S2061
Physics - Image acquisition and processing
ESTRO 2026
practice. Results:
The Pix2Pix model trained with the combination of in- phase and water-only MR (Figure 1) demonstrated the best performance, achieving a mean MAE of 330.9 HU within the implant region, 104.5 HU outside the implant region, and 119.4 HU across the entire body, with strong agreement in dose distribution. The DVH mean dose discrepancies (Figure 2) were below 1% for both the target volumes and OARs, with no outliers above 2%. Visual analysis confirmed that the DL models not only reliably generated soft tissue but were also capable of generating metal implant regions from MR areas with signal loss.
Digital Poster 1554
Widening applicability of MR-only radiotherapy: MR-based synthetic CT generation for patients with hip implants using deep learning methods Nico Camillo Zala 1,2 , Mariia Lapaeva 1,2 , Manuel Günther 2 , Nicolaus Andratschke 1 , Matthias Guckenberger 1 , Stephanie Tanadini-Lang 1 , Riccardo Dal Bello 1 1 Department of Radiation Oncology, University Hospital Zurich and University of Zurich, Zurich, Switzerland. 2 Department of Informatics, Artificial Intelligence and Machine Learning Group, University of Zurich, Zurich, Switzerland Purpose/Objective: Most MR-based synthetic CT (sCT) generation studies exclude patients with metal implants due to the challenges posed by signal loss and artefacts1. This study evaluates the feasibility of deep learning (DL) methods to generate sCT images from T1 Dixon MR images in patients with hip implants, which may widen
the applicability of MR-only radiotherapy to a population that can reach 10-15% in prostate treatments. Material/Methods:
A retrospective analysis was performed on 22 patients with pelvic tumors and hip implants treated at our institution. The cohort included two implant types: total hip replacements and intramedullary nails. The data was divided into a training set (18 patients, 1950 slices) and a test set (4 patients, 500 slices). A Pix2Pix generative adversarial network2 was trained on 2D axial MR–CT pairs using different combinations of MR reconstructions (In-phase, Opposed-phase, Water- only, Fat-only) to enhance sCT quality. The model performance was evaluated using mean absolute error (MAE) in three regions: (1) the implant region (implant mask, obtained by thresholding voxels above 2000 HU on CT, plus 2 cm margin); (2) the outside- implant region; and (3) the entire body. Additionally, the dosimetric accuracy was assessed by comparing photon treatment plans on sCT and CT using Dose- Volume Histogram (DVH) parameters for target volumes and organs-at-risk (OAR). The entrance dose through the implants was avoided to reflect clinical
Conclusion: These findings suggest that DL models, specifically Pix2Pix, are capable of generating accurate sCT images for patients with metal implants in the pelvic area. The results motivate further research into robust MR-only radiotherapy planning workflows that include anatomies previously considered challenging or unsuitable for DL-based sCT generation. References: 1. Fusella M, Alvarez Andres E, Villegas F, et al. Results of 2023 survey on the use of synthetic computed tomography for magnetic resonance Imaging-only radiotherapy: Current status and future steps. Phys Imaging Radiat Oncol. 2024;32:100652. doi:10.1016/j.phro.2024.100652 2. Isola P, Zhu JY, Zhou T, Efros AA. Image-to-Image Translation with Conditional Adversarial Networks. In: 2017 IEEE Conference on Computer Vision and Pattern
Made with FlippingBook - Share PDF online