2.1.

Deep Learning Architectures

DL derives its versatility from the available cell/nodal operations. These operations include linear transformations, filters, and gates that have been inspired by other domain-specific tools, which compute abstract features in the hidden layers. In general, the choice of these operations defines the application of a network. The three most commonly used types of networks used in fNIRS studies are shown in Fig. 1.

Fig. 1

Illustrations of the three most common classes of network architectures used in the reviewed articles are shown in the figure above. (a) An MLP has all nodes fully connected. (b) A convolutional NN (CNN) with a kernel size $2\times 2$, with subsequent pooling layers. (c) An LSTM architecture where final hidden states are used. For illustrative purposes, the output layer is constructed for binary classification problems.

Dense networks such as multilayer perceptrons (MLPs) use linear transformation as cell operations and are synonymous to ANNs. Convolutional neural networks (CNNs) use convolution operations, where a fixed-size filter is used for convolution over the input image or feature map. These are shown to be highly capable of recognizing digits,¹³ objects,¹⁴ and images in general. Similarly, long-short-term-memory (LSTM) networks can be unfurled in the time domain to learn time series data, including handwriting¹⁵ and semantics for language translation.¹⁶ LSTM cells use gates to maintain a cell-state memory.¹⁷ Appropriate data are passed through the cells in successive timesteps, avoiding the vanishing gradient problem (see Sec. 1.2.1) for long time series data.¹⁸

A fundamental attribute of neural networks is introducing nonlinearities using so-called “activation functions”—based on the idea of the firing of biological neurons. The most common types of activation functions are listed in Table 1. The sigmoid function compresses values in the range of 0 and 1, activating large positive values while zeroing out large negative ones. The sigmoid function can also be used at the output for binary classifiers, for example, in Refs. 19 and 20. Dolmans et al.²¹ used the sigmoid output for seven-class classification, although the softmax activation is widely used for multiclass classification. The softmax activation uses exponentiation followed by normalization to assign probabilities to the outputs. Higher softmax outputs at the output layer may be interpreted as confidence in the prediction.²²

Table 1

Activated outputs of input “x” from each of the activation functions is shown in the table. Also shown are the bounds/range of activated values.

The rectified linear unit (ReLU) activation is most widely used for the hidden nodes since it is computationally inexpensive and helps resolve the vanishing gradient problem.²² It has proven to be efficient and effective for CNNs. It deactivates all nodes with negative outputs, which, although effective, deactivates those nodes throughout the training. This issue, also known as dead-ReLU problem, is solved by a recent activation function called exponential linear unit (ELU), which has also been used by Mirbagheri et al.,¹⁹ Saadati et al.,²³ Ortega and Faisal,²⁴ whereas some also used another variant called leaky ReLU.^25,26

The final ingredient of neural networks is the loss function, which depends upon the output type. For classification tasks, the cross-entropy loss function²⁷ measures the difference in the probability distributions between ground truths and network predictions. Multiclass classifications^28–31 use categorical cross-entropy as the loss function, whereas binary classification problems use binary cross-entropy. For regression³² or reconstruction³³ problems, the mean squared error loss function is used. These loss functions may be modified to implement constraints or regularization, e.g., a linear combination of MSE, variance, and two other metrics for a denoising autoencoder (DAE) in Ref. 34.

2.2.

Practical Considerations in Deep Learning

2.2.3.

Inputs to deep networks

After the preprocessing steps (see Sec. 2.1), changes in oxy ($\mathrm{\Delta }\mathrm{HbO}$) and deoxy ($\mathrm{\Delta }\mathrm{HbR}$) hemoglobin, as well as total change ($\mathrm{\Delta }\mathrm{HbT}$) are obtained in time-series format. From these, data samples are segmented based on task settings. Data are segmented trialwise for task-based experiments, and for resting state data or long trials, data are segmented using sliding windows as mentioned previously.

Many studies opt for the discrete probability distribution of concentration changes and extract statistical features such as mean, slope, variance, skewness, kurtosis, max, and range in the form of manual features. In other cases, where the network is allowed to learn and extract features itself, the data are fed in the forms of either spatial maps^20,28,50 or time series themselves. In some studies, segments of data are converted to other forms such as Gramian angular fields³² or spectrogram maps.³⁰ A schematic in Fig. 2 summarizes the sample extraction procedure employed in the studies. While the general trends and techniques used for DL fNIRS studies have been examined, the applications and some application-dependent techniques are further discussed in the next section.

