Loss odyssey in medical image segmentation.

Area of Science:

• Deep Learning
• Medical Image Analysis
• Computer Vision

Background:

• Loss functions are critical for deep learning segmentation methods.
• Numerous loss functions have been developed for diverse segmentation tasks.
• A systematic evaluation of their comparative utility is lacking.

Purpose of the Study:

• To provide a comprehensive review of segmentation loss functions.
• To conduct a large-scale analysis of 20 general loss functions on 3D segmentation tasks.
• To establish a benchmark for loss function performance and guide future development.

Main Methods:

• Systematic review of segmentation loss functions.
• Large-scale empirical analysis of 20 general loss functions.
• Evaluation across four 3D segmentation tasks using six public datasets from multiple medical centers.

Main Results:

• No single loss function consistently outperformed others across all tasks.
• Compound loss functions, including Dice with TopK, focal loss, Hausdorff distance loss, and boundary loss, demonstrated superior robustness.
• Publicly available code and results serve as a benchmark for the research community.

Conclusions:

• Compound loss functions offer the most reliable performance for 3D medical image segmentation.
• The findings provide valuable insights for selecting and developing effective loss functions.
• This work establishes a benchmark for future research in segmentation loss functions.