An Image Analysis Algorithm for Dendritic Spines

Abstract

The structure of neuronal dendrites and their spines underlie the connectivity of neural networks. Dendrites, spines, and their dynamics are shaped by genetic programs as well as sensory experience. Dendritic structures and dynamics may therefore be important predictors of the function of neural networks. Based on new imaging approaches and increases in the speed of computation, it has become possible to acquire large sets of high-resolution optical micrographs of neuron structure at length scales small enough to resolve spines. This advance in data acquisition has not been accompanied by comparable advances in data analysis techniques; the analysis of dendritic and spine morphology is still accomplished largely manually. In addition to being extremely time intensive, manual analysis also introduces systematic and hard-to-characterize biases. We present a geometric approach for automatically detecting and quantifying the three-dimensional structure of dendritic spines from stacks of image data acquired using laser scanning microscopy. We present results on the measurement of dendritic spine length, volume, density, and shape classification for both static and time-lapse images of dendrites of hippocampal pyramidal neurons. For spine length and density, the automated measurements in static images are compared with manual measurements. Comparisons are also made between automated and manual spine length measurements for a time-series data set. The algorithm performs well compared to a human analyzer, especially on time-series data. Automated analysis of dendritic spine morphology will enable objective analysis of large morphological data sets. The approaches presented here are generalizable to other aspects of neuronal morphology.

Neuron tracing in perspective

“…To obtain a more compact description of the neuronal tree, a common next step is to extract the centerlines of the segmented areas (25,35,36,38,41,42,(45)(46)(47)(48)51,(53)(54)(55)(56)(57), for which various skeletonization algorithms have been proposed. Neurite centerlines may alternatively be obtained directly from the grayscale images, by applying Hessian (58)(59)(60)(61) or Jacobian (62) based analysis of critical points, or by nonmaximum suppression (37,63).…”

“…Since the size of these membranous protrusions, in particular their connections to the neuron, are close to the optical resolution limit, deconvolution is often an important preprocessing step (49,55,61,81). But even after deconvolution, spines may appear disconnected (Fig.…”

Image Informatics Strategies for Deciphering Neuronal Network Connectivity

“…Although the nomenclature and classification of automated neuron tracing algorithms is not consistent in literature, from an image informatics perspective, we discern global image processing methods, local tracing methods, and algorithms that use a combination of both. Early attempts to automate the neurite reconstruction process are based on a global intensity threshold, followed by voxel thinning or a medial axis transform to obtain the neurite skeleton 46,47 . As a result of the global threshold, these methods experience difficulties in the presence of signal inhomogeneities, and the iterative nature of the voxel thinning process is computationally intensive.…”

“…A similar method is then used to detect the centrelines of dendritic spines. After segmentation and skeletonization, most dendritic spines are usually identified as protrusions ( Figure 4A) 46,79,80 . Some spines, however, become detached in the segmentation process and should be reassigned, e.g.…”

“…Therefore, most centreline-based algorithms estimate the spine density from maximum intensity projected images, which leads to a substantial underestimation of spine densities 79,80,82 . While variations in the skeletonization algorithm have led to increased accuracy of spine detection in 3D 46,83 , these algorithms are computationally expensive. Model-based algorithms, such as voxel clustering 81 and the marching cubes algorithm 84 , are faster alternatives that identify spines based on a trained classifier.…”