Fig. 2

Extraction of samples to be used for training NNs. Time-series data denoting hemodynamic concentration changes are obtained from the raw data after initial analysis and changed to appropriate formats (shown in right in the form of dimensions of input data samples), based on chosen architecture. $N$, $Ns$, number of samples; $T$, $Tw$, number of timepoints; $n_{\mathrm{ch}}$, number of channels; $h$, $w$, height and width of spatial map images; $n_{\mathrm{stat}}$, number of statistical moments/features.

3.1.

Preprocessing

Like most noninvasive neuroimaging modalities, raw fNIRS signals typically contain confounding physiological signals and other noise that originate from outside of the cerebral cortex, such as the hemodynamics of the scalp and changes in blood pressure and heart rate.⁹⁵ Most studies use some method of preprocessing to try to address this. While many of the papers presented here use band pass filtering or butterworth filtering, various other methods are employed throughout the literature. Independent component analysis (ICA) denoising,¹² wavelet filtering,²⁹ and correlation-based signal improvement filtering⁵¹ are all methods used to remove some of the undesired physiological trends in fNIRS signals. Another recommended method is short separation regression, a method in which signals with only confounding physiological signals are simultaneously collected alongside fNIRS signals and the trends in these signals are removed from the fNIRS data.⁹⁶ Other algorithms such as Savitsky–Golay filtering⁹⁷ and temporal derivative distribution repair⁹⁸ are methods used to correct motion artifacts and baseline drifts in fNIRS signals. Many of these techniques are commonly used in fNIRS studies to improve the quality of the signal as well as ensure that the signal being assessed originates from the brain. Other recommended techniques involve prewhitening the data or decorrelating via PCA to remove any correlation between fNIRS signals prior to analysis,⁹⁶ further facilitating the analysis of signals originating in a region of interest.

3.2.

Feature Extraction and Data Augmentation

One of the most notable problems with fNIRS data is the extensive manual feature extraction and artifact removal typically required for data to be analyzed, preventing many fNIRS studies from being applied in real time. As a result, more effective methods of preprocessing fNIRS data are being explored. One of the most promising benefits of DL is the ability to quickly and automatically learn and extract relevant features in fNIRS data. Some studies have already explored this benefit of DL. Tanveer et al.²⁰ reported the use of a DNN to extract the features that were fed to a K-nearest neighbors (KNN) classifier to detect the drowsiness of subjects during a virtual driving task. Using the features extracted from the DNN, the KNN was able to achieve a classification accuracy of 83.3%. Despite feature extraction typically being computationally expensive, even taking hours with a powerful GPU, the DNN exhibited a mean computation time of 0.024 s for 10 s time windows, a speed that would allow for feature extraction of a 30-min signal to take $<5\mathrm{s}$ with a NVidia 1060 GTX GPU. On top of computational speed, the ability of DL to automatically learn and extract features may reduce bias and errors during feature extraction, allowing for an increase in classification accuracy. One study by Liu et al.⁵² used an echo state network autoencoder (ESN AE) to extract the features that were fed to an MLP, achieving a four-class classification accuracy of 52.45%, outperforming the accuracy of the convolutional autoencoder (CAE) + CNN and manually extracted features fed to an MLP, which achieved accuracies of 47.21% and 37.94%, respectively.

While there have been other papers interested in using DL to extract features to feed into another classifier, there have also been papers that take raw fNIRS data and use the same neural network for feature extraction and classification. Despite end-to-end neural networks being seen as a more ideal solution than manual feature extraction, difficulties with low generalizability make them less commonly used. One study by Dargazany et al.²⁵ used an MLP (with two hidden layers) with raw EEG, body motion and fNIRS data to achieve a reported classification accuracy of 78% to 80% on a motor task with four classes, despite no denoising or preprocessing of data being done. Another study by Fernandez Rojas et al.⁴¹ used raw fNIRS data as the input for an LSTM network, achieving a classification accuracy of 90.6%. To assess generalizability, a 10-fold CV was used, with the classifier achieving an accuracy of 93.1%. Both studies have provided evidence toward the claim that DL techniques are capable of removing the need for manually extracted features from fNIRS data, further progressing toward the real-time end-to-end BCI. Despite the fact that the previously mentioned studies were able to achieve high accuracies without denoising or motion artifact removal, some studies are performed when a subject’s head is in motion. In such studies, fNIRS data can be heavily compromised by specific artifacts in the raw data that could bias any classification task. While many algorithms are used to try to remove motion artifacts, Lee et al. attempted to use a CNN to recognize and remove motion artifacts without relying on the parameters that must be defined to use many of the popular motion artifact removal algorithms.⁵³ In this study, the raw fNIRS time series and the estimated canonical response were used as inputs to the network and the resulting CNR of the DL output was 0.63, outperforming wavelet denoising, which achieved a mean CNR of 0.36. Another study done by Gao et al.³⁴ used subjects who were performing a precision cutting surgical task based on the fundamentals of laparoscopic surgery (FLS) program, which required a large range of motion. With a DAE and the process shown in Fig. 4, 93% motion artifact removal in simulated data and 100% artifact removal in real data were reported, outperforming all comparable artifact removal techniques, including wavelet filtering and principal component analysis. Another study to look at DL for the removal of motion artifacts, Kim et al.⁵⁴ compared a CNN to the performance of wavelet denoising and an autoregressive denoising method. Using simulated data in a manner similar to Gao et al. to determine the ground truth, Kim et al. found that the CNN resulted in a mean square error (MSE) of $\sim 0.004$ to 0.005, while the next best method, the combination of wavelet and autoregressive denoising, resulted in an MSE of $\sim 0.009$. Despite these studies all using different metrics to measure the effectiveness of the network, there is clearly an interest in finding an effective method of removing motion artifacts from fNIRS data.

