Deep learning-based cross-modal MR-CT registration for brain metastases radiotherapy with multi-scale feature refinement and brainstem guidance.

Accurate multimodal deformable registration between magnetic resonance (MR) and computed tomography (CT) images is essential for precise target delineation in brain metastases radiotherapy. However, substantial modality discrepancies and complex anatomical deformations make robust alignment challenging. This study proposes an enhanced Transformer-based registration framework to improve cross-modal spatial correspondence modeling. We introduce MSFRTransMorph, a multi-scale feature refinement(MSFR) extension of the TransMorph architecture. The proposed refinement module is designed to strengthen hierarchical feature interactions and enhance fine-grained deformation modeling. Brainstem segmentation was used as anatomical guidance during training, while no tumor annotations were involved. The model was trained on 141 MR-CT image pairs from patients with brain metastases and evaluated. Registration performance was assessed using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and Jacobian-based deformation metrics. MSFRTransMorph achieved the highest registration accuracy for both the brainstem and gross tumor volume (GTV), with an GTV DSC of 74.25 ± 10.61% and HD95 of 16.92 ± 22.03 mm. However, the increase in volumetric overlap was accompanied by a higher proportion of local deformation folding, indicating reduced topological regularity. The proposed multi-scale refinement mechanism enhances cross-modal feature representation and improves volumetric alignment accuracy in MR-CT registration. The observed trade-off between registration precision and deformation stability highlights the necessity of explicitly balancing alignment accuracy and topological plausibility in high-capacity Transformer-based registration frameworks.