The automatic detection of dendritic spines is still a challenging and yet not fully resolved problem with regard to multi-photon microscopy. The emergence of convolutional neural networks (CNN) like U-Nets enabled the development of deep learning based segmentation pipelines for biomedical images in general and for dendritic spines in particular. While these pipelines are most suitable for in-vitro confocal image data, they provide lower prediction accuracy when applied to volumetric in-vivo two-photon images that have a lower signal-to-noise ratio and larger movement artifacts. Thus, researchers of this field still tend to analyze dendritic spines manually, which is time-consuming and contains the risk of human bias. We therefore developed a pipeline for multi-class semantic image segmentation based on a fully convolutional neural network, that specifically targets 3D in-vivo multi-photon image data. By choosing U-Net as the underlying network architecture, only a few labeled training images are required. Also, the U-Net is applied in a 2D manner to further reduce the computation time. A post-hoc 3D connectivity analysis merges the classified spine pixels and reconstructs the 3D morphology. Our pipeline is capable to segment spines from its associated dendrite with satisfying accuracy and enables the further analysis of, e.g., spine morphology and spine density.