Fig. 4

The illustration of the fNIRS data simulation process and the designed DAE model. (a) The green lines are the experimental fNIRS data, including noisy HRF and resting fNIRS data, while the blue and red lines are simulated ones. (b) DAE model: The input data of the DAE model are the simulated noisy HRF, and the output is the corresponding clean HRF without noise. The DAE model incorporates nine convolutional layers, followed by max-pooling layers in the first four layers and upsampling layers in the next four layers, with one convolutional layer before the output. The parameters are labeled in parentheses for each convolutional layer, in the order of kernel size, stride, input channel size, and kernel number. (c), (d) number of residual motion artifacts for the simulated and experimental data sets, respectively. Adapted from Ref. 34.

It is clear that DL techniques have shown promise for the ability to process and extract features from fNIRS data, but due to a lack of large variety of open-source fNIRS datasets, there has been a recent interest in using DL to generate more fNIRS data for training models. Data augmentation is a technique in which new data are generated to reduce the need for large labeled datasets for many machine learning and DL algorithms that require a lot of labeled data. To be useful, this generated data must not be identical to any of the training data, but must also be realistic, i.e., it must remain within the distribution of the original dataset.⁵⁵ While this is a challenge, some DL techniques such as generative adversarial networks (GANs), have been used to accomplish this. Wickramaratne and Mahmud⁵⁵ have used a GAN to augment their fNIRS dataset to increase the classification accuracy of finger- and foot-tapping tasks. Without augmented data, a classification of 70.4% was achieved with an SVM classifier, and a CNN classifier achieved an accuracy of 80% when trained only on real data. Using a GAN to generate training data, the accuracy of the CNN classifier increased to 96.67% when trained on real data as well as 110% generated data. Similarly, Woo et al.⁵⁶ used a GAN to produce activation t-maps, which, when used to augment the training dataset of a CNN, increased the classification accuracy of a finger tapping task from 92% to 97%, showing that a network that is already performing well may benefit from data augmentation. While the previous studies have used GANs to generate an image representation of fNIRS data, only one study directly used a GAN to augment the dataset with raw time series fNIRS data. Nagasawa et al.²⁶ used a GAN to generate fNIRS time series data to increase the classification accuracy of motor tasks, as shown in Fig. 5. They reported that when augmenting the 16 original datasets with 100 generated datasets, the accuracy of the SVM classifier increased from around 0.4 to 0.733 while the accuracy of the neural network classifier increased from around 0.4 to 0.746. While still a relatively recent development in fNIRS studies, generated datasets using GANS have demonstrated the ability to increase the classification accuracy of commonly used classifiers such as CNN or SVM classifiers, once again demonstrating the versatility of DL techniques in fNIRS research.

Fig. 5

Framework of GAN and data augmentation. (a) The generator creates the data from the random variables $z$, and the critic evaluates the generated and original (measured) data. (b) After the training process, the data generated by the generator (referred to as generated data) are combined with the original fNIRS data as augmented data. (c) Trial-averaged waveforms for the four tasks considered in a CV hold. The red lines denote measured original data and the blue lines denote generated data using WGANs. The shaded area represents 95% confidence intervals. Adapted from Ref. 26.

3.3.

