AbdomenAtlas: A large-scale, detailed-annotated, & multi-center dataset for efficient transfer learning and open...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Developing large-scale, annotated datasets is crucial for advancing AI in medical imaging.
• Existing datasets often lack the diversity and volume required for robust AI model training and validation.
• Abdominal CT analysis presents unique challenges due to complex anatomy and variability in patient populations.

Purpose of the Study:

• To introduce AbdomenAtlas, the largest publicly available abdominal CT dataset with extensive anatomical annotations.
• To provide a benchmark for evaluating the performance and generalizability of AI algorithms in abdominal image segmentation.
• To facilitate the development of large pre-trained models for AI in radiology, reducing expert annotation burden.

Main Methods:

• Compilation of 20,460 3D CT volumes from 112 diverse hospitals.
• Creation of 673,000 high-quality masks for 22 abdominal anatomical structures.
• Utilized a semi-automatic annotation pipeline combining expert radiologist revision and AI-driven prediction refinement.

Main Results:

• Established AbdomenAtlas, the largest annotated abdominal CT dataset to date.
• Demonstrated the utility of the dataset in an ISBI & MICCAI challenge (BodyMaps) for benchmarking AI algorithms.
• The dataset's scale and diversity support the development of robust, generalizable AI models for clinical applications.

Conclusions:

• AbdomenAtlas serves as a vital resource for AI research and development in medical imaging.
• The dataset enables rigorous benchmarking of AI algorithms, ensuring reliable performance in clinical settings.
• AbdomenAtlas is poised to accelerate clinical trials and foster innovation in the medical imaging community.