Domain Generalization With Amplitude-Based Data Generation and Feature Random Suppression.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Domain generalization for image segmentation remains challenging, with existing methods struggling to match single-domain performance.
• Current approaches often fall short in achieving satisfactory segmentation results across different data distributions.

Purpose of the Study:

• To develop advanced techniques for segmenting unknown domains by improving model generalization.
• To enhance segmentation accuracy for medical imaging datasets with varying data characteristics.

Main Methods:

• Proposed a data augmentation method using amplitude perturbation to broaden data distribution and cover target domains.
• Introduced a feature suppression strategy to mitigate over-reliance on source domain features, boosting generalization.
• Designed a luminance contrast consistency (LCC) module for inter-domain style harmonization.
• Developed a multiscale convolutional attention (MSCA) module to improve perception of small objects and overall segmentation.

Main Results:

• Achieved state-of-the-art (SOTA) results on the ATLAS2.0 and Prostate public datasets.
• Demonstrated significant improvements in segmentation performance through the proposed augmentation and attention mechanisms.
• Validated the effectiveness of LCC and MSCA modules in enhancing model robustness and accuracy.

Conclusions:

• The proposed methods effectively address the challenges of domain generalization in image segmentation.
• The combination of data augmentation, feature suppression, LCC, and MSCA modules leads to superior segmentation performance on unseen domains.
• The developed approach offers a promising solution for robust medical image segmentation across diverse datasets.