Research On Brain Neuron Reconstruction Algorithm Based On 3D Cross-over Structure Separation

Studying the morphology and structure of neuronal cells is crucial to understand the working mode and mechanisms of the brain.The key step is the reconstruction(tracing)of brain neurons in 3D optical microscopy images.In the past few decades,researchers in neuroscience and computing have developed many related methods,techniques and tools to study the digital morphological reconstruction of neurons.Several brain neuron reconstruction algorithms have been able to complete the reconstruction task in terms of coverage.However,due to a large number of cross-over structures among neuronal fibers,many reconstruction results may contain topological errors and incomplete reconstruction phenomenon.It is an urgent task to study how to detect the 3D cross-over structure of neurons,and separate them in the neuron images.Therefore,a brain neuron reconstruction algorithm based on 3D cross-over structure separation is proposed to improve the performance of neuron reconstruction.The main research contents are as follows:Firstly,3D cross-over points are detected in neuron image stacks.An anisotropic filtering based segmentation method is employed to extract the foreground of the neuron image,and then a 3D skeletonization algorithm is used to obtain the neuronal skeleton.All 3D skeleton points are recognized as neuron cross-over candidate points.Next,a spherical patch extraction method is used to extract the intensity distribution feature of voxels near the neuron cross-over candidate points,and then the features are fed into a 2D multi-stream convolutional neural network for learning and the network could detect the 3D cross-over points of the neuron.Secondly,3D neuronal cross-over structure features are extracted in the neuron image stacks,and the morphological reconstruction of brain neurons is realized based on the separation of 3D cross-over structure.3D cross-over points are used to locate the 3D cross-over structures in the neuron image stacks.To extract the morphological features of the cross-over structure,we proposed a multiscale upgraded ray-shooting model,through multi-scale extraction of the voxel intensity distribution around the candidate points,with high confidence to obtain robust cross structure features,such as the angle and length between neuronal fibers.Aiming at the detected cross-over structure,a 3D cross-over structure separation method is proposed to eliminate the false connection of the cross-over structure.Based on the extracted features,deformed and separated neuron fiber signal is generated to replace the original neuronal fiber signal of the cross-over structure.After the cross-over structure separation,a new neuron image will be obtained,which could be used for brain neuron reconstruction with the widely-used neuron morphology reconstruction algorithms.Finally,by testing the proposed algorithm on Big Neuron and SWMB neuron dataset,the experimental results clearly show that the proposed method can effectively improve the performance of neuron reconstruction.The results of quantitative analysis experiment and visual comparison both show that the proposed method is helpful to improve the topological performance of neuron reconstruction and generate more faithful reconstruction results.