Frequency Specificity of Regional Homogeneity in the Resting-State Human Brain

Resting state-fMRI studies have found that the inter-areal correlations in cortical networks concentrate within ultra-low frequencies (0.01–0.04 Hz) while long-distance connections within subcortical networks distribute over a wider frequency range (0.01–0.14 Hz). However, the frequency characteristics of regional homogeneity (ReHo) in different areas are still unclear. To examine the ReHo properties in different frequency bands, a data-driven method, Empirical Mode Decomposition (EMD), was adopted to decompose the time series of each voxel into several components with distinct frequency bands. ReHo values in each of the components were then calculated. Our results showed that ReHo in cortical areas were higher and more frequency-dependent than those in the subcortical regions. BOLD oscillations of 0.02–0.04 Hz mainly contributed to the cortical ReHo, whereas the ReHo in limbic areas involved a wider frequency range and were dominated by higher-frequency BOLD oscillations (>0.08 Hz). The frequency characteristics of ReHo are distinct between different parts of the striatum, with the frequency band of 0.04–0.1 Hz contributing the most to ReHo in caudate nucleus, and oscillations lower than 0.02 Hz contributing more to ReHo in putamen. The distinct frequency-specific ReHo properties of different brain areas may arise from the assorted cytoarchitecture or synaptic types in these areas. Our work may advance the understanding of the neural-physiological basis of local BOLD activities and the functional specificity of different brain regions.     

Flexible Analytic Wavelet Transform Based Features for Physical Action Identification Using sEMG Signals

ObjectivesElectromyography (EMG) is recording of the electrical activity produced by skeletal muscles. The classification of the EMG signals for different physical actions can be useful in restoring some or all of the lost motor functionalities in these individuals. Accuracy in classifying the EMG signal indicates efficient control of prosthesis.Material and methodsThe flexible analytic wavelet transform (FAWT) is used for classification of surface electromyography (sEMG) signals for identification of physical actions. FAWT is an efficient method for decomposition of sEMG signal into eight sub-bands, features namely neg-entropy, mean absolute value (MAV), variance (VAR), modified mean absolute value type 1 (MAV1), waveform length (WL), simple square integral (SSI), Tsallis entropy, integrated EMG (IEMG) are extracted from the sub-bands. Extracted features are fed into an extreme learning machine (ELM) classifier with sigmoid activation function.ResultsComprehensive experiments are conducted on the input sEMG signals and the accuracy, sensitivity and specificity scores are used for performance measurement. Experiments showed that among all sub-bands, the seventh sub-band provided the best performance where the recorded accuracy, sensitivity and specificity values were 99.36%, 99.36% and 99.93%, respectively. The comparison results showed best efficiency of proposed method as compared to other methods on the same dataset.ConclusionThis paper investigates the usage of the FAWT and ELM on sEMG signal classification. The results show that the proposed method is quite efficient in classification of the sEMG signals. It is also observed that the seventh sub-band of the FAWT provides the best discrimination property. In the future works, recent wavelet transform methods will be used for improving the classification performance.     

MRI Markers for Mild Cognitive Impairment: Comparisons between White Matter Integrity and Gray Matter Volume Measurements

The aim of the study was to evaluate the value of assessing white matter integrity using diffusion tensor imaging (DTI) for classification of mild cognitive impairment (MCI) and prediction of cognitive impairments in comparison to brain atrophy measurements using structural MRI. Fifty-one patients with MCI and 66 cognitive normal controls (CN) underwent DTI and T1-weighted structural MRI. DTI measures included fractional anisotropy (FA) and radial diffusivity (DR) from 20 predetermined regions-of-interest (ROIs) in the commissural, limbic and association tracts, which are thought to be involved in Alzheimer''s disease; measures of regional gray matter (GM) volume included 21 ROIs in medial temporal lobe, parietal cortex, and subcortical regions. Significant group differences between MCI and CN were detected by each MRI modality: In particular, reduced FA was found in splenium, left isthmus cingulum and fornix; increased DR was found in splenium, left isthmus cingulum and bilateral uncinate fasciculi; reduced GM volume was found in bilateral hippocampi, left entorhinal cortex, right amygdala and bilateral thalamus; and thinner cortex was found in the left entorhinal cortex. Group classifications based on FA or DR was significant and better than classifications based on GM volume. Using either DR or FA together with GM volume improved classification accuracy. Furthermore, all three measures, FA, DR and GM volume were similarly accurate in predicting cognitive performance in MCI patients. Taken together, the results imply that DTI measures are as accurate as measures of GM volume in detecting brain alterations that are associated with cognitive impairment. Furthermore, a combination of DTI and structural MRI measurements improves classification accuracy.     

