Author: Ming Chao, Karyn A Goodman, Yang Lei, Tian Liu, Jing Wang, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai 🌍
Purpose: Real-time volumetric MRI is essential for motion management in MRI-guided radiotherapy (MRIgRT), yet acquiring high-quality 3D images remains challenging due to time constraints and motion artifacts. This study proposes a generalized deep learning framework to reconstruct dynamic 3D volumetric MRI from 2D Cine MRI, incorporating spatial priors and motion correction to ensure robust reconstruction across multi-parametric MRI sequences.
Methods: The proposed method utilizes implicit neural representation (INR) learning combined with spatial coordinate priors and motion correction embeddings to reconstruct dynamic 3D volumetric MRI. Multi-parametric abdominal MRIs (T1-weighted, LAVA, VIBE, SSFSE) were collected at multiple time points to capture patient-specific motion patterns. 2D Cine MRI sequences, simulated from these MR datasets, mimicked real-time imaging conditions. The motion-aligned INR model reconstructed the 3D MRI from 2D cine. Five-fold cross-validation was performed with multi-parametric MRIs in the training folds used for training and the first two daily MRIs in the testing fold for fine-tuning. Quantitative metrics, including normalized mean absolute error (NMAE), normalized cross-correlation (NCC), and target error (TE), were used to evaluate performance.
Results: The framework demonstrated robust reconstruction across all tested MRI modalities. Ground truth 3D MRI data collected from various modalities enhanced the model’s generalizability. Quantitative analysis showed mean NMAE = 0.15±0.08, NCC = 0.89±0.15, and TE = 2.0±1.3 mm. Visual comparisons of reconstructed volumetric MRIs with ground truth images revealed strong spatial coherence and preserved intensity profiles. Line profiles and histograms further confirmed the accuracy of reconstructed images. The model achieved reconstruction speeds of 10–20 frames per second, comparable to 2D Cine MRI acquisition rates.
Conclusion: This study demonstrates a generalized framework for real-time volumetric MRI reconstruction from 2D Cine MRI. By incorporating spatial priors and motion correction, the proposed method offers robust, accurate, and clinically applicable dynamic MRI for improved motion management in radiotherapy.