Add concordia README + pipeline(s) from accepted_submissions

concordia/README.md

@@ -0,0 +1,292 @@
+---
+pretty_name: "OpenH-RF — SynthUS-FSA"
+license: cc-by-4.0
+task_categories:
+  - image-to-image
+  - image-segmentation
+tags:
+  - ultrasound
+  - rf
+  - openh-rf
+  - synthetic
+  - field-ii
+language:
+  - en
+size_categories:
+  - 1K<n<10K
+---
+
+## Dataset Description
+
+SynthUS-FSA is a fully synthetic corpus of pre-beamformed full-synthetic-aperture (FSA)
+ultrasound channel-data captures, generated with the Field II simulator to model a
+128-element L11-5v linear array. Each capture fires one transmit event per element
+(128 transmits, receive on all 128 elements), so every element-to-element combination
+is retained and any receive/transmit beamforming scheme (focused B-mode, multi-angle
+plane-wave compounding, diverging-wave, adaptive/compressive schemes) can be
+retrospectively synthesized from the same channel data. The dataset is packaged in
+[zea](https://zea.readthedocs.io) format — the OpenH-RF reference Python library for
+ultrasound file I/O and beamforming — which every `.hdf5` file and `reconstruct.py`
+depend on. The dataset targets three
+OpenH-RF task categories: (1) generalized reconstruction (FSA channel data paired with
+a synthesized full synthetic-aperture B-mode as ground truth), (2) compressed sensing /
+adaptive transmit (retrospective sub-selection of the 128 transmit events — derived
+from the same captures, no additional files), and (3) segmentation (binary
+echogenicity masks on a subset of phantoms). All data is simulated — no clinical,
+phantom-hardware, or animal acquisition is involved.
+
+A similar data-generation approach was used in Sharifzadeh et al. (2024). Users of
+this dataset are kindly requested to cite that work:
+
+> M. Sharifzadeh, S. Goudarzi, A. Tang, H. Benali, and H. Rivaz, "Mitigating
+> aberration-induced noise: A deep learning-based aberration-to-aberration approach,"
+> *IEEE Transactions on Medical Imaging*, vol. 43, no. 12, pp. 4380–4392, 2024.
+
+## Dataset Contributor(s)
+
+Mostafa Sharifzadeh (mostafa.sharifzadeh@mail.concordia.ca) and Hassan Rivaz
+(hassan.rivaz@concordia.ca), IMPACT Lab, Concordia University.
+
+## Dataset Creation Date
+
+07/01/2026
+
+## License / Terms of Use
+
+CC BY 4.0.
+
+- **Simulated channel data, phantom parameters, and reconstruction/segmentation
+  labels** are generated by the contributors (Field II simulation outputs) and
+  released under CC BY 4.0.
+- **Segmentation mask shapes** (anechoic/hypoechoic/hyperechoic classes, 750
+  captures) are sourced from [Open Images V7](https://storage.googleapis.com/openimages/web/factsfigures_v7.html)'s
+  animal-class segmentation annotations (Google LLC). Only the binary mask is used — never the underlying source photograph. Open Images V7
+  licenses its annotations, including segmentation masks, under CC BY 4.0; the
+  source photographs themselves carry a separate CC BY 2.0 license but are not
+  used anywhere in this dataset.
+- **Diverse-echogenicity amplitude-weight maps** (1,000 captures) are photographs
+  from Wikimedia Commons, restricted to the `CC-BY-4.0` category and verified
+  per-image against each file's `CC BY 4.0` license field before acceptance.
+  `wikimedia_commons_metadata.csv`, at the dataset root, credits each one: a row
+  per capture, keyed by `DatasetFile` for `image_0751.hdf5` through
+  `image_1750.hdf5`, giving the source URL, Commons file page, author, title and
+  license. No photograph appears in the dataset in its original form. Each was
+  converted to grayscale, resampled to the 400 × 450 weight-map grid,
+  histogram-equalized and rescaled to `[0, 1]`. Values above 0.9 were then set to
+  1 and values below 0.1 to 0, producing the weight map stored as
+  `data/diverse_source_image`. This conversion is implemented in
+  `wikimedia_commons_postprocess.py`.