Simultaneous BOLD fMRI and fiber-optic calcium recording in rat neocortex

Functional magnetic resonance imaging (fMRI) based on blood oxygen level-dependent (BOLD) contrast is widely used for probing brain activity, but its relationship to underlying neural activity remains elusive. Here, we combined fMRI with fiber-optic recordings of fluorescent calcium indicator signals to investigate this relationship in rat somatosensory cortex. Electrical forepaw stimulation (1-10 Hz) evoked fast calcium signals of neuronal origin that showed frequency-dependent adaptation. Additionally, slower calcium signals occurred in astrocyte networks, as verified by astrocyte-specific staining and two-photon microscopy. Without apparent glia activation, we could predict BOLD responses well from simultaneously recorded fiber-optic signals, assuming an impulse response function and taking into account neuronal adaptation. In cases with glia activation, we uncovered additional prolonged BOLD signal components. Our findings highlight the complexity of fMRI BOLD signals, involving both neuronal and glial activity. Combined fMRI and fiber-optic recordings should help to clarify cellular mechanisms underlying BOLD signals.     

Automatic ROI Selection in Structural Brain MRI Using SOM 3D Projection

This paper presents a method for selecting Regions of Interest (ROI) in brain Magnetic Resonance Imaging (MRI) for diagnostic purposes, using statistical learning and vector quantization techniques. The proposed method models the distribution of GM and WM tissues grouping the voxels belonging to each tissue in ROIs associated to a specific neurological disorder. Tissue distribution of normal and abnormal images is modelled by a Self-Organizing map (SOM), generating a set of representative prototypes, and the receptive field (RF) of each SOM prototype defines a ROI. Moreover, the proposed method computes the relative importance of each ROI by means of its discriminative power. The devised method has been assessed using 818 images from the Alzheimer''s disease Neuroimaging Initiative (ADNI) which were previously segmented through Statistical Parametric Mapping (SPM). The proposed algorithm was used over these images to parcel ROIs associated to the Alzheimer''s Disease (AD). Additionally, this method can be used to extract a reduced set of discriminative features for classification, since it compresses discriminative information contained in the brain. Voxels marked by ROIs which were computed using the proposed method, yield classification results up to 90% of accuracy for controls (CN) and Alzheimer''s disease (AD) patients, and 84% of accuracy for Mild Cognitive Impairment (MCI) and AD patients.     

Resting State Brain Function Analysis Using Concurrent BOLD in ASL Perfusion fMRI

The past decade has seen astounding discoveries about resting-state brain activity patterns in normal brain as well as their alterations in brain diseases. While the vast majority of resting-state studies are based on the blood-oxygen-level-dependent (BOLD) functional MRI (fMRI), arterial spin labeling (ASL) perfusion fMRI can simultaneously capture BOLD and cerebral blood flow (CBF) signals, providing a unique opportunity for assessing resting brain functions with concurrent BOLD (ccBOLD) and CBF signals. Before taking that benefit, it is necessary to validate the utility of ccBOLD signal for resting-state analysis using conventional BOLD (cvBOLD) signal acquired without ASL modulations. To address this technical issue, resting cvBOLD and ASL perfusion MRI were acquired from a large cohort (n = 89) of healthy subjects. Four widely used resting-state brain function analyses were conducted and compared between the two types of BOLD signal, including the posterior cingulate cortex (PCC) seed-based functional connectivity (FC) analysis, independent component analysis (ICA), analysis of amplitude of low frequency fluctuation (ALFF), and analysis of regional homogeneity (ReHo). Consistent default mode network (DMN) as well as other resting-state networks (RSNs) were observed from cvBOLD and ccBOLD using PCC-FC analysis and ICA. ALFF from both modalities were the same for most of brain regions but were different in peripheral regions suffering from the susceptibility gradients induced signal drop. ReHo showed difference in many brain regions, likely reflecting the SNR and resolution differences between the two BOLD modalities. The DMN and auditory networks showed highest CBF values among all RSNs. These results demonstrated the feasibility of ASL perfusion MRI for assessing resting brain functions using its concurrent BOLD in addition to CBF signal, which provides a potentially useful way to maximize the utility of ASL perfusion MRI.     

