# CheXmask-U: Chest X-ray Lung and Heart Segmentation Masks with Landmark Uncertainty

## Facts

- Provider: Cosarinsky, Gaggion, Ferrante et al. (PhysioNet)
- License: CC-BY-4.0 (https://physionet.org/content/chexmask-u/1.0.0/)
- Format: CSV (in sampled files)
- Stored at: Amazon S3, s3://physionet-open/chexmask-u (us-east-1)
- Access: public, readable without an account
- DOI: https://doi.org/10.13026/6vn3-3j51
- Category: Health & Biomedical
- Source: https://physionet.org/content/chexmask-u/1.0.0/
- Page: https://scigantic.com/archive/cb70b193-96a3-50fc-aad7-40c748361ab7
- Keywords: Health & Biomedical, image segmentation, graph learning, CSV, AWS Open Data

## About

CheXmask-U (PhysioNet v1.0.0, 2026) extends the CheXmask database with node-wise uncertainty for anatomical segmentation of chest radiographs (CXR). It covers images from five public sources, ChestX-ray8 (NIH), CheXpert, MIMIC-CXR-JPG, PadChest and VinDr-CXR, and for each image gives landmark-based contours of the left lung, right lung and heart predicted by the HybridGNet graph neural network. Uncertainty comes from Monte Carlo sampling of the HybridGNet latent space: 50 stochastic landmark predictions per image, summarised as per-landmark mean [x,y] positions and standard deviations. Pixel masks are filled from the mean contours and stored run-length encoded (RLE), with mask height and width, plus Reverse Classification Accuracy (RCA) quality scores (Dice RCA max and mean; the authors suggest Dice RCA mean >= 0.7). Uncertainty was validated against 255 doubly annotated radiographs, synthetic noise and occlusion, and out-of-distribution tests. Layout: OriginalResolution/ holds one large CSV per source (CheXpert_1/2, ChestX-Ray8, MIMIC-CXR-JPG_1..3, Padchest, VinDr-CXR, 0.9 to 4.4 GB each) in original image coordinates, Preprocessed/ holds masks at the standardised 1024x1024 resolution; 39.9 GB uncompressed, read with pandas. The images themselves are not included and must be obtained from each source under its own terms. Masks released under CC BY 4.0; cite the CheXmask-U PhysioNet record and Gaggion et al. (CheXmask, Scientific Data 2024).

## Load the data

```python
import boto3
from botocore import UNSIGNED
from botocore.client import Config
s3 = boto3.client('s3', region_name='us-east-1', config=Config(signature_version=UNSIGNED))
s3.list_objects_v2(Bucket='physionet-open', Prefix='chexmask-u/', Delimiter='/')
```

Lists the files with anonymous access. Runs on any machine, with no cloud account.

## Data structure

File types in sampled files: CSV

Example file: `chexmask-u/1.0.0/OriginalResolution/CheXpert_1.csv`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/cb70b193-96a3-50fc-aad7-40c748361ab7
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/cb70b193-96a3-50fc-aad7-40c748361ab7
- Over MCP: `get_schema_card` with id `cb70b193-96a3-50fc-aad7-40c748361ab7` at https://api.scigantic.com/mcp

## Related topics

- [CSV datasets](https://scigantic.com/topics/csv)
- [Image segmentation datasets](https://scigantic.com/topics/image-segmentation)
- [Health & Biomedical datasets](https://scigantic.com/topics/health-and-biomedical)
- [Graph learning datasets](https://scigantic.com/topics/graph-learning)
