Improving CT-CBCT deformable image registration for cervical cancer adaptive radiotherapy using a deep learning approach.

Purpose: To improve the robustness and anatomical alignment accuracy of CT-CBCT deformable image registration for cervical cancer adaptive radiotherapy. Methods: A deep learning-based registration framework was developed by enhancing a transformer-based model (UTSRMorph) with a normalized gradient field (NGF) constraint. The network integrates convolutional and transformer modules to capture multi-scale anatomical features. A composite loss function combining mutual information, deformation regularization, and gradient-based similarity was used for training. The method was evaluated on internal and external CBCT-CT datasets using Dice, HD95, and Jacobian determinant. Results: On the internal dataset, the proposed method achieved Dice scores comparable to the baseline across all structures (bowel: 82.84%), while improving deformation regularity, with low non-positive Jacobian determinant values (%|J|≤0: 0.13), comparable to UTSRMorph (0.12) and lower than VoxelMorph (0.19) and TransMorph (0.21). Slight improvements in Dice and HD95 were observed in bony structures, particularly the hip bones. On the external dataset, the method demonstrated improved generalization, achieving higher Dice for the bowel (83.40% vs. 82.99% and 82.90%) and reduced HD95 (11.70 mm vs. 11.91 mm and 11.85 mm). Improvements were more evident in boundary alignment, especially in high-contrast regions, while maintaining comparable deformation smoothness. Conclusion: The proposed method (NGF-UTSRMorph) improves the robustness of CT-CBCT registration and enhances boundary alignment without compromising deformation smoothness. Improvements were more evident on the cross-scanner test cohort and in boundary-sensitive metrics, suggesting improved robustness under intensity inconsistencies.