README.md

---
pretty_name: SSL Ultrasound Representation (Stanford RF MultiFocal)
license: other
task_categories:
- feature-extraction
tags:
- ultrasound
- self-supervised
- rf-data
- mae
- jepa
size_categories:
- n<1K
---

# SSL Ultrasound Representation Dataset

IQ-demodulated multi-focal ultrasound channel data from the Stanford Ultrasound RF
Channel dataset, aggregated for self-supervised pre-training (MAE / JEPA /
contrastive). The signals are **not** beamformed — each frame retains its
per-transducer channel data after sub-aperture combination and IQ demodulation.

## Files

| File | Description |
|------|-------------|
| `ultrasonic_dataset.zarr/` | 6D float32 zarr array, Blosc-LZ4 compressed, stored as an unzipped directory tree (one file per chunk, ~28.6 GB total). One chunk per frame on the leading axis → consumers can fetch just the first N frames. |
| `ultrasonic_dataset.metadata.json` | Per-frame provenance (subject, file, subfolder, indices). |
| `README.md` | This card (auto-generated by `push_to_hf.py`). |

## Array layout

- **Name:** `rf_iq`
- **Shape:** `(168, 192, 192, 3, 2, 215)`
- **Axes:** `['frame', 'transducer', 'scanline', 'focal_zone', 'iq', 'sample']`
- **Chunks:** `(1, 192, 192, 3, 2, 215)` (one frame per chunk on the leading axis — aligned with DataLoader workers).
- **Dtype:** `float32`
- **IQ index:** `0` = in-phase, `1` = quadrature.

## Processing pipeline

1. Source: Verasonics Vantage `.mat` files with `MultiFocal` in the filename
   (each file holds a 2-frame cine via `Resource.RcvBuffer.numFrames = 2`).
2. Sub-aperture combination + IQ demodulation via `pymust.rf2iq`.
3. Fast-time axis truncated at **860 samples**
   (everything beyond is noise/zeros), then decimated by
   **4×** → **215** output samples per signal.
4. Stacked into a single 6D array with chunks optimised for 3D MAE patching.

See `scripts/dataset_scripts/build_mae_dataset.py` in the source repo for the
exact pipeline; `scripts/dataset_scripts/visualize_beamformed_pymust.py` beamforms
the same channel data for *visualisation only* (the stored array stays unbeamformed).

## Frame counts

By subject:

| Subject | Frames |
|---------|--------|
| `Rat1` | 12 |
| `Rat10` | 6 |
| `Rat11` | 6 |
| `Rat12` | 6 |
| `Rat13` | 6 |
| `Rat14` | 6 |
| `Rat15` | 6 |
| `Rat16` | 6 |
| `Rat18` | 6 |
| `Rat19` | 6 |
| `Rat2` | 12 |
| `Rat20` | 6 |
| `Rat3` | 12 |
| `Rat4` | 12 |
| `Rat5` | 12 |
| `Rat6` | 12 |
| `Rat7` | 12 |
| `Rat8` | 12 |
| `Rat9` | 6 |
| `VerasonicsAcq` | 6 |

By subject × subfolder:

| Subject | Subfolder | Frames |
|---------|-----------|--------|
| `Rat1` | `—` | 6 |
| `Rat1` | `ExposedLiver` | 6 |
| `Rat10` | `—` | 6 |
| `Rat11` | `—` | 6 |
| `Rat12` | `—` | 6 |
| `Rat13` | `—` | 6 |
| `Rat14` | `—` | 6 |
| `Rat15` | `—` | 6 |
| `Rat16` | `—` | 6 |
| `Rat18` | `—` | 6 |
| `Rat19` | `—` | 6 |
| `Rat2` | `—` | 6 |
| `Rat2` | `ExposedLiver` | 6 |
| `Rat20` | `—` | 6 |
| `Rat3` | `—` | 6 |
| `Rat3` | `ExposedLiver` | 6 |
| `Rat4` | `—` | 6 |
| `Rat4` | `ExposedLiver` | 6 |
| `Rat5` | `—` | 6 |
| `Rat5` | `ExposedLiver` | 6 |
| `Rat6` | `—` | 6 |
| `Rat6` | `ExposedLiver` | 6 |
| `Rat7` | `—` | 6 |
| `Rat7` | `ExposedLiver` | 6 |
| `Rat8` | `—` | 6 |
| `Rat8` | `ExposedLiver` | 6 |
| `Rat9` | `—` | 6 |
| `VerasonicsAcq` | `Rat1` | 6 |

## Loading

```python
from src.dataset import build_split_datasets

splits, meta = build_split_datasets(
    "benbarkow/us-ssl-mf",
    split_strategy="file",  # or "subject"
    cache_dir="/tmp/hf_cache",
    max_frames=16,          # optional: fetch only the first 16 frames' chunks
)
train_ds = splits["train"]
```

`max_frames` exploits the per-frame chunking: only the chunks for the first N
frames are downloaded, so you can iterate on a small box without pulling the whole
array. Omit it to fetch everything.

The split helpers (`split_by_subject`, `split_by_file`) avoid cross-split leakage:
both frames of the same 2-frame cine always end up in the same split, and (for
`split_by_subject`) no subject appears in two splits.