ESTRO 2025 - Abstract Book

S2952

Physics - Image acquisition and processing

ESTRO 2025

893

Digital Poster Enhancing Mid-Field MR Images for Upright Treatments Using Diffusion Schrödinger Bridge Models Muheng Li 1,2 , Damian Manetsch 3 , Luisa Raimondo 4 , Niels Oesingmann 4 , Antony John Lomax 1,2 , Ye Zhang 1 1 Center for Proton Therapy, Paul Scherrer Institute, Villigen, Switzerland. 2 Department of Physics, ETH Zürich, Zürich, Switzerland. 3 Department of Information Technology and Electrical Engineering, ETH Zürich, Zürich, Switzerland. 4 ASG Superconductors spa, Genoa, Liguria, Italy Purpose/Objective: Upright radiotherapy presents numerous potential benefits, especially for proton and particle treatment [1]. Although MR provides superior soft-tissue contrast for imaging patients, existing upright MR scanners operate at mid-field strengths (0.5/0.6T), resulting in images with substantial noise that compromises the quality of treatment planning. Building on our prior work with Diffusion Schrödinger Bridge Models (DSBM) [2], this study focuses on enhancing mid-field MR image by utilizing the DSBM framework. This approach holds significant promise for denoising, especially for upright MR imaging where maintaining high image quality is crucial for accurate treatment planning. Material/Methods: We acquired paired MR images of the head-neck region for 6 volunteers (ethical approved by EKNZ202300674) in supine position at 0.5T (MROpen Evo scanner, ASG Superconductors Spa, Genova, Italy) and 1.5T (MAGNETOM Aera, Siemens Healthineers, Germany), along with the corresponding mid-field upright MR images. The workflow consists of three main stages (Fig. 1): First, the training data set was prepared by registering mid/high-field supine MR image pairs using either conventional rigid registration or our in-house developed CPT-DIR approach [3] for data alignment (Fig. 1a). Second, the DSBM model was trained and validated on these aligned supine image-pairs through k-fold cross validation, with the best-performing model selected based on MAE scores (Fig. 1b). To ensure fair quality assessment, the resulting synthetic high-field supine MR images were registered to ground-truth (GT) high-field supine MR using CPT-DIR, thereby eliminating residual misalignments. Third, the validated model was applied to mid-field upright MR images, generating synthetic high-field upright MR images. (Fig. 1c). SSIM (structural similarity index measure) and MAE (mean absolute error) were used as quantitative evaluation for the supine scenarios within body-masked regions and full images. The DSBM model was then applied to the upright mid-field MR images.