EDRL: Entropy-guided disentangled representation learning for unsupervised domain adaptation in semantic...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning models struggle with domain shift in medical imaging.
• Unsupervised domain adaptation (UDA) aims to bridge this gap without target domain annotations.

Purpose of the Study:

• To develop a robust cross-domain segmentation method for medical images.
• To address the challenge of domain shift in deep learning models.

Main Methods:

• Proposed Entropy-guided Disentangled Representation Learning (EDRL) for UDA.
• Integrated image and feature alignment using disentangled representation and adversarial learning.
• Implemented a dynamic feature selection mechanism via soft gating.

Main Results:

• Achieved state-of-the-art results on CT-MR and multi-sequence cardiac MR datasets.
• Demonstrated high performance in cross-domain segmentation tasks (e.g., 84.8% DSC on CT-MR).

Conclusions:

• The EDRL model effectively enhances cross-domain medical image segmentation.
• Experimental results validate the proposed method's efficacy in handling domain shift.