+- **Point-target phantoms** (250 captures) use no external asset — purely
+  synthetic point scatterers.
+
+All channel data was generated with Field II (Jensen/DTU), distributed for free academic use. Field II's terms do not
+restrict redistribution or licensing of simulation *output* data. Per Field II's terms, any use of this dataset should cite:
+J.A. Jensen, "Field: A Program for Simulating Ultrasound Systems," *Med. Biol.
+Eng. Comp.*, 1996; and J.A. Jensen and N.B. Svendsen, "Calculation of Pressure
+Fields from Arbitrarily Shaped, Apodized, and Excited Ultrasound Transducers,"
+*IEEE Trans. Ultrason., Ferroelec., Freq. Contr.*, 1992.
+
+## Intended Usage
+
+- **Generalized reconstruction** — learning a direct mapping from raw FSA channel
+  data to a B-mode image, and/or from FSA to arbitrary retrospective beamforming
+  targets (focused, multi-angle compounded, diverging-wave, adaptive).
+- **Compressed sensing / adaptive transmit** — sparse-aperture and adaptive transmit
+  design research using retrospective sub-selection of the 128 FSA transmit events.
+- **Segmentation** — pixel-aligned echogenicity-region segmentation (anechoic /
+  hypoechoic / hyperechoic) from raw channel data.
+- **Scatterer-level analysis** — `data/scatterers` exposes the exact Field II
+  point-scatterer cloud (position + amplitude) used to simulate each capture,
+  for tasks that want ground truth finer than a pixel grid (e.g. quantitative
+  ultrasound, scatterer-density estimation).
+
+## Dataset Characterization
+
+- **Data Collection Method:** synthetic (Field II full-synthetic-aperture
+  simulation).
+- **Labeling Method:** synthetic ground truth.
+  - Segmentation masks (anechoic/hypoechoic/hyperechoic) are Open Images V7
+    animal-class segmentation annotations, used for shape only.
+  - Diverse-echogenicity amplitude weighting uses a Wikimedia Commons photograph
+    as a continuous grayscale weight map, preserved as `data/diverse_source_image`
+    so an end user can see what pattern the capture was generated from.
+  - The reconstruction-target B-mode (`data/image`) is synthesized from the same
+    FSA capture (see *Data Validation* below), not an independent acquisition.
+- **Acquisition system (simulated):** 128-element linear array (model: L11-5v),
+  center frequency 5.208 MHz, element width 0.27 mm, kerf 0.03 mm, pitch 0.3 mm,
+  element height 5 mm, elevation focus (Rfocus) 20 mm. Simulated at 104.16 MHz (a
+  high rate required for Field II's numerical precision), then decimated ×5 to a
+  delivered sampling rate of 20.832 MHz. Sound speed 1540 m/s.
+  Receive dynamic-focus reconstruction uses F-number 1.75 (baked into
+  `reconstruct.py`'s `F_NUMBER` constant — zea's own default is 1.0, so this must be
+  supplied explicitly rather than relying on the file alone to reproduce the
+  reference images).
+- **Transmit pulse:** Hann-windowed 2.5-cycle tone burst at the 5.208 MHz center
+  frequency, sampled at the native 104.16 MHz simulation rate. −6dB fractional
+  bandwidth ≈ 76.5% (stored as `probe/probe_bandwidth_percent`).
+- **Coordinate convention:** x = lateral, y = elevation (always 0 for this 2D
+  acquisition), z = axial/depth — the standard zea convention, applied throughout
+  `probe_geometry`, `transmit_origins`, and every map's `coordinates` field.
+- **Frame timing:** each file is a single static frame (`n_frames=1`) with no
+  simulated motion or pulse-repetition-frequency concept, so `scan/time_to_next_transmit`
+  is intentionally omitted rather than populated with a fabricated value. Likewise, no
+  time-gain-compensation was applied to this synthetic data, so `scan/tgc_gain_curve`
+  is omitted.
