ESTRO 2025 - Abstract Book

S3031

Physics - Image acquisition and processing

ESTRO 2025

Keywords: MRI, distortion, stereotactic frame

References: [1] Pappas EP, Seimenis I, Kouris P, et al. Target localization accuracy in frame-based stereotactic radiosurgery: Comparison between MR-only and MR/CT co-registration approaches. J Appl Clin Med Phys 2022;23:e13580. https://doi.org/10.1002/acm2.13580. [2] Pappas EP, Seimenis I, Dellios D, et al. Assessment of sequence dependent geometric distortion in contrast enhanced MR images employed in stereotactic radiosurgery treatment planning. Phys Med Biol 2018;63:135006. https://doi.org/10.1088/1361-6560/aac7bf. [3] Dellios D, Pappas EP, Seimenis I, et al. Evaluation of patient-specific MR distortion correction schemes for improved target localization accuracy in SRS. Med Phys 2021;48:1661–72. https://doi.org/10.1002/mp.14615.

2984

Digital Poster Development of a deep learning deformable image registration algorithm for adaptive radiotherapy of head and neck cancer Donghoon Lee 1 , TreeChariusame Teeradon 2,3 , Yu-Chi Hu 1 , Peng Zhang 1 , Jung Hun Oh 1 , Michalis Aristophanous 1 , Laura Cerviño 1 , Pengpeng Zhang 1 1 Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. 2 Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, New York, USA. 3 Department of Radiology, Faculty of Medicine Siriraj Hospital, Bankok, Thailand Purpose/Objective: Deformable Image Registration (DIR) is vital for adaptive radiotherapy (ART), yet its often poor performance and lengthy computation time limit its clinical application. We developed a DIR algorithm that integrates longitudinal analysis with deep learning techniques, enhancing both the accuracy and efficiency of DIR and enabling timely adjustments during treatment. Material/Methods: We utilized data from sixty head and neck cancer patients treated under an ART protocol, with fifty patients designated for training and cross-validation and ten for testing. Each patient dataset included planning CTs and weekly CBCTs. We employed a deep learning framework, named Seq2Morph, which integrates Convolutional Neural Networks and Recurrent Neural Networks to analyze the spatiotemporal characteristics of the longitudinal imaging data. Seq2Morph, sequentially processes each weekly CBCT to capture deformation patterns throughout the radiotherapy course. The final outputs are sequential deformation vector fields. We utilized a loss function that maximized normalized cross-correlation and enforced inverse consistency and regularization. For evaluation, we measured the structural similarity index (SSIM) by comparing the deformed planning CT to CBCT, assessed deformation accuracy using dice scores and Hausdorff distances (both average, HD_avg, and 95%, HD_95) between propagated and manual contours for the primary GTV, parotid, and submandibular glands, and compared the performance of Seq2Morph against the widely used B-Spline algorithm .