strasbourg-basel: sync data card and figures with GitHub

strasbourg-basel/README.md

@@ -1,5 +1,6 @@
 ---
-pretty_name: "OpenH-RF — BoneSRF Robot-Tracked Fractured-Femur Phantom Dataset"
+name: strasbourg-basel
+pretty_name: "BoneSRF Robot-Tracked Fractured-Femur Phantom Dataset"
 license: cc-by-4.0
 task_categories:
   - other
@@ -22,257 +23,116 @@ size_categories:
   - n<1K
 ---
 
-# BoneSRF — Robot-Tracked Fractured-Femur Phantom Channel Data
+# BoneSRF
 
-**BoneSRF** (Bone Surface Reflection) is an ultrasound channel-data dataset built
-around a simple question: can pre-beamformed RF data recovered from a handheld,
-point-of-care scanner support bone-surface / fracture-reflection research? It
-contributes three 3D-printed, fractured femur phantoms — scanned, beamformed-RF-
-inverted, and packaged in the OpenH-RF `zea` format — to the OpenH-RF initiative.
+<p align="center">
+  <img src="assets/reference_bmode.png" alt="BoneSRF B-mode of phantom 2, whole-bone sweep" width="200">
+</p>
 
-Nine scans in total: each of the three phantoms was swept three times (`distal`,
-`proximal`, `wholebone`), giving nine `zea` HDF5 files in [`data/`](data/). Every
-scan is **robot-tracked** — the probe was mounted on a robotic arm and its pose
-independently recorded — and every file also carries that phantom's **CT scan and
-multi-label segmentation** inside it.
+*B-mode of a fractured-femur phantom, frame 45 of [`data/phantom2_wholebone.hdf5`](https://huggingface.co/datasets/nvidia/OpenH-RF/blob/main/strasbourg-basel/BoneSRF/data/phantom2_wholebone.hdf5), beamformed from the recovered channel data.*
 
 ## Dataset Description
 
-The channel data in this dataset is **not a direct per-element sensor recording**.
-A Clarius handheld probe does not expose its raw per-element channel data, only its
-own internally beamformed RF output. Every `raw_data` array here is therefore a
-numerical estimate: the per-element channel data consistent with the probe's known
-per-scanline focused acquisition geometry (transmit delays, apodization, walking
-sub-aperture) that, if beamformed the same way, would reproduce the real Clarius
-output. This estimate is recovered by solving a conjugate-gradient least-squares
-(CGLS) inversion of a zea `DASOperator` built from that acquisition geometry,
-against the real beamformed phantom scans as the inversion target.
-
-This is phantom data — not simulated, clinical, or in-vivo — intended for
-full-matrix-capture-style beamforming and image-reconstruction research at a
-bone-tissue interface, and as a worked example of recovering pre-beamformed data
-from beamformed-only ultrasound exports.
+<p align="center">
+  <img src="assets/bonesrf_logo.png" alt="BoneSRF" width="200">
+</p>
 
-### How the data was generated
-
-1. **Real acquisition.** A Clarius handheld-probe-class linear array (L20HD3) was
-   used to scan each phantom, producing the probe's own beamformed RF output — this
-   is the *real*, physically acquired data, not simulated.
-2. **Inversion.** The beamformed RF is inverted back into pre-beamformed,
-   per-element channel data using
-   [`das-inverse` (`clarius` branch)](https://github.com/sankethvedula/das-inverse/tree/clarius)
-   — a CGLS solver over a `zea.inverse.DASOperator` forward model of the probe's
-   focused, walking-sub-aperture transmit sequence (`invert_clarius_beamformed.py`).
-   The result is what this dataset calls "simulated" channel data: not
-   sensor-captured, but numerically consistent with the real beamformed acquisition
-   it was inverted from.
-3. **Packaging.** The inverted channel data, transmit-sequence metadata, probe
-   geometry, per-frame probe pose, and the phantom's CT + segmentation are written
-   out as one `zea` HDF5 file per scan, matching the OpenH-RF format spec.
-
-### Probe tracking
-
-Every scan is **tracked**: the probe was mounted on a robotic arm, and its pose was
-independently recorded via a trakSTAR electromagnetic tracking system with a fixed
-fCal image-to-probe calibration. The tracking stream is packaged **per-frame
-indexed** inside each file's metadata (`metadata/probe_pose`: translation, rotation,
-timestamps) — `metadata/probe_pose[i]` corresponds directly to `raw_data[i]`, index
-for index. It does not need a separate sidecar file, and the raw tracking capture is
-not shipped — only the recovered, aligned pose stream.
-
-### The three phantoms
+**BoneSRF** (Bone Surface Reflection) is an ultrasound channel-data dataset for bone-surface and fracture-reflection research. It contains three 3D-printed, fractured femur phantoms, scanned with a handheld point-of-care probe, inverted back to pre-beamformed RF and packaged in the OpenH-RF `zea` format.
 