Use of supervised discretization with PCA in wavelet packet transformation-based surface electromyogram classification

This paper describes a preprocessing stage for nonlinear classifier used in wavelet packet transformation (WPT)-based multichannel surface electromyogram (EMG) classification. The preprocessing stage named sdPCA, which consists of supervised discretization coupled with principal component analysis (PCA), was developed for improving surface EMG classifier generalization ability and training speed on overlap segmented signals. The sdPCA outperforms the fast correlation-based filter (FCBF), PCA, supervised discretization, and their combinations in terms of the highest generalization ability, fast training speed, the small feature size, and an ability to reduce the risks of developing oscillation and being trapped in nonlinear classifier training. The experiments were conducted on a data set consisting of 4-channel surface EMG signals measured from 6 hand and wrist gestures of 12 subjects. The experimental results indicate that the classification system using sdPCA has the highest generalization ability along with the second fastest training speed. The classification accuracy in 12 subjects of the system using sdPCA is 93.30 ± 2.42% taking 400 epochs for training by overlap segmented signals within 100 s. This result is very attractive for further development because we can achieve high-classification accuracy for large data sets by means of the proposed sdPCA without the application of additional algorithms such as local discriminant bases (LDB), majority voting (MV), or WPT sub-bands clustering.     

Frequency Dependent Topological Patterns of Resting-State Brain Networks

The topological organization underlying brain networks has been extensively investigated using resting-state fMRI, focusing on the low frequency band from 0.01 to 0.1 Hz. However, the frequency specificities regarding the corresponding brain networks remain largely unclear. In the current study, a data-driven method named complementary ensemble empirical mode decomposition (CEEMD) was introduced to separate the time series of each voxel into several intrinsic oscillation rhythms with distinct frequency bands. Our data indicated that the whole brain BOLD signals could be automatically divided into five specific frequency bands. After applying the CEEMD method, the topological patterns of these five temporally correlated networks were analyzed. The results showed that global topological properties, including the network weighted degree, network efficiency, mean characteristic path length and clustering coefficient, were observed to be most prominent in the ultra-low frequency bands from 0 to 0.015 Hz. Moreover, the saliency of small-world architecture demonstrated frequency-density dependency. Compared to the empirical mode decomposition method (EMD), CEEMD could effectively eliminate the mode-mixing effects. Additionally, the robustness of CEEMD was validated by the similar results derived from a split-half analysis and a conventional frequency division method using the rectangular window band-pass filter. Our findings suggest that CEEMD is a more effective method for extracting the intrinsic oscillation rhythms embedded in the BOLD signals than EMD. The application of CEEMD in fMRI data analysis will provide in-depth insight in investigations of frequency specific topological patterns of the dynamic brain networks.     

Abnormal cortical networks in mild cognitive impairment and Alzheimer's disease

Recently, many researchers have used graph theory to study the aberrant brain structures in Alzheimer's disease (AD) and have made great progress. However, the characteristics of the cortical network in Mild Cognitive Impairment (MCI) are still largely unexplored. In this study, the gray matter volumes obtained from magnetic resonance imaging (MRI) for all brain regions except the cerebellum were parcellated into 90 areas using the automated anatomical labeling (AAL) template to construct cortical networks for 98 normal controls (NCs), 113 MCIs and 91 ADs. The measurements of the network properties were calculated for each of the three groups respectively. We found that all three cortical networks exhibited small-world properties and those strong interhemispheric correlations existed between bilaterally homologous regions. Among the three cortical networks, we found the greatest clustering coefficient and the longest absolute path length in AD, which might indicate that the organization of the cortical network was the least optimal in AD. The small-world measures of the MCI network exhibited intermediate values. This finding is logical given that MCI is considered to be the transitional stage between normal aging and AD. Out of all the between-group differences in the clustering coefficient and absolute path length, only the differences between the AD and normal control groups were statistically significant. Compared with the normal controls, the MCI and AD groups retained their hub regions in the frontal lobe but showed a loss of hub regions in the temporal lobe. In addition, altered interregional correlations were detected in the parahippocampus gyrus, medial temporal lobe, cingulum, fusiform, medial frontal lobe, and orbital frontal gyrus in groups with MCI and AD. Similar to previous studies of functional connectivity, we also revealed increased interregional correlations within the local brain lobes and disrupted long distance interregional correlations in groups with MCI and AD.     

