S2057
Physics - Image acquisition and processing
ESTRO 2026
Reducing the number of projections can substantially lower imaging dose, benefiting long-term survivors. We hypothesized that generative DL models can synthesize high-quality CBCTs (sCBCT) from sparse- projection CBCTs (spCBCT). This proof-of-concept study evaluated a 3D generative adversarial network (GAN) and a diffusion model for this task. Material/Methods: Raw data from an in-house clinical CBCT dataset were used. spCBCTs were created by reducing projections from 100% to 10% and reconstructed using the open- source TIGRE framework. The 100% projection reconstructions served as ground truth. Two DL architectures, Vox2Vox (GAN-based) and medDDPM (diffusion-based), were trained to generate sCBCTs from spCBCTs. Model performance was evaluated using mean squared error (MSE), structural similarity index (SSIM), and segmentation accuracy. To assess clinical potential, sCBCTs were registered to planning CTs using automatic image registration. Results: The in-house dataset included 2163 CBCTs (964 Head, 571 Thorax, 440 Pelvis, 188 Pelvis Large). Reconstruction of spCBCTs with 10% projections was feasible. The test set for this pilot project comprised 213 scans. Using the Vox2Vox model, the MSE of spCBCT inputs relative to full-dose CBCTs was reduced by 42.9% . SSIM improved from from 0.7796 to 0.8140 . The Vox2Vox model achieved superior visual quality (figure 1) and segmentation performance compared to medDPPM, which failed to converge. However, registration of sCBCTs to planning CT was poor, while original CBCTs registered successfully (figure 2).
blue = 1/5. Higher fractions indicate greater inter- observer agreement; lower values reflect single-observer inclusion. Conclusion: This DICOM-conformant, Python based workflow converts multi-observer RTSTRUCTs into FoI maps encoded as RTDOSE, which can be opened directly in routine radiotherapy systems without additional plugins. By providing an immediate, spatial visualisation of areas of agreement and disagreement, it complements conventional overlap and surface metrics and helps to focus discussion on clinically relevant anatomical differences. In multicentre settings, it can support protocol harmonisation, credentialing and quality assurance, and the development of consensus guidance. As a command- line Python utility, it could be scripted for multi-case use, although this was not assessed in the present work. References: 1. Kelly SM, Turcas A, Corning C, et al. Radiotherapy quality assurance in paediatric clinical trials: first report from six QUARTET-affiliated trials. Radiotherapy and Oncology. 2023;182. 2. Zöllner SK, Amatruda JF, Bauer S, et al. Ewing sarcoma- diagnosis, treatment, clinical challenges and future perspectives. J Clin Med.MDPI. 2021;10(8). 3. Van Der Walt S, Schönberger JL, Nunez-Iglesias J, et al. Scikit-image: Image processing in python. PeerJ. 2014;2014(1). 4. Harris CR, Millman KJ, van der Walt SJ, et al. Array programming with NumPy. Nature.Nature Research. 2020;585(7825):357-362. 5. Cleveland (OH): MIM Software. MIM Software Inc. MIM MaestroTM.6. Palo Alto (CA): Varian. Varian Medical Systems. VelocityTM. Keywords: RTSTRUCT, RTDOSE, consensus mapping, DICOM Reducing CBCT Imaging Dose in Image-guided Radiotherapy Using Generative Deep Learning Models Jeppe Fræhr Linderød 1 , Daniella Elisabet Østergaard 2 , Ivan Richter Vogelius 2,3 , Jens Petersen 1,2 1 Department of Computer Science, University of Copenhagen, Copenhagen, Denmark. 2 Department of Oncology, Copenhagen University Hospital – Rigshospitalet, Copenhagen, Denmark. 3 Department of Clinical Medicine, University of Copenhagen, Copenhagen, Denmark Purpose/Objective: Deep learning (DL)-based image generation offers new opportunities for low-dose cone-beam CT (CBCT) imaging in image-guided radiotherapy (IGRT). Digital Poster Highlight 1366
Made with FlippingBook - Share PDF online