Frequency-aware domain randomization for single-source domain generalization in medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning models for medical image segmentation struggle with domain shift, where test data differs from training data.
• Single-source domain generalization (DG) trains models on one domain to generalize to unseen domains, but current methods suffer from texture bias and limited style diversity.

Purpose of the Study:

• To propose Random Domain Generalization (RandDG), a frequency-aware method for single-source DG in medical image segmentation.
• To enhance generalization ability via coordinated input and feature space perturbations using a lightweight frequency-domain architecture.

Main Methods:

• Introduced a Global U-Shape Network (GUNet) for efficient long-range dependency modeling using Fourier transforms.
• Employed a Uniform Low Frequency spectrum Transform (ULoFT) filter for feature-space perturbation, mixing source statistics with uniform values.
• Utilized a dual-space randomization framework with input-space augmentation (GIN filters) and feature-space perturbation (ULoFT), regularized by a consistency loss.

Main Results:

• Achieved superior generalization on abdominal CT-MRI and cross-center prostate datasets, significantly outperforming competitive methods.
• Demonstrated an average DSC of 87.96% and HD of 4.82 mm on the abdominal dataset, and 75.95% DSC and 8.36 mm HD on the prostate dataset.
• Showed substantial improvements over the UNet baseline, with up to 19.36% higher average DSC and reduced average HD.

Conclusions:

• RandDG effectively addresses texture bias and limited style diversity in single-source DG for medical image segmentation using a frequency-aware dual-space randomization framework.
• The method shows promise for practical clinical deployment where target domain data is unavailable.