Electron Microscopy Image Segmentation with Graph Cuts Utilizing Estimated Symmetric Three-Dimensional Shape Prior

Understanding neural connectivity and structures in the brain requires detailed three-dimensional (3D) anatomical models, and such an understanding is essential to the study of the nervous system. However, the reconstruction of 3D models from a large set of dense nanoscale microscopy images is very challenging, due to the imperfections in staining and noise in the imaging process. To overcome this challenge, we present a 3D segmentation approach that allows segmenting densely packed neuronal structures. The proposed algorithm consists of two main parts. First, different from other methods which derive the shape prior in an offline phase, the shape prior of the objects is estimated directly by extracting medial surfaces from the data set. Second, the 3D image segmentation problem is posed as Maximum A Posteriori (MAP) estimation of Markov Random Field (MRF). First, the MAPMRF formulation minimizes the Gibbs energy function, and then we use graph cuts to obtain the optimal solution to the energy function. The energy function consists of the estimated shape prior, the flux of the image gradients, and the gray-scale intensity. Experiments were conducted on synthetic data and nanoscale image sequences from the Serial Block Face Scanning Electron Microscopy (SBFSEM). The results show that the proposed approach provides a promising solution to EM reconstruction. We expect the reconstructed geometries to help us better analyze and understand the structure of various kinds of neurons.