-Each phantom is a 3D-printed femur, modeled from a CC BY 4.0–licensed femur bone
-dataset, with a fracture pattern simulated differently for each of the three. Each
-printed femur is immersed in ultrasound-coupling gel and scanned by a robotic arm,
-which gives repeatable, controlled probe trajectories instead of a freehand scan.
-Each phantom is scanned three times, at three positions along the bone:
+Each phantom was swept three times (`distal`, `proximal`, `wholebone`), giving nine `zea` HDF5 files. Every scan is robot-tracked: the probe was mounted on a robotic arm and its pose recorded separately. Each file also carries that phantom's CT scan and multi-label segmentation.
 
-- **`distal`** — a sweep over the distal region of the femur
-- **`proximal`** — a sweep over the proximal region of the femur
-- **`wholebone`** — a sweep covering the full length of the femur
+The channel data here is not a direct per-element sensor recording. A Clarius handheld probe does not expose its raw per-element channel data, only its own internally beamformed RF output. Every `raw_data` array is therefore a numerical estimate: the per-element channel data consistent with the probe's known per-scanline focused acquisition geometry (transmit delays, apodization, walking sub-aperture) that would reproduce the real Clarius output if beamformed the same way. It is recovered by a conjugate-gradient least-squares (CGLS) inversion of a zea `DASOperator` built from that geometry, using the real beamformed phantom scans as the inversion target.
 
-**Fracture design:** `REQUIRES_CONTRIBUTOR` — the specific fracture pattern
-simulated in each phantom (location; type: transverse / oblique / comminuted /
-hairline; displacement) has not yet been documented. The CT segmentation carried in
-each file is ground truth for the physical phantom geometry in the meantime.
+This is phantom data, not simulated, clinical or in-vivo. It is intended for full-matrix-capture-style beamforming and image-reconstruction research at a bone-tissue interface, and as an example of recovering pre-beamformed data from beamformed-only ultrasound exports.
 
-## Folder structure
-
-```
-BoneSRF/
-├── README.md              ← this file (dataset overview + data card for all nine scans)
-├── LICENCE                (CC BY 4.0)
-├── pipeline.yaml          (saved zea.Pipeline — one pipeline, shared by all nine scans)
-├── reconstruct.py         (runs pipeline.yaml on any scan → reference_bmodes/<scan>.png)
-├── data/
-│   ├── phantom1_distal.hdf5      (zea channel data + per-frame probe pose + CT)
-│   ├── phantom1_proximal.hdf5
-│   ├── phantom1_wholebone.hdf5
-│   ├── phantom2_*.hdf5           (same three sweeps)
-│   └── phantom3_*.hdf5           (same three sweeps)
-└── reference_bmodes/
-    └── <scan>.png                (one reference reconstruction per scan)
-```
+### How the data was generated
 
-Every file in `data/` stands alone: it holds the (CGLS-recovered) pre-beamformed
-channel data, the full transmit-sequence and probe metadata needed to beamform it,
-the per-frame tracked probe pose, and the CT + segmentation of the phantom it
-depicts.
+1. **Acquisition.** Each phantom was scanned with a Clarius L20HD3 linear array, producing the probe's own beamformed RF output. This is the physically acquired data, not simulated.
+2. **Inversion.** The beamformed RF is inverted back into pre-beamformed, per-element channel data with [`das-inverse` (`clarius` branch)](https://github.com/sankethvedula/das-inverse/tree/clarius), a CGLS solver over a `zea.inverse.DASOperator` forward model of the probe's focused, walking-sub-aperture transmit sequence (`invert_clarius_beamformed.py`). The result is not sensor-captured, but numerically consistent with the real beamformed acquisition it was inverted from.
+3. **Packaging.** The inverted channel data, transmit-sequence metadata, probe geometry, per-frame probe pose, and the phantom's CT + segmentation are written out as one `zea` HDF5 file per scan, matching the OpenH-RF format spec.
 
