Denoising MRSI Data Using Subspace Model-Assisted Langevin Dynamics with Side Information

Motivation: Enhancing the signal-to-noise (SNR) of MRSI signals will improve detection sensitivity for clinical and research applications. Goal(s): To leverage recent advances in machine learning to improve the SNR of MRSI. Approach: A physics-based machine learning method was proposed for denoising MRSI data, incorporating side information such as PD, T1, T2, QSM, and MWF. A deep conditional generative model was trained to capture the relationship between water and metabolite signals. Denoised signals were obtained via subspace-assisted Langevin dynamics. Results: The method was validated using both simulated and in vivo data obtained from healthy subjects and patients with tumor or stroke, demonstrating substantial noise reduction. Impact: This proposed method provides an effective way to enhance the sensitivity of MRSI using physics-based machine learning incorporating side information. The method may further enhance the practical utility of MRSI for research and clinical applications.