ESTRO 2025 - Abstract Book

S2951

Physics - Image acquisition and processing

ESTRO 2025

Results: The average SSIM and PSNR values of the best model for training, validation, and test datasets are (0.89, 32.98), (0.87, 30.06), (0.87, 30.90), respectively. The skip-and-residual DL-model provided the most clinically relevant anatomy representation (Figure 2) with inference times below 7 ms. The remaining models could not capture anatomical changes effectively.

Conclusion: In this work, we developed a novel DL-model that has potential for daily 3D anatomical verification using information from planning CT and daily 2D images. Our results could be applied to decrease pre-treatment verification time, by enabling intrafractional monitoring of anatomical changes while reducing radiation exposure to patients.

Keywords: Deep learning, Synthetic, Motion management