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

PURPOSE: The 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. METHODS: Noise removal based on supervised DL with U-nets was implemented using simulated 1 H 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. RESULTS: Visually 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. CONCLUSION: The 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.

Quantification of MR spectra by deep learning in an idealized setting: Investigation of forms of input, network architectures, optimization by ensembles of networks, and training bias.

PURPOSE: 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.

Deep Learning Classification of Lake Zooplankton.

Plankton are effective indicators of environmental change and ecosystem health in freshwater habitats, but collection of plankton data using manual microscopic methods is extremely labor-intensive and expensive. Automated plankton imaging offers a promising way forward to monitor plankton communities with high frequency and accuracy in real-time. Yet, manual annotation of millions of images proposes a serious challenge to taxonomists. Deep learning classifiers have been successfully applied in various fields and provided encouraging results when used to categorize marine plankton images. Here, we present a set of deep learning models developed for the identification of lake plankton, and study several strategies to obtain optimal performances, which lead to operational prescriptions for users. To this aim, we annotated into 35 classes over 17900 images of zooplankton and large phytoplankton colonies, detected in Lake Greifensee (Switzerland) with the Dual Scripps Plankton Camera. Our best models were based on transfer learning and ensembling, which classified plankton images with 98% accuracy and 93% F1 score. When tested on freely available plankton datasets produced by other automated imaging tools (ZooScan, Imaging FlowCytobot, and ISIIS), our models performed better than previously used models. Our annotated data, code and classification models are freely available online.

Microstructural plasticity in nociceptive pathways after spinal cord injury.

OBJECTIVE: To track the interplay between (micro-) structural changes along the trajectories of nociceptive pathways and its relation to the presence and intensity of neuropathic pain (NP) after spinal cord injury (SCI). METHODS: A quantitative neuroimaging approach employing a multiparametric mapping protocol was used, providing indirect measures of myelination (via contrasts such as magnetisation transfer (MT) saturation, longitudinal relaxation (R1)) and iron content (via effective transverse relaxation rate (R2*)) was used to track microstructural changes within nociceptive pathways. In order to characterise concurrent changes along the entire neuroaxis, a combined brain and spinal cord template embedded in the statistical parametric mapping framework was used. Multivariate source-based morphometry was performed to identify naturally grouped patterns of structural variation between individuals with and without NP after SCI. RESULTS: In individuals with NP, lower R1 and MT values are evident in the primary motor cortex and dorsolateral prefrontal cortex, while increases in R2* are evident in the cervical cord, periaqueductal grey (PAG), thalamus and anterior cingulate cortex when compared with pain-free individuals. Lower R1 values in the PAG and greater R2* values in the cervical cord are associated with NP intensity. CONCLUSIONS: The degree of microstructural changes across ascending and descending nociceptive pathways is critically implicated in the maintenance of NP. Tracking maladaptive plasticity unravels the intimate relationships between neurodegenerative and compensatory processes in NP states and may facilitate patient monitoring during therapeutic trials related to pain and neuroregeneration.

Simultaneous voxel-wise analysis of brain and spinal cord morphometry and microstructure within the SPM framework.

To validate a simultaneous analysis tool for the brain and cervical cord embedded in the statistical parametric mapping (SPM) framework, we compared trauma-induced macro- and microstructural changes in spinal cord injury (SCI) patients to controls. The findings were compared with results obtained from existing processing tools that assess the brain and spinal cord separately. A probabilistic brain-spinal cord template (BSC) was generated using a generative semi-supervised modelling approach. The template was incorporated into the pre-processing pipeline of voxel-based morphometry and voxel-based quantification analyses in SPM. This approach was validated on T1-weighted scans and multiparameter maps, by assessing trauma-induced changes in SCI patients relative to controls and comparing the findings with the outcome from existing analytical tools. Consistency of the MRI measures was assessed using intraclass correlation coefficients (ICC). The SPM approach using the BSC template revealed trauma-induced changes across the sensorimotor system in the cord and brain in SCI patients. These changes were confirmed with established approaches covering brain or cord, separately. The ICC in the brain was high within regions of interest, such as the sensorimotor cortices, corticospinal tracts and thalamus. The simultaneous voxel-wise analysis of brain and cervical spinal cord was performed in a unique SPM-based framework incorporating pre-processing and statistical analysis in the same environment. Validation based on a SCI cohort demonstrated that the new processing approach based on the brain and cord is comparable to available processing tools, while offering the advantage of performing the analysis simultaneously across the neuraxis.

Quality of clinical brain tumor MR spectra judged by humans and machine learning tools.

PURPOSE: To investigate and compare human judgment and machine learning tools for quality assessment of clinical MR spectra of brain tumors. METHODS: A very large set of 2574 single voxel spectra with short and long echo time from the eTUMOUR and INTERPRET databases were used for this analysis. Original human quality ratings from these studies as well as new human guidelines were used to train different machine learning algorithms for automatic quality control (AQC) based on various feature extraction methods and classification tools. The performance was compared with variance in human judgment. RESULTS: AQC built using the RUSBoost classifier that combats imbalanced training data performed best. When furnished with a large range of spectral and derived features where the most crucial ones had been selected by the TreeBagger algorithm it showed better specificity (98%) in judging spectra from an independent test-set than previously published methods. Optimal performance was reached with a virtual three-class ranking system. CONCLUSION: Our results suggest that feature space should be relatively large for the case of MR tumor spectra and that three-class labels may be beneficial for AQC. The best AQC algorithm showed a performance in rejecting spectra that was comparable to that of a panel of human expert spectroscopists. Magn Reson Med 79:2500-2510, 2018. © 2017 International Society for Magnetic Resonance in Medicine.