-## Reconstructing a B-mode
+### Probe tracking
 
-`reconstruct.py` loads the saved `zea.Pipeline` from `pipeline.yaml` and runs it on
-one frame of one scan:
+The probe was mounted on a robotic arm and its pose recorded separately by a trakSTAR electromagnetic tracking system with a fixed fCal image-to-probe calibration. The pose stream is stored per frame in each file's metadata (`metadata/probe_pose`: translation, rotation, timestamps), so `metadata/probe_pose[i]` corresponds to `raw_data[i]`. The raw tracking capture is not shipped, only the recovered, aligned pose stream.
 
-```bash
-python reconstruct.py                          # all nine, at their reference frames
-python reconstruct.py phantom1_distal          # one scan, at its reference frame
-python reconstruct.py phantom1_distal --frame 40 --device cpu
-```
+### The three phantoms
 
-The pipeline is `cast` → `apply_window` → `beamform` (delay-and-sum with a
-physically-motivated per-transmit `pfield` weighting, since this is a per-scanline
-focused walking-sub-aperture acquisition rather than full synthetic aperture) →
-`keras.ops.abs` → axial-only Gaussian blur → `normalize` → `log_compress`, with
-display parameters (dynamic range, p-field settings) also read from `pipeline.yaml`.
-Acquisition geometry comes from each file's own `scan`/`probe` groups.
+Each phantom is a 3D-printed femur, modeled from a CC BY 4.0–licensed femur bone dataset, with a different simulated fracture pattern. Each printed femur is immersed in ultrasound-coupling gel and scanned by a robotic arm, which gives repeatable probe trajectories rather than a freehand scan. Each phantom is scanned three times, at three positions along the bone:
 
-This intentionally avoids `zea.inverse`: that module is for the CGLS inversion that
-*produced* these files, not for reconstructing from them. Runtime is ~30 s per frame
-on CPU. Each file is a full multi-frame sweep of 20.24 GB to 24.49 GB, but only the requested
-frame is read.
+- **`distal`**: a sweep over the distal region of the femur
+- **`proximal`**: a sweep over the proximal region of the femur
+- **`wholebone`**: a sweep covering the full length of the femur
 
 ## Dataset Contributor(s)
 
-Sidaty El Hadramy, Philippe C. Cattin, Juan Verde — IHU Strasbourg and the
-Department of Biomedical Engineering, University of Basel.
+- Sidaty El Hadramy (IHU Strasbourg; Department of Biomedical Engineering, University of Basel)
+- Philippe C. Cattin (IHU Strasbourg; Department of Biomedical Engineering, University of Basel)
+- Juan Verde (IHU Strasbourg; Department of Biomedical Engineering, University of Basel)
 
 ## Dataset Creation Date
 
-Original Clarius acquisitions: 19/06/2026 (`phantom1_distal`) and 25/06/2026
-(`phantom1_proximal`, acquisition ID `20260625-ihu-04_BoneSRF-01_proximal_robot_r3`).
-The acquisition date of the other seven sweeps was not separately recorded — see
-[Known Issues](#known-issues). Converted to `zea` format 21/07/2026–05/08/2026 and
-re-converted 13/08/2026–14/08/2026 to align probe tracking to `raw_data`
-frame-by-frame. CT and segmentation embedded into the files 09/09/2026.
+Clarius acquisitions: 19/06/2026 (`phantom1_distal`) and 25/06/2026 (`phantom1_proximal`). The other seven sweeps carry no recorded acquisition date, see [Known Issues](#known-issues). Converted to the `zea` format in 2026.
 
 ## License / Terms of Use
 
-CC BY 4.0 — see [`LICENCE`](LICENCE). Confirmed by the contributor as cleared for
-CC BY 4.0 release (phantom data; no patient consent or third-party IP encumbrance
-applies). The CT and segmentation data carried inside the files is released under
-the same terms. The femur geometry underlying the 3D-printed phantoms is itself
-sourced from a CC BY 4.0–licensed bone model dataset.
+[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
 
-Full-matrix-capture-style beamforming research on recovered (not directly sensed)
-channel data at a bone-tissue interface: delay-and-sum reconstruction, adaptive or
-aberration-correction beamforming benchmarking, robot/EM-tracked probe-pose fusion
-research, and as a reference example for recovering pre-beamformed data from
-beamformed-only ultrasound exports (e.g. other handheld/point-of-care scanners with
-the same limitation).
+Full-matrix-capture-style beamforming research on recovered (not directly sensed) channel data at a bone-tissue interface: delay-and-sum reconstruction, adaptive and aberration-correction beamforming benchmarks, and robot/EM-tracked probe-pose fusion. It also serves as a reference for recovering pre-beamformed data from other beamformed-only scanners.
 
