Deep Learning Methods For Spatial Frequency Domain Optical Topography

Spatial frequency domain(SFD)imaging is a burgeoning optical imaging technology,which can realize the quantitative measurement of optical properties in a large field of view.With the advantages of fast,noninvasive and high cost performance,it has attracted extensive attention and has been applied to biomedical imaging fields in recent years.However,SFD imaging still faces some problems in the actual measurement.Firstly,the traditional SFD imaging is time-consuming in the process of demodulation of measured images and look-up table reconstruction of optical properties,which limits the requirement of real time imaging;Secondly,the profile of the tissue will distort the modulation characteristics of the measured images and then affect the accuracy of optical properties reconstruction.In addition,the current SFD imaging usually takes biological tissue as a monolayer tissue for two-dimensional optical topological(OT),but ignores the non-uniformity of tissue along the depth direction(approximate multilayer distribution of optical properties).In recent years,deep learning has developed rapidly and gradually applied to the field of traditional imaging,which brings many advantages such as imaging speed,accuracy and so on.Aiming at the above problems,this paper combines SFD imaging with deep learning,developed a deep learning method for SFD-OT,in order to improve the real-time and quantitative of SFD imaging,and improve its clinical practicability.For SFD-OT with monolayer tissue,fast amplitude demodulation and optical properties reconstruction are developed by generative adversarial network and back propagation neural network.Phantom and in vivo experiments were carried out based on the model performance evaluation.The results show that the developed method has good accuracy and generalization for different scenes.Compared with traditional imaging methods,the time required for demodulation and reconstruction is greatly reduced,which is expected to realize real-time imaging application of SFD-OT.For the physiological tissue with complex profile,a SFD measurement image correction method based on target profile is developed in this paper,which corrects the height and angle related intensity of SFD diffuse reflection image from the incident and reflected directions respectively,and further corrects the frequency of the target region by using the spatial frequency interpolation method.Phantom and in vivo experimental results show that this method can significantly improve the reconstruction accuracy of optical properties of tissues with different profiles,which provides a feasible scheme for clinical measurement of SFD imaging.Based on the above methods,this paper further studies the deep learning method for SFD-OT with layered tissue.Firstly,skin-subcutaneous and epidermal-dermis models are constructed for human skin structure.The optical properties of multi-layer structures can be reconstructed by selecting appropriate spatial frequencies based on the penetration depth difference of modulated light at different spatial frequencies.Simulation and experimental results show that the optical properties of the top layer with a thickness of 2mm can be accurately reconstructed at the selected spatial frequency.Finally,the SFD-OT deep learning method for the reconstruction of absorption coefficients of four-layer tissue is explored in this paper,which provides ideas for the SFD imaging of multi-layer tissue.