Functional near-infrared spectroscopy (fNIRS) is currently widely applied in clinical research on functional brain activity states because of the advantages of fNIRS over conventional in vivo brain function detection techniques. fNIRS is a non-invasive and non-radiative technique that is resistant to electromagnetic interferences, provides a reasonable temporal/spatial resolution, and facilitates direct detection of blood oxygen metabolism. As an emerging reconstruction strategy for fNIRS, diffuse optical tomography (DOT) can complete the 3D reconstruction of optical parameters based on accurate photon transport models and can significantly improve the quantitative accuracy and spatial resolution of typical optical tomography techniques. Owing to the reflection measurement geometry of DOT, the detection data are affected by superficial physiological interferences (cardiac pulsation, respiration, and low-frequency oscillations) and random noises (photon-shot and instrumental noises) that originate from the scalp-skull layer; these interferences and noises affect the accuracy and precision of the reconstruction results. In addition, owing to limited boundary measurements, the inverse problem of the DOT has a non-negligible ill-posedness. Thus, handling the ill-posedness of the DOT inverse problem and suppressing physiological interferences and random noise are critical tasks in fNIRS-DOT neuroimaging. In this study, a model-based reconstruction-informed and deep learning approach, composed of a semi-three-dimensional (S3D) DOT and deep convolutional encoder–decoder neural network (DCNN), is developed to improve the reconstruction accuracy and suppress physiological interferences and random noises.

First, an S3D-DOT model is developed based on the properties of near-infrared light activation information distribution in the depth direction and reasonable assumptions about the structural characteristics of the brain. The S3D-DOT model can help in reducing the number of unreconstructed parameters, handle the ill-posedness of the DOT inverse problem, and preliminarily discriminate perturbation maps corresponding to the surface and cerebral-cortex (CC) layer. The preliminary reconstructed image is then used as an input to the subsequent DCNN model, which is composed of two parts, viz. a decoder network and an encoder network. The DCNN model can collect the spatial feature information of the image, effectively separate the activation and interference information, and accurately reconstruct the activation feature in the CC layer map. In general, the proposed model-informed deep-learning architecture is supported by physical models, exploits the spatial-information-extraction capability of convolution and encoding-decoding networks, and can provide highly quantitative and accurate reconstruction results in different application scenarios.

The structural design of the network, parameter selection process, and training process are described in detail. To verify the effectiveness of the proposed method, numerical simulations and phantom experiments are conducted using the fNIRS-DOT system. The final reconstructed images of the proposed method are compared with those obtained using the algebraic reconstruction technique (ART), and appropriate quantitative evaluation indices are selected for the computational analysis. The results of the numerical simulation experiments at specific time points show that the DCNN can effectively suppress the effects of physiological interference and random noise and improve the reconstruction accuracy, with a mean structure similarity index (SSIM) value of >0.998 (Fig.3). DCNN is more advantageous than the ART at a weak excitation time point, and the corresponding time required for reconstruction is significantly less. Subsequently, the performance of the DCNN model is examined under strong noise interferences. The corresponding results demonstrate that the conventional method cannot accurately reconstruct the excitation distribution under these conditions, whereas the proposed algorithm can still guarantee the validity of the reconstruction results (Fig.4). Additionally, the reconstruction capability of the DCNN in complex scenarios is verified through dual-target simulations (Fig.5). Furthermore, practical applicability of the proposed method is preliminarily examined through phantom experiments. The results indicate that the method can accurately filter random noise; however, the reconstructed image is still affected by physiological interferences when its relative intensity is large (Fig.7). Finally, a 3D deep convolutional encoder-decoder neural network (3D-DCNN) model is proposed to enhance the network’s ability to utilize temporal–spatial information and reasonably predict the changes in the excitatory brain regions. The results of the numerical simulation experiments prove that the 3D-DCNN model is more sensitive to small absorption changes and can accurately reconstruct the complete time courses of the average absorption perturbation in the activated region (Fig.9).

In this study, a model-based reconstruction-informed and deep learning approach is developed for enhancing the fNIRS-DOT performance. This proposed approach adopts the S3D-DOT model and DCNN to reduce image artifacts induced by physiological interferences and random noise. This method requires less hardware devices and provides an explicit physical explanation, an excellent accuracy and generalization for different scenes, and a fast reconstruction speed. To assess the effectiveness of the proposed method, a series of preliminary numerical simulations and phantom experiments are conducted, and the results are compared with those of the traditional reconstruction method. The results show that this method can significantly improve the quantification of images, greatly reduce the reconstruction time, and facilitate an excellent generalization, thereby providing an important new reference for dynamic fNIRS-DOT imaging.