ubc/module_C/README.md

metadata

pretty_name: OpenH-RF — UBC Module C Fetal Phantom
license: cc-by-4.0
task_categories:
  - image-to-image
tags:
  - ultrasound
  - rf
  - b-mode
  - beamforming
  - fetal-imaging
  - phantom
  - synthetic
  - multimodal
  - timeseries
  - video
  - openh-rf
language:
  - en
size_categories:
  - 1K<n<10K

UBC Module C

Dataset Description

This dataset contains 3,845 two-dimensional ultrasound frames from two sessions with the Kyoto Kagaku Fetus Ultrasound Examination Phantom “SPACE FAN-ST.” Each frame contains synthetic RF channel data, preserved beamformed IQ, a source-derived B-mode reference, and relative-time links to NDI tracking, BK scanner viewport video, and Intel RealSense video.

The RF channels were simulated from already beamformed and compounded BK3500 IQ using a custom scatterer model and assumed transmit/receive geometry. They are not measured pre-beamforming channels. Modalities have temporal synchronization only; no validated spatial registration is supplied.

Dataset Contributors

• Organization: The University of British Columbia (UBC), Department of Electrical and Computer Engineering and School of Biomedical Engineering.
• Dataset-preparation contact: Zongze Li, zongze@student.ubc.ca.
• Contributors: Septimiu E. Salcudean, Robert Rohling, Qi Zeng, Tajwar Abrar Aleef, Hamid Moradi, Mohammad Honarvar, Wanwen Chen, Zijian Wu, Yuxin Chen, Yu Chung Lee, Zongze Li, Patrick Boyan Chen, and Michael Frew.

Dataset Creation Date

• Physical phantom acquisition: 06/12/2026.
• Release preparation: 09/08/2026.

License / Terms of Use

The dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0) (https://creativecommons.org/licenses/by/4.0/). The HDF5 metadata/credit field carries the following attribution with the contributor names and contact listed above:

University of British Columbia OpenH-RF contributors (2026). UBC Module C: Synthetic Channel Proxy and Multimodal Fetal Phantom Ultrasound. OpenH-RF. CC BY 4.0.

Intended Usage

Research uses include synthetic RF-to-B-mode reconstruction, denoising, super-resolution, fetal-phantom scan-plane assessment, and temporally aligned multimodal learning. The dataset is for research and technical evaluation, not clinical diagnosis. Source-derived reference images are not independent physical ground truth for the simulated channels.

Dataset Characterization

• Collection: Phantom ultrasound and simultaneous room-camera and tracking recordings, followed by deterministic synthetic-channel generation.
• Acquisition system: BK3500 ultrasound; NDI Tool-2 tracked-marker poses; NDI Tool-1 reference-tool poses; Intel RealSense RGB, left/right infrared, and encoded depth video.
• Labels: Source indices, synchronization validity flags, and simulation provenance. No manual fetal-structure segmentations or verified phantom configuration labels are supplied.
• Simulation framework: Custom NumPy/SciPy scatterer forward model using four-point fractional delays, pulse filtering, frame-specific noise, and int16 quantization. Session 1 preserves its original amplitudes and shared random draws, with scatterer depths evaluated on each frame's source IQ axial grid. Session 2 uses fixed deterministic positions and reflectivity draws for the session, with frame amplitudes derived from its IQ envelope. No two-channel proxy data are included as equivalent RF measurements.

Ultrasound and synthetic-channel parameters

Parameter	Value	Status
Frames	3,845	Two sessions; one per matched BK/RealSense frame
Focused transmit events per frame	248	Stored in each zea file
Receive elements	192	Synthetic assumption
Axial RF samples per trace	3,153	Includes 300 zero-tail samples
RF components	1 real component	Stored as int16
Sampling frequency	15 MHz	Stored zea scan parameter
Center/demodulation frequency	3.75 MHz	Stored zea scan parameter
Assumed sound speed	1,540 m/s	Stored zea scan parameter
Assumed element pitch	0.2 mm	Synthetic geometry assumption
Assumed nominal aperture	38.4 mm	192 elements × 0.2 mm
Element width / height	0.18 mm / 5.0 mm	Synthetic model
Assumed transmit focus	98 mm	Copied from source scan metadata
RF trace duration	210.2 µs	Stored raw-data attribute
Zero-amplitude tail	20 µs / 300 samples	Stored raw-data attributes
Target maximum magnitude	26,000 counts	Below int16 full scale

The element geometry and dimensions are simulation assumptions, not measured probe specifications. The relationship between the recorded 98-mm focus and the proprietary scanner transmit law is unknown.

