Current deep learning-based brain medical image registration often relies primarily on intensity information, underutilizes anatomical semantic knowledge, and struggles to balance long-range dependencies with fine-grained feature details. To address these challenges, we propose FCReg, a novel coarse-to-fine registration network, incorporating an Image-Text Feature Coupling (ITFC) module to enrich image features with semantic priors derived from textual descriptions of anatomy, and incorporating KAN-Attention Hybrid (KAH) blocks to enhance the decoder’s perceptual sensitivity to salient features by combining Kolmogorov-Arnold Networks with convolutional attention mechanisms. FCReg investigates textual modality contributions to medical image registration tasks while employing efficient hybrid modules to reconcile information dependencies with fine-grained representation learning. The synergistic combination of ITFC and KAH enables FCReg to demonstrate outstanding performance on the LPBA40 and Mindboggle brain MRI datasets, achieving an optimal balance among registration accuracy, anatomical fidelity, and physically plausible deformation fields without a substantial computational burden.