A frequency-weighted method combined with Common Spatial Patterns for electroencephalogram classification in brain–computer interface

Common Spatial Patterns (CSP) has been proven to be a powerful and successful method in the detection of event-related desynchronization (ERD) and ERD based brain–computer interface (BCI). However, frequency optimization combined with CSP has only been investigated by a few groups. In this paper, a frequency-weighted method (FWM) is proposed to optimize the frequency spectrum of surface electroencephalogram (EEG) signals for a two-class mental task classification. This straightforward method computes a weight value for each frequency component according to its importance for the discrimination task and reforms the spectrum with the computed weights. The off-line analysis shows that the proposed method achieves an improvement of about 4% (averaged over 24 datasets) in terms of cross-validation accuracy over the basic CSP.     

A Shared Neural Substrate for Mentalizing and the Affective Component of Sentence Comprehension

Using event-related fMRI in a sample of 42 healthy participants, we compared the cerebral activity maps obtained when classifying spoken sentences based on the mental content of the main character (belief, deception or empathy) or on the emotional tonality of the sentence (happiness, anger or sadness). To control for the effects of different syntactic constructions (such as embedded clauses in belief sentences), we subtracted from each map the BOLD activations obtained during plausibility judgments on structurally matching sentences, devoid of emotions or ToM. The obtained theory of mind (ToM) and emotional speech comprehension networks overlapped in the bilateral temporo-parietal junction, posterior cingulate cortex, right anterior temporal lobe, dorsomedial prefrontal cortex and in the left inferior frontal sulcus. These regions form a ToM network, which contributes to the emotional component of spoken sentence comprehension. Compared with the ToM task, in which the sentences were enounced on a neutral tone, the emotional sentence classification task, in which the sentences were play-acted, was associated with a greater activity in the bilateral superior temporal sulcus, in line with the presence of emotional prosody. Besides, the ventromedial prefrontal cortex was more active during emotional than ToM sentence processing. This region may link mental state representations with verbal and prosodic emotional cues. Compared with emotional sentence classification, ToM was associated with greater activity in the caudate nucleus, paracingulate cortex, and superior frontal and parietal regions, in line with behavioral data showing that ToM sentence comprehension was a more demanding task.     

Online Classification of Multiple Motor Imagery Tasks Using Filter Bank Based Maximum-a-Posteriori Common Spatial Pattern Filters

ObjectiveThe main objective of this paper is to propose a novel technique, called filter bank maximum a-posteriori common spatial pattern (FB-MAP-CSP) algorithm, for online classification of multiple motor imagery activities using electroencephalography (EEG) signals. The proposed technique addresses the overfitting issue of CSP in addition to utilizing the spectral information of EEG signals inside the framework of filter banks while extending it to more than two conditions.Materials and methodsThe classification of motor imagery signals is based upon the detection of event-related de-synchronization (ERD) phenomena in the μ and β rhythms of EEG signals. Accordingly, two modifications in the existing MAP-CSP technique are presented: (i) The (pre-processed) EEG signals are spectrally filtered by a bank of filters lying in the μ and β brainwave frequency range, (ii) the framework of MAP-CSP is extended to deal with multiple (more than two) motor imagery tasks classification and the spatial filters thus obtained are calculated for each sub-band, separately. Subsequently, the most imperative features over all sub-bands are selected and un-regularized linear discriminant analysis is employed for classification of multiple motor imagery tasks.ResultsPublicly available dataset (BCI Competition IV Dataset I) is used to validate the proposed method i.e. FB-MAP-CSP. The results show that the proposed method yields superior classification results, in addition to be computationally more efficient in the case of online implementation, as compared to the conventional CSP based techniques and its variants for multiclass motor imagery classification.ConclusionThe proposed FB-MAP-CSP algorithm is found to be a potential / superior method for classifying multi-condition motor imagery EEG signals in comparison to FBCSP based techniques.     

Intrinsic Patterns of Coupling between Correlation and Amplitude of Low-Frequency fMRI Fluctuations Are Disrupted in Degenerative Dementia Mainly due to Functional Disconnection

