Author: Kimberly Chan, Anke Henning, Mahrshi Jani, Andrew Wright, Xinyu Zhang 👨🔬
Affiliation: Advanced Imaging Research Center (AIRC), UT Southwestern Medical Center 🌍
Purpose: To evaluate the performance of multiple deep learning architectures for MRSI reconstruction and determine their effectiveness in maintaining high-resolution metabolite mapping while reducing scan times.
Methods: Data were acquired from a healthy volunteer on a 7T Philips DSync MRI system with a 2-channel transmit/32-channel receive head coil. Fully sampled 2D 1H FID MRSI data were collected with FoV = 220 × 220 mm, TE = 1.2 ms, TR = 320 ms, flip angle = 33°, acquisition time = 256 ms, spectral bandwidth = 4000 Hz, spatial resolution = 4.4 × 4.4 × 12 mm, and a 50 × 50 voxel matrix. A GRE anatomical image (256 × 256 × 4) was acquired to generate 20 semi-synthetic training datasets. Neural networks (MLP, CNN1, CNN2, U-Net1, U-Net2, U-Net16, U-Net++, Attention U-Net, Residual U-Net) were trained to predict missing k-space data using anatomical images, with the Adam optimizer, a batch size of 1024, and OneCycle policy.
Results: U-Net2 achieved the highest creatine SNR (19.41 ± 7.13) for 2x accelerated GRAPPA
1H MRSI reconstruction, closely matching the ground truth (22.19 ± 8.16) with a training time of 44 minutes on BioHPC (cloud-based high-performance computing resource). Across acceleration rates (R = 2 to 9), U-Net2 maintained high SNR at lower rates but faced challenges with lipid aliasing at higher rates. At 5x acceleration, reconstructions preserved peak intensity and low noise levels, while metabolite maps demonstrated strong similarity to fully sampled data.
Conclusion: This study demonstrates the effectiveness of U-Net-based models for accelerated MRSI reconstruction, with U-Net2 achieving high SNR while balancing quality and computational efficiency at lower acceleration factors. Simpler models, though computationally efficient, fall short in delivering the high-fidelity reconstructions required for clinical use. The findings underscore the trade-offs between speed and quality, offering valuable insights for optimizing MRSI reconstruction in clinical applications.