Reconstructing damaged fNIRS signals with a generative deep learning model

Abstract

Functional near-infrared spectroscopy (fNIRS) imaging offers a promising avenue for measuring brain function in both healthy and diseased cohorts. However, signal quality in fNIRS data frequently encounters challenges, such as low signal-to-noise ratio or substantial motion artifacts in one or multiple measurement channels, impeding the comprehensive exploitation of the data. Developing a valid method to improve the quality of damaged fNIRS signals is crucial, particularly given the extensive use of wearable fNIRS devices in natural settings where noise issues are even more unavoidable. Here, we proposed a generative deep learning approach to recover damaged fNIRS signals in one or more measurement channels. The model captured spatial and temporal variations in the time series of fNIRS data by integrating multiscale convolutional layers, gated recurrent units (GRUs), and linear regression analyses. We trained the model on a resting-state fNIRS dataset from healthy elderly individuals and evaluated its performance in terms of reconstruction accuracy and functional connectivity matrix similarity. Collectively, the proposed model exhbited an excellent performance for the reconstruction of damaged fNIRS time series. In individual channel-level, the model can accurately reconstruct damaged fNIRS time series (mean correlation = 0.80 ± 0.14) while preserving intervariable relationships (correlation = 0.93). In multiple channel-level, the model maintained robust reconstruction accuracy and consistency in terms of functional connectivity. Our findings underscore the potential of generative deep learning techniques in reconstructing damaged fNIRS signals, providing a novel perspective for the efficient utilization of data in clinical diagnosis and brain research.