 ## Dataset Characterization
 
-- **Data Collection Method:** phantom (3D-printed, bone-mimicking femur, immersed in
-  ultrasound-coupling gel); scanned with a Clarius handheld-probe-class linear array
-  mounted on and moved by a robotic arm; raw channel data recovered via CGLS
-  inversion of the probe's beamformed RF output (not a direct per-element
-  recording); probe pose independently tracked via a trakSTAR EM tracking system,
-  fCal-calibrated.
-- **Labeling Method:** a CT scan of each 3D-printed phantom and a multi-label
-  segmentation of it (authored in 3D Slicer) are carried **inside each of that
-  phantom's three files**, under `custom/ct/` and `custom/ct_segmentation/` — see
-  [CT reference imaging](#ct-reference-imaging). There are no annotations on the RF
-  data itself.
-- **Acquisition system:** Clarius L20HD3, 192-element linear array, 0.130 mm pitch
-  (24.8 mm aperture), 10 MHz center frequency, 30 MHz sampling frequency, 1540 m/s
-  sound speed, ~5.1 cm imaging depth (1984–2016 axial samples depending on scan),
-  single fixed transmit focus at 25.3–25.8 mm (verified: `focus_distances` is
-  constant across all 192 transmits within each scan), 192 focused transmits per
-  frame (one per lateral scanline, no steering), walking sub-aperture per scanline —
-  Hanning-windowed, 47–97 of 192 elements active per transmit (mean ~84, i.e.
-  roughly a quarter to a half of the array, narrowest at the array edges).
+- **Data Collection Method:** phantom (3D-printed, bone-mimicking femur, immersed in ultrasound-coupling gel); scanned with a Clarius handheld-probe-class linear array mounted on and moved by a robotic arm; raw channel data recovered via CGLS inversion of the probe's beamformed RF output (not a direct per-element recording); probe pose independently tracked via a trakSTAR EM tracking system, fCal-calibrated.
+- **Labeling Method:** a CT scan of each 3D-printed phantom and a multi-label segmentation of it (authored in 3D Slicer) are stored inside each of that phantom's three files, under `custom/ct/` and `custom/ct_segmentation/`. See [CT reference imaging](#ct-reference-imaging). There are no annotations on the RF data itself.
+- **Acquisition system:** Clarius L20HD3, 192-element linear array, 0.130 mm pitch (24.8 mm aperture), 10 MHz center frequency, 30 MHz sampling frequency, 1540 m/s sound speed, ~5.1 cm imaging depth (1984–2016 axial samples depending on scan), single fixed transmit focus at 25.3–25.8 mm (`focus_distances` is constant across all 192 transmits within a scan), 192 focused transmits per frame (one per lateral scanline, no steering), Hanning-windowed walking sub-aperture per scanline with 47 to 97 of 192 elements active per transmit (mean 84, narrowest at the array edges).
 
-## CT reference imaging
+The CT and segmentation data inside the files are released under the same CC BY 4.0 terms as the channel data. The femur geometry behind the 3D-printed phantoms comes from a CC BY 4.0–licensed bone model dataset.
 
-Each phantom's CT scan and its multi-label 3D Slicer segmentation are carried inside
-**every one** of that phantom's three `zea` files, under `custom/ct/` and
-`custom/ct_segmentation/`. They are not shipped as separate `.nrrd` sidecars, so no
-file depends on another.
+## Processing the Dataset
 
-| Dataset | Contents |
-|---|---|
-| `custom/ct/volume` | CT volume, `int16`, stored `(k, j, i)` (slice, row, column) |
-| `custom/ct/spacing`, `origin`, `direction`, `affine` | Grid geometry in SI metres; `affine` maps voxel index `(i, j, k, 1)` to an LPS position |
-| `custom/ct/nrrd_header` | Verbatim header of the source NRRD (distances in millimetres) |
-| `custom/ct_segmentation/labelmap` | Layered binary labelmap on the same grid, `uint8` |
-| `custom/ct_segmentation/segment_*` | Per-segment name, label value, layer, colour, bounding box and 3D Slicer ID |
+The acquisitions can be processed with the `reconstruct.py` [script](https://github.com/open-h/OpenH-RF/blob/main/datasets/strasbourg-basel/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.
 
