weillcornell: restructure the data card to the common layout

weillcornell/README.md

@@ -1,6 +1,7 @@
 ---
+name: weillcornell
 license: cc-by-4.0
-pretty_name: OpenH-RF QUS Phantom Dataset
+pretty_name: "Weill Cornell QUS Phantom Dataset"
 task_categories:
   - image-to-image
   - feature-extraction
@@ -17,142 +18,105 @@ size_categories:
   - n<1K
 ---
 
-# OpenH-RF QUS Phantom Dataset
+# Weill Cornell QUS Phantom Dataset
 
-Pre-beamformed RF channel data from three homogeneous tissue-mimicking
-phantoms. Each zea HDF5 acquisition is self-contained and includes raw RF,
-model-derived theoretical backscatter coefficient (BSC), direct per-frame
-Nakagami maps, and 20-frame pooled Nakagami references.
+![Weill Cornell QUS phantom dataset: B-mode, pooled Nakagami maps, and theoretical BSC curves](assets/hero.png)
 
-## OpenH-RF Release Inventory
+*B-mode (frame 0) and 20-frame pooled Nakagami maps from [`data/ac10_15m_SK.hdf5`](https://huggingface.co/datasets/nvidia/OpenH-RF/blob/main/weillcornell/data/ac10_15m_SK.hdf5). The BSC curves are material-model references for the three phantoms; pooled Nakagami maps are statistical references.*
 
-**Current OpenH-RF release:** 120 HDF5 files; 1.00 GB (1,002,700,800 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.
+## Dataset Description
 
-## Contributors
+Pre-beamformed RF channel data from three homogeneous tissue-mimicking phantoms, acquired for quantitative ultrasound (QUS). Next to the raw RF, each acquisition carries the phantom's model-derived theoretical backscatter coefficient (BSC), direct per-frame Nakagami maps, and 20-frame pooled Nakagami references, so single-frame QUS estimators can be trained and evaluated against targets stored in the same file.
 
-Shangke Liu, Tipu Sultan, Cameron Hoerig, and Jonathan Mamou;
-Biomedical Ultrasound Research Laboratory (BURL), Weill Cornell Medicine.
+## Dataset Contributor(s)
 
-Contact: Shangke Liu, shl4035@med.cornell.edu
+- Shangke Liu <shl4035@med.cornell.edu> (contact)
+- Tipu Sultan
+- Cameron Hoerig
+- Jonathan Mamou
+- Biomedical Ultrasound Research Laboratory (BURL), Weill Cornell Medicine
 
-## Dataset Summary
+## Dataset Creation Date
 
-- Verasonics Vantage 256 and GE9LD linear array
-- 192 elements, 0.23 mm pitch, 5.2083 MHz center frequency
-- One normal-incidence plane wave per frame
-- 20.8333 MHz RF sampling; 1540 m/s sound-speed metadata
-- 3 phantoms x 2 operators x 20 probe placements x 20 frames
-- 120 acquisitions and 2400 frames
-- Phantom-only data; no human or animal subjects, clinical metadata, or PHI
-- No predefined train/validation/test split
+Not specified by the contributors.
 
-Keep all frames from one acquisition in the same split.
+## License / Terms of Use
 
-## Files
+[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.
 
-```text
-data/<scan_id>.hdf5                         self-contained zea acquisitions
-reconstruct.py                              B-mode and QUS example
-pipeline.yaml                               zea reconstruction pipeline
-LICENSE                                     CC BY 4.0 license
-ac1_15m_SK_pipeline.png                     reference zea reconstruction
-```
+## Intended Usage
 
-Scan IDs use `ac<number>_<phantom>_<operator>`, for example
-`ac1_15m_SK`. Phantom IDs are `15m`, `18m`, and `60m`; operator IDs are
-`SK` and `TP`; acquisition numbers run from 1 to 20.
+1. Reconstruct B-mode from one frame of raw RF.
+2. Predict the phantom's theoretical BSC curve from one RF frame. This is a material-model target, not a local experimentally estimated BSC curve.
+3. Predict pooled Nakagami `m` or `omega` from one RF frame. Use the supplied direct per-frame map as the single-frame baseline.
 
-## HDF5 Contents
+For Nakagami predictions, report MAE/RMSE and map correlation against the pooled reference. For BSC predictions, report the frequency range and unit and state whether evaluation uses linear BSC or a specified dB conversion.
 
