README.md

---
license: apache-2.0
task_categories:
- image-segmentation
modality:
- CT
language: []
tags:
- medical-imaging
- whole-body-CT
- anatomy-segmentation
- autopet
- multi-organ
pretty_name: DAP Atlas
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: sample_id
    dtype: string
  - name: subject_id
    dtype: string
  - name: study_uid_last5
    dtype: string
  - name: num_slices
    dtype: int32
  - name: ct_middle_slice
    dtype: image
  - name: mask_middle_slice
    dtype: image
  - name: overlay_middle_slice
    dtype: image
  splits:
  - name: train
    num_bytes: 178521349
    num_examples: 533
  download_size: 178530023
  dataset_size: 178521349
---

# DAP Atlas — Dense Anatomical Prediction Atlas

Whole-body CT dataset with **142 anatomical structures** segmented across 533 volumes.
The CT images come from the AutoPET FDG-PET-CT-Lesions cohort; the segmentation masks
were produced by Jaus et al. (2023) via knowledge aggregation across 14 source datasets,
nnU-Net pseudo-labelling, and post-processing using anatomical guidelines.

## Dataset Summary

| Field | Details |
|---|---|
| Modality | CT (whole-body) |
| Body Part | Full body |
| Subjects | 482 unique patients (533 CT volumes — some patients have multiple studies) |
| Labels | 142 anatomical structures + background + unknown_tissue (144 entries, gap at ID 11) |
| Volume Shape | typically 512×512×~390 (varies per study) |
| Spacing | ~0.8 × 0.8 × 2.5 mm |
| Total Size | ~55 GB |
| Mask License | Apache-2.0 |
| CT License | TCIA Restricted (free-use; TCIA fully public since 2025-07-07) |

## Data Structure

```
DAP_Atlas/
├── images/
│   └── AutoPET_<subjectID>_<studyUID5>.nii.gz   # 533 CT volumes
├── masks/
│   └── AutoPET_<subjectID>_<studyUID5>.nii.gz   # 533 mask volumes (paired by filename)
└── labels.json                                   # ID → anatomical structure name
```

CT and mask filenames are identical — pair each `images/X.nii.gz` with `masks/X.nii.gz`.
Mask voxels are uint8 with values in {0, 1, ..., 144} (with a gap at 11). See `labels.json`
for the canonical ID → name mapping (sourced from Table 2 of the paper).

## Splits

The released dataset has no official train/val/test split. All 533 cases form a single
pool. Downstream papers carve their own subsets.

## Notes

- **Mask source:** the published 533 NIfTI files are the V1 expert-validated masks
  (`Atlas_final_dataset_V1_533/`). The repo also distributes nnU-Net model weights
  (Task901 / Task902) for inference on new CTs; those are not redistributed here.
- **Costa numbering** is reversed vs. TotalSegmentator (DAP `costa_1` = lowest rib).
  See https://github.com/alexanderjaus/AtlasDataset/issues/7 for the mapping.
- **Sex-conditional labels** (prostate, uterus, etc.) appear only for matching sex.
- ID 11 is intentionally absent.

## Citation

```bibtex
@article{jaus2023towards,
  title   = {Towards Unifying Anatomy Segmentation: Automated Generation of a
             Full-body CT Dataset via Knowledge Aggregation and Anatomical Guidelines},
  author  = {Jaus, Alexander and Seibold, Constantin and Hermann, Kelsey and
             Walter, Alexandra and Giske, Kristina and Haubold, Johannes and
             Kleesiek, Jens and Stiefelhagen, Rainer},
  journal = {arXiv preprint arXiv:2307.13375},
  year    = {2023}
}

@article{gatidis2022whole,
  title   = {A whole-body FDG-PET/CT dataset with manually annotated tumor lesions},
  author  = {Gatidis, Sergios and Hepp, Tobias and Früh, Marcel and others},
  journal = {Scientific Data},
  volume  = {9},
  number  = {1},
  pages   = {601},
  year    = {2022},
  doi     = {10.1038/s41597-022-01718-3}
}
```

## Sources

- Original masks: https://github.com/alexanderjaus/AtlasDataset
- AutoPET CTs (TCIA): https://www.cancerimagingarchive.net/collection/fdg-pet-ct-lesions/
- Paper: https://arxiv.org/abs/2307.13375