S2055
Physics - Image acquisition and processing
ESTRO 2026
to align CBCT with planning CT geometry, and five cases were reserved for independent testing. Both datasets were resampled to a uniform voxel spacing and normalized before training. Two deep learning architectures, U-Net++ and SwinUNETR, were trained to perform CBCT-to-CT conversion and generate synthetic CT (sCT) images. Image quality was evaluated using MS-SSIM, MAE, and PSNR metrics. To assess clinical applicability, SBRT treatment plans were recalculated on sCTs in Eclipse and compared with the reference CT in terms of dose MAE, DVH differences, and 3D gamma index (2 mm/2%). Results: Both models successfully generated anatomically consistent sCT images from CBCT scans. Image quality and dosimetric results are summarized in Table 1. Both SwinUNETR and U-Net++ achieved comparable image quality (MAE < 80 HU, PSNR ≈ 35 dB, SSIM ≈ 0.93).Dosimetric evaluation showed good agreement with planning CT. SwinUNETR had slightly lower dose MAE (0.55 ± 0.40 Gy) and DVH deviations (0.056 ± 0.021) than U-Net++ (0.75 ± 0.51 Gy, 0.088 ± 0.048). Gamma pass rates (2 %/2 mm) exceeded 97 % for both models.Table 1:Quantitative evaluation of synthetic CTs generated from CBCT
both visual image quality and clinical relevance of CT reconstructions for prostate CTV delineation (Figure 1). The scale enables systematic benchmarking of CT images and supports the identification of optimal DECT and PCCT parameters that may offer an alternative to MRI for prostate delineation.
Conclusion: An expert-derived Likert scale was successfully developed for evaluating CT image quality in radiotherapy planning for prostate cancer patients. This framework offers a robust, reproducible method for assessing the clinical utility of DECT and PCCT reconstructions in prostate cancer. While developed for prostate RT, the approach is generalizable across tumor sites and imaging modalities, supporting structured evaluation of emerging imaging technologies in radiotherapy. Keywords: PCCT, Imaging, Delineation Poster Discussion 1327 Generation of Synthetic CT from CBCT for Adaptive Lung SBRT Aliaksandr Miadzvetski Department of Medical Physics, N.N. Alexandrov National Cancer Centre of Belarus, Liasny, Belarus Purpose/Objective: Stereotactic body radiotherapy (SBRT) for lung tumours requires high geometric and dosimetric accuracy. Cone-beam CT (CBCT) is routinely acquired for patient setup verification, but its limited image quality and inaccurate Hounsfield units restrict its use for adaptive plan recalculation. This study evaluates the performance of two neural network architectures, U-Net++ and SwinUNETR, for CBCT-to-CT conversion in lung SBRT, including both image-quality and
Figure 1 shows visual comparison of sCTs generated by both models for a representative lung SBRT case, with HU difference maps relative to the reference CT.Figure 1: Visual comparison of CBCT-to-CT synthesis using U-Net++ and SwinUNETR for a representative lung SBRT case. Bottom row: HU difference maps between sCT and reference CT.
dosimetric validation. Material/Methods:
Two datasets were used in this study: the SynthRad 2025 dataset [1], comprising 258 lung cases used for initial model training, and an internal clinical dataset of 38 CBCT–CT pairs from lung SBRT patients used for model fine-tuning and evaluation. The clinical cases underwent rigid registration in the Varian Eclipse TPS
Made with FlippingBook - Share PDF online