Dataset Format

Each zea HDF5 file contains one frame and is written with zea 0.1.5. The writer-version field describes serialization, not processing speed. Source IQ and its display reference are under /custom/source_provenance.

Package layout

Path	Contents
acquisitions/session_01/session_01_fNNNN.hdf5	Selected source frames 0055–2586; gaps are intentional
acquisitions/session_02/session_02_fNNNN.hdf5	Selected source frames 0001–5017; gaps are intentional
source_bmode/session_XX/	BK viewport MP4 video
external_camera/session_XX/	RealSense MP4 video
pipeline.yaml	Shared reconstruction configuration for both sessions

session_XX denotes session_01 or session_02. All public identifiers are neutral session labels. HDF5 source_file values identify private source container aliases, not downloadable package files; the source IQ is embedded in each sample. This HF release includes HDF5 files, this card, the pipeline configuration, and the source BK viewport and RealSense MP4 videos. Composite review videos, standalone scripts, timing/tracking tables, calibration sidecars, and validation reports are not included. References inside unchanged HDF5 metadata may identify sidecars from the source package that are not available in this release.

Timestamp conventions

All acquisition timestamps are session-relative. Absolute wall-clock times are intentionally withheld for de-identification. IQ, BK viewport, NDI reference-tool, and RealSense host times are relative to the first source IQ frame of their own session, even when that frame is excluded from the release. RealSense device time is relative to its first camera frame. Negative host-relative camera times can precede the first IQ frame. Each one-frame zea probe-pose timestamp is locally zero, with session-relative IQ time stored separately in /custom/synchronization.

Use alignment indices and validity flags, not video frame rate alone, to link modalities. The HDF5 Tool-2 previous/next/recorded-last indices address the status-zero subset of the source tracking records. Tool-1 reference indices address the full source reference-tool records. These standalone tables are not included in this HF release. MP4 playback time is an encoded presentation timeline, not an acquisition clock; use the per-frame synchronization metadata retained in HDF5.

Dataset Quantification

Quantity	session_01	session_02	Total
Retained ultrasound frames / HDF5 files	1,339	2,506	3,845
Original ultrasound frames	2,613	5,019	7,632
Excluded unmatched or repeated video associations	1,274	2,513	3,787
Retained source frame range (with gaps)	55–2586	1–5017	Original frame numbers preserved
Retained IQ time span (s)	1.063–52.681	0.018–93.757	Separate source-session origins
BK viewport video frames	1,582	2,798	4,380
Frames per RealSense stream	1,591	3,677	5,268
Valid BK / NDI Tool-2 / NDI Tool-1 / RealSense links	1,339 each	2,506 each	3,845 each

Frame selection: Every retained ultrasound frame has a valid BK video link, RealSense video link, interpolated NDI Tool-2 pose, and NDI Tool-1 reference link. Each BK frame and each RealSense frame is associated with at most one retained ultrasound frame. Selection maximizes the number of chronological matches, then minimizes the sum of absolute BK/RealSense timing offsets; source frame order breaks remaining ties. Original frame indices, relative times, and noise seeds are preserved. HDF5 files contain only retained ultrasound frames; videos retain their full temporal context.

HDF5 storage: session_01: 205,236,518,547 bytes (191.14 GiB); session_02: 388,206,776,234 bytes (361.55 GiB). Total: 593,443,294,781 bytes (552.69 GiB). Files use lossless compression.

