ESTRO 2025 - Abstract Book

S3273

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

2628

Poster Discussion Enhancing MRI-Only Radiotherapy: Overcoming Dental Artifacts with 3D GAN-Based Synthetic CT Generation Blanche Texier 1 , Saoussen Gharsallah 1 , Songyue Han 1 , Cédric Hémon 1 , Caroline Lafond 2,1 , Renaud de Crevoisier 3,1 , Joël Castelli 3,1 , Anaïs Barateau 2,1 , Jean-Claude Nunes 1 1 LTSI-UMR 1099, University of Rennes, Rennes, France. 2 Dept. of Medical physics, Centre Eugène Marquis, Rennes, France. 3 Dept. of Radiation Oncology, Centre Eugène Marquis, Rennes, France Purpose/Objective: MRI-only workflows are increasingly adopted in radiation therapy for brain and head-and-neck cancers to improve soft tissue contrast for a more accurate radiotherapy. Since MRI intensities do not directly correspond to Hounsfield Units (HU) required for dose calculation, synthetic CT (sCT) generation via Deep Learning (DL) has emerged as a solution. Most of them are based on Generative Adversarial Networks (GANs) [1]. However, this sCT generation can be challenging in patients with dental artifacts. They can disrupt MR signals, leading to inaccurate sCT generation in these areas and compromised dose planning accuracy. Addressing this issue is critical to allow MRI-only radiotherapy for these patients. The aim of this study was to improve sCT generation in the presence of dental artifacts by comparing the performance of three supervised GAN-based architectures. Material/Methods: This study used 112 paired MR/CT datasets from brain cancer patients, with T1-weighted MRIs acquired on a 1.5 T GE Optima MR450 and CTs on a Siemens SOMATOM Confidence. Over half exhibited dental work, presenting as high HU on CT and signal voids on MRI, variably distributed across one or both sides. sCTs were generated using supervised GAN-based architectures: a 2D conditional GAN (cGAN) with rigid registration for training data, and two 3D cGANs [2], with training data processed via rigid or deformable registration [3] respectively. Models were trained via 4-fold cross validation with a 78/6/28 patient split for training, validation, and testing. The synthesis accuracy was assessed by computing the mean absolute error (MAE) between sCTs and original CTs HU in the dental artifacts, brain, dilated brain, soft tissues, and high densities (above 600 HU). The dilated brain contour was defined as the brain with a 3 cm extension, inside the body contour. Results: Figure 1 and Table 1 summarize the results, showing that the 2D network produced the poorest synthesis (higher MAE and larger “low signal areas”) compared to 3D networks. 3D networks either with rigid or deformable registration achieved similar results.