Accurate Fine-Grained Segmentation of Human Anatomy in Radiographs via Volumetric Pseudo-Labeling

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 DatasetsThis 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 SegmentationExplore 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++.