+
+## Dataset Format
+
+The dataset is 2,000 individual zea HDF5 files (one acquisition per file) under
+`data/` (zea format; OpenH-RF spec floor v0.1.0a3). Reference figures are in
+`examples/`; `reconstruct.py`, `visualize.py`, `pipeline.yaml`, and this card sit
+at the repository root. Per file:
+
+- `data/raw_data` — full FSA channel data, `int16`, quantized from the native
+  simulated (float32) values by per-file peak-scaling into the int16 range
+  (~5e-5 relative quantization step — well below any physically meaningful signal
+  feature, equivalent to how real ultrasound hardware ADCs store raw channel data).
+- `data/image` — the reference B-mode: a synthetic transmit aperture (STA)
+  reconstruction that delay-and-sums all 128 single-element FSA transmits with
+  zea's default DAS pipeline over the intended imaging FOV (see *Subject
+  Metadata*). This is exactly what `reconstruct.py` reproduces from `data/raw_data`.
+- `data/segmentation` — boolean mask with labels `["background", <class>]`,
+  present only for anechoic/hypoechoic/hyperechoic phantoms (750 captures) —
+  the Open Images V7 mask directly. Not produced for diverse or point-target
+  phantoms (no defined region to segment in either case).
+- `data/echogenicity_multiplier` — present alongside `data/segmentation`
+  (same 750 captures): the scalar amplitude multiplier Field II applied to
+  every scatterer inside that region, e.g. "hypoechoic at 0.23×," not just
+  "hypoechoic." A single number per file, so it's stored without a
+  `coordinates` grid (optional per zea's `Map` spec — this isn't spatial).
+- `data/diverse_source_image` — present only for diverse-class phantoms
+  (1,000 captures): the grayscale natural-image echogenicity weight map
+  (continuous `[0, 1]` values) each capture's scatterer amplitudes were
+  generated from. Not a mask — just the reference image, so an end user can
+  see what pattern produced the capture.
+- `data/scatterers` — the exact Field II point-scatterer cloud used to
+  simulate the capture (every capture, all classes): `values` = per-scatterer
+  amplitude (a.u.), `coordinates` = per-scatterer `(x, y, z)` position in
+  metres. ~275,000 scatterers per capture, stored via zea's custom-data
+  extension mechanism as a `(1, n_scatterers, 1)` / `(n_scatterers, 1, 3)`
+  pair rather than a regular pixel grid (a point cloud has no grid to speak
+  of). This is the ground truth Field II actually simulated from — finer than
+  any pixel-grid label derived from it.
+
+Pre-processing applied before packaging: anti-alias FIR decimation (factor 5,
+104.16 MHz → 20.832 MHz) and int16 quantization, both described above. zea's
+`File.create` applies lossless `lzf` HDF5 compression by default.
+
+## Dataset Quantification
+
+This dataset comprises 2,000 FSA captures across five echogenicity classes:
+
+| Class | Count | File indices | Segmentation label | Other class-specific data |
+|---|---|---|---|---|
+| Anechoic | 250 | 0001–0250 | Yes (Open Images V7 mask) | `data/echogenicity_multiplier` (always 0) |
+| Hypoechoic | 250 | 0251–0500 | Yes (Open Images V7 mask) | `data/echogenicity_multiplier` (random, 0.07–0.50) |
+| Hyperechoic | 250 | 0501–0750 | Yes (Open Images V7 mask) | `data/echogenicity_multiplier` (random, 2–8) |
+| Diverse (natural-image-derived) | 1,000 | 0751–1750 | No | `data/diverse_source_image` |
+| Point-target | 250 | 1751–2000 | No | — |
+
+`data/scatterers` (the raw point-scatterer cloud) is present for all 2,000
+captures regardless of class.
+
+Each capture's class label is stored in its file metadata
+(`metadata/annotations/label`).
+
+- **Size on disk:** ≈ 50.3 MB per sample (int16, post-decimation, lzf-compressed), = 100.6 GB for the full 2,000-capture corpus.
