Color-Quality Invariance for Robust Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Single-source domain generalization (SDG) for medical image segmentation is challenging due to variations in image color distribution and quality.
• Existing models trained on high-quality images often fail to generalize to low-quality test images, leading to performance degradation.

Purpose of the Study:

• To develop novel techniques for enhancing domain generalization in medical image segmentation.
• To address the limitations of current methods in handling color and quality shifts between source and target domains.

Main Methods:

• Proposed a dynamic color image normalization (DCIN) module with global (GRIS) and local (LRIS) reference image selection strategies.
• Introduced a color-quality generalization (CQG) loss to enforce invariance to color and quality variations.
• Evaluated the methods on medical image segmentation tasks trained on a single source domain.

Main Results:

• The proposed DCIN module and CQG loss significantly improved segmentation performance over the baseline on target domain datasets.
• Achieved up to a 32.3-point increase in Dice score compared to the baseline.
• Demonstrated robust and usable segmentation results even under substantial domain shifts.

Conclusions:

• The novel DCIN module and CQG loss effectively enhance generalization capabilities in medical image segmentation.
• The proposed methods contribute to developing more robust models that perform well across unseen domains with varying image characteristics.
• The techniques show promise for real-world applications requiring reliable medical image analysis across diverse datasets.