-Grid geometry differs per phantom:
+The script loads the saved `zea.Pipeline` from `pipeline.yaml`, runs it on one frame of one scan, and writes `assets/<scan>.png`.
 
-| Phantom | CT grid | Stored array | Source spacing (mm) |
-|---|---|---|---|
-| phantom1 | `512 × 512 × 594` | `(594, 512, 512)` | `0.546875 × 0.546875 × 0.6` |
-| phantom2 | `512 × 512 × 574` | `(574, 512, 512)` | `0.50390625 × 0.50390625 × 0.6` |
-| phantom3 | `512 × 512 × 594` | `(594, 512, 512)` | `0.5625 × 0.5625 × 0.6` |
+The constants at the top of the script select what is reconstructed:
 
-Each segmentation has three segments, corresponding directly to the three RF sweeps
-of that phantom. **Label values repeat across layers** — 3D Slicer keeps segments on
-separate internal labelmap layers — so read a segment's mask as
-`labelmap[..., segment_layers[s]] == segment_label_values[s]` rather than treating
-the array as one flat labelmap:
+- `SCAN`: the scan to reconstruct, as a local path or an `hf://` URI.
+- `FRAME`: the frame to beamform. `None` uses that scan's reference frame, the one its reference image was rendered from.
+- `DEVICE`: where to run, e.g. `"cpu"`, `"cuda:0"` or `"auto:1"`.
+- `CT`: also plot the CT carried in the file, to `assets/ct_<scan>.png`.
 
-| Segment | `segment_label_values` | `segment_layers` | Corresponds to |
-|---|---|---|---|
-| `BoneSRF-1_Proximal` | 1 | 0 | `phantom1_proximal` |
-| `BoneSRF-1_Distal` | 2 | 0 | `phantom1_distal` |
-| `BoneSRF-1_Complete` | 1 | 1 | `phantom1_wholebone` |
-| `BoneSRF-2_Complete` | 1 | 0 | `phantom2_wholebone` |
-| `BoneSRF-2_Distal` | 1 | 1 | `phantom2_distal` |
-| `BoneSRF-2_Proximal` | 2 | 1 | `phantom2_proximal` |
-| `BoneSRF-3_Proximal` | 2 | 0 | `phantom3_proximal` |
-| `BoneSRF-3_Distal` | 3 | 0 | `phantom3_distal` |
-| `BoneSRF-3_Complete` | 1 | 1 | `phantom3_wholebone` |
+These values reconstruct `phantom2_wholebone` at its reference frame (45) on the CPU:
+
+```python
+SCAN = "hf://nvidia/OpenH-RF/strasbourg-basel/BoneSRF/data/phantom2_wholebone.hdf5"
+FRAME = None
+DEVICE = "cpu"
+```
+
+and produce the B-mode image shown at the top of this card.
+
+The pipeline is `cast` → `apply_window` → `beamform` (delay-and-sum with a per-transmit `pfield` weighting, since this is a per-scanline focused walking-sub-aperture acquisition rather than full synthetic aperture) → `keras.ops.abs` → axial-only Gaussian blur → `normalize` → `log_compress`. Display parameters (dynamic range, p-field settings) also come from `pipeline.yaml`; acquisition geometry comes from each file's own `scan` and `probe` groups.
 
-The CT is reference imaging of the physical phantom in scanner (LPS) space. It is
-**not spatially registered** to the RF frames or to the tracked probe poses; no
-CT↔ultrasound registration is provided with this submission.
+The reconstruction does not use `zea.inverse`. That module is for the CGLS inversion that produced these files, not for reading them back. Runtime is about 30 s per frame on CPU. Each file is a full sweep of 20 to 25 GB, but only the requested frame is read.
 
 ## Dataset Format
 
-Submitted in the [`zea` file format](https://zea.readthedocs.io/en/openh-rf-latest/)
-as nine HDF5 files in [`data/`](data/), blosc-compressed.
+[zea v0.1.6](https://github.com/tue-bmd/zea)
 
-The channel data was recovered from the probe's real, beamformed RF output by
-CGLS-inverting a `zea.inverse.DASOperator` built from the known acquisition
-geometry. `t0_delays`, `tx_apodizations`, `focus_distances`, `transmit_origins`,
-`polar_angles`, and `waveforms_two_way` are copied directly from the values that
-inversion's DAS operator was built with — not re-derived or guessed. The source
-`.npz` had no explicit `demodulation_frequency`; it was substituted with
-`center_frequency` per the standard convention for RF (non-IQ) sources (verified:
-`demodulation_frequency` = `center_frequency` = 10 MHz in every file).
+Submitted in the [`zea` file format](https://zea.readthedocs.io/en/openh-rf-latest/) as nine HDF5 files.
 
