Deep learning approaches for detection and removal of ghosting artifacts in MR spectroscopy

PurposeTo make use of deep learning (DL) methods to detect and remove ghosting artifacts in clinical magnetic resonance spectra of human brain. MethodsDeep learning algorithms, including fully connected neural networks, deep-convolutional neural networks, and stacked what-where auto encoders, were implemented to detect and correct MR spectra containing spurious echo ghost signals. The DL methods were trained on a huge database of simulated spectra with and without ghosting artifacts that represent complex variations of ghost-ridden spectra, transformed to time-frequency spectrograms. The trained model was tested on simulated and in vivo spectra. ResultsThe preliminary results for ghost detection are very promising, reaching almost 100% accuracy, and the DL ghost removal methods show potential in simulated and in vivo spectra, but need further refinement and quantitative testing. ConclusionsGhosting artifacts in spectroscopy are problematic, as they superimpose with metabolites and lead to inaccurate quantification. Detection and removal of ghosting artifacts using traditional machine learning approaches with feature extraction/selection is difficult, as ghosts appear at different frequencies. Here, we show that DL methods perform extremely well for ghost detection if the spectra are treated as images in the form of time-frequency representations. Further optimization for in vivo spectra will hopefully confirm their ghostbusting capacity. Magn Reson Med 80:851-863, 2018. (c) 2018 International Society for Magnetic Resonance in Medicine.