Add weizmann-sampl README + pipeline(s) from accepted_submissions

weizmann-sampl/README.md

@@ -0,0 +1,173 @@
+---
+license: cc-by-4.0
+pretty_name: OpenH-RF Thyroid In-Vivo L11-5v Scan (D1)
+task_categories:
+  - other
+tags:
+  - ultrasound
+  - channel-data
+  - beamforming
+  - thyroid
+  - clinical
+  - zea
+---
+
+# OpenH-RF Thyroid Clinical Scans
+
+## Dataset Description
+
+The data set consists of clinical ultrasound channel data acquired
+with a Verasonics Vantage 128 system and an L11-5v linear array probe, imaging
+the thyroid gland of 30 adult healthy volunteers. The scans were carried out by a senior radiologist
+who specializes in ultrasound thyroid imaging. The purpose of the scans is to provide a complete set of channel data of a thyroid gland scan of the full anatomy.
+
+## Dataset Contributor(s)
+
+Prof. Yonina Eldar's SAMPLLAB Group, Faculty of Mathematics and Computer Science, Weizmann Institute of Science, Rehovot, Israel.
+
+
+## Dataset Creation Date
+
+07/07/2026
+
+## License / Terms of Use
+
+CC-BY-4.0. Acquired under a Weizmann Institute IRB-approved research protocol, with informed
+consent obtained from each subject prior to scanning.
+
+## Intended Usage
+
+General-purpose ultrasound channel-data foundation model pretraining and
+evaluation — in particular DAS beamforming/reconstruction, inverse speed of sound imaging.
+Suitable as a base for future downstream tasks (e.g. thyroid
+segmentation, nodule detection) if paired with additional annotations from the B-mode images.
+
+## Dataset Characterization
+
+- **Data Collection Method:** clinical
+- **Labeling Method:** N/A — no annotations included
+- **Acquisition system:** Verasonics Vantage 128 system, L11-5v linear array
+  probe, center frequency 7.6 MHz (76.8% fractional
+  bandwidth), element width 0.27 mm, aperture width 38.1 mm; RF sampling
+  frequency 31.25 MHz, demodulation frequency 7.8125 MHz, assumed sound
+  speed 1540 m/s.
+
+## Dataset Format
+
+zea file format, one acquisition HDF5 file per subject. Before packaging,
+the frames prior to workspace parameter freezing were removed from the raw channel data frames.
+
+## Dataset Quantification
+- Number of samples / acquisitions / channels: ~1600 / 128 / 128
+- Number of frames:  average 250 (dependent on subject)
+- No train/validation/test split
+- Total size on disk: 374 GB
+- Per-sample feature table:
+
+| Field | Shape                        | Dtype | Units | Description |
+|---|------------------------------|---|---|---|
+| `tracks/track_0/data/raw_data` | (nframes, 128, 4096, 128, 1) | int16 | ADC counts | Raw per-element RF channel data: (frames, transmits, axial samples, elements, 1) |
+| `tracks/track_0/scan/t0_delays` | (128, 128)                   | float32 | s | Per-transmit, per-element transmit delay |
+| `tracks/track_0/scan/tx_apodizations` | (128, 128)                   | float32 | unitless (0-1) | Per-transmit, per-element transmit apodization weight |
+| `tracks/track_0/scan/polar_angles` | (128,)                       | float32 | rad | Per-transmit polar steering angle |
+| `tracks/track_0/scan/azimuth_angles` | (128,)                       | float32 | rad | Per-transmit azimuth steering angle (0 for this 1D linear array) |
+| `tracks/track_0/scan/focus_distances` | (128,)                       | float32 | m | Per-transmit focal distance |
+| `tracks/track_0/scan/transmit_origins` | (128, 3)                     | float32 | m | Per-transmit virtual source origin (x, y, z) |
+| `tracks/track_0/scan/initial_times` | (128,)                       | float32 | s | Per-transmit acquisition start time offset |
+| `tracks/track_0/scan/time_to_next_transmit` | (nframes, 128)               | float32 | s | Per-frame, per-transmit inter-transmit timing |
+| `tracks/track_0/scan/center_frequency` | scalar                       | float32 | Hz | Transmit center frequency (7.6 MHz) |
+| `tracks/track_0/scan/sampling_frequency` | scalar                       | float32 | Hz | RF sampling frequency (31.25 MHz) |
+| `tracks/track_0/scan/demodulation_frequency` | scalar                       | float32 | Hz | Demodulation frequency used downstream (7.8125 MHz) |
+| `tracks/track_0/scan/sound_speed` | scalar                       | float32 | m/s | Assumed speed of sound (1540 m/s) |
+| `tracks/track_0/scan/tgc_gain_curve` | (2048,)                      | float32 | Verasonics TGC units | Time-gain-compensation curve applied during acquisition |
+| `probe/probe_geometry` | (128, 3)                     | float32 | m | Element (x, y, z) positions, L11-5v linear array, ±19.05 mm aperture |
+| `probe/element_width` | scalar                       | float32 | m | Element width (0.27 mm) |
+| `probe/probe_center_frequency` | scalar                       | float32 | Hz | Probe nominal center frequency (7.6 MHz) |
+| `probe/probe_bandwidth_percent` | scalar                       | float32 | % | Probe fractional bandwidth (76.8%) |
+| `probe/lens_thickness` | scalar                       | float32 | m | Acoustic lens thickness (0.390 mm), derived from the Verasonics lens correction scalar; consumed automatically by zea's per-element/per-pixel lens correction during reconstruction (`apply_lens_correction: true` in `pipeline.yaml`) |
+| `probe/lens_sound_speed` | scalar                       | float32 | m/s | Assumed speed of sound in the lens material (1000 m/s) used to derive `lens_thickness` above |
+| `custom/lens_correction` | scalar                       | float64 | wavelengths | Raw Verasonics one-way lens correction delay (2.961 wl), kept for provenance |
+| `metadata/subject/id` | scalar                       | str | - | De-identified subject code |
+| `metadata/subject/type` | scalar                       | str | - | `human` |
+
+
+## Subject Metadata
+- number of subjects: 30,
+- age range: 18-65,
+- 20 Female , 10 Male
+- Thyroid gland
+- healthy volunteers with incidental pathology (e.g. cysts, nodules) in some cases
+- Verasonics vantage 128 system, L11-5v linear array probe
+
+## Data Validation
+
+`reconstruct.py` runs the following `zea.Pipeline` (saved as `pipeline.yaml`,
+alongside this README) to reconstruct a B-mode image from the raw channel
+data: cast → demodulate → DAS beamform (with native, per-element/per-pixel
+lens correction) → envelope detect → normalize → log compress.
+
+```yaml
+# Beamforming config for the L11-5v thyroid scans (focused ray-line imaging).
+# RF data is real (not IQ/baseband, sampleMode=NS200BW) so it is demodulated
+# before beamforming. grid_size_x/z fix the output resolution; the actual
+# depth mapped onto it comes from this file's own N_AX (D1: 2048 samples,
+# ~50mm one-way in 1540 m/s tissue).
+
+parameters:
+  grid_size_x: 300
+  grid_size_z: 400
+  dynamic_range: [-60, 0]
+
+pipeline:
+  operations:
+    - name: keras.ops.cast
+      params:
+        dtype: float32
+    - demodulate
+    - name: beamform
+      params:
+        beamformer: delay_and_sum
+        num_patches: 200
+    - envelope_detect
+    - normalize
+    - log_compress
+```
+
+Example reconstruction, run via `python reconstruct.py --data-dir subjects
+--n-scans 3 --n-frames 3 --seed 0 --output three_patients_grid.png`
+(3 subjects, 3 randomly sampled frames each):
+
+![3 frames from 3 different subjects](three_patients_grid.png)
+
+All 9 sampled frames, across 3 different subjects, show consistent diffuse
+in-vivo tissue speckle with no reconstruction artifacts, confirming the
+acquisition geometry, timing metadata, and native lens correction are
+correctly recorded/applied across the released dataset, not just a single
+acquisition.
+
+## Known Issues
+
+- `decimSampleRate`, `quadDecim`, and `demodFrequency` are absent from the
+  raw Verasonics workspace (`BaselineWorkspace.mat`) and were instead
+  sourced from this session's recalibration checkpoint files,
+  where they were confirmed constant.
+- The default beamforming grid in `pipeline.yaml` (`grid_size_x=300`,
+  `grid_size_z=400`) is coarser than the half-wavelength Nyquist rate for
+  this probe/frequency; this only affects the resolution of the example
+  reconstruction images, not the released raw channel data.
+- The number of frames for each subject is variable, depending on the subject's
+  anatomy and the radiologist's scanning protocol. It may also be affected by
+  bottlenecks in the Verasonics system's data transfer rate,
+  which can cause "dropped frames" when the system cannot keep up with the
+  acquisition speed.
+
+## Ethical Considerations
+
+This acquisition was performed under an IRB-approved research protocol,
+with informed consent obtained from the subject prior to scanning. The
+subjects are identified only by a de-identified code  with no directly
+identifying information (name, exact date of birth, medical record number)
+stored in the released file. No `acquisition_time` timestamp is embedded in
+the zea file, consistent with HIPAA Safe Harbor de-identification guidance
+for human-subject data.
+

