Joint Optimization of Class-Specific Training- and Test-Time Data Augmentation in Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Existing data augmentation techniques often fail to address class imbalance and distribution shifts effectively.
• The alignment of training and testing data distributions is key to improving model generalization.

Purpose of the Study:

• To propose an effective and general data augmentation framework for medical image segmentation.
• To improve segmentation performance by addressing class imbalance and distribution mismatch.
• To integrate training-time and test-time data augmentation strategies.

Main Methods:

• A computationally efficient, gradient-based meta-learning scheme is employed to align training and validation data distributions.
• Class-specific training-time data augmentation (TRA) is learned to increase training subset heterogeneity and mitigate class imbalance.
• Training-time data augmentation (TRA) and test-time data augmentation (TEA) are jointly optimized.

Main Results:

• The proposed framework significantly and consistently improves segmentation performance across four diverse medical image segmentation tasks.
• The method demonstrates effectiveness when integrated with state-of-the-art segmentation models like DeepMedic and nnU-Net.
• Experimental results show superior performance compared to existing data augmentation solutions.

Conclusions:

• The developed data augmentation framework offers a significant advancement in medical image segmentation.
• Joint optimization of TRA and TEA provides a more robust approach to distribution alignment.
• The publicly available code facilitates further research and application in the field.