ubc: sync module_A and module_C cards, module_A pipeline and figures with GitHub

ubc/module_A/README.md

@@ -1,5 +1,6 @@
 ---
-Name: "OpenH-RF UBC Module A Corrected Synthetic-Channel S-WAVE Phantom Dataset"
+name: ubc-module-a
+pretty_name: "UBC Module A Corrected Synthetic-Channel S-WAVE Phantom Dataset"
 license: cc-by-4.0
 task_categories:
   - image-to-image
@@ -18,99 +19,86 @@ size_categories:
   - 1K<n<10K
 ---
 
-# Module A
+# UBC Module A — Synthetic-Channel S-WAVE Liver Fibrosis Phantom
+
+![Example reconstruction](./assets/module_A.png)
+
+*Scan-converted scanline B-mode of a CIRS 039 liver fibrosis phantom, reconstructed from the synthetic channel data in [`acquisitions/`](https://huggingface.co/datasets/nvidia/OpenH-RF/tree/main/ubc/module_A/acquisitions).*
 
 ## Dataset Description
 
-This submission contains eight 3-D Shear-Wave Absolute Vibro-Elastography
-(S-WAVE) acquisitions of CIRS model 039 Shear Wave Liver Fibrosis Phantom
-samples. The source scanner export contains beamformed 64-line RF rather than
-measured pre-beamforming channel capture. Accordingly, every submitted zea
-`raw_data` frame (stored at `/tracks/track_0/data/raw_data`) is explicitly
-labeled **in-silico/synthetic**: a dense
-point-scatterer field is estimated from one corrected source line-RF frame and
-forward-simulated through a documented 64-transmit, 128-receive-element model.
+This submission contains eight 3-D Shear-Wave Absolute Vibro-Elastography (S-WAVE) acquisitions of CIRS model 039 Shear Wave Liver Fibrosis Phantom samples. The source scanner export contains beamformed 64-line RF rather than measured pre-beamforming channel capture. Accordingly, every submitted zea `raw_data` frame (stored at `/tracks/track_0/data/raw_data`) is explicitly labeled **in-silico/synthetic**: a dense point-scatterer field is estimated from one corrected source line-RF frame and forward-simulated through a documented 64-transmit, 128-receive-element model.
 
-The files are physically delay-consistent and reconstruct with zea-native
-scanline delay-and-sum. They are not recovered or measured scanner channel
-data, and they must not be represented as such.
+The files are physically delay-consistent and reconstruct with zea-native scanline delay-and-sum. They are not recovered or measured scanner channel data, and they must not be represented as such.
 
-## Dataset Contributors
+## Dataset Contributor(s)
 
-University of British Columbia. Primary contact: Zongze Li
-(`zongze@student.ubc.ca`). Proposed contributors retained from the original
-Module A data card: 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.
+- Zongze Li <zongze@student.ubc.ca> (dataset-preparation contact)
+- Yu Chung Lee
+- Qi Zeng
+- Wanwen Chen
+- Zijian Wu
+- Yuxin Chen
+- Patrick Boyan Chen
+- Michael Frew
+- Tajwar Abrar Aleef
+- Hamid Moradi
+- Mohammad Honarvar
+- Septimiu E. Salcudean
+- Robert Rohling
+- The University of British Columbia (UBC), Department of Electrical and Computer Engineering and School of Biomedical Engineering
 
 ## Dataset Creation Date
 
-09/01/2026 for this corrected zea-beamformable synthetic-channel package. The source
-phantom acquisitions predate this packaging.
+09/01/2026 for this corrected zea-beamformable synthetic-channel package. The source phantom acquisitions predate this packaging.
 
 ## License / Terms of Use
 
-The dataset package is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
+[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
 
-The package supports raw-to-B-mode reconstruction tests, ultrasound
-beamforming research, synthetic-channel modeling, per-plane RF analysis, and
-3-D placement using the retained motor-plane geometry. The original proposal
-described derived S-WAVE elasticity estimates or maps; those products are not
-available in this release due to missing data. `/custom/labels/factory_elasticity_kpa` is the CIRS
-factory-rated nominal stiffness, not a derived S-WAVE estimate.
+The package supports raw-to-B-mode reconstruction tests, ultrasound beamforming research, synthetic-channel modeling, per-plane RF analysis, and 3-D placement using the retained motor-plane geometry. The original proposal described derived S-WAVE elasticity estimates or maps; those products are not available in this release due to missing data. `/custom/labels/factory_elasticity_kpa` is the CIRS factory-rated nominal stiffness, not a derived S-WAVE estimate.
 
