A Disentangled Representations based Unsupervised Deformable Framework for Cross-modality Image Registration.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Cross-modality magnetic resonance image (MRI) registration is crucial for analyzing diverse MRI datasets.
• Domain shift between different MRI modalities presents a significant challenge for accurate registration.
• Existing methods often require paired data or struggle with unsupervised adaptation.

Purpose of the Study:

• To develop a fully unsupervised deformable framework for cross-modality MRI registration.
• To address the domain shift problem using image disentangling techniques.
• To improve the accuracy and robustness of registration across different MRI modalities.

Main Methods:

• Proposed a multi-modal unsupervised image-to-image translation approach to decompose MRIs into content and style spaces.
• Developed an unsupervised deformable network leveraging the preserved intrinsic information in the content space.
• Introduced a novel loss function with metrics in both original and content image spaces.

Main Results:

• The proposed framework demonstrated superior registration performance compared to two state-of-the-art methods on two validation datasets.
• Image disentangling effectively separated domain-invariant content from modality-specific style.
• The novel loss function contributed to enhanced registration accuracy in the content space.

Conclusions:

• The fully unsupervised deformable framework offers a robust solution for cross-modality MRI registration.
• Image disentangling is a powerful technique for mitigating domain shift in medical image analysis.
• This method can serve as a valuable auxiliary tool in clinical practice for cross-modality MRI analysis.