-Probe pose (`metadata/probe_pose`: translation, rotation, timestamps) was recovered
-from the trakSTAR tracking capture via a fixed fCal image-to-probe calibration,
-resampled onto each `raw_data` frame's own acquisition time before conversion, and
-packaged as a **per-frame indexed signal** (`metadata/probe_pose[i]` ↔
-`raw_data[i]`).
+The channel data was recovered from the probe's real, beamformed RF output by CGLS-inverting a `zea.inverse.DASOperator` built from the known acquisition geometry. `t0_delays`, `tx_apodizations`, `focus_distances`, `transmit_origins`, `polar_angles`, and `waveforms_two_way` are copied directly from the values that inversion's DAS operator was built with, not re-derived. The source `.npz` had no explicit `demodulation_frequency`; it was substituted with `center_frequency` per the standard convention for RF (non-IQ) sources.
 
-CT and segmentation (an addition beyond the original proposal) were copied verbatim
-out of the `.nrrd` files that previously shipped alongside the RF data, so that
-every file is self-contained; the verbatim source NRRD headers are preserved with
-them. Grid geometry is converted from the NRRD's millimetres to zea's SI metres.
+Probe pose (`metadata/probe_pose`: translation, rotation, timestamps) was recovered from the trakSTAR tracking capture via a fixed fCal image-to-probe calibration, resampled onto each `raw_data` frame's own acquisition time before conversion, and stored per frame, so `metadata/probe_pose[i]` corresponds to `raw_data[i]`.
 
-**Reading these files requires `h5py` built against HDF5 ≥ 2.0** (e.g. `h5py` ≥
-3.16) — see [Known Issues](#known-issues).
+CT and segmentation were copied verbatim out of the `.nrrd` files that previously shipped alongside the RF data, so that every file is self-contained; the source NRRD headers are preserved with them. Grid geometry is converted from the NRRD's millimetres to zea's SI metres.
+
+Reading these files requires `h5py` built against HDF5 ≥ 2.0 (e.g. `h5py` ≥ 3.16), see [Known Issues](#known-issues).
 
 ### Fields
 
-Every file has the same field structure; `n_frames` and `n_ax` vary per scan (see
-[the nine scans](#the-nine-scans)).
+Every file has the same field structure; `n_frames` and `n_ax` vary per scan (see [Dataset Quantification](#dataset-quantification)).
 
 | Field | Shape | dtype | Units | Description |
 |---|---|---|---|---|
@@ -299,48 +159,102 @@ Every file has the same field structure; `n_frames` and `n_ax` vary per scan (se
 | `custom.ct_segmentation.segment_label_values` | `(3,)` | uint8 | — | Label value of each segment within its own layer |
 | `custom.ct_segmentation.segment_layers` | `(3,)` | uint8 | — | Index into the last axis of `labelmap` holding each segment |
 
