Collaborative Learning of Augmentation and Disentanglement for Semi-Supervised Domain Generalized Medical Image...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised domain generalization (SSDG) faces challenges from label scarcity and domain shifts.
• Existing SSDG methods fail to differentiate intra-domain shifts (IDS) from cross-domain shifts (CDS) and overlook label scarcity's impact.
• Current approaches often combine semi-supervised learning (SSL) and domain generalization (DG) without addressing their interplay.

Purpose of the Study:

• To propose a novel framework, CausalAD, for semi-supervised domain generalized medical image segmentation.
• To address limitations in existing SSDG methods by decomposing the problem into unsupervised domain adaptation (UDA) and DG.
• To improve the robustness and accuracy of medical image segmentation across diverse domains.

Main Methods:

• CausalAD framework combines unsupervised domain adaptation (UDA) and domain generalization (DG).
• Employs a causal augmentation process using disentangled style factors for UDA.
• Features a disentanglement process to separate domain-invariant content from domain-variant features for DG.
• Utilizes a proxy-based self-paced training strategy (ProSPT) for efficient learning with pseudo-labels.
• Introduces a hierarchical structural causal model (HSCM) for conceptual explanation.

Main Results:

• CausalAD demonstrates significant effectiveness in semi-supervised domain generalized medical image segmentation.
• Achieved superior performance compared to state-of-the-art methods across cross-sequence, cross-site, and cross-modality settings.
• The proposed ProSPT strategy effectively guides training by selecting high-quality pseudo-labels.

Conclusions:

• CausalAD offers a novel and effective approach to semi-supervised domain generalization in medical imaging.
• The framework successfully handles both label scarcity and domain shift challenges.
• The method shows strong potential for real-world applications requiring robust medical image segmentation.