A well-known problem in order to analyze electron microscopy images is that the corresponding segment extraction typically is a tedious and time-consuming process. This is due to the currently used image processing techniques which are hampered by large time or memory consumption or simplistic underlying models. Advanced techniques that obtain high segmentation quality exist, but suffer from practical limitations such as difficult algorithm interpretability or parameter dependencies. We therefore propose a trainable segmentation technique that is based on convolutional neural networks and extracts mitochondrial membranes from EM images. The training phase depends on a limited number of manually annotated samples. Once the model is trained, segmentation becomes completely parameter-independent.