Fast and Low-GPU-memory abdomen CT organ segmentation: The FLARE challenge

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automatic segmentation of abdominal organs in CT scans is crucial for clinical applications.
• Existing benchmarks often overlook model efficiency and cross-center generalizability.
• There is a need for comprehensive evaluation of segmentation methods considering speed and memory usage.

Purpose of the Study:

• To introduce the Fast and Low GPU memory Abdominal oRgan sEgmentation (FLARE) challenge.
• To benchmark abdominal organ segmentation methods on accuracy, inference speed, and GPU memory consumption.
• To encourage the development of efficient and accurate AI models for clinical use.

Main Methods:

• Organized the FLARE challenge, evaluating methods on accuracy, speed, and memory.
• Collected testing cases from diverse medical centers to assess generalizability.
• Encouraged simultaneous optimization of accuracy, inference speed, and low GPU memory.

Main Results:

• The winning method demonstrated a 19x increase in inference speed compared to state-of-the-art.
• The winning method reduced GPU memory consumption by 60% while maintaining comparable accuracy.
• Top methods were summarized, with code and containers made publicly available.

Conclusions:

• The FLARE challenge successfully benchmarked efficient and accurate abdominal organ segmentation models.
• Practical suggestions were provided for developing high-performing segmentation models.
• The FLARE challenge platform remains open for ongoing research and submissions.