Recent advances in cellular and subcellular microscopy demonstrated its potential towards unravelling the mechanisms of various diseases at the molecular level. The biggest challenge in both human- and computer-based visual analysis of micrographs is the variety of nanostructures and mitochondrial morphologies. The state-of-the-art is, however, dominated by supervised manual data annotation and early attempts to automate the segmentation process were based on supervised machine learning techniques which require large datasets for training. Given a minimal number of training sequences or none at all, unsupervised machine learning formulations, such as spectral dimensionality reduction, are known to be superior in detecting salient image structures. This thesis presents three major contributions developed around the spectral clustering framework which is proven to capture perceptual organization features. Firstly, we approach the problem of mitochondria localization. We propose a novel grouping method for the extracted line segments which describes the normal mitochondrial morphology. Experimental findings show that the clusters obtained successfully model the inner mitochondrial membrane folding and therefore can be used as markers for the subsequent segmentation approaches. Secondly, we developed an unsupervised mitochondria segmentation framework. This method follows the evolutional ability of human vision to extrapolate salient membrane structures in a micrograph. Furthermore, we designed robust non-parametric similarity models according to Gestaltic laws of visual segregation. Experiments demonstrate that such models automatically adapt to the statistical structure of the biological domain and return optimal performance in pixel classification tasks under the wide variety of distributional assumptions. The last major contribution addresses the computational complexity of spectral clustering. Here, we introduced a new anticorrelation-based spectral clustering formulation with the objective to improve both: speed and quality of segmentation. The experimental findings showed the applicability of our dimensionality reduction algorithm to very large scale problems as well as asymmetric, dense and non-Euclidean datasets.