A Realistic Framework for Investigating Decision Making in the Brain With High Spatiotemporal Resolution Using Simultaneous EEG/fMRI and Joint ICA.

Human decision making is a multidimensional construct, driven by a complex interplay between external factors, internal biases, and computational capacity constraints. Here, we propose a layered approach to experimental design in which multiple tasks-from simple to complex-with additional layers of complexity introduced at each stage are incorporated for investigating decision making. This is demonstrated using tasks involving intertemporal choice between immediate and future prospects. Previous functional magnetic resonance imaging (fMRI) and electroencephalographic (EEG) studies have separately investigated the spatial and temporal neural substrates, respectively, of specific factors underlying decision making. In contrast, we performed simultaneous acquisition of EEG/fMRI data and fusion of both modalities using joint independent component analysis such that: 1) the native temporal/spatial resolutions of either modality is not compromised and 2) fast temporal dynamics of decision making as well as involved deeper striatal structures can be characterized. We show that spatiotemporal neural substrates underlying our proposed complex intertemporal task simultaneously incorporating rewards, costs, and uncertainty of future outcomes can be predicted (using a linear model) from neural substrates of each of these factors, which were separately obtained by simpler tasks. This was not the case for spatial and temporal features obtained separately from fMRI and EEG, respectively. However, certain prefrontal activations in the complex task could not be predicted from activations in simpler tasks, indicating that the assumption of pure insertion has limited validity. Overall, our approach provides a realistic and novel framework for investigating the neural substrates of decision making with high spatiotemporal resolution.

Investigation of True High Frequency Electrical Substrates of fMRI-Based Resting State Networks Using Parallel Independent Component Analysis of Simultaneous EEG/fMRI Data.

Previous work using simultaneously acquired electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) data has shown that the slow temporal dynamics of resting state brain networks (RSNs), e.g., default mode network (DMN), visual network (VN), obtained from fMRI are correlated with smoothed and down sampled versions of various EEG features such as microstates and band-limited power envelopes. Therefore, even though the down sampled and smoothed envelope of EEG gamma band power is correlated with fMRI fluctuations in the RSNs, it does not mean that the electrical substrates of the RSNs fluctuate with periods <100 ms. Based on the scale free properties of EEG microstates and their correlation with resting state fMRI fluctuations in the RSNs, researchers have speculated that truly high frequency electrical substrates may exist for the RSNs, which would make resting fluctuations obtained from fMRI more meaningful to typically occurring fast neuronal processes in the sub-100 ms time scale. In this study, we test this critical hypothesis using an integrated framework involving simultaneous EEG/fMRI acquisition, fast fMRI sampling (TR = 200 ms) using multiband EPI (MB EPI), and EEG/fMRI fusion using parallel independent component analysis (pICA) which does not require the down sampling of EEG to fMRI temporal resolution. Our results demonstrate that with faster sampling, high frequency electrical substrates (fluctuating with periods <100 ms time scale) of the RSNs can be observed. This provides a sounder neurophysiological basis for the RSNs.

Forecasting the quality of water-suppressed 1 H MR spectra based on a single-shot water scan.

PURPOSE: To investigate whether an initial non-water-suppressed acquisition that provides information about the signal-to-noise ratio (SNR) and linewidth is enough to forecast the maximally achievable final spectral quality and thus inform the operator whether the foreseen number of averages and achieved field homogeneity is adequate. METHODS: A large range of spectra with varying SNR and linewidth was simulated and fitted with popular fitting programs to determine the dependence of fitting errors on linewidth and SNR. A tool to forecast variance based on a single acquisition was developed and its performance evaluated on simulated and in vivo data obtained at 3 Tesla from various brain regions and acquisition settings. RESULTS: A strong correlation to real uncertainties in estimated metabolite contents was found for the forecast values and the Cramer-Rao lower bounds obtained from the water-suppressed spectra. CONCLUSION: It appears to be possible to forecast the best-case errors associated with specific metabolites to be found in model fits of water-suppressed spectra based on a single water scan. Thus, nonspecialist operators will be able to judge ahead of time whether the planned acquisition can possibly be of sufficient quality to answer the targeted clinical question or whether it needs more averages or improved shimming. Magn Reson Med 78:441-451, 2017. © 2016 International Society for Magnetic Resonance in Medicine.

Anterior-posterior dissociation of the default mode network in dogs.

The default mode network (DMN) in humans has been extensively studied using seed-based correlation analysis (SCA) and independent component analysis (ICA). While DMN has been observed in monkeys as well, there are conflicting reports on whether they exist in rodents. Dogs are higher mammals than rodents, but cognitively not as advanced as monkeys and humans. Therefore, they are an interesting species in the evolutionary hierarchy for probing the comparative functions of the DMN across species. In this study, we sought to know whether the DMN, and consequently its functions such as self-referential processing, are exclusive to humans/monkeys or can we also observe the DMN in animals such as dogs. To address this issue, resting state functional MRI data from the brains of lightly sedated dogs and unconstrained and fully awake dogs were acquired, and ICA and SCA were performed for identifying the DMN. Since anesthesia can alter resting state networks, confirming our results in awake dogs was essential. Awake dog imaging was accomplished by training the dogs to keep their head still using reinforcement behavioral adaptation techniques. We found that the anterior (such as anterior cingulate and medial frontal) and posterior regions (such as posterior cingulate) of the DMN were dissociated in both awake and anesthetized dogs.