CTSC-Reg: Unsupervised multimodal medical image registration via cross-task structural consistency

Multimodal medical image registration is pivotal for integrating complementary information across distinct imaging modalities to facilitate precise diagnosis and treatment planning. However, existing translation-based methods typically decouple translation from registration, suffering from unidirectional dependency and reliance on implicit constraints that often compromise anatomical boundary precision. To address these limitations, we propose CTSC-Reg, an unsupervised framework that establishes Cross-Task Structural Consistency for collaborative network optimization. At its core, we introduce an Anatomy-Texture Synergistic Regularization (ATSR) mechanism, which projects heterogeneous features from translation and registration networks into a shared latent anatomical manifold to enforce mutual structural anchors instead of naive feature mimicry. Furthermore, the framework incorporates task-specific differentiated attention mechanisms to optimize global style transfer and local spatial correspondence respectively, alongside an explicit edge consistency loss to maintain boundary integrity. Extensive experiments on MR-CT and T1-T2 benchmark datasets demonstrate that CTSC-Reg consistently outperforms state-of-the-art methods, yielding significant improvements in boundary-sensitive metrics.