Invariant Content Representation for Generalizable Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Domain generalization (DG) in medical imaging is crucial for privacy-preserving, single-source learning.
• Existing DG methods often use global data augmentation, limiting diversity and potentially causing style overfitting.
• This leads to suboptimal robustness on unseen target domains for medical image segmentation.

Purpose of the Study:

• To develop a novel method for enhancing invariant content learning and suppressing style variability in medical image segmentation.
• To improve the robustness and generalization capabilities of models trained on a single source domain to unseen target domains.

Main Methods:

• Proposed an Invariant Content Representation Network (ICRN) incorporating local style augmentation (LSA) using gamma correction.
• Introduced invariant content learning (ICL) to learn generalizable features from augmented and source-domain samples.
• Implemented style adversarial learning (SAL) with domain-specific batch normalization (DSBN) to mitigate source-domain style bias.

Main Results:

• Achieved significant improvements in segmentation performance on cross-domain datasets (Fundus and Prostate).
• Demonstrated an 8.74% and 11.33% increase in overall Dice coefficient (Dice).
• Showed a reduction of 15.88 mm and 3.87 mm in overall average surface distance (ASD) compared to state-of-the-art methods.

Conclusions:

• The proposed ICRN method effectively enhances domain generalization for medical image segmentation.
• The approach successfully balances invariant content learning with style suppression, leading to superior performance on unseen domains.
• This work offers a promising direction for developing more robust and generalizable medical image segmentation models.