Add vanderbilt README + pipeline(s) from accepted_submissions

vanderbilt/README.md

@@ -0,0 +1,80 @@
+---
+pretty_name: "OpenH-RF — Vanderbilt / Multi-Frame Focused Transmit Echocardiography Channel Dataset"
+license: cc-by-4.0
+task_categories:
+  - image-reconstruction
+tags:
+  - ultrasound
+  - rf
+  - openh-rf
+  - echocardiography
+  - cardiac
+  - harmonic-imaging
+  - phased-array
+language:
+  - en
+size_categories:
+  - n<1K
+---
+
+# Multi-Frame Focused Transmit Echocardiography Channel Dataset
+
+## Dataset Description
+
+This multi-frame focused transmit echocardiography channel dataset contains over 2000 frames of fundamental and harmonic data acquired with the P4-2v probe on a Verasonics Vantage 128. This dataset was originally acquired to visualize the left atrial appendage in patients following transesophageal echocardiography. Some patients have atrial fibrillation, which can cause blood clots to form in the appendage.
+
+## Dataset Contributor(s)
+
+Brett Byram (PI), Christopher Khan, Ying-Chun (Preston) Pan, Zoe Marshall
+Vanderbilt University
+
+## Dataset Creation Date
+
+07/11/2026
+
+## License / Terms of Use
+
+CC BY 4.0.
+
+## Intended Usage
+
+This dataset could be useful for training a domain-adaptive network, as it captures a wide range of in vivo image quality. It might also be useful for training a model that converts fundamental to harmonic data — note that the harmonic and fundamental data are not matched (see Dataset Characterization). It contains view names provided by the physician at the time of acquisition, but these views were often slightly modified to accommodate the patient (see Subject Metadata). It could also be useful for identifying thrombus in the left atrial appendage, as ~10% of the dataset came from patients with a thrombus (see Subject Metadata).
+
+## Dataset Characterization
+
+- **Data Collection Method:** clinical
+- **Labeling Method:** human-annotated
+- **Acquisition System:** Verasonics Vantage 128, P4-2v (64 elements). Fundamental: 2.72 MHz; Harmonic: 2.08 MHz transmit / 4.16 MHz receive. Sampling rate: 10.88 MHz.
+- **Frame rate:** Fundamental: 25 Hz (32 frames). Harmonic: 10 Hz (32 frames).
+- Note that the fundamental and harmonic sequences are not matched or interleaved: the harmonic sequence was executed immediately after the fundamental sequence within the same Verasonics setup file.
+
+## Dataset Format
+
+All acquisitions are submitted in the *zea* file format as raw RF channel data, with no preprocessing applied.
+
+| Group / field | Shape | Description |
+|---|---|---|
+| `data/raw_data` | `[n_frames, n_tx, n_ax, n_el, 1]` | Raw RF channel data (fundamental or harmonic, one file per cineloop) |
+
+## Dataset Quantification
+
+- **Samples:** 82 cineloops (32 frames each) from 29 patients — 2,624 fundamental frames
+- Plus a matching set of 2,624 harmonic frames (same cineloops, harmonic sequence)
+- **Train / validation / test split:** N/A
+
+## Subject Metadata
+
+- `imaging_view_name`: one of `{'apical two chamber', 'apical four chamber', 'parasternal short axis', 'parasternal long axis', 'subxiphoid', 'N/A'}`. Because acquisition occurred while patients were recovering from anesthesia, sonographers often deviated from standard view geometry to maximize image quality; as a result, these datasets should not be used to train a view classification network. The `'N/A'` label marks acquisitions that didn't approximate any standard view at all — in these cases, abandoning the standard geometry entirely was the only way to obtain reasonable image quality.
+- `Notes`: unstructured free-text on patient condition and acquisition quality. Common content includes left atrial appendage occlusion devices (Watchman, Amulet, or Conformal), presence of thrombus, comorbidities such as COPD, and patient motion during acquisition.
+
+## Data Validation
+
+A `zea.Pipeline` (cast → demodulate → DAS beamforming → envelope detection → normalization → log compression → scan conversion) reconstructs the B-mode image from the raw channel data and is defined in [pipeline.yaml](pipeline.yaml). Run [reconstruct.py](reconstruct.py) to reproduce it, e.g. `python reconstruct.py --input my_file.hdf5 --n_frames 32`.
+
+## Known Issues
+
+As mentioned in Subject Metadata, view names should only be used as a reference for interpreting the image and should not be used to train a view classification task.
+
+## Ethical Considerations
+
+The study was approved by Vanderbilt's IRB.

vanderbilt/pipeline.yaml

@@ -0,0 +1,33 @@
+pipeline:
+    operations:
+    -   name: keras.ops.cast
+        params:
+            dtype: float32
+    - demodulate
+    -   name: beamform
+        params:
+            num_patches: 400
+            enable_pfield: true
+    - envelope_detect
+    -   name: normalize
+        params:
+            output_range:
+            - 0.0
+            - 1.0
+    - log_compress
+    - scan_convert
+parameters:
+    grid_type: polar
+    polar_limits:
+    - -0.7853981633974483
+    - 0.7853981633974483
+    grid_size_x: 1084
+    grid_size_z: 636
+    dynamic_range:
+    - -60
+    - 0
+    zlims:
+    - 0
+    - 0.18
+    apply_lens_correction: false
+    f_number: 0