Objective Functional near-infrared spectroscopy (fNIRS) has several advantages, such as noninvasiveness, free radiation, and reasonable temporal/spatial resolution. This enables fNIRS-based technologies to be used as an alternative to conventional technologies, such as functional magnetic response imaging (fMRI) and electroencephalogram (EEG), and the technologies are increasingly used in clinical practice to complete neuroimaging. However, because of the reflection geometry used in fNIRS, light travels from a source, through the scalp-skull layer, into the brain, and back out through the scalp-skull layer to be measured by a detector, which decays significantly as depth increases. Therefore, the reconstructed activation using fNIRS is usually contaminated by superficial physiological signals (cardiac pulsation, respiration, and low-frequency oscillations, etc.). Besides, the random interferences induced by the photon-shot and instrumental noises, etc., also have blurring effects on the activation reconstruction because of faint activated hemodynamics in the brain. Thus, suppressing the irritating physiological interferences and random noises has been a critical task in fNIRS-based neuroimaging. In this work, we propose a long-short-term-memory (LSTM) based recurrent neural network (RNN), including a prediction and a classification layer, to suppress physiological interferences and random noises, respectively, to improve reconstruction performance with less repetitive or even individual stimulation. This has some advantages, including shorter measurement time, more subjects, and the ability to examine responses to single stimulation.
Methods The proposed LSTM-based RNN, which is purely data-driven without an auxiliary measurement process, comprises two layers: First, the prediction layer is used to estimate the absorption perturbation induced by the physiological interferences during task stimulation. Then, the estimated time series is used as the reference to adaptively filter the reconstructed absorption perturbation for the removal of the interferences from the physiological signals. Second, the classification layer is applied to reduce the remaining artifacts induced by the random noises in measurements for acquiring a better space-localized solution, converting the filtering procedure to a binary classification problem. Notably, the combination of the space-time filtered results from the predication layer is used as the input to the classification layer, ensuring the robustness and efficiency of the proposed method.
Results and Discussions The numerical simulations and in-vivo experiments are implemented based on fNIRS--DOT (diffuse optical tomography) to describe the network design, training, and filtering process in detail , and the effectiveness of the proposed method is compared with the reference filtering and cycle averaging method (RFCA). The results show that the proposed LSTM-based model improves reconstruction performance for the numerical simulations (Fig. 6) by effectively suppressing the physiological interferences and random noises rather than using more measurement cycles. Furthermore, we examine the effectiveness of the proposed method to deal with the potential time lags of superficial interferences compared with those in the cerebral cortex layers, and the results show that the proposed method performs better under the mentioned condition (Fig. 10). As for the in-vivo experiments, the results from the predication layer show comparable performance as the RFCA, whereas the results from the classification layer show a more concentrated activated region (Fig. 9). Because other modality imaging techniques have not been used to cross verify the activated region, determining whether the proposed model is over-optimized for the activated region is difficult. Thus, training the filtering model to avoid this problem will be an important direction for future work.
Conclusions In this paper, we propose a two-layer LSTM-based RNN that utilizes the prediction and classification of the RNN model to reduce the image artifacts induced by the physiological interferences and random noises in fNIRS-based neuroimaging. The proposed method has a clear physical explanation and needs no additional hardware cost. To evaluate the proposed method, a series of preliminary numerical simulations and in-vivo experiments were implemented, and the results show that it has a promising future for achieving reasonable enhancements, providing a practical approach for the fNIRS-based brain-computer interface application.
