Latent Accelerated Diffusion-based Deformation Estimation for Real-time Volumetric Imaging.
Authors
Affiliations (7)
Affiliations (7)
- Stanford University, 875 Blake Wilbur Dr, Palo Alto, Palo Alto, California, 94304-2205, United States.
- Department of Bioengineering, University of Illinois at Urbana-Champaign, 1406 W Green St, Urbana, Illinois, 61801, United States.
- Department of Radiation Oncology, The University of Texas Southwestern Medical Center, 2280 Inwood Rd., Dallas, 75390, United States.
- Radiation Oncology, UT Southwestern Medical Center, 2280 Inwood Rd, Dallas, Texas, 75390, United States.
- Radiation Oncology, University of Texas Southwestern Medical Center at Dallas, 6363 Forest Park Rd, Dallas, Texas, 75390-9315, United States.
- Department of Radiation Oncology, The University of Texas Southwestern Medical Center, 2280 Inwood Rd, Dallas, Texas, 75390, United States.
- Department of Radiation Oncology, Stanford University, 875 Blake Wilbur Dr., Stanford, California, 94304, United States.
Abstract
We aim to address the technical limitations in 3D respiratory estimation and image reconstruction from ultra-sparse views, overcoming data-acquisition constraints that often hinder conventional deformable image registration and volumetric imaging in real-time clinical applications. 
We propose a Latent Accelerated Diffusion framework for Deformation Estimation enabling Real-time volumetric imaging (LADDER) framework, which integrates: (1) a deformation network (VoxelMorph) that generates a patient specific baseline deformation vector field (DVF) from pre treatment imaging, and (2) a latent diffusion model (LDM)-based DIR model that estimates DVF scaling factors and residual corrects to generate intra-treatment real-time DVF and dynamic volumetric images. The LDM compresses the baseline DVF into a compact latent manifold, enabling fast, projection conditioned refinement guided by anatomical cues. A physics informed loss enforces anatomical regularity and consistency with measured projections. LADDER was trained on the Learn2Reg dataset and evaluated on 10 DIR Lab lung datasets . 
Main Results: With a dual projection input, an end-inhale to end-exhale baseline DVF and a compression and down-sampling factor of 8, on 10 test cases, LADDER achieves a mean target registration error (TRE) of 0.87±0.33 mm, and high volumetric structure similarity (3D SSIM > 0.95) and low volumetric reconstruction error (3D NMSE < 0.006), while maintaining real-time inference (~0.11-0.12s). Further analysis shows the range of DVF span impacts deformation accuracy, and the dual-projection input improves deformation fidelity and reduced variability across breathing phases compared to the single-projection. 
Significance: LADDER enables real time 3D motion estimation and volumetric reconstruction from ultra sparse X ray views by conditioning diffusion in a patient specific DVF latent space. Its submillimeter TRE and real time speed indicate strong potential for next generation motion management and image guidance, supporting on the fly anatomical modeling to enhance the safety and efficacy of lung SBRT.