Author: Mingli Chen, Xuejun Gu, Hao Jiang, Mahdieh Kazemimoghadam, Weiguo Lu, Gregory Szalkowski, Qingying Wang, Kangning Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose: Respiratory motion management is crucial for accurate radiation delivery to moving targets while protecting healthy tissue, relying on time-resolved volumetric imaging and continuous deformable image registration (DIR). To address the challenges in dynamic imaging and continuous DIR, we propose DAIRE-VI, a Diffusion-powered AI framework for continuous deformable image REgistration and time-resolved Volume Image generation.
Methods: The proposed DAIRE-VI model requires only two 3D CT volumes as inputs, one at the end of inhalation ( ) and the other at end of exhalation ( ) phases, and simultaneously generates continuous deformation fields and time-resolved CT. The continuous deformation fields are derived from a learnable series of scaling factors applied to the conditional score function between the inputs, which is estimated by a U-Net-based diffusion network. Leveraging the continuous deformation property, the CT at is progressively warped to the CT at under the physics constraints in each breathing phase during end-to-end training. The VAIRE-VI model was trained with 8 set of 4D-CT lung data (10 phases/set) acquired by DIR-Lab. The model was internally validated on the remaining 2 set of 4D-CT from DIR-Lab and externally 30 pairs of two end of respiratory phases lung CT from Learn2Reg 2022 dataset.
Results: Two metrics, normalized-mean-square error (NMSE) and structural similarity index measure (SSIM), which comparing generated images to ground-true images, were utilized to evaluate DAIRE-VI’s performance. DAIRE-VI achieves an NMSE of 0.159 ± 0.038 and SSIM of 0.759 ± 0.018 in internal validation and NMSE of 0.127 ± 0.020, SSIM of 0.792 ± 0.013 in external validation separately, surpassing the performance of both GroupRegNet and Cascade-VoxelMorph.
Conclusion: VAIRE-VI provides accurate and efficient continuous deformations and dynamic volumetric images generation, surpassing the performance of state-of-the-art methods of cascaded VoxelMorph and GroupRegNet, promising for real-time motion management application.