High-Resolution Magnetic Resonance Spectroscopic Imaging using a Multi-Encoder Attention U-Net with Structural and...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Spatial resolution in Magnetic Resonance Spectroscopic Imaging (MRSI) is limited by signal-to-noise ratio (SNR).
• Multi-parametric MRI scans offer valuable spatial information that can potentially improve MRSI resolution.

Purpose of the Study:

• To develop and evaluate a deep learning method for enhancing the spatial resolution of 1H-MRSI.
• To leverage multi-parametric MRI as spatial priors for MRSI resolution enhancement.

Main Methods:

• A Multi-encoder Attention U-Net (MAU-Net) architecture was designed to process MRSI data and three MRI modalities.
• Spatial attention modules were integrated to learn modality-specific spatial weights.
• The MAU-Net was trained using in vivo brain imaging data from high-grade glioma patients with a combined loss function.

Main Results:

• The proposed MAU-Net successfully reconstructed high-resolution (64x64) MRSI metabolic maps from low-resolution (16x16) input.
• The method demonstrated superior performance compared to existing baseline techniques.
• Enhanced metabolic maps retained high quality and detailed spatial information.

Conclusions:

• Deep learning, specifically the MAU-Net architecture, can effectively enhance MRSI spatial resolution by integrating multi-parametric MRI priors.
• This approach offers a promising solution for improving the diagnostic utility of MRSI in clinical neuroimaging.
• The method shows potential for reconstructing high-quality metabolic maps from undersampled or low-resolution MRSI data.