+- **Train / validation / test split:** none predefined — the corpus is released
+  as a single set for users to partition as their task requires (class membership
+  and index ranges are given above).
+- **Per-sample feature table:**
+
+| Field | Shape | Dtype | Units | Description |
+|---|---|---|---|---|
+| `data/raw_data` | `(1, 128, n_ax, 128, 1)` | int16 | a.u. | FSA channel data: 1 frame, 128 transmits (one per element), `n_ax` axial samples post-decimation (depth-dependent round-trip timing means this varies slightly per capture — read `.shape` at load time rather than assuming a fixed value), 128 receive elements, 1 channel (RF) |
+| `data/image/values` | `(1, 542, 384)` | float32 | dB | Synthetic-aperture (128-transmit) reference B-mode, log-compressed |
+| `data/image/coordinates` | `(542, 384, 3)` | float32 | m | Per-pixel (x, y, z) for `image/values` |
+| `data/segmentation/values`* | `(1, 400, 450, 2)` | bool | – | One-hot echogenicity mask (`background`, class); *anechoic/hypoechoic/hyperechoic only* |
+| `data/segmentation/coordinates`* | `(400, 450, 3)` | float32 | m | Per-pixel (x, y, z) for the segmentation grid |
+| `data/echogenicity_multiplier/values`* | `(1, 1, 1)` | float32 | – | Scalar amplitude multiplier applied inside the segmentation region; *anechoic/hypoechoic/hyperechoic only* |
+| `data/diverse_source_image/values`* | `(1, 400, 450)` | float32 | a.u. | Grayscale echogenicity weight map, `[0, 1]`; *diverse class only* |
+| `data/diverse_source_image/coordinates`* | `(400, 450, 3)` | float32 | m | Per-pixel (x, y, z) for the weight-map grid |
+| `data/scatterers/values` | `(1, ~275000, 1)` | float32 | a.u. | Per-scatterer amplitude (one row per Field II scatterer, not a pixel grid) |
+| `data/scatterers/coordinates` | `(~275000, 1, 3)` | float32 | m | Per-scatterer (x, y, z) position |
+| `probe/probe_geometry` | `(128, 3)` | float32 | m | Element positions |
+| `probe/probe_bandwidth_percent` | scalar | float32 | % | −6dB fractional bandwidth of the transmit pulse (≈76.5%) |
+| `scan/t0_delays`, `scan/tx_apodizations` | `(128, 128)` | float32 | s, – | Per-transmit, per-element delay / apodization (identity apodization — one element fires per transmit) |
+| `scan/azimuth_angles` | `(128,)` | float32 | rad | All zero — 2D acquisition, no azimuthal steering |
+
+## Subject Metadata
+
+Not applicable — no human or animal subjects. Each "subject" is a simulated
+phantom of ≈275,000 scatterers. The reconstructed image FOV is 45 mm (lateral) ×
+40 mm (axial), starting 10 mm from the transducer face. The scatterer field is
+deliberately larger than this FOV (≈49.5 mm wide, extending to 54 mm deep) so
+that beamforming at the FOV edges is fully supported by surrounding scatterers
+and free of edge-truncation artifacts; only the reconstruction grid is cropped
+to the FOV, while `data/raw_data` and `data/scatterers` retain the full extent.
+Per-class amplitude weighting inside the mask/weight-map region:
+
+- **Anechoic:** scatterer amplitude zeroed.
+- **Hypoechoic:** amplitude × U[0.07, 0.50]; the exact per-capture draw is
+  preserved in `data/echogenicity_multiplier`.
+- **Hyperechoic:** amplitude × U[2, 8]; same per-capture preservation via
+  `data/echogenicity_multiplier`.
+- **Diverse:** continuous grayscale weight map from a Wikimedia Commons
+  photograph (histogram-equalized, normalized to [0, 1], clipped at the
+  extremes) — preserved as `data/diverse_source_image`.
