S2449
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
ICC of 0.63 (MRPD=0.6%, MARPD=31.2%), highlighting strong agreement between ground-truth and synthetic PET. Visual analysis of synthetic PET generated from treatment planning CTs (Fig. 1B) showed that synthetic PET-active areas are embedded within the GTV for every patient.
Digital Poster Highlight 2760
Proof of concept: synthetic FDG-PET generation from standard CT imaging for patients with head and neck cancer Maksym Fritsak 1 , Hubert S. Gabry ś 1 , Riccardo Dal Bello 1 , Sebastian M. Christ 1 , Panagiotis Balermpas 1 , Martin Huellner 2 , Matthias Guckenberger 1 , Stephanie Tanadini-Lang 1 1 Department of Radiation Oncology, University Hospital Zurich, Zurich, Switzerland. 2 Department of Nuclear Medicine, University Hospital Zurich, Zurich, Switzerland Purpose/Objective: While FDG-PET/CT is becoming a standard modality for head and neck cancer (HNC) diagnosis, global PET scanner availability remains limited, hindering optimal clinical practices from developing countries. To democratize PET access, we propose a model that generates synthetic PET from CT images. Synthetic PET can guide lesion delineation and provide imaging biomarkers like total lesion glycolysis (TLG) or metabolic tumor volume (MTV). This project expands the work of Salehjahromi et al. [1], by introducing synthetic FDG-PET generation from CT for head and
neck malignancies. Material/Methods:
The proposed deep learning model is a 3D adaptation of the pix2pix GAN [2], with a ResNet++ [3] generator and a PatchGAN [2] discriminator. Training was performed for 300 epochs with a batch size of 1 and a learning rate of 2e-4 for both generator and discriminator. For model training, we used 519 pre- treatment FDG-PET/CT scans of HNC patients from the publicly available HECKTOR 2022 dataset [4], collected across seven hospitals. For validation, we used 91 FDG-PET/CT scans comprising 224 cancer lesions, including both primary tumors and nodal metastases, from the Department of Nuclear Medicine of the University Hospital Zurich. Lesion-level validation metrics included mean SUV, TLG, MTV above 1.5 SUV (MTV1.5) and above 2.5 SUV (MTV2.5). Agreement was assessed using the intra-class correlation coefficient (3,1) (ICC), median relative percentage difference (MRPD), and median absolute relative percentage difference (MARPD) for each parameter. We also tested model performance on 10 treatment planning CTs to evaluate whether the model is capable of a more generalized CT-to-PET conversion. Results: Examples of synthetic PET generation are shown on Fig. 1A. Agreement (Fig. 2) was highest for MTV1.5, with an ICC of 0.98 (MRPD=–0.1%, MARPD=7.9%). MTV2.5 showed an ICC of 0.95 (MRPD=–1.7%, MARPD=29.6%), followed by TLG with an ICC of 0.93 (MRPD=0.6%, MARPD=31.2%), and mean SUV with an
Made with FlippingBook - Share PDF online