The observation of small biological particles in 3D fluorescence microscopy faces challenges due to resolution anisotropy and partial fluorescent labeling. A solution to address these issues consists in combining multiple acquisitions of individual particles. The detection of these single particles within each acquired volume is of utmost importance. A significant drawback of deep learning methods in this context is the substantial requirement for manually annotated training data. In response, we propose a particle detection pipeline for 3D data based on few-shot learning. It generalizes a 2D method based on the principle of positive-unlabeled learning. Our experiments with synthetic and real fluorescence-microscopy data demonstrate that our approach achieves high detection rates even with a limited amount of training data.