Research On Human Brain Fiber Orientation Estimation Method Based On Deep Learnin

Diffusion weighted magnetic resonance imaging(d MRI)has become an important means of studying living tissue in the field of neuroscience.It is currently the only noninvasive technology for studying white matter fibers in the living brain,and has played an important role in the clinical diagnosis of neurological diseases and related scientific research.Among them,the reconstruction of fiber orientation distribution function(f ODF)is of great significance in solving complex multi fiber intersection situations to complete reliable brain connectivity maps,and is helpful in understanding brain connectivity and its neural functions.However,in traditional non deep learning methods,a large amount of high-quality data is usually required to obtain significant f ODF reconstruction results.Under actual clinical conditions,a large amount of highquality data is difficult to obtain,and excessive acquisition time can generate motion artifacts.This article aims to optimize the estimation process of human brain fiber orientation based on deep learning technology to overcome the shortcomings of traditional methods.We will analyze and improve existing deep learning models.Through the application of deep learning technology,we can reduce data requirements,improve computational efficiency,and reduce the impact of motion artifacts.Our research mainly includes the following aspects: analyzing the performance of existing deep learning models to explore their advantages and disadvantages;A novel f ODF reconstruction method based on deep learning is proposed to improve the shortcomings of traditional methods;Design and implement an efficient deep learning network to improve the accuracy and computational efficiency of f ODF reconstruction;Conduct experimental validation to evaluate the performance and feasibility of the new method.We hope that this study can provide more efficient and accurate d MRI data analysis methods for the field of neuroscience,and provide better support for the clinical diagnosis and treatment of neurological diseases.The main research contents are as follows:(1)In response to the issue of unreasonable dataset construction,this article uses spherical harmonic functions to resample the original data,in order to better simulate actual clinical situations and improve the accuracy and practicality of the model.The spherical harmonic function has low error in characterizing spherical data such as d MRI data and is widely used.The publicly available HCP dataset on the internet contains data with three different b-values,each with 90 gradient directions.However,in order to approach clinical data conditions,it is necessary to extract only one b-value of the data and perform reasonable downsampling to reduce the number of gradient directions.The traditional downsampling method randomly selects different numbers of gradient directions in the dataset for sampling,but this method can easily lead to gradient directions not meeting the principle of electron exclusion.By transforming the discrete spherical function into spherical harmonic function,and then re selecting reasonable sampling points,and then transforming it into discrete spherical function,we can get lower quality data closer to clinical conditions.(2)For the analysis and design of deep learning networks,this article mainly studies the analysis and design of deep learning networks,and attempts to apply them to the reconstruction of f ODF.To achieve this goal,we replicated and analyzed the mainstream network structures in the field,and innovatively used spherical convolutional networks to reconstruct f ODF.When designing a network structure,we need to consider the magnitude and performance of the parameters of the model.Due to the sampling characteristics of d MRI data,we use a spherical convolutional neural network(S-CNN)to better utilize the spatial structure of the input d MRI data.We use the vertex coordinates of the icosahedron and its fractal body as sampling coordinates to achieve the goal of uniform sampling.This method can effectively improve the performance of the network and can be applied to various types of d MRI data.Under the condition of single shell data,our network exhibits better results than traditional non learning methods.This indicates that our network can better utilize the characteristics of input data and can be used for f ODF reconstruction tasks.(3)For the evaluation and analysis of f ODF reconstruction results,this paper comprehensively considers the mainstream evaluation methods within the neighborhood.While conducting a comprehensive,objective and scientific analysis of the results,we conducted a more detailed analysis of the results in combination with the downstream fiber tracing process.By analyzing the results,we found some key and difficult points in the f ODF reconstruction process,such as the impact of data noise on the results,the impact of different parameters on the results,and so on.These analysis results provide insights into key research directions in the research field,enabling researchers to gain a deeper understanding of the process and methods of f ODF reconstruction,and providing guidance for future research.