info string	licenses sequence	images list	categories list	annotations list
PAX-Ray++, v2	[]	[{"id":1,"file_name":"0215a_frontal.png","height":512,"width":512,"date_captured":""},{"id":3,"file_(...TRUNCATED)	[{"supercategory":"spine","id":0,"name":"spine"},{"supercategory":"cervical spine","id":1,"name":"ce(...TRUNCATED)	[{"id":178,"image_id":1,"category_id":0,"segmentation":{"size":[512,512],"counts":"kTV3:R;J[H4XM<:^O(...TRUNCATED)

PAX-Ray++ Dataset

The PAX-Ray++ Dataset is a high-quality dataset designed to facilitate segmentation tasks for anatomical structures in chest radiographs. By leveraging pseudo-labeled thorax CT scans projected onto a 2D plane, this dataset provides fine-grained annotations resembling traditional X-ray imaging. This enables the development and evaluation of models tailored to anatomical segmentation in medical imaging.

Key Features

Large Dataset: Contains 7,377 frontal and lateral view images, each carefully pseudo-labeled.
Fine-Grained Annotation: Offers annotations for 157 distinct anatomical classes, ensuring comprehensive coverage of thoracic anatomy.
Extensive Instances: Includes over 2 million annotated instances, providing a robust foundation for training and evaluation.
2D Projection of 3D Data: Combines the richness of 3D CT data with the accessibility of 2D radiographic images.

Applications

The PAX-Ray++ dataset is designed to support:

Anatomical segmentation in chest X-rays.
Development of machine learning models for medical imaging tasks.
Research on transfer learning between CT-derived and true radiographic images.

Related Repositories

1. Dataset Dataloaders

2D Anatomy Datasets
This repository provides dataloaders for PAX-Ray++ and other datasets, making it easy to integrate the dataset into your machine learning pipelines.

2. Model Development and Applications

Chest X-Ray Anatomy Segmentation
Explore pre-trained models and pipelines designed specifically for the PAX-Ray++ dataset and other similar datasets. This repository demonstrates how to apply segmentation models trained on PAX-Ray++.

Getting Started

Download the Dataset

git lfs install
git clone [email protected]:datasets/cmseibold/PAX-RayPlusPlus

or follow the instructions in 2D Anatomy Datasets :

sh src/prepare_data/prepare_paxraypp/get_paxraypp_full.sh

Dataset Structure

The dataset consists of:

Frontal View Images: Annotated radiographs projected from thorax CT scans.
Lateral View Images: Corresponding annotations for lateral projections.

Annotations are provided in a compatible COCO format for easy integration with common machine learning frameworks such as MMDetection and Detectron2.

Requirements

Make sure to install the required libraries by following the instructions in the 2D Anatomy Datasets repository.

Citation

If you use the PAX-Ray++ dataset in your research or projects, please consider citing the dataset.

@inproceedings{Seibold_2023_CXAS,
author    = {Constantin Seibold, Alexander Jaus, Matthias Fink,
Moon Kim, Simon Reiß, Jens Kleesiek*, Rainer Stiefelhagen*},
title     = {Accurate Fine-Grained Segmentation of Human Anatomy in Radiographs via Volumetric Pseudo-Labeling},
year      = {2023},
}

Feedback and Contributions

Contributions, feedback, or questions are welcome! Feel free to open an issue or submit a pull request in the relevant repositories.