A Bayesian Image-Domain Approach to Virtual Non-Contrast Imaging In Dual-Energy Hypersight CBCT
Abstract
Purpose
The new generation of cone-beam CT (CBCT) systems, such as HyperSight (Varian Medical Systems), provide CT-like image quality with improved HU accuracy. Dual-energy Cone beam CT (DE-CBCT)–based virtual non-contrast (VNC) imaging avoids additional non-contrast CT scan required for treatment planning. This work proposes a Bayesian, image-domain VNC algorithm for HyperSight CBCT, with the goal of robust iodine removal while preserving underlying anatomy.
Methods
Comprehensive calibration measurements were performed using soft tissue–, fat–, lung–, bone–equivalent materials and calcium and iodine inserts with various densities and concentrations. Low (80 kVp) and high-energy (140 kVp) scans were acquired using the HyperSight CBCT system. A material classification knowledge base algorithm characterized by the mean and covariance matrix of each material was developed. For each voxel, soft probabilistic material classification was performed based on Mahalanobis distance to these manifolds. A Bayesian model, which represents each voxel as a baseline material component with an additive iodine contribution, was selectively applied to highly ambiguous voxels. Classification probabilities were incorporated as Bayesian priors, and a Markov chain Monte Carlo (MCMC) solver was employed.
Results
The novel VNC algorithm successfully generated VNC CBCT images in which iodine-enhanced regions were substantially suppressed across the tested concentration range, while the underlying Rando phantom anatomy including bone regions were well preserved. No obvious anatomical distortion or over-subtraction was observed in non-enhanced regions. The method additionally produced voxel-wise iodine maps and uncertainty estimates.
Conclusion
A novel Bayesian VNC algorithm was developed that leverages statistical dual energy material classification priors for HyperSight CBCT. The results show promise for robust iodine removal while preserving anatomical fidelity. Further validation and quantitative evaluation are ongoing.