-Each file follows zea 0.1.6 and contains one track. Raw RF is stored as `int16`
-ADC counts without demodulation, decimation, or resampling.
+## Dataset Characterization
 
-| Field under `tracks/track_0/data/` | Shape | dtype | Unit | Meaning |
-|---|---|---|---|---|
-| `raw_data` | `(20,1,n_ax,192,1)` | `int16` | ADC counts | Raw RF input |
-| `theoretical_bsc/values` | `(20,z,x,157)` | `float32` | `m^-1 sr^-1` | Model-derived BSC target |
-| `nakagami_m_per_frame/values` | `(20,z,x)` | `float32` | `1` | Direct single-frame shape estimate |
-| `nakagami_omega_per_frame/values` | `(20,z,x)` | `float32` | `a.u.^2` | Direct single-frame spread estimate |
-| `nakagami_m_pooled/values` | `(20,z,x)` | `float32` | `1` | 20-frame pooled shape reference |
-| `nakagami_omega_pooled/values` | `(20,z,x)` | `float32` | `a.u.^2` | 20-frame pooled spread reference |
+- **Data collection method:** phantom — three homogeneous tissue-mimicking phantoms (`15m`, `18m`, `60m`), each scanned by two operators (`SK`, `TP`) at 20 probe placements
+- **Labeling method:** derived — BSC from the phantom material model; Nakagami maps estimated from the RF
+- **Acquisition system:** Verasonics Vantage 256, GE9LD linear array, 192 elements, 0.23 mm pitch, 5.2083 MHz center frequency, 20.8333 MHz RF sampling
+- **Transmit sequence:** one normal-incidence plane wave per frame; 1540 m/s sound speed
 
-The five QUS products are zea Map fields with `values`, Cartesian
-`coordinates` in `[x,y,z]` order and metres, `description`, `unit`, `min`, and
-`max`. The BSC field also contains 157 `labels` of the form
-`frequency_hz=<value>`, spanning 3.3162435-6.4900716 MHz.
+## Processing the Dataset
 
-Map grids are `(z,x)=(26,35)` for `15m` and `18m`, and `(53,35)` for `60m`.
-The analysis windows are approximately 2.96 mm axial by 4.37 mm lateral with
-75% overlap. Nakagami maps use normal-incidence delay-and-sum beamforming,
-Hann receive apodization, and intensity-domain maximum-likelihood fitting,
-with `omega = E[A^2]`.
+The acquisitions can be processed with the `pipeline.yaml` definition in this folder and the [zea library](https://github.com/tue-bmd/zea).
 
-The theoretical BSC is a homogeneous phantom material property: each phantom
-has one curve, broadcast over space and all 20 frames. The pooled Nakagami maps
-are also repeated along the frame axis so every single-frame input has the same
-20-frame reference. The pooled estimate includes the selected input frame.
+`zea` streams the data from the Hugging Face Hub and processes it according to the pipeline. You can try it out with the following command:
 
-Metadata uses `metadata.subject.type="phantom"` and a phantom ID. Anatomical
-annotations and anatomical-view annotations are intentionally omitted.
-Plane-wave acquisition is recorded in the file description, and the linear
-array geometry is recorded in the probe fields.
+```bash
+zea process \
+  --dataset hf://nvidia/OpenH-RF/weillcornell/data/ac10_15m_SK.hdf5 \
+  --config hf://nvidia/OpenH-RF/weillcornell/pipeline.yaml
+```
 
-## Suggested Tasks
+Alternatively, you can use the `reconstruct.py` [script](https://github.com/open-h/OpenH-RF/blob/main/datasets/weillcornell/reconstruct.py) as provided in the [OpenH-RF GitHub repository](https://github.com/open-h/OpenH-RF). Besides the B-mode, it prints the selected frame's BSC band and unit, the QUS map shapes, and the error of the direct per-frame Nakagami maps against the pooled references.
 
-1. Reconstruct B-mode from one frame of raw RF.
-2. Predict the phantom's theoretical BSC curve from one RF frame. This is a
-   material-model target, not a local experimentally estimated BSC curve.
-3. Predict pooled Nakagami `m` or `omega` from one RF frame. Use the supplied
-   direct per-frame map as the single-frame baseline.
+## Dataset Format
 
-For Nakagami predictions, report MAE/RMSE and map correlation against the
-pooled reference. For BSC predictions, report the frequency range and unit and
-state whether evaluation uses linear BSC or a specified dB conversion.
+[zea v0.1.6](https://github.com/tue-bmd/zea)
 
-All QUS maps are public reference targets, not hidden competition labels.
+Each file contains one track. Raw RF is stored as `int16` ADC counts without demodulation, decimation, or resampling. The five QUS targets are zea Map fields next to `raw_data`, with Cartesian coordinates in metres. The BSC field has 157 frequency `labels` (`frequency_hz=<value>`) spanning 3.3162435-6.4900716 MHz.
 
-## Quick Start
+Map grids are `(z,x)=(26,35)` for `15m` and `18m`, and `(53,35)` for `60m`. The analysis windows are approximately 2.96 mm axial by 4.37 mm lateral with 75% overlap. Nakagami maps use normal-incidence delay-and-sum beamforming, Hann receive apodization, and intensity-domain maximum-likelihood fitting, with `omega = E[A^2]`.
 
-For the current release, use an OpenH-RF environment with `zea==0.1.6` and a supported Keras backend:
+The theoretical BSC is a homogeneous phantom material property: each phantom has one curve, broadcast over space and all 20 frames. The pooled Nakagami maps are also repeated along the frame axis so every single-frame input has the same 20-frame reference. The pooled estimate includes the selected input frame.
 
