AbdomenAtlas provides a large-scale, multi-center dataset of 5,195 annotated CT volumes for abdominal organ segmentation. It aims to facilitate efficient transfer learning and open algorithmic benchmarking in medical imaging. The dataset is suitable for researchers and developers working on medical image segmentation tasks.

How It Works

The project offers pre-trained U-Net and Swin UNETR models trained on a combination of 14 public CT datasets. It also provides a framework for data preparation, including generating datalists and preprocessing nii.gz files. The core functionality involves using these models to generate pseudo-labels for new datasets, which can then be assembled with original or revised annotations, prioritizing higher-quality labels.

Quick Start & Requirements

• Install by cloning the repository: git clone https://github.com/MrGiovanni/AbdomenAtlas
• Requires Python environment setup and specific dependencies (details in installation instructions).
• Pre-trained models (U-Net, Swin UNETR) are available for download.
• Data must be in .nii.gz format.
• GPU is required for model inference.
• See: Installation Instructions

Highlighted Details

• AbdomenAtlas 1.0 contains 5,195 CT volumes with 9 annotated abdominal organs.
• AbdomenAtlas 1.1 (linked via SuPreM) expands to 25 classes.
• The dataset was annotated efficiently using a human-in-the-loop approach, reducing annotation time significantly.
• Includes links to associated papers from NeurIPS 2023, Medical Image Analysis, and RSNA 2023.

Maintenance & Community

• Associated with Johns Hopkins University.
• Supported by the Lustgarten Foundation and Patrick J. McGovern Foundation.
• Active development indicated by multiple associated papers and dataset versions.
• Google Groups mailing list available: https://groups.google.com/u/2/g/bodymaps

Licensing & Compatibility

• The README does not explicitly state a license for the code or dataset.
• Associated papers are published in academic journals.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

• Currently, only .nii.gz data format is supported, with a stated TODO to support DICOM.
• The full dataset (8,448 volumes) is not fully released, with only 3,410 volumes committed for release.