Neuron Image Segmentation Based On Multi-level 3D Fully Convolutional Networks And Hessian Matrix

The brain has always been a focus of biomedical imaging research,and neurons are the basic components of the brain.To understand how the brain works,researchers devote themselves to studying the structure and function of neurons and their connections.Noises and weak signal structures are inevitably produced in the process of acquiring 3D neuron images,which are limit to the follow-up study of neuron images,such as feature points detection and neuron reconstruction(tracing).Therefore,a neuron image segmentation algorithm,based on multi-level 3D fully convolutional networks and Hessian matrix,is proposed to segment the neuron images,which is of benefit to remove noise and enhance weak signal.The main research contents are as follows:Firstly,V-Net,a fully convolutional network used in medical image segmentation,is improved to construct the multi-level 3D fully convolutional networks to segment the neuron images.1.Appropriate number of network layers are used to adapt the input size of the network and ensure the size of the feature map at the bottom of the network,so as to learn the image features better.2.In the neuron images,considering the difference of the resolution and the size in X direction,Y direction and Z direction,the anisotropic convolution kernel is used in the network.3.Due to the differences in structure and size of different neuron species,inception model is added in the network to increase the applicability of various neuron images.4.In the neuron images,the number of positive pixels is much smaller than the number of negative pixels.In order to solve this imbalance problem,the weighted cross entropy loss is used in the multi-level 3D fully convolution networks.Secondly,due to the discontinued structures in neuron segmentation results which were obtained by the multi-level 3D fully convolution networks,the repair model based on Hessian matrix analysis is built to repair the discontinued structures.The first step of this model is to identify the location that needs to be repaired,which is called break point detection.Considering that the break point can be regarded as the termination point of the neuron segment,the break point is detected by the multi-scale ray-shooting model.Meanwhile,the ray-burst model is used to estimate the local diameter and detect the direction of the neuron segment,thus the region to be repaired is determined.The next step is to calculate the eigenvalues of Hessian matrix.According to the value of the eigenvalues and the relationship between them,each point in that region is identified as a foreground or background point.Then the neuron segment is repaired and the repaired neuron segmentation results are obtained.Finally,the neuron images in BigNeuron dataset were segmented by the proposed segmentation method,the feature points were detected and the neuron structure was reconstructed on the segmented images.Experimental results show that the segmentation algorithm proposed in this paper has high accuracy,and significantly improves the feature point detection accuracy and neuron reconstruction performance.That demonstrates the effectiveness and feasibility of the proposed segmentation algorithm.