Colormap augmentation: a novel method for cross-modality domain generalization.

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Computer Vision

Background:

• Domain generalization is critical for medical image analysis across diverse data sources.
• Current deep learning segmentation models require significant computational resources for generalization.
• Cross-modality segmentation faces challenges due to domain shift.

Purpose of the Study:

• To evaluate a novel, simple, and effective method for enhancing deep learning model generalization in medical image segmentation.
• To address the domain shift problem in cross-modality segmentation.
• To reduce the computational cost associated with training generalized segmentation networks.

Main Methods:

• Applied eight augmentation methods individually to a source domain dataset.
• Tested generalized models on unseen target domain datasets from different imaging modalities.
• Leveraged standard augmentation, intensity augmentations, and color transformations (CmapAug).

Main Results:

• The CmapAug method, combined with standard augmentations, substantially improved the Dice Score compared to a baseline.
• Achieved Dice scores as high as 83.2% for liver structure segmentation.
• Outperformed the baseline, which struggled with segmentation in some cases.

Conclusions:

• Augmentation methods effectively address domain generalization and mitigate cross-modality domain shift.
• The proposed CmapAug strategy offers a simple, powerful solution for segmentation tasks.
• This approach has significant potential in clinical settings with limited target domain data.