Deep learning can accelerate and quantify simulated localized correlated spectroscopy.

Area of Science:

• Biophysics
• Medical Imaging
• Artificial Intelligence

Background:

• Nuclear magnetic resonance spectroscopy (MRS) is crucial for determining biochemical structures and concentrations.
• Overlapping peaks in 1D MRS spectra hinder the identification of similar metabolites.
• Localized correlated spectroscopy (L-COSY) offers improved resolution but is time-consuming and complex to quantify.

Purpose of the Study:

• To develop a deep learning model for accelerating L-COSY acquisition.
• To utilize deep learning for accurate quantification of L-COSY spectra.
• To compare the performance of deep learning against compressed sensing for L-COSY reconstruction.

Main Methods:

• Simulated metabolite spectra for human biochemicals were used for training and testing.
• A deep learning model was trained to reconstruct accelerated L-COSY spectra.
• The model's accuracy in spectral quantification was evaluated under varying signal-to-noise ratios (SNR).

Main Results:

• Deep learning significantly outperformed compressed sensing in reconstructing L-COSY spectra at higher acceleration factors.
• At four-fold acceleration, the deep learning model achieved <5% normalized mean squared error, compared to 20% for compressed sensing.
• For quantification at low SNR (25% noise), the deep learning model demonstrated <8% normalized mean squared error.

Conclusions:

• Deep learning shows promise for accelerating L-COSY experiments, reducing acquisition times.
• The developed AI model enables accurate quantification of L-COSY spectra, even in noisy conditions.
• These findings suggest potential for improved efficiency and accuracy in MRS-based metabolic pathway analysis.