Deep Learning–Based Reconstruction: Model Comparison for Variable-Density GRAPPA ¹ H MRSI

Proton magnetic resonance spectroscopic imaging ( ¹ H MRSI) enables quantitative mapping of brain metabolites, but its clinical use remains limited by long acquisition time. The goal of this work to improve the applicability of high-resolution ¹ H FID-MRSI at 7T by enhancing GRAPPA-based acceleration through deep learning–driven k-space reconstruction. In particular, compared with conventional GRAPPA, MultiNet PyGRAPPA enables substantially higher in-plane acceleration while suppressing residual lipid aliasing and preserving metabolite map fidelity in non–lipid-suppressed MRSI. Building on the MultiNet PyGRAPPA framework, we introduce a comprehensive comparison of advanced machine-learning models for predicting missing k-space points. Multiple architectures—including multilayer perceptrons, convolutional neural networks, and several U-Net variants—were trained within a variable-density k-space undersampling scheme to support acceleration factors of R = 4, 6, and 7. The proposed U-Net model extends the MultiNet concept by leveraging nonlinear hierarchical feature extraction, thereby improving reconstruction fidelity while maintaining robustness to noise.The methods were evaluated in vivo using retrospectively undersampled 7T ¹ H FID-MRSI datasets from healthy volunteers and patients. Quantitative analyses demonstrate that the U-Net outperforms the original MultiNet approach, offering improved SNR retention rate, reduced lipid RMSE, and higher structural similarity of major metabolites. Metabolite maps reconstructed with the U-Net showed reduced lipid artifacts and improved anatomical consistency. In conclusion, integrating deep convolutional networks into GRAPPA-based k-space prediction provides a more reliable and higher-fidelity reconstruction pipeline. When combined with variable-density undersampling, this approach enables faster acquisition of high-resolution ¹ H MRSI without compromising spectral quality or metabolite quantification.