Network-Based Scatter Suppression for Robust and Accurate Mutual-Information-Based Image Registration In Image-Guided Radiation Therapy
Abstract
Purpose
In image-guided radiation therapy (IGRT), patient positioning is typically estimated by rigid registration between cone-beam CT (CBCT)–derived digitally reconstructed radiographs (DRRs) and treatment kV X-ray images acquired at the same gantry angle. However, treatment kV X-ray images are heavily degraded by scattered radiation, leading to reduced contrast and degraded image similarity, which can compromise registration accuracy. This study proposes a network-based scatter suppression approach for treatment kV X-ray imaging to enhance image quality and improve the robustness of rigid 2D–2D image registration for patient positioning in IGRT, while preserving real-time clinical feasibility.
Methods
Scatter suppression was applied to treatment kV X-ray images prior to image registration. A transmission estimation network was employed to predict a pixel-wise transmission map representing scattering-induced attenuation. This transmission map was integrated into a radiographic scattering model to suppress low-frequency scatter components and enhance structural contrast. Experimental validation was performed using a spine phantom imaged on a Varian TrueBeam system. Rigid 2D–2D image registration between DRRs and kV X-ray images was conducted before and after scatter suppression, and registration performance was quantitatively assessed using a mutual-information (MI)–based similarity metric.
Results
Scatter suppression substantially improved visual consistency between CBCT-derived DRRs and treatment kV X-ray images, resulting in clearer anatomical depiction during image registration. While ground-truth positional deviations were unavailable, scatter-suppressed images yielded more stable estimates of in-plane translation and rotation. MI values consistently increased after scatter suppression, and MI maps exhibited enhanced spatial coherence across anatomical regions, indicating improved robustness of image registration.
Conclusion
The proposed network-based scatter suppression approach enhances image visibility and similarity in treatment kV X-ray imaging, leading to more robust rigid 2D–2D image registration in IGRT. These results demonstrate the potential of using learned transmission estimation for real-time scatter mitigation and reliable patient positioning under clinical on-treatment imaging conditions.