| Functional near infrared spectroscopy(fNIRS)can infer the neuronal activity by measuring the absorption perturbation induced by hemodynamic responses in the brain.It is becoming a highly popular modality for implementing neuroimaging application,due to its safety,inexpensive instrumentation,fewer limitations on the subjects and reasonable temporal-spatial resolution.In fNIRS-based neuroimaging,prior to reaching the cortical-layer,the near-infrared lights must travel through superficial layers(e.g.,scalp,skull),and then travel out from the scalp before measuring by the detectors.Due to the absorption and scattering,the activation in the fNIRS signals is seriously attenuated and exists complicated non-linear relationship.In addition,the interferences caused by heartbeat,breathing,etc.and random noises caused by measurement will also induce perturbation.These factors will seriously limit the sensitivity,quantitation and temporal-spatial resolution of the fNIRS-based neuroimaging,and undermine the feasibility,depth and breadth of clinical application.This thesis focuses on the sensitivity,quantitation and temporal-spatial resolution improvement of fNIRS-based neuroimaging,and the main contents are as follows:(1)A 3-wavelength(785nm,808 nm and 830nm),240-channel(20 sources and 12detectors)fNIRS instrument is developed,which features the square-wave modulation lock-in photon-counting scheme that combines multi-channel parallelization of lock-in detection,ultra-high sensitivity and large dynamic-range of photon-counting technique.Phantom experiments are conducted to investigate specifications of the instrument,and the results show the instrument can operate steadily with large dynamic range(>106d B),and detect the absorption perturbation in the deep tissue under high temporal resolution.(2)In addition to high performance instrument,the application of a reconstruction scheme,which can effectively suppress physical interferences and random noises,is also an important way to improve fNIRS-based neuroimaging.Herein,a new scheme is explored by combining semi three-dimensional(S3D)reconstruction and image fusion,referred to S3D-Fusion.Firstly,the physiological interferences are estimated using S3 D reconstruction,and then the topography map covering the cortical-layer is adaptively filtered using the estimated signals.Secondly,the filtered reconstruction is decomposed by wavelet transform,and fused to generate two-dimensional template.Finally,the reconstruction is weighted based on the template for resolution improvement.Numerical simulation and phantom experiments prove the ability to suppress physiological interferences and random noises without repeated measurements,which significantly improve the reconstruction performance.(3)To cope with the 3D information loss in the S3D-Fusion,a new scheme is described by combining superficial reconstruction inversion filtering(IF)and long short-term memory(LSTM)classification(LC),referred to IF-LC.In IF process,the superficial perturbation is firstly pre-generated using a S3 D scheme,and the perturbation is inversed into measurement space to estimate the changes induced by the physiological interferences,which are served as reference for adaptively filtering the raw measurements.Then the 3D absorption perturbation is generated by diffuse optical tomography.In LC process,a LSTM classification network is trained based on the taskrelated hemodynamic response,with which to generate 3D template by classifying the reconstruction.Finally,this 3D template is applied to weight the filtered reconstruction.Numerical simulation proves the method’s ability to improve the resolution,and in-vivo experiments prove the ability to improve the classification accuracy.(4)Due to simplicity of probe configuration,lattice array is widely used in practice.Yet,both S3D-Fusion and IF-LC need overlapping array to ensure accuracy.To break this limit,a scheme with application in lattice array is investigated by combining LSTM predication and Kalman filtering,referred to LSTM-Kalman.Firstly,a LSTM network is trained with the perturbation during baseline for interferences profiling in the task stimulation.Then,based on the prior information from LSTM predication network,a Kalman filtering process is applied to process the measurements for real-time optimal estimation of the task-related perturbation.This method is validated using numerical simulation and in-vivo experiments,showing the comprehensively improved performance in reconstruction and classification achieved in a purely data-driven way.Overall,the sensitivity and quantitation of fNIRS-based neuroimaging is improved using high-sensitivity instrumentation and advanced reconstruction scheme.Effectiveness of the developed instrument and schemes is investigated by numerical simulation,phantom and in-vivo experiments.The results in this thesis provide powerful system support and significant theoretical guidance for fNIRS-based neuroimaging. |
