Automatic segmentation is a core technology for dendritic spines detection, identification and reconstruction. A novel level set segmentation method is proposed for neuron object in 3D fluorescence confocal images. In the fist step, 3D image are smoothed and enhanced by curvature anisotropic diffusion filter, and then the level set speed images are computed. In the second step, the seed points of neuron object are computed automatically located at the extreme local iso-surface curvature, which correspond to the local ridge or valley points on neuron object. Then in the third step, the fast marching method is used to produce the initial level set shape images. In the last step, the initial level set shape images are passed as input to the shape detection based level set algorithm to compute the final 3D neuron object. This method reduces the computation time by minimizing level set propagation, which converges at the optimal object within a fixed iteration number. Experiments on 2-photon lasers or 3D fluorescence confocal images demonstrate that this method is effective and efficient.