+## Folder structure
+
+```
+BoneSRF/
+├── README.md              ← this file (dataset overview + data card for all nine scans)
+├── LICENCE                (CC BY 4.0)
+├── pipeline.yaml          (saved zea.Pipeline, shared by all nine scans)
+├── reconstruct.py         (runs pipeline.yaml on any scan, and plots its CT)
+├── assets/
+│   ├── reference_bmode.png       (the B-mode shown below)
+│   └── ct_<scan>.png             (CT slices of the scan's phantom)
+├── data/
+│   ├── phantom1_distal.hdf5      (zea channel data + per-frame probe pose + CT)
+│   ├── phantom1_proximal.hdf5
+│   ├── phantom1_wholebone.hdf5
+│   ├── phantom2_*.hdf5           (same three sweeps)
+│   └── phantom3_*.hdf5           (same three sweeps)
+└── reference_bmodes/
+    └── <scan>.png                (one reference reconstruction per scan)
+```
+
+Every file in `data/` is self-contained: it holds the CGLS-recovered pre-beamformed channel data, the transmit-sequence and probe metadata needed to beamform it, the per-frame probe pose, and the CT and segmentation of the phantom it shows.
+
+## CT reference imaging
+
+Each phantom's CT scan and its multi-label 3D Slicer segmentation are stored inside all three of that phantom's `zea` files, under `custom/ct/` and `custom/ct_segmentation/`. They are not shipped as separate `.nrrd` sidecars, so no file depends on another.
+
+<p align="center">
+  <img src="assets/ct_phantom2_wholebone.png" alt="CT slices of phantom2" width="800">
+</p>
+
+Three slices of phantom2's CT, written by `reconstruct.py` with `CT = True`, with the `BoneSRF-2_Complete` segment outlined in red. The printed femur is hollow, so it reads dark against the bright coupling gel, and the coronal view shows the fracture: the bone is in separate, displaced pieces.
+
+| Dataset | Contents |
+|---|---|
+| `custom/ct/volume` | CT volume, `int16`, stored `(k, j, i)` (slice, row, column) |
+| `custom/ct/spacing`, `origin`, `direction`, `affine` | Grid geometry in SI metres; `affine` maps voxel index `(i, j, k, 1)` to an LPS position |
+| `custom/ct/nrrd_header` | Verbatim header of the source NRRD (distances in millimetres) |
+| `custom/ct_segmentation/labelmap` | Layered binary labelmap on the same grid, `uint8` |
+| `custom/ct_segmentation/segment_*` | Per-segment name, label value, layer, colour, bounding box and 3D Slicer ID |
+
+Grid geometry differs per phantom:
+
+| Phantom | CT grid | Stored array | Source spacing (mm) |
+|---|---|---|---|
+| phantom1 | `512 × 512 × 594` | `(594, 512, 512)` | `0.546875 × 0.546875 × 0.6` |
+| phantom2 | `512 × 512 × 574` | `(574, 512, 512)` | `0.50390625 × 0.50390625 × 0.6` |
+| phantom3 | `512 × 512 × 594` | `(594, 512, 512)` | `0.5625 × 0.5625 × 0.6` |
+
+Each segmentation has three segments, one per RF sweep of that phantom. Label values repeat across layers, because 3D Slicer keeps segments on separate internal labelmap layers, so read a segment's mask as `labelmap[..., segment_layers[s]] == segment_label_values[s]` rather than treating the array as one flat labelmap:
+
+| Segment | `segment_label_values` | `segment_layers` | Corresponds to |
+|---|---|---|---|
+| `BoneSRF-1_Proximal` | 1 | 0 | `phantom1_proximal` |
+| `BoneSRF-1_Distal` | 2 | 0 | `phantom1_distal` |
+| `BoneSRF-1_Complete` | 1 | 1 | `phantom1_wholebone` |
+| `BoneSRF-2_Complete` | 1 | 0 | `phantom2_wholebone` |
+| `BoneSRF-2_Distal` | 1 | 1 | `phantom2_distal` |
+| `BoneSRF-2_Proximal` | 2 | 1 | `phantom2_proximal` |
+| `BoneSRF-3_Proximal` | 2 | 0 | `phantom3_proximal` |
+| `BoneSRF-3_Distal` | 3 | 0 | `phantom3_distal` |
+| `BoneSRF-3_Complete` | 1 | 1 | `phantom3_wholebone` |
+
+The CT is reference imaging of the physical phantom in scanner (LPS) space. It is not spatially registered to the RF frames or to the tracked probe poses; no CT-to-ultrasound registration is provided.
+
 ## Dataset Quantification
 
 **Current OpenH-RF release:** 9 HDF5 files; 200.75 GB (200,745,025,536 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.
 