Brain–Computer Interface

One of the most promising applications for fNIRS research is BCI. Many studies on BCI use traditional machine learning or DL techniques to identify a certain task based on cortical activation. Hennrich et al.⁵⁷ used a DNN to classify when subjects were performing mental arithmetic, word generation, mental rotation of an object and relaxation. The reported accuracy of the DNN with two hidden layers was 64.1%, which is comparable to the 65.7% accuracy of the shrinkage LDA despite the LDA using custom-built features while the input for the DNN was denoised and normalized fNIRS data. Kwon and Im⁵⁸ used similar inputs for their CNN, denoised and baseline corrected $\mathrm{\Delta }\mathrm{HbR}$ and $\mathrm{\Delta }{\mathrm{HbO}}_2$ data, to classify mental arithmetic from an idle fixation task. This CNN achieved a classification accuracy of 71.20%, which surpassed the 65.74% accuracy of the shrinkage LDA classifier that used feature vectors as inputs. Wickramaratne and Mahmud⁴² also used a CNN to classify mental arithmetic from an idle fixation task in 2021. Like Kwon and Im, a shrinkage LDA with feature vectors was used as the input. Unlike the previous study however, the inputs for the CNN were Gramian angular summation fields (GASF), which are a type of image that is constructed from time series data that maintains some temporal correlation between points. With this CNN and GASF inputs, a classification accuracy of 87.14% was achieved, once again outperforming the shrinkage LDA, which achieved an accuracy of 66.08%. Similarly, Ho et al.⁵⁹ also used a two-dimensional (2D) representation of fNIRS data to try and discriminate between differing levels of mental workload. In this study, Ho et al. found that using a CNN with spectrograms generated from fNIRS data achieved a classification accuracy of 82.77%. This was outperformed by a deep belief network, which is a type of network similar to an MLP. The deep belief network, using manually extracted features from the HbR and ${\mathrm{HbO}}_2$ signals achieved an accuracy of 84.26%. In a similar mental workload task, Asgher et al.⁶⁰ found that using an LSTM with similarly extracted features from HbR and ${\mathrm{HbO}}_2$ time signals resulted in a mental workload classification accuracy of 89.31%. While many of the studies presented compared the performance of DL techniques to that of traditional machine learning or other algorithms, Naseer et al.⁶¹ compared the classification accuracy of an MLP to that of a kNN, Naive Bayes, SVM, LDA, and QDA algorithm. On a two-class mental workload task, the MLP achieved an accuracy of 96% while the QDA, Naive Bayes, and SVM classifiers all achieved similar accuracies of about 95% and the LDA and kNN algorithms performed much worse, with accuracies of 80% and 65%, respectively. All classifiers in this study used manually extracted features such as skewness and kurtosis of the fNIRS as inputs into the algorithms. Hakimi et al.⁶² also used manually extracted features from fNIRS time series to perform two-class classification between stress and relaxation states and with a CNN, achieved an accuracy of 98.69%, further reinforcing that DL can give very high classification accuracies with mental workload tasks.

While the classification of mental tasks such as arithmetic or word generation is commonly used, many forms of BCI are designed to help those who have restricted or reduced motor function. Because of this, another popular type of experiment found in BCI studies involves executing motor tasks. Trakoolwilaiwan et al.²⁹ were able to achieve an accuracy of 92.68% with a CNN in a three-class test to distinguish the finger tapping of the left hand, right hand, and both hands at rest despite the CNN being used as both a feature extractor and classifier. The CNN outperformed the SVM and ANN classifiers, which reported accuracies of 86.19% and 89.35%, respectively, which were given the extracted feature inputs commonly used in BCI fNIRS studies (signal mean, variance, kurtosis, skewness, peak, and slope). Since much of the interest in BCI applications involve aiding those who lack the ability to move, some studies focus not on the cortical activations of movement but rather on the cortical activations of imagining movement. In one of these motor imagery studies, Janani et al.³⁰ had subjects perform a hand clenching or foot tapping task, and shortly after, imagine themselves performing the same task. Two different types of input images were used to see from which method a CNN classifier would more effectively extract features. The first type of input image turned all of the data points within a 20-s window into an $M\times N$ matrix, where $M$ is the number of data points and $N$ is the number of channels. The second type of input used a short-time Fourier transform to turn the one-dimensional data into 2D time-frequency maps of each channel that were stacked on top of each other to form an input image. One study by Erdoⓖan et al.⁶³ similarly performed classification between motor imagery versus motor execution using an MLP with a classification accuracy of 96.3% between finger tapping and rest and 80.1% accuracy between finger tapping and imagined finger tapping when using manually extracted features from fNIRS data. Hamid et al.⁶⁴ attempted to distinguish between a treadmill walking task and rest using bandpass filtered fNIRS data and an LSTM. The LSTM achieved an accuracy of 78.97%. When compared with traditional classifiers that had statistical features manually extracted, the kNN achieved the next best performance of 68.38% accuracy. The SVM and LDA were also outperformed by DL, achieving accuracies of 66.63% and 65.96%, respectively, despite using manually extracted features as inputs. This demonstrates the ability for DL to perform well on fNIRS data without requiring the extensive processing or feature extraction typically used in conjunction with other classifiers for fNIRS data. For many of these BCI motor tasks, being used for prosthetics would be more applicable. In these situations, more fine motor control using fNIRS would need to be assessed. Khan et al.⁶⁵ addressed this by performing six-class classification between rest and each finger on the right hand of the subjects, achieving an accuracy of 60%. Ortega and Faisal²⁴ attempted to distinguish between a left- and right-hand gripping task using a PCA to reduce dimensionality of the denoised time series data before feeding the segmented time series into a CNN-based architecture. The resulting accuracy of this study was 77%. To further investigate this, Ortega et al.³³ used a CNN with attention and simultaneously recorded EEG signals to try to reconstruct the grip force of each hand during the task. This resulted in an average fraction of variance accounted for of 55% when reconstructing the discrete grip force profiles, demonstrating that not only can the DL techniques distinguish which hand was performing a motor task, they also display progress toward using fNIRS and EEG signals to determine the amount of force exerted during that motor task. Ortega and Faisal⁶⁶ then attempted to use this architecture to determine force onset and which hand was providing more force. This resulted in a force onset detection of 85% but only a hand disentanglement accuracy of 53%, showing that there is still progress to be made toward the complex decoding and reconstruction of motor activities.