Splits: None supplied. Neighboring frames and the two sessions share a phantom and acquisition setting. Split by session for temporal separation; this does not establish generalization to different phantoms or patients.

Per-sample feature table

Name	Shape	Dtype	Units	Description
/tracks/track_0/data/raw_data	(1, 248, 3153, 192, 1)	int16	scaled ADC-like counts	Synthetic real-RF channel tensor ordered as frame, transmit, axial sample, receive element, RF component
/tracks/track_0/data/image/values	(1, 704, 248, 1)	uint8	display gray level	Stored source-derived reference image; not used as a reconstruction input
/tracks/track_0/data/image/coordinates	(704, 248, 3)	float32	m	Cartesian coordinate for each stored reference-image pixel
/probe/probe_geometry	(192, 3)	float32	m	Assumed synthetic receive-element positions
/tracks/track_0/scan/tx_apodizations	(248, 192)	float32	unitless	Assumed transmit apodization per event and element
/tracks/track_0/scan/t0_delays	(248, 192)	float32	s	Assumed transmit delay per event and element
/tracks/track_0/scan/transmit_origins	(248, 3)	float32	m	Transmit-beam origins
/tracks/track_0/scan/focus_distances	(248,)	float32	m	Focus distance for each transmit event
/tracks/track_0/scan/azimuth_angles	(248,)	float32	rad	Azimuth angle for each scanline
/tracks/track_0/scan/initial_times	(248,)	float32	s	ADC start time for each transmit event
/tracks/track_0/scan/tgc_gain_curve	(3153,)	float32	unitless	Stored time-gain-compensation curve
/metadata/probe_pose/translation	(1, 3)	float32	m	Interpolated NDI Tool-2 marker translation; not a calibrated probe-tip pose
/metadata/probe_pose/rotation	(1, 4)	float32	unitless	NDI Tool-2 marker quaternion in xyzw order
/metadata/probe_pose/timestamps	(1,)	float32	s	Locally rebased one-frame pose time; session-relative IQ time is stored separately
/custom/source_provenance/source_beamformed_iq	(N_IQ, 248, 2)	float32	arbitrary	Preserved source beamformed and compounded IQ; axial sample count varies by frame
/custom/source_provenance/source_bmode_normalized	(704, 248)	float32	normalized intensity	Source IQ envelope displayed over −50 to 0 dB
/custom/source_provenance/imaging_depth_m	()	float32	m	0.14-m depth read by reconstruction
/custom/simulation/scatterer_positions	(95232, 3)	float32	m	Synthetic coordinates; source-grid-dependent in session 1, fixed across session 2
/custom/simulation/scatterer_magnitudes	(95232,)	float32	arbitrary	Signed synthetic scatterer magnitudes
/custom/simulation/simulation_settings	(11,)	float64	mixed	Element, transmit, scatterer, aperture, focus, timing, scaling, and seed settings; ordering is documented in the dataset attribute
/custom/synchronization/*	scalar or small vector	mixed	s, ms, rad, m, index, or unitless	Relative-time frame links, validity flags, NDI poses, source-frame fields, and deterministic simulation seeds
/custom/session_calibration/image_t_tracker	(4, 4)	float64	translation in mm; linear-block units unconfirmed	Source-supplied transform; direction and spatial registration unconfirmed

N_IQ is 1,824, 1,840, or 1,888, as recorded per frame in /custom/synchronization/source_valid_axial_samples. The synthetic RF tensor has a fixed 3,153 axial samples regardless of source IQ length.

Scalar zea fields additionally store the 15-MHz sampling frequency, 3.75-MHz center and demodulation frequencies, and 1,540-m/s sound speed. Dataset-level attributes provide descriptions and units for individual fields.

Camera representation and calibration

RGB and left/right IR are privacy crops, independently reviewed for each session: x=320, y=200, width=720, height=520 from 1280 × 720 source frames. All source frames and their order are retained at 30 encoded fps. Crop-adjusted principal points satisfy cx_crop = cx_source − 320 and cy_crop = cy_source − 200; focal lengths, distortion coefficients, and inter-camera extrinsics are unchanged. Depth remains an uncropped encoded preview. Standalone calibration and crop-check sidecars are not included in this HF release.