 ## Dataset Characterization
 
-- Data Collection Method: physical phantom source acquisition followed by a
-  deterministic in-silico channel simulation.
+- Data Collection Method: physical phantom source acquisition followed by a deterministic in-silico channel simulation.
 - Labeling Method: CIRS factory-rated nominal phantom elasticity.
 - Acquisition System: Ultrasonix/Sonix with 4DEC9-5/10 end-firing 3-D probe.
 - Submitted Data Tier: phantom / simulation.
-- Simulation Framework: custom deterministic NumPy/SciPy forward simulator;
-  files and reconstruction were validated with zea 0.1.4.
+- Simulation Framework: custom deterministic NumPy/SciPy forward simulator; files and reconstruction were validated with zea 0.1.4.
 
 Acquisition and simulation details:
 
 - Source system: Ultrasonix/Sonix with 4DEC9-5/10 end-firing 3-D probe.
-- Source content: 20 motor planes × 25 retained temporal RF frames × 64
-  beamformed scan lines × 5,196 axial samples per case.
-- Corrected RF ordering: the 108 stored 16-line events per plane contain two
-  complete setup images at events 0-3 and 4-7, followed by four sector-major
-  groups of 25 temporal frames. Temporal frame `k` is assembled from raw events
-  `[8+k, 33+k, 58+k, 83+k]`.
-- Temporal frames: all corrected source frames 0-24 are retained; none are
-  deleted or wrapped.
-- Cases / nominal stiffnesses: 1.83, 2.8, 6.2, 6.39, 7.79, 11.63, 17.47,
-  and 21.23 kPa.
+- Source content: 20 motor planes × 25 retained temporal RF frames × 64 beamformed scan lines × 5,196 axial samples per case.
+- Corrected RF ordering: the 108 stored 16-line events per plane contain two complete setup images at events 0-3 and 4-7, followed by four sector-major groups of 25 temporal frames. Temporal frame `k` is assembled from raw events `[8+k, 33+k, 58+k, 83+k]`.
+- Temporal frames: all corrected source frames 0-24 are retained; none are deleted or wrapped.
+- Cases / nominal stiffnesses: 1.83, 2.8, 6.2, 6.39, 7.79, 11.63, 17.47, and 21.23 kPa.
 - Source line-RF sampling frequency: 40 MHz.
-- Source temporal sampling: 250 Hz (4 ms between retained RF frames within a
-  plane); 20 motor planes span approximately 27.78° from first to last centre.
+- Source temporal sampling: 250 Hz (4 ms between retained RF frames within a plane); 20 motor planes span approximately 27.78° from first to last centre.
 - Simulated channel sampling frequency: 20 MHz.
 - Transmit/demodulation center frequency: 4 MHz.
 - Sound speed used by simulation and reconstruction: 1,540 m/s.
-- Probe geometry: 128 elements on a 10.0-mm-radius convex arc with 0.17-mm
-  pitch and 6.5-mm elevation height.
-- Sector: 64 focused lines spanning 124.68°; the manufacturer-listed 149°
-  value is the extended-sector capability, not the normal sector used here.
+- Probe geometry: 128 elements on a 10.0-mm-radius convex arc with 0.17-mm pitch and 6.5-mm elevation height.
+- Sector: 64 focused lines spanning 124.68°; the manufacturer-listed 149° value is the extended-sector capability, not the normal sector used here.
 - Transmit focus: 30 mm; modeled transmit aperture: F/1.5.
-- Simulated inter-transmit interval: 250 microseconds; the final interval in
-  each one-frame file is zero because no later transmit is stored in that file.
-- Scatterers: 256 per source line (16,384 per output frame), with deterministic
-  frame-specific sub-line and sub-sample jitter.
-- Channel trace: 3,546 samples (177.3 microseconds), including a final exact-zero
-  tail of 400 samples (20 microseconds).
-- Quantization: `int16`, scaled to an absolute peak of 26,000; no sample reaches
-  int16 full scale.
+- Simulated inter-transmit interval: 250 microseconds; the final interval in each one-frame file is zero because no later transmit is stored in that file.
+- Scatterers: 256 per source line (16,384 per output frame), with deterministic frame-specific sub-line and sub-sample jitter.
+- Channel trace: 3,546 samples (177.3 microseconds), including a final exact-zero tail of 400 samples (20 microseconds).
+- Quantization: `int16`, scaled to an absolute peak of 26,000; no sample reaches int16 full scale.
+
+## Processing the Dataset
+
+The acquisitions can be processed with the `reconstruct.py` [script](https://github.com/open-h/OpenH-RF/blob/main/datasets/ubc/module_A/reconstruct.py) as provided in the [OpenH-RF GitHub repository](https://github.com/open-h/OpenH-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.
+
+`reconstruct.py` reconstructs one frame (default `case_1.83/ubc_swave_cirs_1.83_p10_f13.hdf5`); [`reconstruct_multiframe.py`](https://github.com/open-h/OpenH-RF/blob/main/datasets/ubc/module_A/reconstruct_multiframe.py) reconstructs the temporal sequence at one motor plane (`CASE_ID`, `PLANE`, `N_FRAMES`) into an animated GIF.
 