-Nine acquisitions, one continuous sweep each; 1428 frames in total, 200.75 GB of stored HDF5 data. No train / val / test split (each file is a single reference acquisition).
-Every value below was read back from the files themselves.
-
-| Scan | Frames | Poses | `n_ax` | Focus | Reference frame | Size on disk |
-|---|---|---|---|---|---|---|
-| `phantom1_distal` | 144 | 144 | 2016 | 25.70 mm | 130 | 20,534,067,200 B (20.53 GB) |
-| `phantom1_proximal` | 174 | 174 | 2000 | 25.55 mm | 100 | 24,492,900,352 B (24.49 GB) |
-| `phantom1_wholebone` | 158 | 158 | 1984 | 25.30 mm | 150 | 21,930,508,288 B (21.93 GB) |
-| `phantom2_distal` | 162 | 162 | 2000 | 25.65 mm | 40 | 22,799,908,864 B (22.80 GB) |
-| `phantom2_proximal` | 142 | 142 | 2016 | 25.75 mm | 40 | 20,236,337,152 B (20.24 GB) |
-| `phantom2_wholebone` | 155 | 155 | 2016 | 25.80 mm | 45 | 21,953,183,744 B (21.95 GB) |
-| `phantom3_distal` | 166 | 166 | 1984 | 25.30 mm | 100 | 23,042,916,352 B (23.04 GB) |
-| `phantom3_proximal` | 167 | 167 | 2016 | 25.80 mm | 0 | 23,672,127,488 B (23.67 GB) |
-| `phantom3_wholebone` | 159 | 159 | 1984 | 25.35 mm | 100 | 22,083,076,096 B (22.08 GB) |
+Nine acquisitions, one continuous sweep each; 1,427 frames in total. No train / val / test split (each file is a single reference acquisition). Every scan has one tracked probe pose per frame.
+
+| Scan | Frames | `n_ax` | Focus | Reference frame | Size on disk |
+|---|---|---|---|---|---|
+| `phantom1_distal` | 144 | 2016 | 25.70 mm | 130 | 20.53 GB |
+| `phantom1_proximal` | 174 | 2000 | 25.55 mm | 100 | 24.49 GB |
+| `phantom1_wholebone` | 158 | 1984 | 25.30 mm | 150 | 21.93 GB |
+| `phantom2_distal` | 162 | 2000 | 25.65 mm | 40 | 22.80 GB |
+| `phantom2_proximal` | 142 | 2016 | 25.75 mm | 40 | 20.24 GB |
+| `phantom2_wholebone` | 155 | 2016 | 25.80 mm | 45 | 21.95 GB |
+| `phantom3_distal` | 166 | 1984 | 25.30 mm | 100 | 23.04 GB |
+| `phantom3_proximal` | 167 | 2016 | 25.80 mm | 0 | 23.67 GB |
+| `phantom3_wholebone` | 159 | 1984 | 25.35 mm | 100 | 22.08 GB |
 
 ## Subject Metadata
 
-3D-printed, bone-mimicking musculoskeletal phantoms ("BoneSRF" — bone surface
-reflection targets); no human or animal subject. Scanned with a robot-mounted
-Clarius L20HD3 linear array at 10 MHz / ~5.1 cm depth / single transmit focus.
+3D-printed, bone-mimicking musculoskeletal phantoms (bone surface reflection targets); no human or animal subject. Scanned with a robot-mounted Clarius L20HD3 linear array at 10 MHz / ~5.1 cm depth / single transmit focus.
 
 ## Data Validation
 
-Every file was validated against the installed `zea` data spec — `File.validate()`
-(structural) and `File.validate_spec()` (full dtype / shape / dimension consistency)
-— and all nine report compliant, with `/data/raw_data` present and non-empty.
-`reconstruct.py` runs end-to-end on every scan and is deterministic across runs; the
-images in [`reference_bmodes/`](reference_bmodes/) are its output.
+All nine files pass the `zea` data spec, both `File.validate()` (structural) and `File.validate_spec()` (dtype, shape and dimension consistency). `reconstruct.py` runs end-to-end on every scan; the images in `reference_bmodes/` are its output.
 
 ## Known Issues
-- **Acquisition dates are incomplete.** Only `phantom1_distal` (19/06/2026) and
-  `phantom1_proximal` (25/06/2026) have a recorded original Clarius acquisition date;
-  the other seven sweeps do not carry one in the file or in the source capture.
-- **No CT↔ultrasound registration.** The CT lives in scanner LPS space and the probe
-  poses in trakSTAR tracker space. Nothing in this submission relates the two.
-- **CT intensity units are unverified.** The source NRRD headers record no unit. The
-  value range (−1024 … ~500) is consistent with Hounsfield units, but this has not
-  been confirmed by the contributor.
+- **Fracture patterns are not documented per phantom.** The location, type (transverse / oblique / comminuted / hairline) and displacement of each phantom's fracture are not recorded. The CT segmentation in each file is ground truth for the physical phantom geometry.
+- **Acquisition dates are incomplete.** Only `phantom1_distal` (19/06/2026) and `phantom1_proximal` (25/06/2026) have a recorded original Clarius acquisition date; the other seven sweeps do not carry one in the file or in the source capture.
+- **No CT-to-ultrasound registration.** The CT is in scanner LPS space and the probe poses in trakSTAR tracker space. Nothing here relates the two.
+- **CT intensity units are unverified.** The source NRRD headers record no unit. The value range (−1024 to about 500) is consistent with Hounsfield units, but this has not been confirmed.
 
 ## Ethical Considerations