF

	<p align="center"><img src="assets/esaf_topview_sweep.gif" width="55%" alt="Top-view MIP: the naive reconstruction (magenta) accumulating into a smear arc as the probe sweeps 180 degrees about the rotation axis, against the eSAF label (cyan)"></p>

	Top view down the rotation axis of a point target at 20 mm depth, 6 mm off-axis, [`data/tall_elev_H12__point_z020_r6.hdf5`](https://huggingface.co/datasets/nvidia/Open-H-RF/blob/main/wpi/data/tall_elev_H12__point_z020_r6.hdf5). Magenta is the naive per-angle reconstruction, filling in as the array (white line) sweeps 180° about the rotation axis (small circle); cyan is the paired eSAF label shipped in the file; white is where both agree. eSAF collapses the rotational smear arc back onto the target (arc-FWHM 1.26 → 0.39 mm).

	## Dataset Description

	Rotational 3D ultrasound acquisitions of point, pair and off-axis targets, captured with an elevation-focused 1D linear array rotated 180° about its axial axis (1° steps). Each acquisition stores the raw per-element channel RF of a single normal plane-wave transmit at every rotation angle, i.e. the data before in-plane beamforming, which enables flexible offline beamforming and elevational Synthetic Aperture Focusing (eSAF). Every file also carries the eSAF-reconstructed 3D volume as a paired label.

	The release is mostly simulated (Field II): a grid of 10 probe types × 19 targets spanning depths where the elevational beam thickness (the artifact eSAF corrects) varies. It is complemented by 5 measured phantom rotational scans with the physical Japan Probe 68-element array (same geometry as the simulation) at 10–45 mm depth. No human or animal subjects.

	## Dataset Contributor(s)

	- Ryo Murakami (contact)
	- Medical FUSION Laboratory, Worcester Polytechnic Institute

	## Dataset Creation Date

	06/15/2026

	## License / Terms of Use

	[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/legalcode.en). Retain attribution and identify modifications when reusing the data.

	## Intended Usage

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

	## Dataset Characterization

	- Data Collection Method: synthetic and phantom.
	- Simulated (190): Field II (MATLAB, `xdc_focused_array` + `calc_scat_multi`), scatterers rotated about the axial axis with the transducer fixed. 10 probe types vary lateral aperture (`n_el` 32/68/128, pitch 0.1/0.2/0.3 mm), elevation height `H` (4/8/12 mm) and elevation focus `R` (25/45/90 mm or unfocused), encoded in the probe name (`tall_elev_H12` = `H` 12 mm; `baseline_R45_H8` = `R` 45 mm, `H` 8 mm) and in `probe/*`.
	- Measured (5): rotational scans of the Japan Probe 68-element array on a wire/point phantom with a Galil-controlled 180° rotation, acquired as 7-angle CPWC, time-tag-synced to the motor angles and reduced to the centre (0°) plane wave per angle to match the simulation schema.
	- Labeling Method: synthetic ground truth. Target positions are exact, and each file's `description` attribute records probe and target, e.g. `probe=tall_elev_H12 target=point_z020_r6 r0=6.0mm z=20.0mm`. The paired eSAF volume (`custom/saf_bmode`) is the label.
	- Acquisition system: Japan Probe JP_Linear_68 (baseline): 68 elements, pitch 0.2 mm, element width 0.15 mm, element height 8 mm, elevational lens focus 45 mm; center frequency 10 MHz; sampling 40 MHz; sound speed 1490 m/s; one normal plane-wave transmit per rotation angle; 180° rotation in 1° steps. Parameters match the paper simulation.

	## Processing the Dataset

	The acquisitions can be processed with the `reconstruct.py` [script](https://github.com/open-h/Open-H-RF/blob/main/datasets/wpi/reconstruct.py) as provided in the [Open-H-RF GitHub repository](https://github.com/open-h/Open-H-RF), together with the `pipeline.yaml` definition in this folder and the [zea library](https://github.com/tue-bmd/zea). The script streams the data from the Hugging Face Hub, beamforms the rotation frame closest to ±90° (where an off-axis target lies in-plane) and plots it beside the per-frame rotation angle from `metadata/probe_pose`. Set `ZEA_FILE` and `FRAME` at the top of the script to pick another acquisition or frame. The default, frame 90 (~−90°) of [`data/experiment__acq_exp_30mm.hdf5`](https://huggingface.co/datasets/nvidia/Open-H-RF/blob/main/wpi/data/experiment__acq_exp_30mm.hdf5), gives:

	<p align="center"><img src="assets/bmode.png" width="70%" alt="B-mode frame of the measured phantom at -90 deg probe rotation, with the per-frame rotation-angle trajectory alongside it"></p>

	The rotational eSAF itself (in-plane DAS → `recon_3d` → `safrot_backproj`) is the published MATLAB implementation; its output is shipped in every file as `custom/saf_bmode`:

	```python
	with zea.File("data/baseline_R45_H8__point_z080_r4.hdf5") as f:
	saf = {e.name: e for e in f.custom} # custom/saf_bmode elements
	volume_db = saf["values"].data # (1, z, x, y) float32 dB
	coordinates = saf["coordinates"].data # (z, x, y, 3) float32 m
	print(saf["values"].description) # axes + eSAF parameters + arc-FWHM
	```

	## Dataset Format

	[zea v0.1.6](https://github.com/tue-bmd/zea)

	One zea HDF5 file per acquisition, single track. The raw channel RF and scan parameters are in the standard data/scan groups (`tracks/track_0`); the probe rotation per frame is the `metadata/probe_pose` trajectory (`euler_xyz`, radians, rotation in the z component, zero translation; `sampling_frequency = 1.0 Hz` is nominal, one pose per frame).

