MrGiovanni/AbdomenAtlas

What it solves

AbdomenAtlas addresses the extreme time and labor costs associated with manually annotating large-scale medical imaging datasets. Specifically, it provides a massive dataset of 8,448 CT volumes (3.2 million slices) with multi-organ annotations, reducing the annotation time for such a volume of data from an estimated 30.8 years to just three weeks.

How it works

The project utilizes a human-in-the-loop annotation method combined with AI models to accelerate the process. It provides pre-trained checkpoints for Swin UNETR and U-Net architectures trained on multiple public CT datasets. Users can generate pseudo-labels using these models, which are then assembled with original annotations and human-revised labels based on a priority system (Original > Revised > Pseudo).

Who it’s for

Medical imaging researchers and AI developers focusing on abdominal organ segmentation and transfer learning in radiology.

Highlights

• Massive Scale: Includes annotations for 8,448 CT volumes across multiple organs (spleen, liver, kidneys, stomach, gallbladder, pancreas, aorta, and IVC).
• Efficient Annotation: Employs a method that accomplishes in three weeks what would normally take decades of manual effort.
• Pre-trained Models: Offers ready-to-use U-Net and Swin UNETR checkpoints for immediate segmentation tasks.
• Active Learning Support: Includes optional functionality to generate attention maps to guide human annotators.