Data Validation

Reconstruction

The contributor reports validation with Python 3.12, zea 0.1.5, and a CUDA 12.8 / PyTorch configuration. The retained pipeline.yaml records the shared reconstruction configuration; the standalone reconstruction entry point and its dependency file are not included in this HF release.

The contributor's reconstruction reads synthetic raw_data, zea scan/probe parameters, and imaging depth. zea's native aligned scanline DAS uses exactly one transmit per scanline, followed by envelope detection, normalization, log compression, and the documented display conversion. Source IQ, stored B-mode, and videos are not reconstruction inputs. Prepared pipelines are reused only for matching scan/probe parameters and precision/device settings.

The contributor's mixed-precision qualification compares against float32, using float32 if scanline RMSE exceeds 0.001 dB or maximum difference exceeds 0.05 dB. This describes the contributor's validation, not a runnable command included in this release. The contributor reports reconstruction of 90 frames across the two sessions, with four endpoints shared with a six-frame precision comparison. This does not imply that every RF file has been reconstructed.

Software references

• Harris, C. R., et al. (2020). Array programming with NumPy. Nature, 585, 357–362.
• Virtanen, P., et al. (2020). SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature Methods, 17, 261–272.
• Stevens, T. S. W., et al. (2026). zea: A Toolbox for Cognitive Ultrasound Imaging. Journal of Open Source Software, 11(121), 9881. See also the zea operations documentation.

Subject Metadata

• Subject type: Phantom; no human patient or animal subject.
• Phantom: Kyoto Kagaku Fetus Ultrasound Examination Phantom “SPACE FAN-ST.”
• Stable HDF5 subject ID: Kyoto-Kagaku-SPACE-FAN-ST in both sessions.
• Sessions: session_01 and session_02, recorded separately from subject ID.
• Anatomical region: Fetal ultrasound training phantom within an abdominal model.
• Demographics/pathology: Not applicable. No verified phantom configuration label.

Known Issues

• Synthetic provenance: Already beamformed IQ cannot recover original element-level receive signals. RF and source-derived B-mode share source information; similarity is not independent physical validation.
• Unverified acquisition geometry: The synthetic geometry is assumed. Scanner transmit law, receive aperture, RF time-zero, and proprietary compound/multibeam processing are unavailable.
• Display approximation: Scan conversion uses a 31.1-mm display aperture, ±14.96° scanline range, virtual apex, and 4% black border. This differs from the 38.4-mm synthetic receive aperture and does not alter DAS delays.
• Spatial registration: Probe/tool poses and ultrasound/cameras are temporally synchronized only, not spatially registered. The source-supplied image_t_tracker matrix has an unconfirmed transform direction. No camera-to-NDI or camera-to-ultrasound transform, calibrated probe-tip mapping, or phantom-anatomical registration is supplied. NDI axis interpretation is unverified. RealSense inter-camera calibration does not establish those mappings.
• Camera encoding: RGB/IR are lossy H.264 crops. Depth is an H.264 display preview, not lossless metric Z16 depth.
• Identical infrared views: Within each session, source IR-left/right are byte-identical, as are their cropped derivatives. They are not independent stereo views; supplied stereo extrinsics do not resolve this source limitation.

Ethical Considerations

Ultrasound content is phantom and simulation data, with no patient records, demographics, or animal subjects. Patient consent, human-subject IRB approval, and animal-welfare review are not applicable to that content.

Room-camera recordings originally captured people. Released RGB/IR crops retain hands/forearms but exclude faces and the wider room. Uncropped RGB/IR, audio, original source identifiers, absolute acquisition clocks, and local filesystem paths are excluded. The camera review fully decoded all ten retained streams, inspected 48 distributed frames per stream, and reviewed every face-detector hit. The contributor reports that the reviewed hits were false positives and no faces were visible in the crops. The underlying review records are not included in this HF release.