Low frequency fluctuations (LFFs) of the BOLD signal are a major discovery in the study of the resting brain with functional magnetic resonance imaging (fMRI). Two fMRI-based measures, functional connectivity (FC), a measure of signal synchronicity, and the amplitude of LFFs (ALFF), a measure of signal periodicity, have been proved to be sensitive to changes induced by several neurological diseases, including degenerative dementia. In spite of the increasing use of these measures, whether and how they are related to each other remains to be elucidated. In this work we used voxel-wise FC and ALFF computed in different frequency bands (slow-5: 0.01-0.027 Hz; slow-4: 0.027-0.073 Hz; and full-band: 0.01-0.073 Hz), in order to assess their relationship in healthy elderly as well as the relevant changes induced by Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI). We found that in healthy elderly subjects FC and ALFF are positively correlated in anterior and posterior cingulate cortex (full-band, slow-4 and slow-5), temporal cortex (full-band and slow-5), and in a set of subcortical regions (full-band and slow-4). These correlation patterns between FC and ALFF were absent in either AD or MCI patients. Notably, the loss of correlation between FC and ALFF in the AD group was primarily due to changes in FC rather than in ALFF. Our results indicate that degenerative dementia is characterized by a loss of global connection rather than by a decrease of fluctuation amplitude.     

Multimodal functional network connectivity: an EEG-fMRI fusion in network space

EEG and fMRI recordings measure the functional activity of multiple coherent networks distributed in the cerebral cortex. Identifying network interaction from the complementary neuroelectric and hemodynamic signals may help to explain the complex relationships between different brain regions. In this paper, multimodal functional network connectivity (mFNC) is proposed for the fusion of EEG and fMRI in network space. First, functional networks (FNs) are extracted using spatial independent component analysis (ICA) in each modality separately. Then the interactions among FNs in each modality are explored by Granger causality analysis (GCA). Finally, fMRI FNs are matched to EEG FNs in the spatial domain using network-based source imaging (NESOI). Investigations of both synthetic and real data demonstrate that mFNC has the potential to reveal the underlying neural networks of each modality separately and in their combination. With mFNC, comprehensive relationships among FNs might be unveiled for the deep exploration of neural activities and metabolic responses in a specific task or neurological state.     

Cortical Network Models of Firing Rates in the Resting and Active States Predict BOLD Responses

Measurements of blood oxygenation level dependent (BOLD) signals have produced some surprising observations. One is that their amplitude is proportional to the entire activity in a region of interest and not just the fluctuations in this activity. Another is that during sleep and anesthesia the average BOLD correlations between regions of interest decline as the activity declines. Mechanistic explanations of these phenomena are described here using a cortical network model consisting of modules with excitatory and inhibitory neurons, taken as regions of cortical interest, each receiving excitatory inputs from outside the network, taken as subcortical driving inputs in addition to extrinsic (intermodular) connections, such as provided by associational fibers. The model shows that the standard deviation of the firing rate is proportional to the mean frequency of the firing when the extrinsic connections are decreased, so that the mean BOLD signal is proportional to both as is observed experimentally. The model also shows that if these extrinsic connections are decreased or the frequency of firing reaching the network from the subcortical driving inputs is decreased, or both decline, there is a decrease in the mean firing rate in the modules accompanied by decreases in the mean BOLD correlations between the modules, consistent with the observed changes during NREM sleep and under anesthesia. Finally, the model explains why a transient increase in the BOLD signal in a cortical area, due to a transient subcortical input, gives rises to responses throughout the cortex as observed, with these responses mediated by the extrinsic (intermodular) connections.     

Assessing the influence of different ROI selection strategies on functional connectivity analyses of fMRI data acquired during steady-state conditions

In blood oxygen level dependent (BOLD) functional magnetic resonance imaging (fMRI), assessing functional connectivity between and within brain networks from datasets acquired during steady-state conditions has become increasingly common. However, in contrast to connectivity analyses based on task-evoked signal changes, selecting the optimal spatial location of the regions of interest (ROIs) whose timecourses will be extracted and used in subsequent analyses is not straightforward. Moreover, it is also unknown how different choices of the precise anatomical locations within given brain regions influence the estimates of functional connectivity under steady-state conditions. The objective of the present study was to assess the variability in estimates of functional connectivity induced by different anatomical choices of ROI locations for a given brain network. We here targeted the default mode network (DMN) sampled during both resting-state and a continuous verbal 2-back working memory task to compare four different methods to extract ROIs in terms of ROI features (spatial overlap, spatial functional heterogeneity), signal features (signal distribution, mean, variance, correlation) as well as strength of functional connectivity as a function of condition. We show that, while different ROI selection methods produced quantitatively different results, all tested ROI selection methods agreed on the final conclusion that functional connectivity within the DMN decreased during the continuous working memory task compared to rest.     

