The Construction Of Neuronal Axons Reconstruction Tool Based On The Gaussian Mixture Model

Neuron morphology reconstruction is to quantify the topological structure and geometric characteristics of neurons from three-dimensional neural images.Neuroscience research shows that the quantitative data of neurons can be directly used for statistical analysis and biological analysis related to neuron morphology.The quantitative data of neuron morphology has become a bridge between neuroimage data and new knowledge in neuroscience.Neurons are composed of cell bodies and nerve fibers.The projection of nerve fibers reveals the process of information transmission.Nerve fiber reconstruction is the most important part of neuron reconstruction.Nowadays,the efficiency of nerve fiber reconstruction tools is low.One of the reasons is that it is difficult to effectively identify the direction of the nerve fiber when separating the entangled nerve fiber,so that the nerve fiber skeleton point cannot be extracted correctly.This will increase the subsequent manual modification tasks and prolong the reconstruction time.Therefore,the development of nerve fiber reconstruction methods and tools that can solve the above problems is an urgent need.In response to the above-mentioned problems,this article developed an automatic nerve fiber reconstruction method,and developed an interactive tool for nerve fiber reconstruction on the basis of this method.The main work is divided into the following points:(1)This paper uses the residual convolutional neural network to segment the threedimensional nerve fiber image,and then obtain the nerve fiber signal point set.In order to be able to adapt to diverse neural images,this paper proposes a training method based on transfer learning.This method fine-tunes the parameters of the original model by adding a small amount of new samples,so that the new model performs better than the original model on the new dataset.(2)This paper proposes a clustering method of neural fiber signal points based on Gaussian mixture model.The algorithm is based on the data of the neural fiber foreground point set,and then constructs the Gaussian mixture model equation,and then divides the foreground point set into several small areas,and finally forms a complete nerve fiber shape by connecting the small areas.This method fits each small area into an ellipsoid shape,and limits the radius of the ellipsoid to solve difficult phenomena such as short distances between fibers and false connections.(3)This paper develops an interactive tool for nerve fiber reconstruction.This tool is to characterize nerve fiber morphology in foreground point set.This tool integrates the deep learning segmentation method and transfer learning training method,and provides a simple and easy-to-operate interface,which can be used directly by non-machine learning researchers.The tool also provides the CPU multi-threading and GPU hardware interface of the automatic reconstruction method.On this basis,the tool provides a visual environment for three-dimensional images and nerve fiber morphology,as well as interactive functions such as manual modification of fiber morphology.The neural fiber automatic reconstruction method proposed in this paper was tested on the f MOST dataset and the public datasets of Big Neuron and DIADEM.The reconstruction results on these data sets prove that the method in this paper achieves a good reconstruction index.On the f MOST dataset,the average reconstruction precision rate reached 0.94,and the average recall rate reached 0.94.The reconstruction index is higher than the existing three automatic reconstruction tools: Vaa3 d,GTree,Neuron Studio.The neural fiber interactive tool developed in this article integrates the deep learning segmentation module and the transfer learning training module.The use of these modules does not need to rely on machine learning experience.The tool extends the automatic reconstruction method to GPU hardware.The GPU method runs 35 times faster than CPU multi-threading,which can efficiently complete the automatic nerve fiber reconstruction work.The tool also developed the interactive function of manual modification of fiber morphology.At the end of this article,the accurate morphological data of nerve fibers that are automatically reconstructed and manually modified using the tools are displayed.