-```bash
-KERAS_BACKEND=jax python reconstruct.py \
-  --input data/ac1_15m_SK.hdf5 \
-  --output ac1_15m_SK_pipeline.png \
-  --frame 0
-```
+## Dataset Quantification
 
-The script defines the pipeline in code, writes `pipeline.yaml`, saves the
-B-mode PNG, then prints the selected frame's RF shape, BSC band and unit, QUS
-map shapes, and direct per-frame-to-pooled Nakagami errors.
+**Current OpenH-RF release:** 120 HDF5 files; 1.00 GB (1,002,700,800 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.
 
-## Validation
+- **Acquisitions / frames:** 120 acquisitions (3 phantoms x 2 operators x 20 probe placements, one HDF5 file each), 20 frames per acquisition — 2,400 frames total
+- **File naming:** `data/ac<n>_<phantom>_<operator>.hdf5`, with `n` from 1 to 20, e.g. `ac10_15m_SK.hdf5`
+- **Splits:** no predefined train/validation/test split. Keep all frames from one acquisition in the same split.
 
-The original 120 HDF5 files passed zea 0.1.3 `File.validate()` and
-`File.validate_spec()`. Raw RF and embedded maps were checked for shape, dtype,
-finite values, coordinates, labels, min/max, and the documented frame-broadcast
-behavior. The original files used zea 0.1.3's default Blosc/Zstd+bitshuffle compression.
+| Field under `tracks/track_0/data/` | Shape | dtype | Unit | Meaning |
+|---|---|---|---|---|
+| `raw_data` | `(20,1,n_ax,192,1)` | `int16` | ADC counts | Raw RF input |
+| `theoretical_bsc/values` | `(20,z,x,157)` | `float32` | `m^-1 sr^-1` | Model-derived BSC target |
+| `nakagami_m_per_frame/values` | `(20,z,x)` | `float32` | `1` | Direct single-frame shape estimate |
+| `nakagami_omega_per_frame/values` | `(20,z,x)` | `float32` | `a.u.^2` | Direct single-frame spread estimate |
+| `nakagami_m_pooled/values` | `(20,z,x)` | `float32` | `1` | 20-frame pooled shape reference |
+| `nakagami_omega_pooled/values` | `(20,z,x)` | `float32` | `a.u.^2` | 20-frame pooled spread reference |
+
+## Subject Metadata
+
+No human or animal subjects. `metadata/subject/type` is `phantom` and `metadata/subject/id` names the phantom (`phantom_15m`, `phantom_18m`, `phantom_60m`).
 
-![Reference zea reconstruction of ac1_15m_SK frame 0](assets/ac1_15m_SK_pipeline.png)
+## Data Validation
 
-The reference uses frame 0 and the common 597 x 300 zea grid at approximately
-3-25 mm depth.
+A `zea.Pipeline` (cast → demodulate → DAS beamforming → envelope detection → normalization → log compression) is defined in `pipeline.yaml`, on a common 301 x 600 grid (isotropic ~0.073 mm pixels) for every acquisition. Its output for frame 0 of `ac10_15m_SK` is shown at the top of this card. Raw RF and the embedded maps were checked for shape, dtype, finite values, coordinates, labels, min/max, and the frame-broadcast behavior described above.
 
-## Limitations
+## Known Issues
 
-- Acquisition timestamps, PRF, explicit TGC curves, and emitted waveforms are
-  unavailable; do not use these frames for calibrated temporal or flow analysis.
+- Acquisition timestamps, PRF, explicit TGC curves, and emitted waveforms are unavailable; do not use these frames for calibrated temporal or flow analysis.
 - Repeated frames are strongly correlated.
-- BSC targets are model-derived references, not independent experimental BSC
-  measurements; there are three unique phantom curves.
-- Pooled Nakagami maps are lower-variance statistical references, not
-  independent physical ground truth.
-- Nakagami `omega` depends on gain, attenuation, beam sensitivity, and RF
-  amplitude scale; it is not system-independent.
-- This release contains Verasonics data only.
-
-## License
-
-CC BY 4.0. See `LICENSE`. Commercial and non-commercial reuse is permitted
-with attribution. The contributors confirm that these phantom data are cleared
-for release under CC BY 4.0.
+- BSC targets are model-derived references, not independent experimental BSC measurements; there are three unique phantom curves.
+- Pooled Nakagami maps are lower-variance statistical references, not independent physical ground truth.
+- Nakagami `omega` depends on gain, attenuation, beam sensitivity, and RF amplitude scale; it is not system-independent.
+
+## Ethical Considerations
+
+Phantom-only data; no human or animal subjects, clinical metadata, or PHI. The contributors confirm that these phantom data are cleared for release under CC BY 4.0.
+

weillcornell/pipeline.yaml

@@ -13,8 +13,8 @@ pipeline:
             - 1.0
     - log_compress
 parameters:
-    grid_size_x: 300
-    grid_size_z: 597
+    grid_size_x: 600
+    grid_size_z: 301
     xlims:
     - -0.021965
     - 0.021965