In the motor imagery tasks, the first input image method achieved an accuracy of 77.58% using ${\mathrm{HbO}}_2$ data, while the second method achieved an accuracy of 80.49% using ${\mathrm{HbO}}_2$ data. It could be noted that HbR and HbT were also tested but for both methods, ${\mathrm{HbO}}_2$ showed consistently higher results. While ${\mathrm{HbO}}_2$ is commonly used in fNIRS studies due to higher SNR, Yücel et al.⁹⁵ and Herold et al.⁹⁶ reported that trends found in ${\mathrm{HbO}}_2$ signals but not in HbR signals may be due to higher sensitivity of ${\mathrm{HbO}}_2$ to systemic signals not originating in the brain. For this reason, it is generally recommended that both HbR and ${\mathrm{HbO}}_2$ signals are assessed in fNIRS studies. Other traditional classifiers were also used and the closest results were achieved by the metacognitive radial basis function network, which achieved a classification accuracy of 80.83%. Another study that focused on motor imagery, Ma et al.³⁵ used a type of time series data that included a one-hot label as the input for the neural networks. Of the DL techniques used, the fully connected network (FCN) and residual network (ResNet) achieved the highest average accuracy of 98.6%. Of the traditional machine learning techniques tested, the SVM had the highest classification accuracy of 94.7%. Interestingly enough, for the DL networks, the mean classification accuracy achieved with ${\mathrm{HbO}}_2+\mathrm{HbR}+\mathrm{HbT}$ data, 98.3%, was the same as the accuracy when only HbR+HbT data were input, which was attributed to the feature extraction capabilities of the DL techniques. This study also evaluated the accuracy of individual channels, with single-channel classification accuracy ranging from 61.0% to 80.1% with the three highest accuracy channels being found in the somatosensory motor cortex and primary motor cortex.

While many studies have found considerable success in distinguishing between certain tasks using cortical activations, most studies use simple tasks in controlled environments and focus on distinguishing tasks from each other and resting state. For a practical BCI, more complex tasks and classification methods will need to be considered. Zhao et al.⁶⁷ addressed this by having participants perform a task in which they would pick up a table tennis ball with chopsticks and lift it about 20 cm in the air, using their nondominant hand. An LSTM was used to try to determine when the task was completed, and $\mathrm{\Delta }{\mathrm{HbO}}_2$ data were used as the input, resulting in an accuracy of 71.70%, outperforming the 66.6% accuracy of the SVM that was given mean, variance, kurtosis, and skew features of the fNIRS data as inputs.

