# CheXmask-U Database v1.0.0
A large-scale collection of anatomical segmentation masks for chest X-rays with node-wise uncertainty estimates.
## Overview
CheXmask-U provides anatomical segmentation masks derived from chest radiographs across multiple public databases:
- ChestX-ray8
- Chexpert
- MIMIC-CXR-JPG
- Padchest
- VinDr-CXR
Each image includes **landmark-based uncertainty estimates** generated using the HybridGNet framework. The dataset also includes quality metrics based on Reverse Classification Accuracy (RCA).
## Dataset Structure
The dataset consists of CSV files for each source database. Each CSV contains:
| Column Name | Description |
|------------|-------------|
| Image ID | Reference to original image in source dataset |
| Dice RCA (Max) | Maximum Dice Similarity Coefficient for RCA |
| Dice RCA (Mean) | Mean Dice Similarity Coefficient for RCA |
| Landmarks (Mean) | Mean positions of anatomical landmarks across stochastic predictions |
| Landmarks (Std) | Standard deviation of landmarks positions, representing node-wise uncertainty |
| Left Lung | Left lung segmentation mask in RLE format |
| Right Lung | Right lung segmentation mask in RLE format |
| Heart | Heart segmentation mask in RLE format |
| Height | Height of segmentation mask |
| Width | Width of segmentation mask |
## Data Processing
All images were processed to maintain consistent quality:
1. Images were preprocessed to 1024x1024 resolution
2. HybridGNet model was applied to generate landmark predictions
3. **N=50 stochastic landmark predictions per image were sampled from the latent space**
4. Landmarks mean and standard deviation were computed
5. Pixel-level masks were derived from mean landmarks
6. RCA scores were calculated for quality assessment
## Usage Guidelines
1. **Source Images**: Users must obtain source images from the original databases and comply with their access requirements, including any ethical training or courses.
2. **Quality Threshold**: For analysis, use only segmentation masks with Dice RCA (Mean) ≥ 0.7
3. **Uncertainty Information**: The `Landmarks (Std)` column provides node-wise uncertainty; users may consider the average landmark standard deviation per image as an additional criterion for filtering uncertain predictions.
4. **Resolution**: Pre-processed masks (1024x1024) are included to ensure consistent resolution across datasets.