ESTRO 2025 - Abstract Book

S2980

Physics - Image acquisition and processing

ESTRO 2025

variance differences between reconstruction algorithms, patient, fraction, and spatial direction variables showing no significance.

Conclusion: The optimized pCBCTs were successfully RED calibrated and validated for IGRT and planning for LARC radiotherapy. The results of this study suggest that the use of pCBCTs for LARC adaptive radiotherapy could be implemented in a clinical workflow although more investigation for their extended use might still be necessary. Further investigation could explore the possibility offered by Polyquant of directly reconstructing in RED, overcoming the pop-CC calibration.

Keywords: CBCT, dose calculation, IGRT

1708

Digital Poster Deep Learning method for Effective Atomic Number Extraction from Photon-Counting Dual-Energy CT in Radiation Treatment Planning Kihong Son 1 , Daehong Kim 2 1 Medical Information Research Section, Electronics and Telecommunications Research Institute, Daejeon, Korea, Republic of. 2 Department of Radiological Science, Eulji University, Seongnam, Korea, Republic of Purpose/Objective: This study aims to enhance the extraction of effective atomic numbers (EAN) from photon-counting detector (PCD)- based dual-energy CT images using a deep learning (DL) method to improve dose accuracy, optimize outcomes, and minimize side effects in radiation treatment planning. Traditional analytical models, such as Rutherford et al. and stoichiometric calibration, are not only time-consuming, requiring minutes to process, but also have limitations in accuracy, occasionally resulting in errors in EAN estimation [1-3]. The proposed DL method focuses on improving accuracy while significantly reducing processing time. Material/Methods: Dual-energy scans of a phantom containing tissue-equivalent materials were performed using a PCD-based CT. The materials were segmented, and a thin plate spline-based nonlinear warping transformation was applied to generate 100,000 synthetic images of various shapes for training. The DL architecture utilized a CNN-based UNET-like model, which effectively integrates spatial and contextual information through an encoder-decoder structure with skip connections. This design enables precise differentiation of tissue materials in CT images by combining fine details with high-level features. Results: The EAN extracted using the proposed DL method was benchmarked against analytical methods (Rutherford et al. and stoichiometric calibration). Figure 1(a) illustrates the phantom setup with eight tissue-equivalent materials. Figure 1(b) compares EAN images obtained using Rutherford et al., stoichiometric calibration, and DL, alongside the reference EAN image. The mean errors for Rutherford et al., stoichiometric calibration, and DL were 3.59%, 4.63%, and 0.72%, respectively. The maximum error was 19.86% (Rutherford et al.) and 24.64% (stoichiometric calibration) in the liver, whereas DL exhibited a maximum error of 1.75% in the inner bone. The structural similarity index measure (SSIM) values were 0.536, 0.534, and 0.950 for Rutherford et al., stoichiometric calibration, and DL, respectively. The DL method significantly reduced computation time, extracting EANs in under 1 second compared to 6 minutes for analytical methods (512×512 pixels, 3.20 GHz CPU).