Author: Lando S. Bosma, Victoria Brennan, Nicolas Cote, ChengCheng Gui, Nima Hassan Rezaeian, Jue Jiang, Sudharsan Madhavan, Josiah Simeth, Neelam Tyagi, Harini Veeraraghavan, Michael J Zelefsky 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, NYU Langone Health, University Medical Center Utrecht, Memorial Sloan Kettering Cancer Center 🌍
Purpose: Deep learning-based deformable image registration (DIR) models often lack robustness when applied to datasets with differing imaging characteristics. We aimed to (1) improve registration network performance by integrating a segmentation subnetwork to regularize training and (2) evaluate cross-domain generalization using prostate MRI data from two MR platforms: a 1.5T MR-Linac (MRL) and a 3T MR-simulation (MRSim) system.
Methods: The study analyzed two types of MR images: 3T T2-weighted MRSim images from 34 patients (320 pairs) and 1.5T MR-Linac images from 95 patients (360 pairs). For same-domain evaluation, the MRSIM dataset was split into 262 training and 58 testing pairs, while the MRL dataset was split into 288 training and 72 testing pairs. For cross-domain testing, we evaluated 42 MRL pairs using the model trained on MRSIM data. Images were preprocessed using standardized preprocessing with intensity normalization and resampling to 192×192×128 dimensions. A previously published ProRSeg method combining segmentation subnetwork to regularize registration network training was used to align these MRI pairs. We evaluated registration performance against SyN algorithm and Evolution methods across both same-domain and cross-domain scenarios.
Results: ProRSeg achieved mean DSC values of 0.88, 0.86, and 0.85 for bladder, rectum, and prostate respectively on MRSIM pairs, with HD95 < 3mm. In cross-domain performance, the MRSIM-trained model maintained statistically similar performance on MRL-to-MRL registration (DSC: 0.87, 0.89, 0.89). Without segmentation supervision, performance dropped significantly (p<0.00001) on cross-domain pairs (DSC: 0.70, 0.69, 0.73; HD95: 4.28-8.35mm). ProRSeg outperformed conventional methods (SyN, Evolution) across all structures, while the MRL-MRL model provided benchmark DSC values of 0.91, 0.88, and 0.88.
Conclusion: Integration of segmentation supervision during DIR network training enables robust cross-domain generalization, achieving median accuracy of 1.3-2.4 mm for prostate, bladder, and rectum across 1.5T and 3T MRI systems. This advances the clinical feasibility of deep learning DIR for MR-guided adaptive radiotherapy.