[Paper Review] Acceleration and Quantitation of Localized Correlated Spectroscopy using Deep Learning: A Pilot Simulation Study

Zohaib Iqbal, Dan Nguyen+2 more|arXiv (Cornell University)|Jun 28, 2018

Advanced MRI Techniques and Applications56 references4 citations

TL;DR

This study proposes a deep learning approach using a densely connected U-Net (D-UNet) to accelerate non-uniformly sampled localized correlated spectroscopy (L-COSY) and accurately quantify metabolite concentrations. The method achieves less than 5% normalized mean squared error at four-fold acceleration and under 8% error at 25% SNR, outperforming compressed sensing in reconstruction and quantitation tasks using simulated MRS data.

ABSTRACT

Nuclear magnetic resonance spectroscopy (MRS) allows for the determination of atomic structures and concentrations of different chemicals in a biochemical sample of interest. MRS is used in vivo clinically to aid in the diagnosis of several pathologies that affect metabolic pathways in the body. Typically, this experiment produces a one dimensional (1D) 1H spectrum containing several peaks that are well associated with biochemicals, or metabolites. However, since many of these peaks overlap, distinguishing chemicals with similar atomic structures becomes much more challenging. One technique capable of overcoming this issue is the localized correlated spectroscopy (L-COSY) experiment, which acquires a second spectral dimension and spreads overlapping signal across this second dimension. Unfortunately, the acquisition of a two dimensional (2D) spectroscopy experiment is extremely time consuming. Furthermore, quantitation of a 2D spectrum is more complex. Recently, artificial intelligence has emerged in the field of medicine as a powerful force capable of diagnosing disease, aiding in treatment, and even predicting treatment outcome. In this study, we utilize deep learning to: 1) accelerate the L-COSY experiment and 2) quantify L-COSY spectra. We demonstrate that our deep learning model greatly outperforms compressed sensing based reconstruction of L-COSY spectra at higher acceleration factors. Specifically, at four-fold acceleration, our method has less than 5% normalized mean squared error, whereas compressed sensing yields 20% normalized mean squared error. We also show that at low SNR (25% noise compared to maximum signal), our deep learning model has less than 8% normalized mean squared error for quantitation of L-COSY spectra. These pilot simulation results appear promising and may help improve the efficiency and accuracy of L-COSY experiments in the future.

Motivation & Objective

• To accelerate the time-consuming 2D L-COSY MRS acquisition using deep learning.
• To enable accurate quantitation of overlapping metabolite signals in L-COSY spectra.
• To overcome limitations of conventional reconstruction and fitting methods in 2D MRS, such as long scan times and complex spectral fitting.
• To validate the performance of deep learning in reconstructing non-uniformly sampled L-COSY data and quantifying fully sampled spectra using simulated metabolite data.
• To establish a foundation for clinical translation by demonstrating feasibility and accuracy under realistic noise and sampling conditions.

Proposed method

• A densely connected U-Net (D-UNet) architecture was trained to reconstruct non-uniformly sampled (NUS) L-COSY spectra from undersampled k-space data.
• The D-UNet model was trained end-to-end on simulated L-COSY data generated with realistic metabolite peak positions, intensities, and exponential line broadening.
• Reconstruction performance was compared against compressed sensing using ℓ₁-norm minimization as a baseline.
• For quantitation, a separate D-UNet model was trained to predict metabolite concentrations directly from fully sampled L-COSY spectra.
• The models were evaluated using normalized mean squared error (NMSE) across varying signal-to-noise ratios (SNR) and acceleration factors.
• Training data included metabolites such as NAA, creatine, glutamate, and 2-hydroxyglutarate (2HG), with prior knowledge of peak positions and shapes incorporated into the simulation.

Experimental results

Research questions

• RQ1 Can a deep learning model reconstruct non-uniformly sampled L-COSY spectra with higher accuracy than compressed sensing at high acceleration factors?
• RQ2 Can a deep learning model accurately quantify metabolite concentrations in fully sampled L-COSY spectra under low SNR conditions?
• RQ3 How does the performance of the D-UNet model compare to traditional reconstruction and quantitation methods in simulated L-COSY data?
• RQ4 To what extent does the model generalize across different SNR levels and sampling patterns in 2D MRS?
• RQ5 Can the same deep learning framework be extended to other 2D MRS techniques such as JPRESS or TOCSY?

Key findings

• At four-fold acceleration, the D-UNet achieved less than 5% normalized mean squared error (NMSE) in reconstructing NUS L-COSY spectra, significantly outperforming compressed sensing, which yielded 20% NMSE.
• Under low SNR conditions (25% of maximum signal), the D-UNet maintained less than 8% NMSE for metabolite quantitation, demonstrating robustness to noise.
• The D-UNet model showed superior reconstruction fidelity compared to compressed sensing, particularly in resolving overlapping cross-peaks in the 2D spectrum.
• The method demonstrated consistent performance across multiple SNR levels and acceleration factors, indicating strong generalization capability on simulated data.
• The framework is extensible to other 2D MRS techniques, including JPRESS, NOESY, and TOCSY, as shown in supplemental results.
• The study highlights the potential for clinical translation by enabling L-COSY scans with acquisition times comparable to 1D single-voxel MRS (3–5 minutes), making it feasible for routine clinical use.