README.md

---
pretty_name: AngioFlow GHD WSS Graphs
tags:
  - graph
  - mesh
  - cfd
  - wall-shear-stress
  - aneurysm
  - medical
---

# AngioFlow GHD WSS Graphs

Per-case graphs of CFD wall shear stress (WSS) on GHD-parameterized cerebral aneurysm
surfaces, for training a graph transformer (`AngioWSSGraphGPS`) to predict WSS time
series from geometry alone. One graph per `(source, case, frame)`; node count varies
case to case (no shared/fixed topology, no common mesh).

This archive (`ghd_graphs.zip`) is a full unzip-in-place copy of the shared local
dataset root `angioflow_ghd_graphs/` — unzip it and the three subfolders below land
next to each other exactly as shown.

## Contents

| subfolder | internal tag | cases | geometry source | CFD source | real vs synthetic |
|---|---|---|---|---|---|
| `AneuGv1/` | `v1` | 634 | legacy canonical GHD reconstruction, ICP'd onto the CFD wall | AngioFlow v1 transient CFD | real patient |
| `Angioflow/` | `v2` | 251 (152 `AneuX` + 99 `ImperialNHS`, see `extraction_manifest.csv`'s `dataset` column) | per-case GHD fit placed into the CFD frame | merged v2 CFD | real patient (both sub-cohorts) |
| `AneuGv2/` | `aneugv2` | 141 | AneuG-*generated* shapes (`dome.obj`), their own CFD runs | `whding123/AneuGv2` on the Hub | synthetic |

Every case currently has **80 WSS frames** (one full cardiac-cycle time series) and
**`aneurysm_type == 0`** only (bifurcated/sidewall type-1 CFD data doesn't exist yet
upstream — the extraction scripts skip and delete any type-1 output).

Total: 1,026 cases x 80 frames = 82,080 `(source, case, frame)` samples, ~18 GB
uncompressed.

## Per-case files

```
<subfolder>/static_<version>_<case>.pt        # one per case -- geometry + static fields
<subfolder>/wss_<version>_<case>_<t>.pt       # one per (case, frame) -- [N,3] WSS target, Pa
<subfolder>/extraction_manifest.csv           # per-case QC row, one file per subfolder
<subfolder>/sanity/*.png                      # pred/GT render samples (training-time renders, not part of the data itself)
```

`static_<version>_<case>.pt` is a dict:

| field | shape | notes |
|---|---|---|
| `pos` | `[N,3]` | GHD-surface vertex coordinates, metres |
| `faces` | `[F,3]` | triangle connectivity |
| `edge_index` | `[2,E]` | **not guaranteed symmetric as stored** — symmetrize before use (see Loading, below) |
| `normals` | `[N,3]` | stored value is a QC artifact only — recompute at load time instead (see Loading) |
| `geo_dist` | `[N,3]` | geodesic distance to each of the 3 openings (type 0), normalized to `[0,1]` per column |
| `cot_eigval` | `[k]` (k=16) | cotangent-Laplacian eigenvalues, **ascending by value, not frequency** — for a connected mesh the smoothest mode is near the END of this array |
| `cot_eigvec` | `[N,k]` | matching eigenvectors (positional encoding) |
| `ghd_eigvec` | `[N,n_ghd]` (n_ghd=8) | canonical GHD basis carried onto this mesh, unit-norm per column |
| `flow_split` | `[2]` | outlet flow ratios |
| `aneurysm_type` | scalar | `0` for every case currently present |
| `node_type` | `[N]` | **convention differs by source**: `AneuGv1`/`Angioflow` use `0=wall,1=focused,2=inlet,3=outlet_0,4=outlet_1`; `AneuGv2` uses `0=wall`, then `1..n` by that case's own `wall_zones` order (no "focused" class) |
| `knn_dist_mm` | `[N]` | distance (mm) from each node to the nearest native CFD wall node — QC signal for extraction fidelity |
| `meta` | dict | the same row as this case's line in `extraction_manifest.csv` |

`wss_<version>_<case>_<t>.pt` is a bare `[N,3]` tensor, Pa, one file per frame `t` (0-79).