 ## Dataset Format
 
-There are 8 × 20 × 25 = **4,000** one-frame HDF5 files. One file per source
-frame keeps full zea schema validation and reconstruction memory-bounded.
-For motion analysis, group files by case and motor plane, then order `f00`
-through `f24`; `/custom/source_provenance/source_timing` retains the source
-clock index, within-plane temporal offset, and relative plane timestamp.
+[zea v0.1.4](https://github.com/tue-bmd/zea)
+
+There are 8 × 20 × 25 = **4,000** one-frame HDF5 files. One file per source frame keeps full zea schema validation and reconstruction memory-bounded. For motion analysis, group files by case and motor plane, then order `f00` through `f24`; `/custom/source_provenance/source_timing` retains the source clock index, within-plane temporal offset, and relative plane timestamp.
 
 ```text
 module_A/
@@ -127,8 +115,7 @@ module_A/
 
 Each HDF5 file contains the following per-sample features:
 
-Numeric arrays use zea 0.1.4's native frame-chunked Blosc/Zstd plus bitshuffle
-compression; reading these arrays requires `hdf5plugin`, which is imported by zea.
+Numeric arrays use zea 0.1.4's native frame-chunked Blosc/Zstd plus bitshuffle compression; reading these arrays requires `hdf5plugin`, which is imported by zea.
 
 | Name | Shape | Dtype | Units | Description |
 |---|---:|---|---|---|
@@ -142,8 +129,7 @@ compression; reading these arrays requires `hdf5plugin`, which is imported by ze
 | `/custom/simulation/scatterer_positions` | `(16384, 3)` | `float32` | m | Point-scatterer positions used for channel simulation. |
 | `/custom/source_provenance/*` | mixed | mixed | documented per field | Source indices, plane transform, timing, and excitation settings. |
 
-Synthetic provenance is written at the HDF5 root and on the raw-data dataset:
-`synthetic=true`, `in_silico=true`, and `data_origin=in-silico/synthetic`.
+Synthetic provenance is written at the HDF5 root and on the raw-data dataset: `synthetic=true`, `in_silico=true`, and `data_origin=in-silico/synthetic`.
 
 ## Dataset Quantification
 
@@ -154,21 +140,15 @@ Synthetic provenance is written at the HDF5 root and on the raw-data dataset:
 - Train / validation / test split: not predefined.
 - Per-sample raw tensor: `1 × 64 × 3546 × 128 × 1`, `int16`.
 - Per-sample source line-RF tensor: `64 × 5196`, `int16`.
-- Exact HDF5 payload size: **108,889,767,936 bytes** (108.890 GB; 101.412
-  GiB).
+- Exact HDF5 payload size: **108,889,767,936 bytes** (108.890 GB; 101.412 GiB).
 
 ## Subject Metadata
 
-The source consists of eight phantom acquisitions only, one for each listed
-factory-rated stiffness. There are no human or animal subjects and no subject
-demographics. The phantom model is CIRS model 039 Shear Wave Liver Fibrosis
-Phantom.
+The source consists of eight phantom acquisitions only, one for each listed factory-rated stiffness. There are no human or animal subjects and no subject demographics. The phantom model is CIRS model 039 Shear Wave Liver Fibrosis Phantom.
 
 ## Data Validation
 
-The contributor reports validation with Python 3.12.3 and zea 0.1.4.
-The contributor's reconstruction reads only the zea `raw_data` field and
-acquisition parameters, not the stored image or preserved source line RF:
+The contributor reports validation with Python 3.12.3 and zea 0.1.4. The contributor's reconstruction reads only the zea `raw_data` field and acquisition parameters, not the stored image or preserved source line RF:
 
