README.md

---
license: cc-by-4.0
task_categories:
  - image-segmentation
  - image-classification
language:
  - en
tags:
  - medical
  - radiology
  - chest-ct
  - anatomy
  - segmentation
  - ct-scan
  - projection
pretty_name: RadGenome-Anatomy
size_categories:
  - 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/*.parquet
  - split: validation
    path: data/validation/*.parquet
dataset_info:
  features:
  - name: volume_id
    dtype: string
  - name: split
    dtype: string
  - name: image_pa
    dtype: image
  - name: image_ll
    dtype: image
---

# RadGenome-Anatomy

**RadGenome-Anatomy** is a large-scale chest radiograph anatomy segmentation dataset
constructed from the [RadGenome-ChestCT](https://huggingface.co/datasets/RadGenome/RadGenome-ChestCT) corpus
(originally based on [CT-RATE](https://huggingface.co/datasets/ibrahimhamamci/CT-RATE)).
It contains **25,692 volumetric studies** (24,128 train / 1,564 validation), yielding paired
postero-anterior (PA) and lateral (LL) projection images at **384 × 384** resolution.

Across the two radiographic views, the dataset provides **10,790,646 fine-grained anatomy masks**
over **210 canonical anatomy classes** and **513,860 region masks** over **10 anatomical groups**,
for a total of **11,304,506 binary mask instances**.

Each row represents one CT study and contains its PA and LL projection images.

---

## Dataset Summary

| Property | Value |
|---|---|
| **Studies** | 25,692 total (24,128 train / 1,564 val) |
| **Views per study** | 2 (PA + LL) |
| **Image resolution** | 384 × 384 |
| **Anatomy classes** | 210 structures (4-level hierarchy) |
| **Region classes** | 10 body-system groups |
| **Anatomy masks** | 10,790,646 (5,395,323 PA + 5,395,323 LL) |
| **Region masks** | 513,860 (256,930 PA + 256,930 LL) |
| **License** | CC-BY-4.0 |
| **Source** | RadGenome-ChestCT / CT-RATE |

### Splits

| Split | Studies | PA projections | LL projections | Anatomy masks | Region masks |
|---|---|---|---|---|---|
| train | 24,128 | 24,129 | 24,129 | 10,133,770 | 482,580 |
| validation | 1,564 | 1,564 | 1,564 | 656,876 | 31,280 |
| **total** | **25,692** | **25,693** | **25,693** | **10,790,646** | **513,860** |

---

## Dataset Structure

### Data Fields

| Column | Type | Description |
|---|---|---|
| `volume_id` | `str` | Unique study identifier, e.g. `train_1_a_1`. |
| `split` | `str` | Dataset split: `train` or `validation`. |
| `image_pa` | `Image` | PA (posteroanterior, front) chest projection image (JPEG, 384×384). |
| `image_ll` | `Image` | LL (lateral, side) chest projection image (JPEG, 384×384). |

### Anatomy Label Universe

The dataset defines **210 canonical anatomy classes** organized as a four-level hierarchy:
*body system → organ → substructure → canonical label*. At the top level, classes are grouped
into **10 body systems**, with a highly non-uniform per-system class count:

| Body system | # classes | Example structures |
|---|---|---|
| Skeletal | 93 | ribs (1–12 L/R), thoracic vertebrae (T1–T12), cervical/lumbar vertebrae, sternum, clavicles, scapulae, humerus, femur |
| Abdominal | 42 | liver (with segments), spleen, pancreas, kidneys, gallbladder, stomach, intestine |
| Mediastinal | 25 | aorta, IVC/SVC, carotid/subclavian arteries, brachiocephalic vessels, iliac/renal vessels |
| Cardiac | 11 | heart, atria (L/R), ventricles (L/R), myocardium, ascending aorta, left auricle, heart tissue |
| Pulmonary | 15 | left/right lung, upper/middle/lower lobes (L/R), lung nodule, tumor, effusion, pulmonary vein, pulmonary embolism |
| Airway | 6 | trachea, bronchi, larynx (glottis, supraglottis), cricopharyngeal inlet |
| Endocrine | 8 | thyroid (L/R + gland), adrenal glands (L/R), thymus |
| Esophageal | 2 | esophagus structures |
| Breast | 3 | breast structures |
| Neural / soft tissue | 5 | spinal cord, skin, muscle |

These same 10 body systems also serve as the **region-mask** label set (10 classes/view).

The full ordered list of canonical labels is in [`label_universe.json`](./label_universe.json) at the repo root.
Use it to map labels to fixed class indices for consistent multi-label training.

---

## Usage

### Load with 🤗 Datasets

```python
from datasets import load_dataset

ds = load_dataset("EvidenceAIResearch/radgenome-anatomy")
print(ds)
```

### Access images

```python
from PIL import Image
import io

row = ds["train"][0]
pa_img = Image.open(io.BytesIO(row["image_pa"]["bytes"]))
ll_img = Image.open(io.BytesIO(row["image_ll"]["bytes"]))
```


## License

[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) — derived from CT-RATE.
Commercial use is not permitted without prior permission from the original data providers.
See the [original dataset terms](https://huggingface.co/datasets/RadGenome/RadGenome-ChestCT) for full conditions.

## Citation

```
@article{ye2026radgenome,
  title={RadGenome-Anatomy: A Large-Scale Anatomy-Labeled Chest Radiograph Dataset via Physically Grounded Volumetric Projection},
  author={Ye, Shuchang and Meng, Mingyuan and Wang, Hao and Naseem, Usman and Kim, Jinman},
  journal={arXiv preprint arXiv:2605.17368},
  year={2026}
}
```