The role of the DLPFC in inductive reasoning of MCI patients and normal agings: An fMRI study

Previous studies of young people have revealed that the left dorsolateral prefrontal cortex (DLPFC) plays an important role in inductive reasoning. An fMRI experiment was performed in this study to examine whether the left DLPFC was involved in inductive reasoning of MCI patients and normal agings, and whether the activation pattern of this region was different between MCI patients and normal agings. The fMRI results indicated that MCI patients had no difference from normal agings in behavior per-formance (reaction time and accuracy) and the activation pattern of DLPFC. However, the BOLD re-sponse of the DLPFC region for MCI patients was weaker than that for normal agings, and the func-tional connectivity between the bilateral DLPFC regions for MCI patients was significantly higher than for normal agings. Taken together, these results indicated that DLPFC plays an important role in inductive reasoning of agings, and the functional abnormity of DLPFC may be an earlier marker of MCI before structural alterations.     

颞叶癫痫脑“默认模式”的fMRI研究

功能磁共振成像(functional magnetic resonance imaging,fMRI)被用于检测静息时脑功能神经网络.作者运用静息fMRI检测海马硬化颞叶癫痫(temporal lobe epilepsy,TLE)脑"默认模式",采用感兴趣区域功能连接分析检测16例TLE患者和16名正常对照静息时脑的"默认模式",并进行组内和组间分析.研究发现,与正常对照相比,TLE静息时海马、颞极、额叶、颞叶、壳核及楔前叶等脑区与后扣带回的功能连接增强.研究结果表明TLE患者的固有脑功能组织模式有可能出现紊乱.这一研究将有助于从脑功能的角度了解癫痫患者某些临床症状的发病机理,为今后癫痫诊治的发展提供一定的帮助.     

Is cross-validation better than resubstitution for ranking genes?

MOTIVATION: Ranking gene feature sets is a key issue for both phenotype classification, for instance, tumor classification in a DNA microarray experiment, and prediction in the context of genetic regulatory networks. Two broad methods are available to estimate the error (misclassification rate) of a classifier. Resubstitution fits a single classifier to the data, and applies this classifier in turn to each data observation. Cross-validation (in leave-one-out form) removes each observation in turn, constructs the classifier, and then computes whether this leave-one-out classifier correctly classifies the deleted observation. Resubstitution typically underestimates classifier error, severely so in many cases. Cross-validation has the advantage of producing an effectively unbiased error estimate, but the estimate is highly variable. In many applications it is not the misclassification rate per se that is of interest, but rather the construction of gene sets that have the potential to classify or predict. Hence, one needs to rank feature sets based on their performance. RESULTS: A model-based approach is used to compare the ranking performances of resubstitution and cross-validation for classification based on real-valued feature sets and for prediction in the context of probabilistic Boolean networks (PBNs). For classification, a Gaussian model is considered, along with classification via linear discriminant analysis and the 3-nearest-neighbor classification rule. Prediction is examined in the steady-distribution of a PBN. Three metrics are proposed to compare feature-set ranking based on error estimation with ranking based on the true error, which is known owing to the model-based approach. In all cases, resubstitution is competitive with cross-validation relative to ranking accuracy. This is in addition to the enormous savings in computation time afforded by resubstitution.     

Classification of Motor Imagery EEG Based on Time-Domain and Frequency-Domain Dual-Stream Convolutional Neural Network

Background and objectiveAn important task of the brain-computer interface (BCI) of motor imagery is to extract effective time-domain features, frequency-domain features or time-frequency domain features from the raw electroencephalogram (EEG) signals for classification of motor imagery. However, choosing an appropriate method to combine time domain and frequency domain features to improve the performance of motor imagery recognition is still a research hotspot.MethodsIn order to fully extract and utilize the time-domain and frequency-domain features of EEG in classification tasks, this paper proposed a novel dual-stream convolutional neural network (DCNN), which can use time domain signal and frequency domain signal as the inputs, and the extracted time-domain features and frequency-domain features are fused by linear weighting for classification training. Furthermore, the weight can be learned by the DCNN automatically.ResultsThe experiments based on BCI competition II dataset III and BCI competition IV dataset 2a showed that the model proposed by this study has better performance than other conventional methods. The model used time-frequency signal as the inputs had better performance than the model only used time-domain signals or frequency-domain signals. The accuracy of classification was improved for each subject compared with the models only used one signals as the inputs.ConclusionsFurther analysis shown that the fusion weight of different subject is specifically, adjusting the weight coefficient automatically is helpful to improve the classification accuracy.