Microstructure Parameters Fitting And Brain White Matter Analysis Based On Residual Network And Transfer Learning

In the field of brain imaging,diffusion tensor imaging(DTI)can provide information on white matter in the brain,which can be used to explore changes in the tissue structure of the brain area,but it lacks the specificity of the microstructure that distinguishes white matter and gray matter.With the development of medical imaging technology,Multi-compartment models for estimating microstructure specificity have emerged.The neurite orientation dispersion and density imaging(NODDI)model is the most popular multi-compartment model method for estimating the microstructure of brain tissue.But the NODDI model requires a large number of diffusion gradients to accurately estimate the microstructure of brain tissue.At the same time,the data is fitted by the maximum likelihood method.The calculation is complicated and takes a long time,which makes it unable to be widely used in clinical practice.In view of this situation,first,we propose an efficient method for fitting microstructure parameters through residual networks.This method replaces the parameter update step of the traditional proximal gradient method in a data-driven manner to solve the convex optimization problem and realize the parameter fitting.Taking root mean square error,structural similarity index,and peak signal-to-noise ratio as evaluation indicators,the method is compared in simulation dataset and real dataset.Then,in view of the dependence of network methods on data,and sometimes the data collection is time-consuming and difficult to obtain,this article takes ADHD dataset as the research object and attempts to use the transfer learning method to relax the network's restrictions on data.Finally,To verify the feasibility of the method proposed in this paper.We use diffusion tensor imaging method and microstructure method(including NODDI model,AMICO model and this method)to process ADHD data.And the TBSS analysis was carried out,which compared the advantages and disadvantages of the statistical analysis results of the three microstructure models used for the single-shell DTI data extraction orientation dispersion OD value,and deepened the research on ADHD disease.Because the single-shell DTI data is non-high-angle resolution.It is difficult to describe the direction of complex nerve fibers and analyze changes in tissue orientation dispersion.This paper adopts the microstructure method to overcome this difficulty.It is the first time to apply the microstructure method to single-shell data.The final experimental results are as follows: In the aspect of neural network microstructure parameter fitting,with different signal-to-noise ratios and different gradients simulation data as the experimental data,the comprehensive performance of the proposed method is superior to the NODDI and AMICO methods;With different gradient real Human Connectome Project(HCP)dataset as experiments data,under different gradients,the proposed method is better than the NODDI and AMICO methods,which breaks the limitation of the gradient,which is of great significance for data collection and post-processing;In the aspect of applied transfer learning,the need for training data and the time of data fitting are reduced without affecting the imaging results;In practical application,the method in this article takes less time and is the same as the statistically significant brain area obtained by the NODDI method,including: corpus callosum,left and right superior longitudinal fasciculus,left and right external capsule,right retrolenticular part of the internal capsule.Only three brain regions were statistically significant using the AMICO method,including:corpus callosum,right external capsule,right retrolenticular part of the internal capsule.it can be seen that it is not suitable for single-shell DTI data.And the brain regions with FA,AD,MD and RD values with statistical significance extracted by diffusion tensor imaging technique did not include the left external capsule and the right superior longitudinal fasciculus.It can be seen that microstructural methods can supplement diffusion tensor imaging method in the study of neurological diseases of the brain.