Unsupervised Multimanifold Cross-Guided Diffusion Deformable Registration for Cardiac MRI.

Abstract

Diffusion networks demonstrate remarkable robustness in extracting complex structural features across various domains of medical image processing. In the task of cardiac image registration, diffusion networks excel at reconstructing intricate structural details, thereby enabling effective representation of cardiac anatomical motion. In this article, we propose an unsupervised diffusion registration framework named MCG-Reg for 3-D cardiac magnetic resonance (MR) image registration, employing a multimanifold cross-fusion strategy. MCG-Reg comprises two components: the multimanifold cross-fusion (MCF) module and the weighted fusion codec (WFC) module. MCF module decouples the cardiac image, leveraging multifrequency and multiscale features for cross-attention (CA) calculation, and fuses with the edge image to enable adaptive focus gathering and edge perception capabilities in the model, thereby enhancing the effective aggregation of local and global features. WFC module further processes cardiac features by utilizing offset attention to capture large displacement information, while employing feature energy maps for residual connections to enhance the model's attention perception ability, thus facilitating better topology maintenance and boundary constraint realization. The registration accuracy and model generalization of the proposed MCG-Reg are validated in publicly available ACDC, M&Ms, and CAP datasets. The experimental results verify that it achieves state-of-the-art performance in comparison to related methods, highlighting the significant potential of the proposed framework in cardiac image analysis applications.