	The paired eSAF label is the zea custom field `custom/saf_bmode`: `values` is `(1, z, x, y)` float32 in dB (0 dB = volume max, empty pixels −inf) over a ±2 mm depth window about the target, and `coordinates` holds the per-pixel `[x, y, z]` in metres. It lives in `custom` rather than the data group because it is a single frame, while `raw_data` has ~180. The per-case arc-FWHM before/after is in the `description` of `custom/saf_bmode/values`.

	Pre-processing: none for the simulations beyond the forward model; for the measured scans, frame-to-angle synchronisation, per-angle dwell averaging and centre-plane-wave selection. Both remain raw per-element RF (not demodulated or decimated).

	## Dataset Quantification

	Current Open-H-RF release: 195 HDF5 files; 1.87 GB (1,870,462,976 bytes) stored; root `zea_version` 0.1.6. Sizes include all HDF5 contents and use decimal units (MB = 10^6 bytes, GB = 10^9 bytes, TB = 10^12 bytes), not decoded-array memory or original-source download sizes.

	- 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 axis is listed above under Data Collection Method; each file's own `description` attribute gives its exact probe/target/r0/depth. (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 `metadata/probe_pose` — z Euler component, radians — not necessarily uniform).
	- Stored HDF5 size: 1.87 GB (1,870,462,976 bytes), including the paired `saf_bmode` label volumes.
	- 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).

	Paths below are inside each `.hdf5`; with `zea.File` use `f.data` / `f.scan` / `f.metadata.probe_pose`, and `f.custom` for the SAF label volume.

	\| field (HDF5 path) \| shape \| dtype \| units \| description \|
	\|-----------------------------------\|--------------------------------\|---------\|-------\|-------------\|
	\| `tracks/track_0/data/raw_data` (`f.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) \|
	\| `probe/probe_geometry` \| (n_el, 3) \| float32 \| m \| element positions (lateral x, 0, 0) \|
	\| `tracks/track_0/scan/sampling_frequency` \| scalar \| float32 \| Hz \| 4.0e7 \|
	\| `tracks/track_0/scan/center_frequency` \| scalar \| float32 \| Hz \| 1.0e7 \|
	\| `tracks/track_0/scan/demodulation_frequency` \| scalar \| float32 \| Hz \| 1.0e7 (= center frequency; used by the reference pipeline's demodulate op) \|
	\| `tracks/track_0/scan/sound_speed` \| scalar \| float32 \| m/s \| 1490 \|
	\| `tracks/track_0/scan/initial_times` \| (1,) \| float32 \| s \| t0 (first-sample time) \|
	\| `tracks/track_0/scan/t0_delays` \| (1, n_el) \| float32 \| s \| transmit delays (0; normal plane wave) \|
	\| `tracks/track_0/scan/polar_angles` \| (1,) \| float32 \| rad \| transmit steering (0) \|
	\| `metadata/probe_pose/rotation` \| (n_frames, 3) \| float32 \| rad \| probe pose per frame, `euler_xyz`; rotation about the axial (z) axis is the z component \|
	\| `metadata/probe_pose/translation` \| (n_frames, 3) \| float32 \| m \| probe tip translation (all zero — pure rotation) \|
	\| `metadata/probe_pose/sampling_frequency` \| scalar \| float32 \| Hz \| 1.0 — nominal one-pose-per-frame rate, not a physical value \|
	\| `metadata/credit` (`f.metadata.credit`) \| scalar \| str \| – \| dataset credit / attribution (lab, contact, citation, license) \|
	\| `metadata/subject/type` (`f.metadata.subject.type`) \| scalar \| str \| – \| `simulated phantom` (Field II sims) or `phantom` (measured `experiment__*` scans) \|
	\| `custom/saf_bmode/values` (custom field, via `f.custom`) \| (1, n_z, n_el, n_el) \| float32 \| dB \| paired label: elevational-SAF reconstructed 3D B-mode volume, log-compressed normalized envelope (0 dB = max, empty pixels −inf); dims = (frame, z=depth, x=lateral, y=elevation) \|
	\| `custom/saf_bmode/coordinates` \| (n_z, n_el, n_el, 3) \| float32 \| m \| per-pixel `[x, y, z]` positions of the label volume (target ± ~2 mm depth window) \|

	## Subject Metadata

	No human or animal subjects / no PHI. Each file stores `metadata/subject/type`: `simulated phantom` for the Field II simulations, `phantom` for the measured `experiment__*` scans. Creator attribution is stored per file in `metadata/credit`.

	## Data Validation

	`reconstruct.py` beamforms a single rotation frame with the `pipeline.yaml` chain (Cast → Demodulate → Beamform(delay_and_sum) → EnvelopeDetect → Normalize → LogCompress) and plots the stored rotation trajectory beside it, as a check on the acquisition parameters and the frame-to-angle mapping.

	The eSAF labels were checked per case with the arc-FWHM of the rotational smear before and after eSAF, shown here annotated for the case in the hero image:

	<p align="center"><img src="assets/esaf_topview_before_after.png" width="60%" alt="Annotated top-view MIP: mm axes, the rotation axis and its 180 degree sweep, a color legend, and the arc-FWHM numbers"></p>

	## 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` label 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`, source) 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`, source).

	## 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 or IRB constraints.

	## Raw Source Data

	The raw pre-conversion outputs are archived under the same CC BY 4.0 license at
	<https://huggingface.co/datasets/RyoMurakami/OpenH-RF-eSAF-raw>: the raw Verasonics per-line
	channel-RF captures (with 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.

	## 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.