---
pretty_name: WSI_Embedding
task_categories:
  - feature-extraction
size_categories:
  - 1K<n<10K
---

# WSI_Embedding

Patch and slide embeddings for whole-slide images, plus the run records used to produce them.

**Original WSIs are not in this repo.** They stay on HPC / the source dataset remotes. Do not expect `.svs` / `.tif` here.

Repo: [`thanminh01/WSI_Embedding`](https://huggingface.co/datasets/thanminh01/WSI_Embedding)

## Top-level layout

| Path | What it is | Download when |
|---|---|---|
| `embeddings/<dataset>/<model>/<mag>x_<patch>px_<overlap>px_overlap/` | Encoder tensors (`.h5` / `.pt` / WSI-LLaVA json) and per-model TRIDENT `_config_` / `_logs_` | You need features |
| `Run-Information/runs-report/<dataset>/` | Job logs, VALIDATION, SLURM, perf, TRIDENT `runs/*.json`, cohort CSVs | You need provenance |
| `Run-Information/prep/<dataset>/` | Contours, geojson, thumbnails, `wsi_states`, patch-coord h5 `(N,2)`, overlay jpgs | You need QC / failed-mask / tiny-slide evidence |
| `Run-Information/parked/<dataset>/` | Encoder files for slides excluded from the official cohort | You need the parked cases |

Selective download: pull only `embeddings/DHMC-LUAD/titan/...` or only `Run-Information/prep/TCGA-LUAD`.

## Datasets

### DHMC-LUAD (encoder cohort 143)

There are **two HPC job directories**. They are **not** duplicates of the same run:

| Hub name | HPC folder | Role |
|---|---|---|
| `DHMC-LUAD` | `runs-information/full-int-DHMC-LUAD/` | Canonical shared coords + later encoders (Virchow/PRISM/KEEP/MUSK/WSI-LLaVA/PRISM2, MUSK 1024-d redo) |
| `DHMC-LUAD-TITAN-CONCH` | `runs-information/full-int-DHMC-LUAD (TITAN CONCH)/` | First encoder pack: TITAN + CONCH v1.5 / v1 at 20×/512 |

Use `DHMC-LUAD` for current coords, masks, and later-model logs. Use `DHMC-LUAD-TITAN-CONCH` for the original TITAN/CONCH job (SLURM 15450381 / 15450382).

Kept encoders (see per-model folders under `embeddings/DHMC-LUAD/`): TITAN 768, CONCH v1.5 768, CONCH v1 512, Virchow 2560, PRISM 1280, Virchow2 2560, Virchow2 CLS 1280, PRISM2, KEEP 768, MUSK **1024-d vision CLS** (`ms_aug=False`, no CLIP head in the stored tensor), WSI-LLaVA captions + case-bank. GigaPath tiles at 512 exist only as the WSI-LLaVA tile encoder, not as an official GigaPath product.

### TCGA-LUAD (encoder cohort 531)

531 diagnostic slides after parking **10 TCGA-05 DX1** cases (tiny GDC canvas + failed tissue mask). Those 10 are **not** in `embeddings/` tensors (except leftovers under `parked/`). Their `wsi_states`, thumbnails, contours, and coords **are** in `Run-Information/prep/TCGA-LUAD/` so the failure stays inspectable.

Tensors for TCGA-LUAD may still be landing on HPC; this `Run-Information` tree is uploaded independently.

### TCGA-LUSC

Not in this `Run-Information` snapshot: prep was still writing on the cluster at upload time. Encoder tensors are not published yet.

## Canonical mag / patch

| Models | Setting |
|---|---|
| TITAN, CONCH v1.5, CONCH v1, WSI-LLaVA | 20× / 512 |
| PRISM, Virchow, Virchow2, Virchow2 CLS, PRISM2 | 20× / 224 |
| KEEP | 20× / 256 |
| MUSK | 20× / 384 |

## How to download

```bash
# tensors only, one model
huggingface-cli download thanminh01/WSI_Embedding \
  --repo-type dataset \
  --include "embeddings/DHMC-LUAD/titan/**"

# QC / logs only
huggingface-cli download thanminh01/WSI_Embedding \
  --repo-type dataset \
  --include "Run-Information/**"
```

## What is intentionally absent

- Original whole-slide images
- HPC `runs-information/full-int-*/features_*` trees (those tensors live under `embeddings/` so they are not uploaded twice)
- `TCGA-LUSC` run snapshot while segmentation/coords were still in flight