Purpose: 3D fluoroscopy estimates the patient volume using a prior volume and a single x-ray image, which enables real-time positional monitoring of the target and organs-at-risks during treatment. Previous studies drove motion models by matching the forward-projected volumes to the x-ray images, which is prone to large errors at views with inferior contrast. To overcome this challenge, we propose a novel approach combining principal component analysis (PCA) motion modelling with diaphragm tracking.
Methods: Prior to treatment, the motion model is constructed by deformable image registration (DIR) between 4D-CT phases followed by a PCA. The diaphragm model is built by segmentation from the end-exhale CT. On each treatment day, the motion model is deformed to match the time-averaged CBCT to account for daily anatomic changes. During treatment, the diaphragm is tracked on each kV image using template matching. The PCA motion model is driven by the diaphragm position and used to estimate the patient volume. Five patients from the TCIA 4D-Lung database with multiple 4D-CT scans were used for evaluation. The first 4D-CT was used as the planning CT. The remaining were used as the ground truth volumes on each treatment day. Using Monte Carlo simulation, 17 sets of kV projections with scatter noise (0-360Â°) were generated from the ground truth. Geometric error was evaluated by registering the estimated volume to the ground truth.
Results: Geometric error in the tumor was 0.1Â±0.5mm, -0.1Â±1.8mm, and 0.2Â±0.9mm in the left-right, superior-inferior, and anterior-posterior directions. Within the thoracic region, the root-mean-squared and 95th percentile 3D errors were 2.5Â±0.4mm and 4.9Â±1.1mm (meanÂ±std across patients), as compared to 6.5Â±1.8mm and 12.2Â±3.4mm if anatomic changes were not accounted for.
Conclusion: Together, PCA motion modelling and diaphragm tracking can enable robust real-time volumetric estimation of the patient anatomy, offering improved tumor targeting, organs-at-risk avoidance, and improved dose reconstruction.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Australian Government National Health and Medical Research Council Early Career Fellowship APP1120333 and Cancer Institute New South Wales Early Career Fellowship CS00481.