One commonly used technology for BCI studies is EEG, due to the high temporal resolution and portability and noninvasiveness. Because it has a high temporal resolution but low spatial resolution, it is common to combine EEG measurements with fNIRS, due to both being portable and noninvasive. In a motor imagery study, Ghonchi et al.⁶⁸ used fNIRS to augment the EEG data being collected. Three types of DL networks were used as classifiers, a CNN for its capability to extract special features, an LSTM for its ability to extract temporal features, and a recurrent CNN (RCNN) for its ability to extract both temporal and spatial features. The RCNN achieved the highest classification accuracy of 99.6% when both EEG and fNIRS data were used. Interestingly, the accuracy of both the CNN and LSTM increased when fNIRS data were added, jumping from 85% and 81% to 98.2% and 95.8%, respectively. This indicates that despite EEG data having higher temporal resolution, fNIRS data still contribute both spatial and temporal information. A 2016 study by Chiarelli et al.⁶⁹ also found an increase in classification accuracy when combining EEG and fNIRS data. When performing two-class classification on a motor imagery task with an MLP, the average accuracies of EEG and fNIRS data alone were 73.38% and 71.92%, respectively, but when using both modalities, accuracy increased to 83.28%, further reinforcing that the simultaneous acquisition of EEG and fNIRS can provide more relevant information than either modality on their own. Cooney et al.⁷⁰ found that when combining fNIRS and EEG data, they were able to distinguish between multiple combinations of overt speech with a CNN classifier, achieving an accuracy of 46.31%. When tested on imagined speech, the classifier achieved an accuracy of 34.29%, which is also higher than the random chance value of 6.25% for 16 possible combinations, which shows promise in the use of EEG and fNIRS for assisting patients who may be unable to verbally communicate. Sun et al.⁷¹ used a CNN to try to distinguish between rest mental arithmetic and motor imagery tasks using both EEG and fNIRS data by generating tensors of the fused EEG and fNIRS data. For the motor imagery tasks, this resulted in an accuracy of 77.53% and for the mental arithmetic tasks, this resulted in an accuracy of 91.83%. Using the same dataset, Kwak et al.⁷² applied a branched CNN architecture that used the fNIRS data to generate spatial feature maps, which were then fed to the EEG maps to try to obtain higher spatial resolution than ordinary EEG and higher temporal resolution than fNIRS. The resulting classification accuracies were 78.97% for motor imagery and 91.96% for mental arithmetic tasks, which are only slight improvements over the tensor fusion methods of Sun et al. Khalil et al.⁷³ also used a fusion of fNIRS and EEG data to distinguish between rest and a mental workload tasks. With a CNN, an accuracy of 68.94% was achieved when trained on data from 5 of the 26 participants. When training on 16 participants then performing transfer learning to an additional five participants, accuracy increased to 94.52%, exemplifying not only how important larger datasets are for DL but also how transfer learning can be used to address this. It may be of interest to explore the use of transfer learning from other fNIRS datasets, which use different hardware systems, different tasks, or different fNIRS channel arrangements, since many studies rely on collecting fNIRS data specific to their own applications. As a result, it would be important to see if transfer learning can help extract meaningful features from fNIRS data independently of the part of the brain being recorded or the hardware being used. While many studies have only recently begun looking to use fNIRS for real-time BCI applications, current studies in the field have found use in DL techniques for increased classification accuracy and automatic feature extraction.

3.4.

Diagnostic Tools

One promising use of DL techniques with fNIRS is in clinical applications, most notably as a diagnostic tool. Xu et al.⁴⁶ recorded resting state fNIRS data from the bilateral frontal gyrus and bilateral temporal lobe of children. Using a single channel in the left temporal lobe, a CGRNN classifier was able to successfully classify autism spectrum disorder (ASD) in children with 92.2% accuracy, 85.0% sensitivity, and 99.4% specificity for 7 s of resting-state HbR data. As shown in Fig. 6, multiple ${\mathrm{HbO}}_2$ and HbR channels showed statistically significant differences between the group with ASD and the control group. Xu et al.⁷⁴ used a CNN classifier with the attention layers and achieved an accuracy sensitivity and specificity of 93.3%, 90.6%, and 97.5%, respectively, using similar resting state data to classify between ASD and typically developing (TD) subjects. Xu et al.⁷⁵ further explored using fNIRS data to detect autism is subjects and found that using a CNN+LSTM classifier and ${\mathrm{HbO}}_2$ data, they were able to achieve a single-channel accuracy, sensitivity and specificity of 95.7%, 97.1%, and 94.3%, respectively, for the detection of ASD. Aside from autism, the use of fNIRS to diagnose other psychological disorders has been studied. With an average classification accuracy of 96.2%, Ma et al.⁷⁶ were able to distinguish bipolar depression from major depressive disorder in adults during a verbal fluency task using an LSTM. Wang et al.⁷⁷ managed to distinguish between healthy subjects and those diagnosed with major depressive disorder with an accuracy of 83.3%. This was accomplished using long recording times of 150 min each from a relatively large sample size of 96 subjects. As can be seen in Table 2, this is a larger sample size than any of the other fNIRS studies presented, which lends to confidence in the generalizability of this neural network to new subjects. Chao et al.⁷⁸ used a cascade forward neural network (a network similar to an MLP) to perform and achieved an average classification accuracy of 99.94% between depressed and healthy subjects when a fear stimulus was presented across 32 subjects. Chou et al.⁷⁹ used an MLP network to classify between subjects with first episode schizophrenia and healthy subjects, achieving a classification accuracy of 79.7% with only about 160 s of recorded data from each subject.

