Significance: In functional near-infrared spectroscopy (fNIRS) research, ensuring signal quality is a critical preprocessing step. However, traditional index-based metrics such as the coefficient of variation (CV) and scalp coupling index (SCI) rely on arbitrary thresholds and often misclassify channels. Aim: We present DL-QC-fNIRS, a deep learning framework for the automated, channel-wise assessment of signal quality. Approach: Our method involves generating continuous wavelet transform scalograms of oxyhemoglobin signals and employing subject-specific cardiac frequency extraction to improve physiological specificity. These inputs are then classified using convolutional neural networks (CNNs). We benchmarked four CNN architectures (GoogLeNet, ResNet-50, SqueezeNet, and EfficientNet-B0) on two independent datasets and one combined heterogeneous dataset. Results: GoogLeNet achieved the highest accuracy ( ) on the combined dataset, demonstrating strong sensitivity and specificity across test sets. Compared with CV and SCI, DL-QC-fNIRS yielded markedly higher F1-scores and a more favorable balance between sensitivity and specificity. Conclusions: DL-QC-fNIRS is provided as an open-source MATLAB-based graphical interface, enabling accessible and standardized integration into fNIRS workflows. These findings highlight DL-QC-fNIRS as a scalable, expert-level tool for improving the reliability and reproducibility of optical neuroimaging data.
Guglielmini et al. (Thu,) studied this question.