weizmann-sampl/pipeline.yaml

@@ -0,0 +1,34 @@
+# Beamforming config for the L11-5v thyroid scans (focused ray-line imaging).
+# Copied unchanged from phantom_test/pipeline.yaml -- same probe.
+#
+# RF data is real (not IQ/baseband, sampleMode=NS200BW) so it is demodulated
+# before beamforming. grid_size_x/z fix the *output* resolution; the actual
+# depth mapped onto it comes from each file's own N_AX and can differ by
+# patient -- convert.py picks N_AX from whichever checkpoint's settings are
+# active for the frames it kept (e.g. D1: N_AX=2048, ~50mm one-way; A3:
+# N_AX=2176, ~53mm one-way, both in 1540 m/s tissue).
+
+parameters:
+  grid_size_x: 377
+  grid_size_z: 525
+  dynamic_range: [-60, 0]
+  apply_lens_correction: true
+  focal_region_length: 0.02
+  # lens_thickness and lens_sound_speed are read from the probe spec in the file
+  # (convert.py derives them from the Verasonics lens_correction custom element
+  # via estimate_lens_probe_params) -- no manual initial_times shift needed here.
+
+pipeline:
+  operations:
+    - name: keras.ops.cast
+      params:
+        dtype: float32
+    - apply_window
+    - demodulate
+    - name: beamform
+      params:
+        beamformer: delay_and_sum
+        num_patches: 200
+    - envelope_detect
+    - normalize
+    - log_compress