SU‐E‐J‐01: 3D Fluoroscopic Image Estimation From Patient‐Specific 4DCBCT‐Based Motion Models

Abstract

Purpose:3D motion modeling derived from 4DCT images, taken days or weeks before treatment, cannot reliably represent patient anatomy on the day of treatment. We develop a method to generate motion models based on 4DCBCT acquired at the time of treatment, and apply the model to estimate 3D time‐varying images (referred to as 3D fluoroscopic images).Methods:Motion models are derived through deformable registration between each 4DCBCT phase, and principal component analysis (PCA) on the resulting displacement vector fields. 3D fluoroscopic images are estimated based on cone‐beam projections simulating kV treatment imaging. PCA coefficients are optimized iteratively through comparison of these cone‐beam projections and projections estimated based on the motion model. Digital phantoms reproducing ten patient motion trajectories, and a physical phantom with regular and irregular motion derived from measured patient trajectories, are used to evaluate the method in terms of tumor localization, and the global voxel intensity difference compared to ground truth.Results:Experiments included: 1) assuming no anatomic or positioning changes between 4DCT and treatment time; and 2) simulating positioning and tumor baseline shifts at the time of treatment compared to 4DCT acquisition. 4DCBCT were reconstructed from the anatomy as seen at treatment time. In case 1) the tumor localization error and the intensity differences in ten patient were smaller using 4DCT‐based motion model, possible due to superior image quality. In case 2) the tumor localization error and intensity differences were 2.85 and 0.15 respectively, using 4DCT‐based motion models, and 1.17 and 0.10 using 4DCBCT‐based models. 4DCBCT performed better due to its ability to reproduce daily anatomical changes.Conclusion:The study showed an advantage of 4DCBCT‐based motion models in the context of 3D fluoroscopic images estimation. Positioning and tumor baseline shift uncertainties were mitigated by the 4DCBCT‐based motion models, while they caused errors when using 4DCT‐based motion models.Partially funded by Varian research grant.