Spectrum intervention based invariant causal representation learning for single-domain generalizable medical image...

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Domain shift due to acquisition variance hinders segmentation model performance.
• Current methods focus on data diversity or domain-invariant representations, modeling superficial statistical dependencies.
• A causal perspective is needed for robust and generalizable segmentation.

Purpose of the Study:

• To propose a Spectrum Intervention based Invariant Causal Representation Learning (SI²CRL) framework.
• To unify data generation and representation learning from a causal viewpoint.
• To achieve domain-robust segmentation by addressing underlying causal factors.

Main Methods:

• Reifying object elements in the frequency domain as phase variables for data generation.
• Employing an amplitude-based intervention module for low-frequency perturbations.
• A two-stage causal synergy modeling process for representation learning: causal decoupling and adversarial causal purification.

Main Results:

• Consistent performance gains across diverse medical imaging tasks: cross-site prostate MRI, cross-modality abdominal CT-MRI, and cross-sequence cardiac MRI segmentation.
• Demonstrated superior robustness compared to state-of-the-art methods.
• Effective filtering of style-sensitive non-causal factors and derivation of causally sufficient information.

Conclusions:

• The proposed SI²CRL framework effectively addresses domain shift in medical image segmentation.
• Causal representation learning provides a more robust and generalizable approach than traditional methods.
• SI²CRL offers a promising direction for improving the reliability of AI in medical imaging.