The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution. Methods: Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF).Results: Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value. Conclusion:MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance-like in the case of MF.

Purpose To develop localization sequences for in vivo MR spectroscopy (MRS) on clinical scanners of 3 T to record spectra that are not influenced by magnetization transfer from water. Methods Image‐selected in vivo spectroscopy (ISIS) localization and chemical‐shift‐selective excitation (termed I‐CSE) was combined in two ways: first, full ISIS localization plus a frequency‐selective spin‐echo and second, two‐dimensional (2D) ISIS plus a frequency‐selective excitation and slice‐selective refocusing. The techniques were evaluated at 3 T in phantoms and human subjects in comparison to standard techniques with water presaturation or metabolite‐cycling. ISIS included gradient‐modulated offset‐independent adiabatic (GOIA)‐type adiabatic inversion pulses; echo times were 8‐10 ms. Results The novel 2D and 3D I‐CSE methods yield upfield spectra that are comparable to those from standard MRS, except for shorter echo times and a limited frequency range. On the downfield/high‐frequency side, they yield much more signal for exchangeable protons when compared to MRS with water presaturation or metabolite‐cycling and longer echo times. Conclusion Novel non‐water‐excitation MRS sequences offer substantial benefits for the detection of metabolite signals that are otherwise suppressed by saturation transfer from water. Avoiding water saturation and using very short echo times allows direct observation of faster exchanging moieties than was previously possible at 3 T and additionally makes the methods less susceptible to fast T2 relaxation.

No abstract

Purpose The detection of nicotinamide‐adenine‐dinucleotide (NAD+) is challenging using standard 1H MR spectroscopy, because it is of low concentration and affected by polarization‐exchange with water. Therefore, this study compares three techniques to access NAD+ quantification at 3 T–one with and two without water presaturation. Methods A large brain volume in 10 healthy subjects was investigated with three techniques: semi‐LASER with water‐saturation (WS) (TE = 35 ms), semi‐LASER with metabolite‐cycling (MC) (TE = 35 ms), and the non‐water‐excitation (nWE) technique 2D ISIS‐localization with chemical‐shift‐selective excitation (2D I‐CSE) (TE = 10.2 ms). Spectra were quantified with optimized modeling in FiTAID. Results NAD+ could be well quantified in cohort‐average spectra with all techniques. Obtained apparent NAD+ tissue contents are all lower than expected from literature confirming restricted visibility by 1H MRS. The estimated value from WS‐MRS (58 μM) was considerably lower than those obtained with non‐WS techniques (146 μM for MC‐semi‐LASER and 125 μM for 2D I‐CSE). The nWE technique with shortest TE gave largest NAD+ signals but suffered from overlap with large amide signals. MC‐semi‐LASER yielded best estimation precision as reflected in relative Cramer‐Rao bounds (14%, 21 μM/146 μM) and also best robustness as judged by the coefficient‐of‐variance over the cohort (11%, 10 μM/146 μM). The MR‐visibility turned out as 16% with WS and 41% with MC. Conclusion Three methods to assess NAD+ in human brain at 3 T have been compared. NAD+ could be detected with a visibility of ∼41% for the MC method. This may open a new window for the observation of pathological changes in the clinical research setting.

PurposeThe inherently poor SNR of MRS measurements presents a significant hurdle to its clinical application. Denoising by machine or deep learning (DL) was proposed as a remedy. It is investigated whether such denoising leads to lower estimate uncertainties or whether it essentially reduces noise in signal‐free areas only.MethodsNoise removal based on supervised DL with U‐nets was implemented using simulated 1H MR spectra of human brain in two approaches: (1) via time‐frequency domain spectrograms and (2) using 1D spectra as input. Quality of denoising was evaluated in three ways: (1) by an adapted fit quality score, (2) by traditional model fitting, and (3) by quantification via neural networks.ResultsVisually appealing spectra were obtained; hinting that denoising is well‐suited for MRS. However, an adapted denoising score showed that noise removal is inhomogeneous and more efficient for signal‐free areas. This was confirmed by quantitative analysis of traditional fit results as well as DL quantitation following DL denoising. DL denoising, although apparently successful as judged by mean squared errors, led to substantially biased estimates in both implementations.ConclusionThe implemented DL‐based denoising techniques may be useful for display purposes, but do not help quantitative evaluations, confirming expectations based on estimation theory: Cramér Rao lower bounds defined by the original data and the appropriate fitting model cannot be circumvented in an unbiased way for single data sets, unless additional prior knowledge can be incurred in the form of parameter restrictions/relations or applicable substates.