S2099
Physics - Image acquisition and processing
ESTRO 2026
consistently better HU accuracy and image quality of NRR-trained model.
Digital Poster Highlight 4889 Evaluating the impact of registration strategies on diffusion-based CBCT-to-sCT synthesis Jiaming Cao, Chelsea A. H. Sargeant, Alan McWilliam, Eliana Vasquez Osorio Division of Cancer Sciences, University of Manchester, Manchester, United Kingdom Purpose/Objective: Cone-Beam CT (CBCT) is important in Adaptive Radiation Therapy (ART) for monitoring daily anatomical changes, but its poor image quality, caused by noise and scatter, limits its wider application in dose recalculation and adaptive replanning. Generating high-quality synthetic CT (sCT) from CBCT can partially overcome these limitations. Deep learning models such as U-Net, GANs, and diffusion- based frameworks have been applied with promising improvements [1], but normally rely on paired CBCT– planning CT (pCT) data, where registration accuracy strongly impacts learning. This study evaluates how rigid (RR) and non-rigid (NRR) registration affect sCT generation based on a conditional diffusion model– conditional Iterative α -(de)Blending (cIADB) model [2], for its deterministic mechanism that improves quality, consistency and computational efficiency compared with stochastic diffusion models. Material/Methods: Paired CBCT–pCT datasets from 168 patients who underwent head-and-neck radiotherapy were registered on 3D volumes, using both RR and NRR by NiftyReg [3]. Axial slices were extracted, resized to 256 × 256, Hounsfield unit (HU) clipped to [-1000, 2000] and normalized to [-1, 1]. Separate cIADB models were trained on RR- and NRR-aligned datasets under same configurations. Evaluation was conducted on NRR- aligned test data using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE) against pCT. Results: As shown in Figure 1(a), the scatter-density plot, most data points lie above the diagonal for PSNR and SSIM and below for MAE, confirming that NRR training generates higher quality sCTs. Quantitatively, NRR achieved higher median PSNR (29.25 dB, IQR 3.19) and SSIM (0.92, IQR 0.03) than RR (22.68 dB, IQR 3.08; 0.82, IQR 0.05), and a lower median MAE (57.29 HU, IQR 36.88) versus RR (161.19 HU, IQR 97.05), with all differences statistically significant (p < 0.01), indicating improved accuracy and reduced variability. Qualitatively, Figure 1(b) illustrates that NRR-based synthesis better reproduces anatomical details consistent with pCT, while RR-trained models exhibit structure error (Figure 1(b) red box, around airway), emphasizing the advantage of NRR for anatomically faithful sCT generation. Table 1 further confirms
Conclusion: Improved registration accuracy enhances voxel-level fidelity and structural consistency in diffusion-based CBCT-to-sCT synthesis. Compared with RR, NRR provides better anatomical correspondence and intensity alignment, improving robustness and reliability. Future work will integrate optimized registration into downstream tasks such as dose recalculation and auto-segmentation within ART workflows. References: [1] Altalib A, McGregor S, Li C, Perelli A. Synthetic CT Image Generation From CBCT: A Systematic Review. IEEE Transactions on Radiation and Plasma Medical Sciences. 2025;9(6):691-707.[2] Cao J, Sargeant C, McWilliam A, Osorio E. Conditional Iterative α - (de)Blending Model for CBCT-to-sCT Synthesis: Towards a Deterministic and Simple Process. 2025. p. 149-58.[3] Modat M, Ridgway GR, Taylor ZA, Lehmann M, Barnes J, Hawkes DJ, et al. Fast free-form deformation using graphics processing units. Computer methods and programs in biomedicine. 2010;98(3):278-84. Keywords: Synthetic CT, Diffusion Model, Image Registration
Made with FlippingBook - Share PDF online