Deep learning-based metabolite quantification in proton magnetic resonance spectroscopy of the brain

Abstract

Proton MR spectroscopy of the brain is a noninvasive technique to extract neurochemical information from the brain and allow the analysis of primary and secondary brain tumors and metabolic diseases. Acquired spectra are inevitably degraded due to many factors, such as line-broadening, low signal-to-noise ratio, overlapping metabolic signals, and variability in spectral baselines. Current software for spectral fitting and metabolite quantifications are based on nonlinear least squares fitting algorithms. These approaches are not suitable for all MR metabolic signals and are, moreover, very time consuming. Thus, magnetic resonance spectroscopic imaging still is rarely used in daily clinical routine regardless of its high potential.This thesis investigates how deep learning can accelerate spectral fitting and metabolite quantification. A convolutional neural network (CNN) called Superfit and consisting of two serialized autoencoders, was developed to remove the baseline from the original spectra and to disassemble the baseline free spectra into their metabolic components. For training, validation, and testing of Superfit, almost 70,000 spectra were simulated based on a basis set of the metabolites of interest and in vivo spectral baselines from earlier studies. The trained Superfit was capable of disassembling the metabolites and quantifying the concentrations of the total metabolites with a mean absolute percentage error of 8.99% plus/minus 7.41%. Furthermore, Superfit could fit and quantify over 10000 spectra in less than 40s implying subminute metabolite quantification via deep learning is possible.