MR spectroscopy frequency and phase correction using convolutional neural networks.

Area of Science:

• Medical Imaging
• Spectroscopy
• Artificial Intelligence

Background:

• Magnetic Resonance Spectroscopy (MRS) is crucial for non-invasive metabolic profiling.
• Frequency- and phase-correction (FPC) is essential for accurate MRS data analysis.
• Current FPC methods can be limited in speed and robustness.

Purpose of the Study:

• To introduce a novel Convolutional Neural Network (CNN)-based approach for automated FPC of MRS data.
• To achieve fast and accurate FPC for single-voxel MEGA-PRESS MRS data.
• To evaluate the CNN model's performance against existing deep learning and conventional methods.

Main Methods:

• Trained and validated two CNNs (frequency and phase) on simulated and in vivo MEGA-PRESS MRS data with artificial offsets.
• Compared CNN performance against Multilayer Perceptrons (MLP) and conventional spectral registration (SR) methods.
• Assessed model robustness by adding random noise and applying varying frequency/phase offsets.

Main Results:

• The CNN model exhibited greater robustness to noise than MLP, with lower mean absolute errors in frequency and phase offset prediction.
• CNN-based FPC outperformed MLP and SR methods on in vivo datasets, even with added offsets.
• Quantitatively improved spectral quality was observed using the CNN approach.

Conclusions:

• CNN-based FPC provides an automated solution for MRS data preprocessing.
• The developed CNN approach significantly enhances MRS spectral quality compared to state-of-the-art methods.
• This technique holds promise for accelerating and improving MRS data analysis in research and clinical settings.