Neural circuitry reconstruction is a central problem in neuroscience and the focus of the field of connectomics, where automatic high-throughput techniques are essential since human analysis is labor intensive and impractical, due to the size of the image volume. This dissertation presents three probabilistic graphical model-based methods for solving the problem of reconstructing the 3D model of neuronal cells in electron microscopy images, where different modeling approaches, optimization algorithms, and feature selections are utilized. In the first part, a 1D hidden Markov model-based contour tracking algorithm is proposed for tracing a single or a few neuronal processes which may have undergone arbitrary deformation, displacement, and topological changes. In this method, uncertain segments with lower likelihoods are detected, and then a few hypothetical arcs are created to perform contour refinement to enable the discovery and corresponding tracing of topology changes. Secondly, an agglomerative framework for grouping superpixels, using a hyper-graph, is proposed as a scalable framework for simultaneously tracing of a large number of cells. Lastly, a two-dimensional hidden Markov model is trained and employed for tracing a large number of cells, by modeling the problem as a pixel labeling task. This method leverages the concept of spatially adaptive states, wherein the state-space at each pixel is locally restricted to be a subset of the full state-space. This local adaptation of states, not only reduces the computational complexity significantly, but also improves the segmentation accuracy. While the first contour tracking algorithm precisely locates cell boundaries, the second hyper-graph-based framework and the third pixel-labeling-based algorithm easily scale to hundreds of cells, and hence represent three complementary techniques that together offer significant advancement on the roblem of 3D reconstruction of neuronal cells.