LDM-Morph: Latent diffusion model guided deformable image registration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deformable image registration is crucial for medical imaging tasks.
• Current deep learning methods (CNNs, Transformers) lack semantic information, limiting performance.
• Pixel-space similarity metrics ignore high-level anatomical features, causing deformation folding.

Purpose of the Study:

• To introduce LDM-Morph, an unsupervised deformable registration algorithm.
• To improve semantic feature representation and anatomical feature matching.
• To enhance topology preservation and registration accuracy.

Main Methods:

• Integrated features from Latent Diffusion Models (LDM) for semantic enrichment.
• Designed a Latent and Global feature-based Cross-Attention module (LGCA).
• Proposed a hierarchical metric evaluating similarity in pixel and latent-feature spaces.

Main Results:

• LDM-Morph outperformed state-of-the-art CNNs and Transformers on 2D cardiac and 3D datasets.
• Achieved comparable topology preservation and computational efficiency.
• Demonstrated improved registration accuracy.

Conclusions:

• LDM-Morph effectively addresses limitations of existing deformable registration methods.
• The integration of LDM features and hierarchical metric significantly improves registration.
• The proposed framework offers a promising solution for medical image registration.