# Data splits (chest CT)

- Split ID lists; download each dataset from its source under its own terms.
- `val` is for model selection; `test` is official unless noted.

| Dataset | Files | ID format | Train / Val / Test | Notes |
|---|---|---|---|---|
| LIDC-IDRI | `splits/lidc_idri/` | `PatientID/StudyUID/SeriesUID` | 705 / 148 / 157 | Patient-disjoint |
| RSNA-PE | `splits/rsna_pe/` | `StudyUID/SeriesUID` | 4,983 / 1,062 / 1,060 | Study-level (Ke et al., MIDL 2024) |
| MSD-Lung | `splits/msd_lung/` | MSD case id | 44 / 6 / 13 | Holdout of 63 cases |
| NSCLC-Radiomics | `splits/nsclc_radiomics/` | `LUNG1-XXX` | 294 / 42 / 84 | LUNG1-035/128 excluded (corrupt) |
| CT-RATE | `splits/ct_rate/` | volume name | 43,400 / 2,994 / 3,002 | `test` = official validation (1,304 patients); `val` = 1,304 patients from the official train |

CT-RATE notes (the lists already exclude CT-RATE's non-chest volumes and 3 unreadable ones):
- **Pretraining:** also excludes `pretrain_exclusions.txt` (ReXGroundingCT overlap) and volumes with an empty report decomposition → 43,337 / 2,990 / 3,002.
- **Fine-tuning:** also excludes 11 unusable volumes (`ft/chest_ct/*/dataset/ct_rate.py`) → 43,390 / 2,993 / 3,002.