Automatic methods for dendrite and dendritic spine segmentation are important for high-throughput neuroanatomical datasets. Assessing their performance compared to human annotators and ideally the real underlying neuroanatomy is crucial to determine overall reliability of those methods. However, most datasources, for example in two-photon microscopy, do not even allow identifying dendritic spines as an ultimate ground truth because of optical resolution limits. Here, we introduce synthetic microscopy to generate neuroanatomy stacks using a combination of random walk dendritic growth and randomly spawning dendritic spines. To convert those to realistic looking microscopy images, we use circular generative adversarial networks (CycleGANs) that use non-paired data to translate one domain into another - in our case, from synthetic to microscopy-like image stacks. We then utilize DeepD3, an open toolbox for dendritic spine quantification, to compute recall and precision of dendritic spine recovery. Our preliminary results suggest that out of 1000 artificially generated dendritic spines, about 8% of the total spines were assigned to multiple ground truth spines. This raises the question, if what we see, track and quantify is actually happening in the sample analyzed.