Deep Learning From Diffuse Optical Oximetry Time-Series: An fNIRS-Focused Review of Recent Advancements and Future Directions

Abstract

Human neuroscience is undergoing a paradigm shift from traditional lab settings to natural environments. Functional Near Infrared Spectroscopy (fNIRS) and its variant, High-Density Diffuse Optical Tomography (HD-DOT) are rapidly evolving techniques that are increasingly adopted across disciplines. The high ease of use of advanced systems can enable continuous brain monitoring and thus the acquisition of large amounts of data. Integrating these data with modern deep learning (DL) promises to offer robust and generalizable solutions to ongoing challenges in fNIRS-related domains. As DL is a rather new field in fNIRS, we conduct a method-focused review, discussing 100 papers in the context of architectures, applications, and learning strategies. Based on the limitations in literature and the research gap between fNIRS and other domains, we conduct a tutorial study with guidelines from the wider DL field. We focus on: straightforward pre-processing pipelines; the trade-off between available data and model complexity of different architectures, including transformers; the generalizability of models for unseen data; and explainability. Finally, we provide a problem-focused discussion, gathering essential problems in the community, and introduce advanced DL solutions. This review serves as a strategic guide for advancing the current methodology for DL approaches in the fNIRS field.