 ```text
 Cast to float32
@@ -182,43 +162,22 @@ Cast to float32
   -> PNG
 ```
 
-In scanline mode the output has 64 angular lines—one per stored focused
-transmit. zea constructs the one-line-per-transmit grid and the one-hot aligned
-transmit mask internally; no custom `flat_pfield` or transmit-weight code is
-used.
+In scanline mode the output has 64 angular lines—one per stored focused transmit. zea constructs the one-line-per-transmit grid and the one-hot aligned transmit mask internally; no custom `flat_pfield` or transmit-weight code is used.
 
-`pipeline.yaml` records the reconstruction configuration.
-`zlims: [0.0001, 0.100023]` are line depths from
-each transmit origin on the curved element surface.
+`pipeline.yaml` records the reconstruction configuration. `zlims: [0.0001, 0.100023]` are line depths from each transmit origin on the curved element surface.
 
-The contributor reports structural checks across all files, plus full zea
-schema validation and raw-only reconstruction on three distributed samples
-from each case (24 samples). For those samples, the preserved source line RF
-was also compared with the original E-scan events.
+The contributor reports structural checks across all files, plus full zea schema validation and raw-only reconstruction on three distributed samples from each case (24 samples). For those samples, the preserved source line RF was also compared with the original E-scan events.
 
-This HF release includes the HDF5 files, this card, and the pipeline
-configuration. Review outputs, standalone reconstruction scripts, and
-validation sidecars are not included.
+This HF release includes the HDF5 files, this card, and the pipeline configuration. Review outputs, standalone reconstruction scripts, and validation sidecars are not included.
 
 ## Known Issues
 
-- The measured source is beamformed line RF. The submitted 128-channel data are
-  deterministic forward simulations conditioned on that source.
-- The simulation preserves source-derived spatial and temporal contrast but
-  cannot recover proprietary scanner receive processing or true measured
-  element-level noise/coupling.
-- The stored image is a reference only and is deliberately excluded from the
-  reconstruction path.
-- Mechanical excitation frequencies (40, 50, and 60 Hz) are simultaneous.
-  Configured phases are retained in source setting units; absolute
-  actuator-to-ultrasound phase synchronization is not claimed.
-- A scanner TGC curve, probe bandwidth, and exact transmit waveform were not
-  available. Reconstruction therefore applies no TGC and uses the documented
-  simulation pulse.
+- The measured source is beamformed line RF. The submitted 128-channel data are deterministic forward simulations conditioned on that source.
+- The simulation preserves source-derived spatial and temporal contrast but cannot recover proprietary scanner receive processing or true measured element-level noise/coupling.
+- The stored image is a reference only and is deliberately excluded from the reconstruction path.
+- Mechanical excitation frequencies (40, 50, and 60 Hz) are simultaneous. Configured phases are retained in source setting units; absolute actuator-to-ultrasound phase synchronization is not claimed.
+- A scanner TGC curve, probe bandwidth, and exact transmit waveform were not available. Reconstruction therefore applies no TGC and uses the documented simulation pulse.
 
 ## Ethical Considerations
 
-The data are phantom acquisitions only. There are no human subjects, no PHI,
-and no patient demographics. The phantom model is CIRS model 039 Shear Wave
-Liver Fibrosis Phantom. Institutional data-release approval was confirmed in
-the original contributor data card.
+The data are phantom acquisitions only. There are no human subjects, no PHI, and no patient demographics.

ubc/module_A/pipeline.yaml

@@ -17,11 +17,13 @@ pipeline:
             - 1.0
             percentile: 99.5
     - log_compress
+    - scan_convert
 parameters:
     enable_scanline: true
     grid_type: polar
     # In scanline mode, zlims are depths along each transmit beam measured
-    # from its origin on the curved element surface.
+    # from its origin on the curved element surface. zea reads distance_to_apex
+    # (10 mm) from the file and adds it to form rho_range for scan conversion.
