Improved V-Net Based Image Segmentation for 3D Neuron Reconstruction

Digital reconstruction (tracing) of neuron morphology in volumetric microscopy images is critical for the analysis and understanding of neuron function. However, most of the existing neuron tracing methods are not applicable in challenging datasets where the neuron structures are contaminated by noises or have discontinued segments. In this paper, we propose a segmentation method, which serves as a preprocessing step prior to applying reconstruction methods, based on deep learning to identify the location of neuronal voxels. This preprocessing step is expected to enhance the neuronal structures and reduce the impact of image noise in the data, which would result in improved reconstruction results. We train 3D fully convolutional networks (FCNs) for segmenting the neuronal microscopy images in an end-to-end manner. Specifically, the V-Net architecture, which is a 3D FCN, is improved by using anisotropic convolution kernels and changing the number of layers to fit our neuron datasets. The improved V-Net takes 3D microscopy images as the inputs and their voxel-wise segmentation maps as the outputs. Experimental results show that this network improved the neuron tracing performance significantly in many challenging datasets when using different reconstruction methods.