Deep Learning Improves Robustness of Voxelwise Kinetic Modeling for Hyperpolarized Carbon-13 MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Hyperpolarized (HP) 13C MRI enables real-time metabolic imaging.
• Kinetic parameter estimation is crucial for quantitative analysis.
• Nonlinear least-squares (NLLS) fitting is a standard but sensitive method.

Purpose of the Study:

• To assess if deep learning enhances the robustness of voxelwise kinetic parameter estimation in HP 13C MRI.
• To compare deep learning (neural network, NN) performance against NLLS fitting.
• To evaluate parameter estimation for pyruvate-to-lactate conversion rate (k_PL), vascular-extravascular exchange rate (k_VE), and vascular volume fraction (v_B).

Main Methods:

• A hybrid neural network (NN) was trained using synthetic HP 13C MRI data.
• The NN estimated kinetic parameters including k_PL, k_VE, and v_B.
• Performance was compared to NLLS across various signal-to-noise ratio (SNR) levels, flip angles, and in vivo conditions.

Main Results:

• NN and NLLS showed comparable k_PL estimation at high SNR.
• The NN outperformed NLLS at low SNR and for weakly identifiable parameters (k_VE, v_B).
• In vivo, NN-derived maps were more spatially coherent than NLLS maps.

Conclusions:

• NLLS is effective for k_PL under ideal conditions.
• The NN offers more stable voxelwise maps, particularly for challenging parameters and low SNR or in vivo scenarios.
• Further validation is required to confirm quantitative accuracy.