ESTRO 2025 - Abstract Book

S2969

Physics - Image acquisition and processing

ESTRO 2025

devices were performed to prepare the training (121 cases) and testing dataset (10 cases). Two deep learning models based on conditional GAN framework were developed: 1) traditional cGAN model built with CNN based 9 block Resnet; 2) improved cGAN model built with Vision transformer block. MRI to CT synthesizing was performed using both the two models. IMPT (Intensity modulated proton therapy) plans were firstly designed using Raystation treatment planning system (non-clinical version) based on the planning CT and then transferred to the two sets of sCT. Proton dose distributions were recalculated based on the sCT and then compared with those on the ground truth planning CT. Results: From the aspects of sCT image quality, compared with the traditional CNN based model, the proposed VIT based model can improve MAE from 41.57±3.71 to 37.23±2.45 ， PSNR from 36.92±0.97 to 37.99±0.73, SSIM from 0.922±0.016 to 0.926±0.015. As to the dosimetric comparison, the global 3D gamma passing rate (criteria of 1mm/1%, dose threshold: 10%) of sCT by VIT based cGAN model was 96.56±1.01, compared with that of 94.70±1.26 for the traditional CNN based cGAN model. Conclusion: The proposed VIT based cGAN model can generate sCT with higher accuracy than the traditional CNN based model, and thus it is feasible for dose calculations in proton therapy treatment planning of nasopharyngeal carcinoma. Digital Poster Motion detection accuracy of dual-energy X-ray imaging for lung cancer tracking: an experimental phantom study Nawal Alqethami 1,2 , Wentao Xie 1 , Prasannakumar Palaniappan 1 , Felix Ginzinger 3 , Philipp Steininger 3 , Marco Riboldi 1 1 Department of Medical Physics, Ludwig Maximilians Universität, Munich, Germany. 2 Research and development, Brainlab AG, Munich, Germany. 3 Research and development, medPhoton GmbH, Salzburg, Austria Purpose/Objective: This study evaluates the accuracy of Dual-Energy (DE) X-ray imaging to measure lung tumour motion, as a function of variable tumour size and different imaging protocols. Material/Methods: We developed an in-house experimental setup using pork ribs mounted on a cylinder replicating the human torso. A stepper motor configured to imitate a breathing cycle with 1.75 cm peak-to-peak amplitude in a 3.6 s period was used, with a holder for artificial tumours. Six spherical tumours made of tissue-equivalent material were moulded in different sizes (0.5 to 3.0 cm). Each underwent DE planar scans with fast kV switching (300 frames at 11.9 fps) using 60 kV/17 ms paired with 120 kV/12 ms at two different mA settings (15 and 20 mA). We applied deformable image registration to correct the tumour motion between the low energy (LE) and high energy (HE) sequential frames. A series of weighting factors (w st ) (0.55 to 0.7), in increments of 0.01, were tested to perform the weighted logarithmic subtraction (WLS), aiming to suppress the bones and enhance the tumour. The contrast-to-noise ratio (CNR) was calculated to assess the image quality by calculating the CNR between the tumour segmentation and the background. Ground truth (GT) motion was established by measuring the distance between two manually segmented tumour positions, corresponding to selected frames at the extreme of the longitudinal motion trajectory. We relied on template matching (TM) to measure the peak-to-peak tumour motion in both single-energy and DE images, by defining a template around the tumour at the first peak. Keywords: Deep learning; Synthetic CT; Proton therapy. 1364