Dendritic spine detection and segmentation for 4D 2-photon microscopy data using statistical models from digitally reconstructed fluorescence images

The brain with its neurons is a complex organ which is not yet fully decoded. Many diseases and human behavior are affected by the brain. In neurobiological experiments neurons are studied and imaging of neurons is a key technology. 2-Photon Microscopy (2PM) enables to image the volume of labeled, living, pyramidal neurons. Moreover, in a second channel a different marker can be used for specific structures. Synapses are not visible in 2PM. However, the size of spines has a relation to the strength of synapses. Therefore, dendritic spines are of interest. Time series imaging of living neurons is possible too. The analysis of fluorescence images is difficult, time consuming and error-prone even for experts. Furthermore, the reproducibility of manual analysis is not given. Therefore, the automatic detection, segmentation and tracking of dendritic spines in 2PM data are required.
We introduce an approach to detect, segment and track spines in time series from 2PM. We train a statistical dendrite intensity and spine probability model with 2D data from Digitally Reconstructed Fluorescence Images (DRFIs). DRFIs are synthetic images computed from geometrical shapes of dendrites and their spines, which can be reconstructed in Electron Microscopy (EM) data. This concept enables us to overcome the issue of expert labeled spines in fluorescence images. We are able to predict the spine probability for 2D slices due to the information transfer from the EM domain to the fluorescence image domain. In combination with further features a robust spine prediction is feasible. The prediction is projected back to the original space of the image. Thus, a prediction and segmentation of spines in 3D is possible.
Imaging time series of dendrite pieces is a challenging task. The handling of the sample between different imaging steps (e.g. storing the samples in an incubator) requires a registration of the different time points. After segmentation of spines in individual time points a tracking of the spine candidates over the registered time points is required because spines can move. Successful tracking of spines enables to trace intensity changes of individual spines. The tracing of intensity changes is possible for multiple image channels and opens the possibility of manifold applications.
We demonstrate the successful detection, segmentation and tracking of spines in single time points and time series in practical applications. We are able to detect spines with a presynaptic bouton in single time point images with multiple channels. Moreover, we demonstrate the successful detection and segmentation of spines in time series. For time series we demonstrate the possibility to track Endoplasmic Reticulum of spines over time. In such experiments the whole complexity of image analysis for fluorescence time series is solved.