S2067
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster Highlight 2140
2 presents representative transversal slices demonstrating improved anatomical consistency when incorporating surface data.
Registration Uncertainty Undermines CBCT-to-CT Conversion: A Case for Synthetic Training Data Lukas Zimmermann 1,2 , Michael Rauter 3 , Maximilian Schmid 1,2 , Dietmar Georg 1,2 , Barbara Knäusl 1,2 1 Medical University of Vienna, Department of Radiooncology, Vienna, Austria. 2 Medical University of Vienna, hristian Doppler Laboratory for Image and Knowledge Driven Precision Radiation Oncology, Vienna, Austria. 3 University of Applied Sciences Wiener Neustadt, Competence Center for Preclinical Imaging and Biomedical Engineering, Wiener Neustadt, Austria Purpose/Objective: Adaptive radiotherapy requires accurate cone-beam CT (CBCT) to CT conversion for treatment plan evaluation and modification. Current deep learning approaches rely on paired clinical CBCT-CT data, thus contain registration errors which in turn might compromise model performance. To overcome this influencing factor a synthetic CBCT simulator was developed. The aim was to generate perfectly aligned training data sets from planning CTs and demonstrate superior geometric fidelity compared to real-data trained models. Material/Methods: A physics-based CBCT simulator was developed, that incorporated following real world factors of influence on image quality: breathing motion (driven by surrogate structures like bowel bag), scatter, and noise. The projection and volume resampling are accelerated by CUDA kernels. Two Elekta XVI data pelvic datasets were used: a clinical dataset with deformably registered pairs (Elastix with Impact Loss) (n=180) and the SynthRAD2023 (https://synthrad2023.grand-challenge.org/) dataset with rigidly registered pairs (n=180), plus 139/180. Additionally, CT-only scans for synthetic data generation for the clinical/SynthRAD dataset were used. Two model configurations were trained using the nnUNet framework modified for regression: synthetic data only and real data only (n=180). For testing, additional 30 paired cases were included for both registration methods. Evaluation metrics included Mean Absolute Error (MAE), Peak-signal-to- noise-ratio (PSNR), Structure Similarity index measure (SSIM) (following SynthRAD2023 guidelines), normalized mutual information (NMI) and correlation coefficient (CC) between CBCT and generated synthetic CT to assess geometric alignment. Results: Synthetic CBCT generation required 10seconds for a 11Mega voxels dataset. Synthetic-trained models showed MAE of 48.5/77.1 HU versus 34.8/53.2 HU for models trained on real-data (clinical/SynthRAD). However, models trained on synthetic data
Figure 1 - Boxplots of the Mean Squared Error (MSE), maximum distance to Agreement (maxDTA), and gamma passing rate ( γ PR) with a tolerance of 2%- 2mm.
Figure 2 - Transversal view with external (green),FoV (red), and surface scan (white) displayed onstan_sCT (left) andsurf_sCT (right). Conclusion: Integrating surface scan information into deformable registration improves sCT generation for limited- FoV CBCT, yielding anatomically more accurate images and enhanced dosimetric precision for breast cases. Future work will explore other superficial targets, and potential applications for adaptive radiotherapyand dose accumulation. References: [1] Thing, R. S., Pirs, H., & Seco, J. (2022). Evaluation of CBCT based dose calculation in the thorax and pelvis using two generic algorithms. Physica Medica: European Journal of Medical Physics, 103, 157–165.[2] Archambault, Y., Boylan, C., Bullock, D., Morgas, T., Peltola, J., Ruokokoski, E., & Vaniqui, A. (2020). Making on-line adaptive radiotherapy possible using artificial intelligence and machine learning for efficient daily re- planning.Medical Physics International Journal, 8(2), 77–86.[3] Shields, B., & Ramachandran, P. (2023). Generating missing patient anatomy from partially acquired cone-beam computed tomography images using deep learning: a proof of concept. Physical and Engineering Sciences in Medicine, 46(3), 1321–1330. Keywords: synthetic CT, SGRT, surface
Made with FlippingBook - Share PDF online