Fig. 6

A DL model combining LSTM and CNN (LAC) was used to accurately identify ASD from a TD subject based on time-varying behavior of spontaneous hemodynamic fluctuations from fNIRS. Adapted from Ref. 75.

EEG technology is commonly used in clinical settings, however, some studies have used EEG alongside fNIRS for diagnostic studies. Sirpal et al.⁴⁴ used an LSTM architecture and fNIRS data to detect seizures with 97.0% accuracy, and with EEG data, achieved 97.6% accuracy, but with combined EEG and fNIRS data, accuracy increased to 98.3%, once again displaying how hybrid EEG-fNIRS recordings can increase classification accuracy, even when accuracy is already high. Rosas-Romero et al.⁸⁰ also combined fNIRS and EEG signals to detect epilepsy. Despite only having recordings from five subjects, a CNN was able to detect pre-ictal segments fNIRS and EEG data with an average accuracy of 99.67% with fivefold CV.

Another use for fNIRS is to help detect mild cognitive impairment (MCI), the prodromal stage of Alzheimer’s disease. Yang et al.⁸¹ employed three different strategies using CNNs to detect MCI. Using an N-back, Stroop, and verbal fluency task (VFT), a CNN trained on concentrations changes of ${\mathrm{HbO}}_2$ achieved accuracies ranging from 64.21% in the N-back task to 78.94% in the VFT. When using activation maps as the inputs for the CNN, the accuracy ranged from 71.59% in the VFT to 90.62% in the N-back task. The final strategy employed, using correlation maps as inputs showed lower accuracies than the activation maps, with the highest accuracy being 85.58% for the N-back task. Yang et al.⁴⁷ used temporal feature maps as inputs for the CNN classifier, resulting in average accuracies of 89.46%, 87.80%, and 90.37% with the N-back, Stroop, and VFT, respectively. In another study done in 2021 by Yang and Hong,⁸² pretrained networks were used to distinguish between subjects with MCI and the healthy control group. Using resting state fNIRS data, the network with the highest accuracy, VGG19, achieved an accuracy of 97.01% when connectivity maps were used as the input, outperforming the conventional machine learning techniques, with LDA classifier reporting the highest accuracy of 67.00%. Ho et al.⁸³ attempted to use fNIRS and DL techniques to distinguish not only between healthy and prodromal Alzheimer’s afflicted subjects but also subjects with asymptomatic Alzheimer’s disease and dementia due to Alzheimer’s disease. Not only did this study try to distinguish between different stages of Alzheimer’s disease, the study used a notably large sample size of 140 participants, which was larger than any other study reported in this review as shown in Table 2. The 86.8% accuracy of the CNN-LSTM network when fivefold cross-validated not only shows the ability of the network to distinguish between a wide range of subjects with and without Alzheimer’s disease but also shows the ability of this network to distinguish between different stages of Alzheimer’s disease, making this a very promising tool for clinical use.

Yet another clinical application for fNIRS measurements with DL techniques was explored by Fernandez Rojas et al.,⁴¹ where raw fNIRS data and an LSTM were used to distinguish between high and low levels of pain as well as whether the pain was caused by a hot or cold stimulus with an achieved accuracy of 90.6%. Being able to assess the intensity of pain as well as the cause of it could be exceptionally useful in instances where patients are unable to communicate, such as with nonverbal patients. This further displays the usefulness of fNIRS with DL techniques as a robust and accurate diagnostic tool.

3.5.

Analysis of Cortical Activations

Outside of BCI and diagnostic tools, fNIRS data still have many uses. Understanding the functional connectivity of the brain is an essential part of understanding the mechanisms behind numerous neurological phenomena. As a result, there is an interest in using neuroimaging to understand the functional connectivity of the brain. Behboodi et al.⁸⁴ used fNIRS to record the resting-state functional connectivity (RSFC) of the sensorimotor and motor regions of the brain. In this study, four methods were used, seed-based, ICA, ANN, and CNN. Unlike the first two methods, very limited preprocessing was used for the ANN and CNN, with both using filtered ${\mathrm{HbO}}_2$ data to form the connectivity maps. Each connectivity map was then compared with the expected activation based on the physiological location of each detector, which was used as the ground truth, using an ROC curve, as shown in Fig. 7, with the CNN achieving an area under the curve (AUC) of the receiver operating characteristic curve (ROC curve) of 0.92, which outperformed the ANN, ICA, and seed-based methods with each reporting an AUC of 0.89, 0.88, and 0.79, respectively. Another study that investigated RSFC, Sirpal et al.,⁸⁵ collected EEG and fNIRS data and attempted to use the EEG data and an LSTM to recreate the fNIRS signals. Using only the gamma bands of the EEG signal was found to have the lowest reconstruction error, below 0.25. The reconstructed fNIRS signals were further validated by comparing the functional connectivity of the signals constructed using only gamma bands and those using the full spectrum EEG signals with the functional connectivity of the experimental fNIRS data. Despite the signals formed using gamma bands only having a lower reconstruction error, the root MSE of the full spectrum EEG signals was consistently lower.

