Study On Deep Neural Network Based Diffuse Optical Tomography

Diffuse optical tomography(DOT)is a non-invasive,safe and convenient biomedical imaging method.It reconstructs the optical parameter distribution inside bio tissues by the measured escaping photon intensity distribution on the boundary.But since the gap between the amount of detected and expected parameters,the algorithm is an underdetermined inversion problem inherited from the physical setup of the system configuration.Therefore,using a priori information acquired from other methods is proved to be feasible to improve the imaging quality and speed.Monte Carlo simulation,as a flexible and accurate calculation model for light propagation,can provide empirical information with the same format as the measured light intensity distribution in a way that do not increase the system complexity.In this paper,we propose a deep convolutional neural network model,which feeds the a priori in Monte Carlo simulations to the calculation of photon propagation,to realize rapid and accurate forward calculation model.And depend on this model,a rapid heterogenous region imaging deep convolutional neural network is implemented.Firstly,to satisfy the demand of training data for neural network applications,we modified the parallel calculating mode for the Monte Carlo simulation tool.Since the discrepancy of light path during the lifetime of the photons,waiting time between the threads causes waste of computational resources.By dividing the lifetime of parallel running threads to part of the whole lifetime of a photon dispersed by scattering number,the modified model increase the computation speed to 1.66 times to non-optimized model and 32.6 times to non-parallel model under the complexity of real human head structure.A case study of bio-optical application,which estimates the effect of follicle to near infrared spectrum measurement,using Monte Carlo simulation dataset is reported then.According to the theory,effect of follicle in various distribution density to directly measured light intensity can be estimated by exponential approximation,while the effect may lead to misinterpretation of hemodynamic change tendency according to the algorithms for hemodynamic parameter calculation.The real human structure based Monte Carlo simulation library provided the parameter needed in the model.At the same time,the process can be used for further digital phantom library production.For the computational model of light transport in bio tissues using artificial neural network,convolution network is proposed to simulate the light propagation process.Experimental results show that the convolution kernels,which represents the point spread function of certain optical parameters,can be acquired by the optimization process for the error between simulation and forward calculated result.And by using two distinguish kernels for forward and backward propagating characteristic,or one 3-D kernel,this model can accurately proceed the propagation for homogeneous medium and the kernels can be further used for calculation of inhomogeneous medium.Based on the transportation model using deep convolutional network,two modified structures are proposed to calculate the position of optically heterogeneous region by the measured outgoing photon intensity distribution and corresponding Monte Carlo simulation library is created.Four input and output types for the two network structures are trained and validated.The validation result shows that deep convolutional network can be used to rapidly image the position and depth of the heterogeneous region.