## Loading

The reference loader is `ood_fix/dataset.py::AngioWSSGraphDataset` in the
`AneuG-Net` repo (not included in this archive). Key behavior to replicate if
loading independently:

```python
from ood_fix.dataset import AngioWSSGraphDataset

ds = AngioWSSGraphDataset(root="<path to this unzipped folder>",
                           versions=None,           # None = all three sources
                           frames="all",            # or e.g. [0] for frame-0 only
                           aneurysm_types=(0,))
data = ds[0]   # torch_geometric.data.Data: pos, normals, geo_dist, cot_eigvec,
               # cot_eigval, ghd_eigvec, edge_index, y (WSS), t (frame idx),
               # aneurysm_type, version, case
```

- **Normals are recomputed at load time** (`pytorch3d.structures.Meshes.verts_normals_packed()`),
  not read from the stored `normals` field — the stored value came from whichever
  renderer wrote it (pyvista for v1/v2, plain numpy for aneugv2) and is not uniform
  across sources.
- **`edge_index` must be symmetrized** before use (`SymmetrizeEdges` in the reference
  pipeline) — the raw stored value is directed/one-sided.
- Apply any positional/value normalization exactly once — a transform chain applied
  both inside a custom `Dataset.get()` and again by PyG's own `__getitem__` will
  double-apply (this bit the reference pipeline once; its `dataset.py` fix is to let
  `get()` return the raw `Data` and apply transforms only through PyG's wrapper).

## Train/test split (as used by the reference training script, `ood_fix/train_wss.py`)

- **Test** = `n_test_aneux` (30 by default) cases drawn from `Angioflow`'s own
  `dataset == "AneuX"` cases — real-patient shapes, held out so evaluation never
  leaks onto a shape family the model trained on.
- **Train** = everything else: all of `AneuGv1`, all of `AneuGv2` (synthetic, never
  eligible for test), all `ImperialNHS` cases in `Angioflow`, and whatever `AneuX`
  cases weren't drawn for test.
- The eventual target metric is generalization to the held-out **AneuX** cases
  specifically (real patients) — `AneuGv1` is the most abundant source but the least
  representative of that target distribution; `Angioflow`/`AneuGv2` are more directly
  relevant. A `--balance_domains` flag in the reference trainer equalizes per-epoch
  sampling across the three source folders above (not just real-vs-synthetic) so
  `AneuGv1`'s raw volume (634 cases x 80 frames) doesn't dominate gradient updates.

## Known caveats

- **Patient independence across sources is unverified.** Splitting by
  `(source, case)` keeps one case's frames together, but does not establish that
  different sources/cases are distinct patients — exact-geometry checks won't catch
  refits, re-cuts, or generated descendants of the same base shape. Worth auditing
  upstream patient IDs / synthetic-generation provenance before assuming no leakage
  across `AneuGv1`/`Angioflow`/`AneuGv2`.
- **Type-1 (two-opening) aneurysms are not represented** in this archive at all —
  every case is `aneurysm_type == 0`. A model trained on this data assumes 3 geodesic
  channels (`geo_dist`'s last dim) and would need a schema change to admit type-1.
- A separate **waveform file** (`waveform_yiying.txt`, the inlet flow waveform used
  to condition the time embedding) is **not included** in this archive — it lives
  outside this dataset root in the source repo
  (`angioflowv2_merged/_meta/waveform_yiying.txt`) and is a required input to the
  reference model's forward pass, not optional.

## Provenance

Produced by `ood_fix/cfd_data_factory/extract_wss_dataset.py` (`AneuGv1`,
`Angioflow`) and `ood_fix/cfd_data_factory/extract_ghd_graphs_aneugv2.py`
(`AneuGv2`) in the `AneuG-Net` repo (private). See that repo's
`ood_fix/PIPELINE.md` for the full upstream pipeline (CFD registration, GHD
fitting/placement, subdivision) if you have access to it.