Unsupervised deep consistency learning adaptation network for cardiac cross-modality structural segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep neural networks excel at medical image segmentation but degrade when applied to new domains with different data distributions.
• Unsupervised domain adaptation is crucial for leveraging annotated source data with unannotated target data, especially when annotations are scarce.
• Existing methods often fail with large domain gaps, necessitating advanced adaptation techniques.

Purpose of the Study:

• To introduce a novel unsupervised deep consistency learning adaptation network for medical image segmentation.
• To address performance degradation in deep neural networks when adapting to new data domains.
• To improve cardiac structural segmentation across different modalities without target domain annotations.

Main Methods:

• A deep consistency learning adaptation network with input and output space consistency learning.
• A domain translation path featuring a symmetric alignment generator with attention to cross-modality features and anatomy.
• A segmentation path utilizing entropy map minimization, output probability map minimization, and segmentation prediction minimization.

Main Results:

• The proposed method demonstrates robust performance on challenging cross-modality cardiac segmentation tasks.
• Experimental results show superior performance compared to existing unsupervised domain adaptation methods.
• Ablation experiments validate the effectiveness of the proposed framework's components.

Conclusions:

• The novel unsupervised deep consistency learning network effectively mitigates domain shift issues in medical image segmentation.
• The framework achieves robust cardiac structural segmentation across modalities, even with significant domain gaps.
• This approach offers a promising solution for unsupervised domain adaptation in medical imaging.