Author: Karyn A Goodman, Yang Lei, Tian Liu, D. Michael Lovelock, Charlotte Elizabeth Read, Jing Wang, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai 🌍
Purpose: Real-time volumetric MRI is essential for precise motion management in MRI-guided radiotherapy (MRIgRT). While 2D Cine MRI offers high temporal resolution for motion tracking, it inherently lacks volumetric information, limiting its ability to capture out-of-plane motion and anatomical changes during treatment. We propose a deep learning framework to reconstruct real-time dynamic volumetric MRI from 2D Cine MRI. The method leverages patient-specific 4D MRI as prior information to incorporate patient-specific motion patterns.
Methods: Patient-specific 4D MRIs were generated by deformably registering diagnostic T1-weighted MRIs to pretreatment 4D CT, capturing patient-specific respiratory motion. 2D Cine MRI datasets, simulated at the tumor isocenter, were extracted from the 4D MRI to mimic real-time imaging. A patient-specific deep learning model utilized 4D MRI to encode motion patterns into spatial priors and motion correction embeddings. These priors guided an implicit neural representation (INR) to estimate volumetric MRI from 2D Cine MRI in real time, enabling precise motion management. The model was trained on 4D MRI data using five-fold cross-validation and fine-tuned on diagnostic MRI before reconstructing volumetric MRI from simulated 2D Cine MRI. The accuracy of reconstructed volumes was evaluated using normalized mean absolute error (NMAE), normalized cross-correlation (NCC), and tumor target registration error (TRE).
Results: The proposed method achieved a mean NMAE = 0.12±0.08, NCC = 0.91±0.09, and TRE = 1.7±0.4 mm, validating the model's capability for accurate and consistent volumetric reconstruction. Visual comparisons of reconstructed volumetric MRIs with ground truth images across different respiratory phases demonstrate its ability to track soft-tissue intra-fractional motion. Line profiles and histograms analysis further confirmed the fidelity of reconstructed images.
Conclusion: This study presents a novel framework for reconstructing dynamic volumetric MRI from 2D Cine MRI using patient-specific 4D MRI motion priors. This method addresses the limitations of 2D Cines, enabling real-time volumetric motion monitoring.