AbdomenCT-1K: Is Abdominal Organ Segmentation a Solved Problem?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models achieve high performance in abdominal organ segmentation on benchmark datasets.
• Existing datasets often lack diversity, limiting the generalizability of state-of-the-art methods to real-world clinical scenarios.

Purpose of the Study:

• To introduce AbdomenCT-1K, a large and diverse dataset for abdominal organ segmentation.
• To evaluate the generalization capabilities of current segmentation methods on diverse data.
• To establish new benchmarks for supervised, semi-supervised, weakly supervised, and continual learning in abdominal organ segmentation.

Main Methods:

• Construction of the AbdomenCT-1K dataset comprising over 1000 multi-center, multi-phase, multi-vendor, and multi-disease abdominal CT scans.
• Large-scale evaluation of state-of-the-art segmentation methods on the AbdomenCT-1K dataset.
• Development of novel methods for four distinct segmentation benchmarks: fully supervised, semi-supervised, weakly supervised, and continual learning.

Main Results:

• State-of-the-art methods demonstrate limited generalization across different medical centers, phases, and unseen diseases.
• Performance gaps highlight the need for more robust segmentation models.
• Proposed methods serve as strong baselines for the established benchmarks.

Conclusions:

• The AbdomenCT-1K dataset is crucial for advancing research in generalizable abdominal organ segmentation.
• The established benchmarks address key challenges in supervised, semi-supervised, weakly supervised, and continual learning.
• This work paves the way for clinically applicable automated abdominal organ segmentation.