---
license: apache-2.0
pretty_name: FLIP Ark+ tutorial chest X-ray splits (synthetic)
viewer: false
tags:
  - medical
  - chest-xray
  - synthetic
  - federated-learning
---

# FLIP Ark+ tutorial chest X-ray classification splits

Synthetic chest X-ray DICOMs (768×768, 8-bit RGB) with Yes/No lesion labels, backing the Ark+ NVFLARE
tutorials in [FLIP](https://github.com/londonaicentre/FLIP) (Federated Learning Interoperability
Platform).

| Folder | Role | DICOMs |
|---|---|---|
| `site1/` | Site-1 training split — `arkplus_fine_tuning` | 1910 |
| `site2/` | Site-2 training split — `arkplus_fine_tuning` | 1872 |
| `site1_holdoff/` | Site-1 hold-out split — evaluation tutorials | 478 |
| `site2_holdoff/` | Site-2 hold-out split — evaluation tutorials | 468 |

Each folder contains:

- `accession-resources/<accession_id>/<accession_id>.dcm` — one DICOM per accession.
- `sample_get_dataframe_response.csv` — cohort dataframe: an `accession_id` column plus the lesion
  label columns (`Effusion`, `Edema`, `Consolidation`, `Infiltration`, `Lung Nodule or Mass`,
  `Pneumothorax`, `Lungs in normal arrangement`), valued `Yes`/`No`.

All images are **fully synthetic** (DeCaf `balanced_synthetic_split`) — no patient data.

## Usage

The FLIP tutorials fetch these splits with `huggingface_hub.snapshot_download` and normalise them into
a gitignored local layout. From the FLIP repo root:

```bash
make -C fl-tutorials download-arkplus-finetuning-data   # site1 + site2  → data/arkplus/site{1,2}/
make -C fl-tutorials download-arkplus-eval-data         # hold-out splits → data/arkplus/site{1,2}_holdoff/
```

Licensed Apache-2.0.