S2106
Physics - Image acquisition and processing
ESTRO 2026
methods are generally either too slow or too simplistic for real-world application, while synthetic CT approaches often lack transparency and are prone to hallucinations. This study aims to develop a novel Sim2real scatter correction model, bridging the gap between simulated and experimental data and improving the clinical reliability of CBCT for ART. Material/Methods: A 2D U-Net model [1] was trained on Monte Carlo (MC) simulated CBCT projections, which model X-ray scattering accurately but may imperfectly represent the CBCT system. For each raw projection, the network, referred to as MC-Simulated in Figure 1, was trained to extract the scatter component, which was subtracted from the projections to reconstruct a scatter-free CBCT.To improve realism and mitigate MC limitations, a second strategy referred to as Sim2Real was evaluated. In this approach, the MC-based network is fine-tuned on experimental CBCT data acquired with a cylindrical calibration phantom to perform domain adaptation from the simulation to experimental setting. Experimental scatter distribution for the calibration phantom was derived using a prior-based method leveraging information from a reference CT of the phantom [2,3].
Conclusion: This study presents a novel data-driven algorithm for selecting reference patients using anatomical factors using principal component analysis. The resulting BestRef-based registration enables consistent spatial normalisation of inter-patient data, providing a vital framework for voxel-wise population-level analysis in radiotherapy research. This methodology supports the identification of spatially localised dose-side effect relationships. Keywords: Image registration, Reference patient selection Digital Poster 5155 A Novel Sim2Real Deep Learning Approach for Scatter Correction in Cone-Beam CT Camille Draguet 1,2 , Hugo Bouchard 1 , Arthur Lalonde 1,2 1 Département de Physique, Université de Montréal, Montreal, Canada. 2 Axes Imagerie et Ingénierie, CRCHUM, Montreal, Canada Purpose/Objective: Adaptive radiotherapy (ART) relies on daily imaging to account for anatomical changes. However, onboard cone-beam CT (CBCT) used in ART is affected by X-ray scatter, which degrades image quality, limits detection of small lesions, and compromises auto-contouring accuracy. Conventional physics-based correction
The models’ performance was evaluated on CBCT images of an anthropomorphic head phantom using image similarity metrics, including mean error (ME), mean absolute error (MAE), and structural similarity index (SSIM), to quantify voxel-wise agreement with a registered reference CT. Dice Similarity Coefficients (DSC) of auto-contours generated by a clinical tool (OrganRT, Varian) were used to evaluate whether scatter correction improves organ delineation. Results: The comparison between scatter correction strategies and the reference CT is shown in Table 1. MC-based and Sim2Real approaches were evaluated against a no-correction strategy, where raw projections are directly reconstructed into a CBCT volume. Sim2Real outperformed both no-correction and MC-based methods in image similarity metrics. For DSC, Sim2Real either surpassed the other strategies or
Made with FlippingBook - Share PDF online