IG-BiReg: implicit-guided bidirectional residual registration framework for cardiac cine MRI images

To achieve accurate registration of cardiac cine magnetic resonance imaging (MRI) sequences by addressing the challenges of large non-rigid deformations during the cyclic movement of the heart. Approach: We propose a new collaborative framework built on the key idea of "explicit motion generation combined with implicit global verification". To capture the continuous dynamics of the heart, the explicit approach utilizes a bidirectional residual fusion network to generate deformation fields based on local spatiotemporal context. To ensure global structural consistency, the implicit supervision is proposed: the warped sequences are mapped into mean intensity and principal component feature spaces. This dual-space constraint acts as a global feedback mechanism to enforce anatomical plausibility and suppress non-anatomical deformations across the entire sequence. Main results: The method achieved average Dice coefficients of 0.843 and 0.842 on the ACDC and M&Ms datasets, respectively. Furthermore, it demonstrated highly accurate contour alignment, yielding average 95% Hausdorff Distance (HD95) and Average Symmetric Surface Distance (ASSD) values of 4.837 and 1.768 on the ACDC dataset, and 4.918 and 1.899 on the M&Ms dataset. Significance: The proposed method successfully addresses the inadequate handling of boundary frames in bidirectional temporal modeling and overcomes the limitations of groupwise registration by adapting to sequences of different lengths. By effectively tracking complex non-rigid cardiac deformations, it provides a robust solution that facilitates reliable quantitative clinical analysis. .