Multi-PhaseMorph: multi-scale unsupervised registration enabling joint alignment of multi-phase CT imaging of the abdomen

Multi-phase imaging protocols, such as CT urography (CTU), allow clinicians to capture dynamic anatomical and physiological changes by acquiring images at multiple time points, typically before and after contrast administration. However, registering images from different phases is challenging due to contrast-induced intensity variations, physiological motion, and limitations of traditional pairwise methods. We propose Multi-PhaseMorph, an unsupervised model designed for joint registration of multi-phase images. Specifically, Multi-PhaseMorph simultaneously registers noncontrast (NCT) and excretory (EXC) phases to the nephrographic (NEP) phase using a unified architecture featuring a shared Swin Transformer-based encoder and dual decoders generating phase-specific deformation fields. The model employs a multi-scale loss combining intensity, deformation, and structural terms across resolutions and utilizes an organspecific spatial weight map to focus alignment on clinically relevant structures such as kidneys. We evaluated Multi- PhaseMorph using a retrospective three-phase CTU dataset of 126 patients (mean age 65 ± 12 years) with manual kidney segmentations, split into training (80%), validation (10%), and testing (10%) sets. Compared to VoxelMorph, TransMorph, and ANTsPy on separate two registration tasks, NCT-to-NEP and EXC-to-NEP, Multi-PhaseMorph achieved higher average Dice scores (0.938 ± 0.014 for NCT, 0.935 ± 0.017 for EXC) and SSIM values (0.924 ± 0.061 for NCT, 0.923 ± 0.032 for EXC). Combined Dice and SSIM scores across phases were 0.936 ± 0.015 and 0.933 ± 0.017, respectively, outperforming comparison methods. These results demonstrate that leveraging shared anatomical context across phases significantly enhances registration accuracy and anatomical consistency, offering substantial benefits for clinical applications like enhancement analysis and lesion tracking.