Fig. 7

(a) The cortical areas of hand-movement in sensorimotor and motor cortices. (b) Anatomical areas of sensorimotor and motor areas that are related to the hand-movement monitored by fNIRS are highlighted in green. (c) Group RSFC map derived from CNN-based resting-state connectivity detection. (d) ROC curves on different methods. Adapted from Ref. 84.

Other types of fNIRS studies have also been done that utilized DL. Some studies have analyzed the cortical activations of subjects to predict emotions. Bandara et al.⁸⁶ used music videos from the DEAP database⁹⁹ to classify the emotional valence and arousal of subjects using a CNN+LSTM architecture and fNIRS data recorded from the prefrontal cortex. Using the subjects’ self-assessments as ground-truth, a classification accuracy of 77% was reported. Another study collecting fNIRS signals from the prefrontal cortex, Qing et al.⁸⁷ used a CNN to determine the preference levels of subjects toward various Pepsi and Coca-Cola ads, achieving an average three-class classification accuracy of 87.9% for 30 s videos. 15 and 60 s videos showed similar accuracies of 84.3% and 86.4%, respectively. Similarly, Ramirez et al.⁹⁰ attempted to decode consumer preference toward 14 different products. With a CNN, Ramirez et al. were able to distinguish between a strong like and strong dislike of the presented product with an accuracy of 68.6% with fNIRS data, 77.98% with just EEG data, and 91.83% with combined fNIRS and EEG data. Hiwa et al.⁸⁸ studied the use of fNIRS and CNNs for the identification of a subject’s gender, achieving an accuracy of $\sim 60\%$ when using the filtered fNIRS data from only five channels.

While most studies have focused on using cortical activations to classify when a specified task is being completed, a few studies have begun looking into predicting the skill level of the subject at a given task. Andreu-Perez et al.⁸⁹ classified the expertise of subjects watching 30 s clips of the video game League of Legends using fNIRS data and facial expressions. The fully connected deep neural network (FCDNN) and deep classifier autoencoder (DCAE) were compared against many traditional machine learning techniques, including SVM and kNN in a three-class test to determine the skill level of the subject watching. Using only fNIRS data, the DL classifiers had the two highest accuracies of 89.84% and 90.70% for the FCDNN and DCAE, respectively, while the most accurate machine learning technique, SVM, only achieved an accuracy of around 58.23%. When the predicted emotion scores were also included, the accuracy of the FCDNN and DCAE improved to 91.44% and 91.43%, respectively. Most of the machine learning techniques also saw minor or no improvements, with the SVM still achieving an accuracy of 58.69%. The XGBoost classifier saw a large increase in accuracy when emotion scores were added, increasing from 50.79% to 71.55%, which was still about 20% lower than the accuracies of the DL techniques used. Another study by Gao et al.³² looked to predict the surgical skill level of medical school students who were being assessed on tasks based on the FLS protocols. In this study, subjects would perform an FLS precision cutting task while fNIRS data were recorded from the prefrontal cortex. The subjects would be assessed and scored in accordance with FLS procedures. A brain-NET architecture was used to predict the FLS scores of each participant based on the features extracted from the recorded fNIRS data. The brain-NET model reported an ROC AUC of 0.91, outperforming the kernel partial least squares, random forest, and support vector regression methods that were also tested when the dataset was sufficiently large. Figure 8 shows the $R^2$ value of each methodology as the size of the dataset increases. The ROC curve can also be seen along with a graphic of the experimental setup in Fig. 8. From the wide range of studies published, there are many different applications of fNIRS currently being researched, and the use of DL techniques has become of interest in recent years.

Fig. 8

(a) Schematic depicting the FLS box simulator where trainees perform the bimanual dexterity task. A continuous-wave spectrometer is used to measure functional brain activation via raw fNIRS signals in real time. (b) Examples of acquired time-series hemoglobin concentration data from six PFC locations while the subject is performing the PC surgical task. (c) The true versus predicted FLS score plots. Each blue dot represents one sample. The red line is the $y=x$ line. The dot-dashed green lines represent the certification pass/fail threshold FLS score value. (d) ROC curves for each model with corresponding AUC values in the legend. Adapted from Refs. 100 and 32.