S2070
Physics - Image acquisition and processing
ESTRO 2026
computed tomography. (arXiv:2507.22953).https://arxiv.org/abs/2507.229534. J Wasserthal et al. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence, 5(5), e230024.https://doi.org/10.1148/ryai.2300245. https:// github.com/mic-dkfz/autopet-3- submission6. https://github.com/hakal104/autoPETIII/ 7. https://github.com/YazdanSalimi/Lesion- Segmentation Keywords: Oligometastatic disease, Longitudinal FDG- PET/CT Feasibility of Deep Learning-Based Synthetic CT Generation for MRI-Only Brain Radiotherapy Planning Sharad Singh 1 , Sumanta Manna 2 , Rumita Singh 1 , Pramod Kumar Gupta 1 , Ragul T 2 1 Radiation Oncology, Kalyan Singh Super Specialty Cancer Institute, Lucknow, India. 2 Medical Physics, Kalyan Singh Super Specialty Cancer Institute, Lucknow, India Purpose/Objective: MRI-only radiotherapy planning eliminates the need for separate CT acquisition by generating synthetic CT (sCT) images from MRI data for dose calculation. This study evaluates the clinical feasibility and dosimetric Digital Poster 3081 accuracy of a commercial deep learning–based Synthetic sCT generation solution for brain radiotherapy by comparing dose distributions between plans computed on sCT and conventional planning CT (pCT). Material/Methods: Thirty brain tumor patients treated with RapidArc were prospectively included. MRI and pCT were acquired for each. MRI was performed on a 3T Siemens Magnetom Vida RT Pro Edition using RT-optimized 3D T1 VIBE- Dixon (in- and opposed-phase) sequences acquired with two 4-channel large flex coils, three-clamp thermoplastic immobilization, flat tabletop setup, and room laser-based positioning. sCT images were generated using a deep learning–based conditional Generative Adversarial Network (cGAN) algorithm integrated in the syngo.via RT Image Suite. The sCT series was rigidly co-registered with the pCT images in the Eclipse treatment planning system. All structures delineated on the pCT were transferred to the sCT dataset, and each patient’s clinical plan was recalculated on the sCT. Dosimetric agreement between pCT- and sCT-based dose plans was evaluated by comparing dose–volume histogram (DVH) parameters for the planning target volume (PTV) and organs at risk (OARs). The mean dose differences
source Python-based tool, supports image preprocessing, filtering, and radiomics extraction based on DICOM and NIfTI formats (CT, PET, MRI, mammography, and 3D dose distribution). A normal anatomy segmentation model, CADS [3], was validated on 2,864 planning CT scans with corresponding ROIs, showed superior segmentation quality and a greater number of segmented structures when compared to TotalSegmentator [4]. Existing lesion segmentation models such as PET-Assisted Reporting System (“PARS”, Siemens Healthineers), the best AutoPET 2024 submissions [5, 6], and other models [7] showed unsatisfactory results when tested on heterogeneous out-of-distribution data, leading to the collaboratively developed model, GLOW-FDG, trained on the diverse dataset, which achieved the highest overall performance across different cancer types among evaluated lesion segmentation models.
Conclusion: The temporal OMD FDG-PET/CT dataset was collected and preprocessed using the presented tools. Further integration of clinical data remains a major task. References: 1. M. Fritsak et al. Technical Note: Vendor-Specific Approach for Standardized Uptake Value Calculation (arXiv:2410.13348). https://doi.org/10.48550/arXiv.241 0.133482. https://github.com/medical-physics-usz/z- rad3. M. Xu et al. CADS: A comprehensive anatomical dataset and segmentation for whole-body anatomy in
Made with FlippingBook - Share PDF online