Topology-preserved transformer-based CBCT-CT registration via normalized gradient field guidance for nasopharyngeal carcinoma.

Accurate deformable registration between planning CT (computed tomography) and cone-beam CT (CBCT) is essential for adaptive radiotherapy in nasopharyngeal carcinoma, but cross-modality intensity differences, low soft-tissue contrast, and CBCT artifacts complicate alignment. We propose NJTransMorph, an unsupervised transformer-based registration framework that integrates normalized gradient field loss and Jacobian determinant regularization to improve multimodal structural alignment and deformation plausibility. NJTransMorph was evaluated against VoxelMorph, UTSRMorph, and TransMorph on an internal cohort (n = 35), an external scanner cohort (n = 27), and an external multi-center cohort (n = 61). Compared with TransMorph, NJTransMorph improved Dice score from 86.44% to 87.35% and reduced Hausdorff distance (HD95) from 2.01 to 1.87 mm on the internal cohort. It also reduced non-positive Jacobian determinants on the external scanner cohort (0.27%-0.23%) and multi-center cohort (0.21%-0.15%). Thus, among the evaluated deep learning methods, NJTransMorph improved registration accuracy and deformation regularity across scanner and center shifts.