K-CMorph: Integrating K-space Consistency and Complex-Valued Processing for Improved MRI Deformable Registration

Xinyu Sun, Hui Ding, Chong Liu, Jiacheng Lu, Feiyan Li, Zhuhong Shao, Jingbing Yang

Abstract: Deformable image registration plays a vital role in medical imaging, particularly for aligning 3D brain MRI scans. Deep learning has achieved remarkable success in spatial domain registration, yet the integration of K-space frequency data—central to MRI—remains largely overlooked. Notably, the phase component of K-space, which carries rich structural information, is underutilized in existing methods. To address these gaps, we propose K-CMorph, a novel deep learning-based registration algorithm that ensures K-space consistency through complex-valued processing and attention fusion. Our method leverages the complementary strengths of spatial-domain features and K-space frequency data. Specifically, we employ complex-valued convolutional networks to process both magnitude and phase components of K-space, preserving subtle structural variations critical for precise registration. An attention fusion mechanism adaptively integrates spatial and frequency-domain features, enhancing robustness across diverse anatomical regions. Additionally, we introduce the K-space Image Consistency Loss (KIC Loss), a novel loss function that enforces global coherence in the frequency domain, minimizing artifacts and ensuring fidelity in the registered output. Experiments on the available IXI dataset, K-CMorph demonstrates significant improvements over state-of-the-art methods.

External IDs:dblp:conf/icic/SunDLLLSY25