Research On Neuron Critical Points Detection And Morphology Reconstruction In 3D Brain Images

Brain science is one of the most cutting-edge science and technology fields in the 21st century.It is important for the prevention and treatment of brain diseases,and the research of brain-inspired intelligence.Using microscopic imaging system to collect neuronal images of the brain and digitally reconstruct the morphology of neurons are helpful in studying how the brain works,which are important approaches for brain science research.Since the robustness and accuracy of the existing automatic neuron reconstruction methods cannot meet the practical needs,neuronal morphology reconstruction still relies mainly on manual or semi-automatic methods.Manual methods trace the morphology of neurons in three-dimensional(3D)images manually with the help of visualization software.Semi-automatic methods usually require manual selection of seed points(e.g.,critical points of neurons)in images,and image analysis algorithms are then used to automatically link the seed points to complete the reconstruction.However,with the rapid development of microscopic imaging technology,the volume of 3D brain images has grown to terabyte-scale,and neither manual nor semi-automatic algorithms can efficiently handle such a huge amount of data.Therefore,it is still an open problem to automatically and accurately reconstruct the morphology of neurons from large-scale brain microscopic imaging data.To address this problem,this thesis investigates neuronal morphology reconstruction methods in 3D images from two aspects.On the one hand,two neuron critical points(i.e.,terminations,branch points and cross-over points)detection methods are proposed,and the detected critical points are combined with existing neuronal morphology reconstruction algorithms to realize critical-points based neuronal morphology reconstruction.On the other hand,this thesis proposes a novel deep-learning based method for neuron morphology reconstruction in 3D images.It is the first attempt to train a deeplearning based neuron reconstruction algorithm using synthetic images,which shows satisfying performance in real test images.The specific research contents are as follows:1.Multiscale Ray-Shooting Model Based Neuron Terminations Detection and Morphology ReconstructionTo achieve robust detection of terminations,this thesis proposes a 3D terminations detection algorithm based on a multiscale ray-shooting model and a termination visual prior.The multi scale ray-shooting model can extract and analyze the voxel intensity distribution features around the termination candidates from multiple scales,adapt to the change of diameters,and effectively reject the small spurs on the surface of neuronal structures to achieve robust detection of 2D terminations.The proposed termination visual prior is based on a key observation-when observing a 3D termination from three orthogonal directions without occlusion,we can recognize it in at least two views.By combining it with the multiscale ray-shooting model,the 3D terminations can be detected.The proposed method can not only detect the terminations of neurons in 3D images,but also detect the terminations of other tree-like structures,such as bronchi in 3D images and retinal blood vessels in 2D images.Experimental results demonstrate that the accuracy of the proposed method is higher than the compared termination detection methods,and prove that the terminations can be used as seed points for the existing neuron reconstruction algorithms.2.Spherical Patches Extraction Based Neuron Critical Points Detection and Morphology ReconstructionTo simultaneously detect three types of critical points,this thesis proposes a neuron critical point detection method in 3D microscopy images based on the Spherical Patches Extraction(SPE)and convolutional neural networks(CNNs),providing sufficient seed points to the existing neuron reconstruction algorithms.The SPE algorithm extracts the voxel intensity distribution features around a critical point candidate with a set of concentric spheres.It then projects the extracted features into 2D space by coordinate transformation to generate a set of 2D spherical patches so that 2D CNNs can be used to detect critical points in 3D images.Then,we design a 2D multi-stream CNN to classify the given critical point candidates into four classes:termination,branch point,cross-over point or non-critical point.Each stream of the 2D multi-stream CNN takes a spherical patch as input.Experimental results show that the accuracy of the proposed neuron critical point detection method is higher than the compared state-of-the-art critical point detection methods.The 3D neuronal morphology reconstruction results based on three types of critical points show that the detected critical points can be used as seed points for neuron reconstruction and the performances are better than the neuron reconstruction method based on only terminations.In addition,to advance the development of automatic analysis methods for brain neuron images,such as 3D neuron critical point detection and automatic seed-based neuron reconstruction,we established a dataset called Whole Mouse Brain Sub-Image(WMBS)for neuron critical point detection and made it publicly available.3.Automatic 3D Neuronal Morphology Reconstruction Based on Deep Learning and Synthetic Training ImagesTo accurately reconstruct the morphology of neurons,the thesis proposes a novel method for automatic neuronal morphology reconstruction,called SPE-DNR.It combines the SPE algorithm and deep learning to achieve automatic neuronal morphology reconstruction in 3D brain images.Based on 2D CNN and the spherical patches extracted by the SPE algorithm,SPE-DNR can determine the local direction of neuronal structures while classifying any voxels into foreground and background.In this way,SPE-DNR can automatically trace the centerline of neuronal structures starting from a set of seed points and determine when to stop tracing,thus completing neuronal morphology reconstruction.In addition,to avoid introducing erroneous training labels to SPE-DNR caused by imperfect manual annotations,we propose an image synthesis method to generate synthetic training images with exact labels.The image synthesis method simulates the noise,gaps and abrupt radius changes of neuronal structures in 3D neuron images,which results in more realistic synthetic images.To demonstrate the robustness and generalization ability of our SPE-DNR,we validated it on real 3D neuron images from three datasets.Experimental results demonstrate that SPE-DNR has better overall performances compared with the proposed critical-points based neuron reconstruction methods and other widely used neuronal morphology reconstruction algorithms.