Sim2Real Deep Learning for Physics-Based Scatter Correction In CBCT-Guided Adaptive Radiotherapy
Abstract
Purpose
Adaptive radiotherapy (ART) requires accurate daily volumetric imaging, yet cone-beam CT (CBCT) is fundamentally limited by X-ray scatter, leading to degraded contrast, CT number inaccuracy, and reduced reliability of automated segmentation. These limitations restrict confident clinical use of CBCT for ART. Existing physics-based correction methods are often impractical, while purely data-driven synthetic CT approaches may lack physical interpretability and risk hallucination. This work proposes a simulation-to-real (Sim2Real) deep learning framework that combines Monte Carlo–based physical modeling with experimental domain adaptation for robust CBCT scatter correction.
Methods
A 2D U-Net was trained on Monte Carlo–simulated CBCT projections to estimate and remove projection-domain scatter prior to reconstruction. To address simulation-to-experiment mismatch, the network was fine-tuned using experimental CBCT data acquired with a cylindrical calibration phantom. Experimental scatter targets were generated using a prior-based method informed by a registered reference CT. Performance was evaluated on CBCT images of an anthropomorphic head phantom using mean error (ME), mean absolute error (MAE), and structural similarity index (SSIM) relative to the reference CT. Clinical relevance was assessed using Dice similarity coefficients (DSC) from automatically generated organ contours.
Results
Sim2Real consistently improved CBCT image quality compared to uncorrected and Monte Carlo–only correction. Lower ME and MAE and higher SSIM demonstrated improved CT number accuracy and structural fidelity. These gains translated into improved downstream performance, with auto-contours generated on Sim2Real–corrected CBCT achieving higher or comparable DSC values, typically within 1% of the best-performing method.
Conclusion
Sim2Real domain adaptation enables accurate, physically grounded CBCT scatter correction that improves both image fidelity and auto-contouring reliability. By integrating simulation accuracy with experimental calibration, this framework advances the clinical readiness of deep learning–based CBCT correction for adaptive radiotherapy.