X-ray projection-guided deformation vector field estimation and dynamic CBCT imaging using patient-specific vector-quantized diffusion.

Authors

Abstract

To develop a conditional vector-quantized diffusion-based dynamic CBCT imaging framework using dual-view X-ray projections for accurate reconstruction of respiratory-induced lung tumor motion, with the potential to enable four-dimensional (4D) target verification during lung radiotherapy. Patient-specific respiratory motion models were constructed from planning 4D-CT images using deformable image registration and principal component analysis (PCA). Synthetic DVFs were generated by sampling PCA coefficients and applied to the reference-phase CT images to produce dynamic 3D volumes, from which paired dual-view DRRs were generated. A conditional vector-quantized diffusion framework was developed, consisting of separate 2D and 3D VQ-VAEs for encoding DRRs and DVFs into discrete latent representations, followed by a conditional diffusion model to learn the mapping from X-ray projection features to DVF latent representations. During inference, dual-view X-ray projections were encoded and processed by the conditional diffusion model to estimate DVFs, which were then used to deform the reference CT image and reconstruct temporally resolved 3D volumes. The simulation study demonstrates that the estimated DVFs effectively captured the respiratory-induced deformation of lung tumors, and the reconstructed dynamic volume CT images demonstrated accurate anatomical motion characteristics across respiratory phases. The proposed patient-specific vector-quantized diffusion framework enables accurate estimation of respiratory-induced lung tumor motion from dual-view X-ray projections and facilitates the high-temporal-resolution dynamic volume CT reconstruction. The results demonstrate its potential as an image-guided solution for tumor motion tracking in lung tumor radiotherapy.

Tags

Topics