Author: Karyn A Goodman, Yang Lei, Tian Liu, Charlotte Elizabeth Read, Jing Wang, Qian Wang, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology, Beth Israel Deaconess Medical Center 🌍
Purpose: Accurate motion management in MRI-guided radiotherapy (MRIgRT) relies on real-time volumetric MRI to track intra-fractional anatomical changes. Dense k-space sampling, while capable of producing high-resolution images, is inherently slow and impractical for real-time monitoring. On the other hand, sparse k-space sampling significantly accelerates acquisition but often results in image artifacts and loss of critical anatomical details due to the reduced data density. To address these challenges, this study introduces an ultra-sparse k-space reconstruction method that combines motion decomposition and sinusoidal representation networks (SIREN). It leverages patient-specific motion patterns to reconstruct high-quality volumetric MRI from minimal sampling, enabling real-time imaging with both speed and accuracy.
Methods: A novel reconstruction framework integrates patient-specific motion priors into SIREN by combining spatial coordinate priors and motion decomposition. High-quality 4D MRIs were generated by deformably registering diagnostic T1-weighted MRIs to pretreatment 4D CTs, capturing patient-specific respiratory motion. Ultra-sparse k-space data (sampling rate 0.05) were simulated to mimic per frame scanning time during treatment. Low-rank motion decomposition extracted dominant abdominal motion components, which, along with spatial priors, were embedded into SIREN for 3D volumetric reconstruction. Performance was assessed on 10 pancreatic cancer patients using normalized mean absolute error (NMAE), normalized cross-correlation (NCC), and tumor target registration error (TRE).
Results: The proposed method demonstrated robust reconstruction with mean NMAE = 0.12±0.11, NCC = 0.90±0.13, and TRE = 1.8±0.6 mm. Ultra-sparse k-space sampling enabled a 20-fold reduction in scanning time compared to full sampling while maintaining clinically acceptable volumetric detail. Line profiles showed strong contrast agreement between reconstructed and full sampling MRIs.
Conclusion: By embedding patient-specific motion information and spatial coordinate priors, this method effectively addresses the challenges of sparse data reconstruction while maintaining accuracy and robustness. It significantly reduces imaging sampling rate, potentially enabling real-time motion tracking during MRIgRT.