Padr-Net: Phase-Aware Dual-Branch 4DCT-4DCBCT Deformable Registration for Lung Cancer Radiotherapy
Abstract
Purpose
Conventional 3D CT-CBCT registration is typically performed independently at a single respiratory phase, limiting the ability to characterize 4D respiratory motion and temporal continuity between planning and treatment. This work develops a phase-aware, modality-cooperative 4D registration method for lung cancer radiotherapy.
Methods
We propose PADR-Net, a phase-aware dual-branch 4DCT-4DCBCT deformable image registration network that jointly models planning 4DCT and treatment 4DCBCT respiratory motion within a unified anatomical reference frame. A phase-aware motion modeling module estimates 3D deformation fields mapping each 4DCT phase to a mid-ventilation reference phase to obtain a consistent planning motion representation. Guided by this representation, a cross-modality alignment module estimates deformation fields mapping each 4DCBCT phase to its corresponding 4DCT phase, enabling phase-to-phase planning-to-treatment registration in the same anatomical frame. A temporal deformation coordination module (CMHM) imposes smooth temporal evolution across phases to suppress high-frequency temporal perturbations and local distortions.
Results
On a clinical lung cancer 4DCT-4DCBCT dataset, affine registration achieved a lung-mask Dice similarity coefficient of 0.952, whereas PADR-Net improved Dice to 0.986. Using the 95th percentile Hausdorff distance (HD95) to assess boundary agreement, PADR-Net achieved a mean HD95 of 2.6 mm.
Conclusion
PADR-Net integrates phase-aware motion modeling, 4DCT-4DCBCT cross-modality alignment, and temporal consistency constraints to jointly register planning and treatment 4D respiratory motion in a unified anatomical reference frame with millimeter-level accuracy. The method provides a spatiotemporally consistent registration basis for 4D dose accumulation and adaptive planning in lung cancer radiotherapy.