BLUE RIBBON POSTER MULTI-DISCIPLINARY: Evaluation of Three-Dimensional Fiducial Marker Position Fidelity In Fbpconvnet-Based Sparse-View Reconstruction Using Treatment Kv Images for Prostate SBRT
Abstract
Purpose
This study aims to quantitatively evaluate the three-dimensional positional fidelity of implanted fiducial markers in deep learning-based sparse-view reconstruction from on-treatment kV images acquired during prostate stereotactic body radiotherapy (SBRT).
Methods
Sparse-view kV projection data with 10, 30, and 60 views were simulated by forward projection of CT images under clinical acquisition geometry and reconstructed using FBP to generate the initial 3D reconstructed images. The images were processed using an FBPConvNet-based convolutional neural network (CNN). The network learns a residual mapping between artifact-corrupted FBP images and artifact-reduced reconstructions, thereby combining physics-based reconstruction with data-driven refinement. The resulting reconstructions were used for fiducial marker localization and quantitative evaluation by extracting marker positions via intensity-based segmentation followed by centroid calculation. Localization accuracy was evaluated for nine test patients using multiple quantitative metrics, including mean positional error, Bland–Altman analysis, mean and worst-case target registration errors (TREₘ and TREw), and cumulative success rate analysis.
Results
Qualitative results showed that fiducial markers remained clearly identifiable in FBPConvNet reconstructions across all sparse-view conditions. Quantitative analysis demonstrated that mean TRE values remained at the submillimeter level for all patients, while worst-case TRE values remained below approximately 0.6 mm. Bland–Altman analysis revealed minimal bias and no systematic directional trends in the left–right and anterior–posterior directions. Each data point corresponds to the patient-level mean marker position averaged across slices that contain fiducial markers. In both directions, the mean bias remains close to zero, and most differences fall within the limits of agreement (LoA).
Conclusion
These results demonstrate that CNN-based sparse-view reconstruction can preserve fiducial marker positional fidelity with submillimeter accuracy under sparse-view acquisition. By quantitatively validating marker localization performance, this study supports the feasibility of using deep learning–reconstructed kV images for post-treatment intrafraction motion assessment and suggests a pathway toward future real-time motion monitoring in radiotherapy workflow.