# Split assignments and excluded subjects

One folder per real X-ray source used in the FleXray paper, including the sources
we cannot redistribute, so the exact partitions and quality-control exclusions can
be rebuilt from the original downloads. Paths are relative to each source's
download directory and contain no image data.

## `<Dataset>/splits.csv`

Every image that FleXray trained or evaluated on.

| column | meaning |
|---|---|
| `sample_id` | FleXray sample key (matches `dataset.yml` for redistributed sources) |
| `subject_id` | patient/subject grouping key; all samples of a subject share one split |
| `split` | `train`, `val`, or `test` as used in the paper |
| `source_path` | image path relative to the original download (`''` when the source has no per-image file) |

All results in the paper are scored on `test`; `val` was used only for model development.

## `<Dataset>/exclusions.csv`

Every image removed before split assignment. An empty file means nothing was excluded.

| column | meaning |
|---|---|
| `subject_id` | FleXray subject key of the excluded image (`''` for source files that were never annotated) |
| `source_path` | image path relative to the original download when known |
| `reason` | one of the reasons below |
| `detail` | human-readable note, source filename, or scope |

| `reason` | applied when |
|---|---|
| `manual_review` | the annotation was flagged as unreliable during manual review (`detail` gives the reason where documented) |
| `duplicate_image` | the pixels exactly duplicate another image; one representative is kept |
| `empty_mask` | the annotation has no foreground after mapping into the FleXray label protocol |
| `missing_local_image` | the annotation references an image absent from the downloadable image set |
| `landmark_out_of_bounds` | AASCE: landmark coordinates fall outside the image |
| `corrupt_zero_filled_png`, `out_of_bounds_femur_polygon` | MTDDH: source image or polygon is unusable |
| `not_referenced_by_dataset2_coco` | MTDDH: PNG present in the archive but absent from the Dataset2 annotations |

All criteria were fixed before any model was scored and apply identically to FleXray
and every baseline. These files are generated by
`scripts/release/export_split_manifests.py` in the OmniAX repository.