     zlims:
     - 0.0001
     - 0.100023
@@ -31,3 +33,10 @@ parameters:
     - 0.0
     f_number: 0.0
     apply_lens_correction: false
+    # The 64 transmits span +/-1.088 rad. The probe's stored polar_limits
+    # (+/-1.4144 rad) describe its full field of view, which would scan-convert
+    # into a sector wider than the one actually acquired.
+    polar_limits:
+    - -1.088
+    - 1.088
+    resolution: 0.0004

ubc/module_C/README.md

@@ -1,5 +1,6 @@
 ---
-pretty_name: "OpenH-RF — UBC Module C Fetal Phantom"
+name: ubc-module-c
+pretty_name: "UBC Module C Fetal Phantom"
 license: cc-by-4.0
 task_categories:
   - image-to-image
@@ -21,29 +22,34 @@ size_categories:
   - "1K<n<10K"
 ---
 
-# UBC Module C
+# UBC Module C — Synthetic-Channel Multimodal Fetal Phantom
+
+![Example reconstruction](./assets/module_C.png)
+
+*Scan-converted B-mode of the Kyoto Kagaku SPACE FAN-ST fetal phantom, reconstructed from the synthetic channel data in [`acquisitions/`](https://huggingface.co/datasets/nvidia/OpenH-RF/tree/main/ubc/module_C/acquisitions).*
 
 ## 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.
+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.
+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
+## Dataset Contributor(s)
 
-- **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.
+- Zongze Li <zongze@student.ubc.ca> (dataset-preparation contact)
+- Yu Chung Lee
+- Qi Zeng
+- Wanwen Chen
+- Zijian Wu
+- Yuxin Chen
+- Patrick Boyan Chen
+- Michael Frew
+- Tajwar Abrar Aleef
+- Hamid Moradi
+- Mohammad Honarvar
+- Septimiu E. Salcudean
+- Robert Rohling
+- The University of British Columbia (UBC), Department of Electrical and Computer Engineering and School of Biomedical Engineering
 
 ## Dataset Creation Date
 
@@ -52,39 +58,18 @@ synchronization only; no validated spatial registration is supplied.
 
 ## 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.
+[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
 
-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.
+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.
+- **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
 
@@ -106,17 +91,19 @@ physical ground truth for the simulated channels.
 | 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.
+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.
+
+## Processing the Dataset
 
+The acquisitions can be processed with the `reconstruct.py` [script](https://github.com/open-h/OpenH-RF/blob/main/datasets/ubc/module_C/reconstruct.py) as provided in the [OpenH-RF GitHub repository](https://github.com/open-h/OpenH-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.
+
+Set `ZEA_FILE` and `FRAME` at the top of the script to pick a frame (default `session_01/session_01_f1306.hdf5`).
 
 ## 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`.
+[zea v0.1.5](https://github.com/tue-bmd/zea)
+
+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
 
@@ -128,33 +115,13 @@ not processing speed. Source IQ and its display reference are under
 | `external_camera/session_XX/` | RealSense MP4 video |
 | [`pipeline.yaml`](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.
+`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.
+**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
 
@@ -169,20 +136,11 @@ synchronization metadata retained in HDF5.
 | 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.
+**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.
+**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
 
@@ -211,48 +169,24 @@ this does not establish generalization to different phantoms or patients.
 | `/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.
+`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.
+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.
+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.
+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
 
@@ -271,38 +205,21 @@ comparison. This does not imply that every RF file has been reconstructed.
 
 ## 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.
+- **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.
+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.
+
+## Citation
+
+The HDF5 `metadata/credit` field carries the following attribution:
+
+> 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.