Résumé
Respiratory motion is a major source of image degradation in liver single-photon emission computed tomography (SPECT), causing lesion blurring and biased activity quantification. Conventional countermeasures such as respiratory gating or deformable image registration (DIR) either suffer from high noise due to count splitting or require multi-phase data rarely available in clinical workflows. To overcome these limitations, DVF-Generator is proposed—a physics-aware conditional generative model that learns to synthesize realistic, patient-specific 3D deformation vector fields (DVFs) for respiratory motion directly from a static attenuation map (μ -map) and compact respiratory parameters. The model employs a FiLM-conditioned 3D U-Net architecture, trained on 4D XCAT phantoms with a composite loss that enforces motion accuracy, image fidelity, smoothness, and Jacobian positivity, ensuring physically plausible and topology-preserving deformations. Quantitative and qualitative evaluations on synthetic phantoms demonstrate that DVF-Generator achieves sub-voxel motion accuracy and anatomically consistent deformation, outperforming classical amplitude-based, rigid, and Demons registration baselines. Beyond synthetic data, an exploratory test on a real patient SPECT/CT case shows that the model generalizes to anatomy, reproducing realistic diaphragm excursions and smooth lung–liver motion patterns consistent with physiological breathing. In summary, integrating physics-based constraints with generative modeling establishes a new foundation for motion-aware quantitative imaging in nuclear medicine.