The Devil is in the Frequency: Constrained and Adaptive Fine-Grained Domain Perturbation for Robust Medical...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Domain generalization (DG) is crucial for reliable medical diagnostics across diverse healthcare systems.
• Domain shifts from variations in imaging protocols and devices impede accurate anatomical identification.
• Current data augmentation methods struggle to bridge domain gaps without distorting anatomical features.

Purpose of the Study:

• To introduce the Adaptive Dual-Space Spectral Perturbation (AdaDSP) framework to address domain generalization challenges in medical imaging.
• To enhance the diversity of training data and capture sensitive frequency bands hindering generalization.
• To encourage learning domain-invariant representations while preserving discriminative capacity.

Main Methods:

• AdaDSP injects learnable spectral perturbations into input images and feature maps for broad-level enhancement.
• A Fine-Grained Spectral Perturbation module uses attention mechanisms to modulate frequency distributions and adaptively perturb sensitive bands.
• A Universal Triple-stage Semantic Constraint Framework promotes domain-invariant learning.

Main Results:

• The proposed AdaDSP framework significantly enhances data diversity and captures critical frequency bands.
• The Fine-Grained Spectral Perturbation module effectively modulates frequency distributions.
• The Universal Triple-stage Semantic Constraint Framework aids in learning robust, domain-invariant representations.

Conclusions:

• AdaDSP outperforms state-of-the-art methods in medical image analysis domain generalization.
• The framework achieves notable improvements of 2.40% and 2.99% in two medical imaging tasks.
• AdaDSP offers a promising solution for improving diagnostic consistency across diverse healthcare settings.