Automatic segmentation of intracellular compartments is a powerful technique, which provides quantitative data about presence, spatial distribution, structure and consequently the function of cells. With the recent development of high throughput volumetric data acquisition techniques in electron microscopy (EM), manual segmentation is becoming a major bottleneck of the process. To aid the cell research, we propose a technique for automatic segmentation of mitochondria and endolysosomes obtained from urinary bladder urothelial cells by the dual beam EM technique. We present a novel publicly available volumetric EM dataset – the first of urothelial cells, evaluate several state-of-the-art segmentation methods on the new dataset and present a novel segmentation pipeline, which is based on supervised deep learning and includes mechanisms that reduce the impact of dependencies in the input data, artefacts and annotation errors. We show that our approach outperforms the compared methods on the proposed dataset. •A novel public volumetric data-set of cellular ultra-structure in electron microscopy volumes.•A new state-of-the-art pipeline for segmentation of mitochondria and endo-lysosomes.•Contrast enhancement with transfer learning improves segmentation of unbalanced EM data.