Classification of Functional Near-Infrared Spectroscopy Based on Gramian Angular Difference Field and a...

Functional near-infrared spectroscopy (fNIRS) is a non-invasive functional neuroimaging technique widely employed in brain-computer interface (BCI) research and diverse clinical applications. The key challenge in fNIRS applications lies in extracting nonlinear structures and complex patterns from one-dimensional time series data. Gramian angular difference field (GADF) transforms one-dimensional time series into two-dimensional images, providing effective feature representation for subsequent signal classification. However, most studies have not explored the combined effects of image features and time series features. In this paper, we propose a deep learning model, VisiTempNet, which integrates both time series and GADF image features in a temporal-spatial fusion approach. The model first performs convolution on time series data based on delayed hemodynamic responses to highlight key features. It then separates the feature extraction process into two parallel modules, and normalizes and fuses these features with learnable weights, assigning greater importance to the most relevant information for classification. Experimental results show that our model achieved an accuracy of 76.65±2.43% on the open access fNIRS2MW dataset, outperforming all baseline models. This validates the effectiveness of combining image and time series features and demonstrates the superiority of the proposed model.