ESTRO 2025 - Abstract Book

S2950

Physics - Image acquisition and processing

ESTRO 2025

planning computed tomography (pCT) images, to generate on-demand synthetic CBCT (sCBCT) images—aimed to facilitate rapid clinical decision making.

Material/Methods: Image data from 392 prostate cancer patients treated at our institution between 2019 and 2024 were retrospectively collected. Patients received either 70 Gy in 35 fractions to the prostate bed or 36 Gy in 6 fractions to the prostate. In total, 3075 pre-treatment CBCT and 392 pCT- images were utilized for method development. Digitally reconstructed radiographs (DRRs) were generated from each CBCT at 0° and 270° to mimic the appearance of X-ray biplanar images acquired on our treatment machines. Three DL-models—a base model, one with an added skip-connection, and one with both a skip-and residual connection—were developed to identify the optimal architecture for our task (Figure 1). Each model, combined DRRs and pCT-images to generate sCBCT-images that mimicked pre-treatment CBCT. The training, validation, and test datasets included 358, 30, and 4 individual patients, respectively. During training, each DL-model utilized a combination of loss functions, including mean square error (MSE) and a perceptual loss (PL), based on a pre-trained VGG19-model. The models’ performance was evaluated using MSE, PL, structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and visual inspection.