Denoising single MR spectra by deep learning: Miracle or mirage?

Area of Science:

• Medical Imaging
• Machine Learning
• Spectroscopy

Background:

• Magnetic Resonance Spectroscopy (MRS) suffers from low signal-to-noise ratio (SNR), limiting clinical utility.
• Machine learning and deep learning (DL) have been proposed to denoise MRS data.

Purpose of the Study:

• To investigate if DL denoising reduces estimate uncertainties in MRS or only removes noise in signal-free areas.
• To evaluate the impact of DL denoising on quantitative analysis of MRS data.

Main Methods:

• Supervised DL with U-nets was applied to simulated 1H MR spectra of the human brain.
• Two approaches were used: time-frequency domain spectrograms and 1D spectra as input.
• Denoising quality was assessed using an adapted fit quality score, traditional model fitting, and neural network quantification.

Main Results:

• DL denoising produced visually appealing spectra, suggesting effectiveness.
• An adapted score revealed inhomogeneous noise removal, more efficient in signal-free regions.
• Quantitative analysis showed DL denoising led to substantially biased estimates, despite low mean squared errors.

Conclusions:

• DL-based denoising is suitable for visual display of MRS but not for quantitative analysis.
• Estimation theory confirms that DL cannot unbiasedly circumvent Cramér Rao lower bounds without incorporating prior knowledge.