+- **Point-target:** 10–20 point targets per phantom (count ~ U[10, 20], rounded to an integer), each target's amplitude an independent U[15, 35] draw.  The targets consist of adjacent groups of high-amplitude scatterers within `data/scatterers`.
+
+All classes except point-target additionally include 2–5 extra bright point
+scatterers (amplitude 18–22) scattered at valid random positions, for realism.
+
+## Data Validation
+
+**Setup:** `reconstruct.py` and `visualize.py` require `zea>=0.1.1`, `matplotlib`,
+and `KERAS_BACKEND=jax` (or `torch`/`tensorflow`) set before import. This submission
+was built and verified against the [OpenH-RF repo](https://github.com/open-h/OpenH-RF)'s environment (`uv sync` inside a clone of that repo installs `zea` and every other
+dependency these scripts need — see that repo's README for the exact commands).
+
+`reconstruct.py` (+ `pipeline.yaml`, zea's default DAS pipeline: Cast → ApplyWindow →
+Demodulate → Beamform → EnvelopeDetect → Normalize → LogCompress; each stage is
+explained in `reconstruct.py`'s own module docstring) reconstructs a synthetic
+transmit aperture (STA) B-mode from `data/raw_data` using all 128 transmits, over
+the same field of view as the stored `data/image` reference. A second helper,
+`visualize.py <file>.hdf5`, renders the stored `data/image`, its class-specific
+label, and the `data/scatterers` cloud side by side. `examples/` contains, for a
+representative capture of each of the five classes, an `<id>_bmode.png` (the STA
+reference reconstruction from `reconstruct.py`, physical mm axes, titled with the
+sample ID) and an `<id>_panels.png` (the `visualize.py` output: stored `data/image`,
+class label — segmentation foreground for anechoic/hypoechoic/hyperechoic,
+`data/diverse_source_image` for diverse, none for point-target — and the
+`data/scatterers` cloud coloured by |amplitude|, on shared equal-aspect mm axes),
+confirming the label, reconstruction, and scatterer field are spatially registered.
+
+## Known Issues
+
+- **`data/image` is a log-compressed dB B-mode, not raw beamformed RF** —
+  This is a minor packaging note, not a coverage
+  gap, precisely because the data is FSA: `data/raw_data` retains every
+  element-to-element transmit/receive combination (128 single-element transmit
+  firings × 128-element receive aperture on the L11-5v), which is a more
+  general representation than any one fixed beamformed product could be: by
+  delay-and-sum with the appropriate per-element transmit delays and apodization
+  (linear superposition over the 128 single-element firings), it is a sufficient
+  basis to retrospectively synthesize other transmit/receive schemes — multi-angle
+  plane-wave compounding at arbitrary steering angles, diverging-wave imaging from
+  an arbitrary virtual source behind the array, conventional focused/multi-line
+  transmit — with no re-acquisition. `reconstruct.py` as shipped only implements
+  one of these (the full 128-transmit STA beamforming used to produce
+  `data/image`, via zea's default DAS pipeline, `pipeline.yaml`); it does not take
+  a scheme argument. Reconstructing a different scheme means modifying
+  `reconstruct.py` accordingly — supplying the corresponding transmit
+  delays/apodization to zea's `Beamform` op.
+
+## Ethical Considerations
+
+Entirely synthetic data. No human or animal subjects are involved, and no IRB/ethics
+approval is required or applicable. The "animal-class" Open Images V7 segmentation
+annotations and Wikimedia Commons photographs referenced elsewhere in this card
+contribute only geometric silhouette shapes and grayscale texture patterns,
+respectively — no live animal or human subject, tissue, or imagery of either is
+used anywhere in this dataset.

concordia/pipeline.yaml

@@ -0,0 +1,17 @@
+pipeline:
+    operations:
+    -   name: keras.ops.cast
+        params:
+            dtype: float32
+    - apply_window
+    - demodulate
+    -   name: beamform
+        params:
+            num_patches: 200
+    - envelope_detect
+    -   name: normalize
+        params:
+            output_range:
+            - 0.0
+            - 1.0
+    - log_compress