S2282
Physics - Machine learning and AI algorithms
ESTRO 2026
methods. Material/Methods:
with the digital twin, the mean absolute percentage error between target and experimental FLASHKNiFE output ranges from 12.07 to 27.12 %. Conclusion: A first version of the digital twin was developed and provides a framework for configuring FLASHKNiFE to meet user-specific requirements. Future enhancements, particularly through enlarging the training dataset, are expected to increase model accuracy. References: [1] Colnot J. et al. Commissioning and performance evaluation of the new electron UHDR FLASHKNiFE® system for FLASH radiation therapy. Physica Medica 135 (2025): 105014. https://doi.org/10.1016/j.ejmp.2025.105014.[2] Kacem H. et al. Modification of the microstructure of the CERN-CLEAR-VHEE beam at the picosecond scale modifies ZFE morphogenesis but has no impact on hydrogen peroxide production. Radiotherapy and Oncology 209 (2025): 110942. https://doi.org/10.1016/j.radonc.2025.110942.[3] Marinelli M. et al. Design, realization, and characterization of a novel diamond detector prototype for FLASH radiotherapy dosimetry. Medical Physics 49(3) (2022): 1902–1910. https://doi.org/10.1002/mp.15473. Keywords: Surrogate model, Particle Accelerator, Dosimetry Digital Poster 607 Missing tissue generation in CT scans using a transformer-based AI model Mojtaba Behzadipour 1 , Lulin Yuan 2 , Ford Sleeman 2 , Ryan Wargo 2 , Mitchell Polizzi 2 , Siyong Kim 2 1 Radiation Oncology, Virginia Commonwealth University Health System, Richmond, USA. 2 Radiation Oncology, Virginia Commonwealth University, Richmond, USA Purpose/Objective: Missing tissue at the periphery of CT images, known as out-of-field artifact or truncation, occurs when patient anatomy extends beyond the scanner’s field of view, leading to incomplete data and inaccuracies in dose calculation. This study aims to reconstruct these missing regions with anatomical fidelity using TFill, a transformer-based completion network. Given the bilateral symmetry commonly observed in axial anatomy, TFill leverages global contextual interactions between contralateral structures through attention mechanisms and combines them with HU-based tissue decomposition to achieve anatomically consistent and visually realistic reconstruction compared with conventional CNN- or GAN-based
The TFill framework comprises two sequential sub- networks: TFill-Coarse, which captures global context via a transformer encoder, and TFill-Refined, which restores high-frequency detail through an attention- aware layer. Initially, the model was trained on full HU images without decomposition. Observing that bone reconstruction was weak, we introduced HU-based decomposition and trained separate models for bone and soft tissue to strengthen learning in bone regions. The NSCLC-Radiomics dataset (422 patients) was used, with 400 for training and 22 for testing. All images underwent couch removal, resizing (512 to 542), and random noise augmentation. Soft tissue masks were applied at lateral edges, while bone masking used a probability distribution derived from training data. Models were optimized with Adam ( β ₁ =0.5, β ₂ =0.9, batch=24) and distinct hyperparameters for bone and soft tissue. Four configurations of Attention-Aware Layers (AAL) were tested in the refinement stage.
Results: Both soft tissue and bone models demonstrated stable learning behavior and anatomically coherent reconstructions. Learning rate trends across coarse and refinement stages showed consistent convergence for all four Attention-Aware Layer (AAL) configurations, confirming effective optimization. Example images from the original, masked, and refinement-stage reconstructed CT slices for both soft tissue and bone are shown in Figure 2.
Made with FlippingBook - Share PDF online