Domain-Agnostic Learning with Anatomy-Consistent Embedding for Cross-Modality Liver Segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning models struggle with generalization across different data domains.
• Existing Domain Adaptation (DA) methods typically focus on single target domains.
• Domain Agnostic Learning (DAL) aims to generalize models to multiple, diverse target domains.

Purpose of the Study:

• To introduce a novel Domain-Agnostic Learning framework with Anatomy-Consistent Embedding (DALACE).
• To enable knowledge transfer for both domain adaptation and domain-agnostic learning tasks.
• To achieve cross-modality liver segmentation invariant to domain variations while preserving anatomical structures.

Main Methods:

• Proposed DALACE framework integrating domain-transfer and task-transfer learning.
• Learned disentangled representations for modality invariance and anatomical consistency.
• Utilized Domain-Agnostic Module (DAM) and Anatomy-Preserving Module (APM) with segmentation consistency supervision.

Main Results:

• DALACE achieved superior performance in Domain Adaptation (DA) tasks, with a Dice Similarity Coefficient (DSC) of 0.847.
• DALACE significantly improved Domain Agnostic Learning (DAL) tasks, reaching a DSC of 0.794.
• Demonstrated successful disentanglement visualization, enhancing interpretability and confirming benefits for downstream tasks.

Conclusions:

• DALACE effectively addresses challenges in cross-modality medical image segmentation.
• The framework shows significant improvements over state-of-the-art methods in both DA and DAL scenarios.
• Disentangled representations and anatomical preservation are crucial for robust domain-agnostic learning.