S2084
Physics - Image acquisition and processing
ESTRO 2026
5 Department of Internal Medicine and Therapeutics, University of Pavia, Pavia, Italy
Purpose/Objective: In-room CBCT plays a central role in radiotherapy for daily assessment of anatomical variations at the treatment isocenter [1]. Its role is even more critical in particle therapy (PT), where steep dose gradients and range sensitivity require precise anatomical accuracy [2]. In this context, generating synthetic CT (sCT) images from CBCT using deep learning (DL) represents a promising approach to enable accurate dose estimation and plan adaptation. This study investigates the generalizability of DL models for pelvic sCT generation from CBCT acquired with a customized, limited Field of View (FOV) CBCT system. Material/Methods: Two DL models were evaluated, i.e. a 2.5D U-Net and a CycleGAN, which were pretrained on a full FOV dataset and subsequently fine-tuned on a limited FOV dataset. The full FOV dataset comprised 202 pelvic CBCT-CT volumes, including 171 from the SynthRAD 2023 challenge [3] and 31 from the Pelvic Reference Dataset [4]. Of these, 181 volumes were used for training and 20 for testing the pretraining performance. The limited FOV dataset consisted of 12 volumes from 9 gynecological patients acquired at National Center for Oncological Hadrontherapy (CNAO), with 9 volumes used for training and 3 for testing (660 axial slices). The test patients were treated with carbon ions. The generated sCTs were evaluated using both intensity and dose based metrics with respect to the correspondent re-evaluation CT. Due to the limited FOV, CBCTs were cropped and embedded in the planning CT for dose analysis. Results: Figure 1 shows a comparison between sCTs generated from the limited FOV dataset and the corresponding CBCT and planning CT, along with the associated dose distribution calculated on the sCT and re-evaluation CT. Quantitative results, for both datasets, are summarized in Table 1. The accuracy of the used test set was in agreement with other studies concerning sCT generation.Gamma analysis of dose distributions generated on re-evaluation CT and sCT yielded pass ratios of 99.79% for the 2.5D U-Net and 99.67% for the CycleGAN under 3%/3 mm criteria. DVH analysis demonstrated differences within clinical tolerances, with mean CTV D95% deviations of 0.04 Gy for the 2.5D U-Net and 0.27 Gy for the CycleGAN.
Conclusion: Our results demonstrate the good performance of DL networks in generating sCT from limited-FOV CBCT of the pelvic site, with the 2.5 U-Net achieving better results. These findings support the potential integration of such models into treatment planning adaptation workflows for PT with carbon ions. References: [1] C. Kurz, C. Hua, and G. Landry, “Conventional x-ray in-room imaging,” in Imaging in Particle Therapy, in 2053-2563. , IOP Publishing, 2024, pp. 5–1 to 5–18. doi: 10.1088/978-0-7503-5117-1ch5.[2] M. Durante, R. Orecchia, and J. S. Loeffler, “Charged-particle therapy in cancer: clinical uses and future perspectives,” Nat Rev Clin Oncol, vol. 14, no. 8, pp. 483–495, 2017, doi: 10.1038/nrclinonc.2017.30.[3] E. M. C. Huijben et al., “Generating Synthetic Computed Tomography for Radiotherapy: SynthRAD2023 Challenge Report,” Mar. 2024, [Online]. Available: http://arxiv.org/abs/2403.08447[4] A. A. Yorke, G. C. McDonald, D. Solis, and T. Guerrero, “Pelvic Reference Data,” The Cancer Imaging Archive. Keywords: Deep Learning, Synthetic CT, Dose evaluation Digital Poster 4043 Impact of SGRT optimization on IGRT rotational corrections in lung radiotherapy Giuseppina Rita Borzì 1 , Nina Cavalli 1 , Elisa Bonanno 1 , Andrea Girlando 2 , Martina Pace 1 , Sara Panebianco 1 , Giuseppe Stella 3 , Lucia Zirone 1 , Carmelo Marino 1 1 Medical Physics, Humanitas - Istituto Clinico Catanese, Misterbianco, Italy. 2 Radiotherapy, Humanitas - Istituto Clinico Catanese, Misterbianco, Italy. 3 Physics and
Made with FlippingBook - Share PDF online