Deep Learning Reconstruction for Accelerated Radial CEST MR Fingerprinting
Abstract
Purpose: To accelerate chemical exchange saturation transfer magnetic resonance fingerprinting (CEST-MRF) using temporal deep learning reconstruction. Methods: Ten healthy volunteers underwent repeated test-retest golden-angle radial CEST-MRF acquisitions with 30 measurements and 34 spokes per measurement. The Movienet network was adapted to reconstruct accelerated radial CEST-MRF data. Separate models were trained using 1, 2, 4, 8, 16, 20, and 34 spokes per measurement. Temporal and coil correlations were exploited by concatenating time and coil dimensions in the input. GRASP-based locally low-rank (GRASP-LLR) reconstruction was used for comparison. Quantitative DRONE fitting generated tissue parameter maps. Movienet-8 was evaluated with leave-one-subject-out analysis. Prospectively acquired 8-spoke datasets were compared with paired 34-spoke acquisitions, and Movienet-8 and Movienet-34 were compared with a 64-spoke acquisition. Results: In the held-out retrospective series, median Pearson correlations between Movienet and GRASP-LLR were 0.795, 0.947, 0.974, 0.977, and 0.981 for 1, 8, 16, 20, and 34 spokes, respectively. Corresponding NRMSE values were 14.7, 8.2, 5.9, 5.3, and 5.1 percent. Across 10 leave-one-subject-out volunteers, median image correlation was 0.965 and NRMSE was 7.5 percent. Reconstruction of a 30-slice series required 1.7 s with Movienet versus approximately 45 min with GRASP-LLR. Conclusion: Movienet enables whole-brain radial CEST-MRF acquisition in 3 min, with image reconstruction in less than 2 s.
Related
- Unimodality and Radial Monotonicity of the Magnetic Resonance Fingerprinting T1/T2 Matching Objective
- Comparison of spectral performance of three dual-energy CT scanners equipped with a deep-learning image reconstruction algorithm and one photon counting CT scanner: A phantom study
- Optimising Geometric Deep Learning for Varied Reconstruction Tasks in High Granularity Calorimeters
- Monocular 3D bubble reconstruction: DNS-informed ray tracing and deep learning
- Prior-Guided Physics-Informed Deep Learning for Fast 4D MRI Reconstruction
BibTeX
@misc{cheema2026deep,
title = {{Deep Learning Reconstruction for Accelerated Radial CEST MR Fingerprinting}},
author = {Karandeep Cheema and Victor Murray and Ricardo Otazo and Ouri Cohen},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.23805},
eprint = {2610.23805},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.23805},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - Deep Learning Reconstruction for Accelerated Radial CEST MR Fingerprinting
AU - Karandeep Cheema
AU - Victor Murray
AU - Ricardo Otazo
AU - Ouri Cohen
PY - 2026
DA - 2026/10/16
PB - arXiv
DO - 10.48550/arxiv.2610.23805
UR - https://arxiv.org/abs/2610.23805
AB - Purpose: To accelerate chemical exchange saturation transfer magnetic resonance fingerprinting (CEST-MRF) using temporal deep learning reconstruction. Methods: Ten healthy volunteers underwent repeated test-retest golden-angle radial CEST-MRF acquisitions with 30 measurements and 34 spokes per measurement. The Movienet network was adapted to reconstruct accelerated radial CEST-MRF data. Separate models were trained using 1, 2, 4, 8, 16, 20, and 34 spokes per measurement. Temporal and coil correlations were exploited by concatenating time and coil dimensions in the input. GRASP-based locally low-rank (GRASP-LLR) reconstruction was used for comparison. Quantitative DRONE fitting generated tissue parameter maps. Movienet-8 was evaluated with leave-one-subject-out analysis. Prospectively acquired 8-spoke datasets were compared with paired 34-spoke acquisitions, and Movienet-8 and Movienet-34 were compared with a 64-spoke acquisition. Results: In the held-out retrospective series, median Pearson correlations between Movienet and GRASP-LLR were 0.795, 0.947, 0.974, 0.977, and 0.981 for 1, 8, 16, 20, and 34 spokes, respectively. Corresponding NRMSE values were 14.7, 8.2, 5.9, 5.3, and 5.1 percent. Across 10 leave-one-subject-out volunteers, median image correlation was 0.965 and NRMSE was 7.5 percent. Reconstruction of a 30-slice series required 1.7 s with Movienet versus approximately 45 min with GRASP-LLR. Conclusion: Movienet enables whole-brain radial CEST-MRF acquisition in 3 min, with image reconstruction in less than 2 s.
ER -