Using a deep learning prior for accelerating hyperpolarized ¹³C MRSI on synthetic cancer datasets

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Deep Learning
• Metabolic Imaging

Background:

• Hyperpolarized carbon-13 (¹³C) Magnetic Resonance Spectroscopic Imaging (MRSI) offers metabolic insights but is limited by long acquisition times.
• Accelerating ¹³C MRSI is crucial for clinical translation and broader applications in cancer research.

Purpose of the Study:

• To develop and validate a deep learning (DL) approach combined with k-space data fidelity for accelerated ¹³C MRSI.
• To demonstrate the method's effectiveness on synthetic cancer datasets, including human brain and prostate tumors, and mouse models.

Main Methods:

• A two-site exchange model based on the Bloch equation was used to simulate synthetic MRSI data.
• A DL prior was trained using singular maps from simulated data and applied to reconstruct undersampled k-space data.
• The method was evaluated on diverse synthetic tumor datasets with varying undersampling factors and added Gaussian noise.

Main Results:

• The DL-based reconstruction achieved improved accuracy (5-8 dB higher peak SNR) compared to traditional compressed sensing methods.
• The method demonstrated robust performance against significant levels of Gaussian noise (up to 10% SD).
• Quantitative evaluations showed reasonable normalized root-mean-square errors.

Conclusions:

• The proposed singular value decomposition (SVD) + iterative DL model provides a general framework for DL-based MRI reconstruction in metabolic imaging.
• The method enables robust measurement of tumor morphology and metabolic images at six-fold acceleration.