JournalNeural Excitability, Synapses, and Glia

An open-source tool for analysis and automatic identification of dendritic spines using machine learning

Steps in image segmentation. Once an image is loaded (A), pixels are converted to grayscale floating-point numbers ranging between 0 to 1. Highest probability background is identified using Otsu’s method of globally thresholding. The resulting binary image (B) includes only regions of neuronal tissue. The location of each feature vector, as well as the individual values of perimeter distance features, were quantified based on a geodesic distance transform of the binary image of the dendrite (C), using the dendrite backbone as a seed location. Identification of potential spine locations by local maxima along perimeter (D).

Synaptic plasticity, the cellular basis for learning and memory, is mediated by a complex biochemical network of signaling proteins. These proteins are compartmentalized in dendritic spines, the tiny, bulbous, post-synaptic structures found on neuronal dendrites. The ability to screen a high number of molecular targets for their effect on dendritic spine structural plasticity will require a high-throughput imaging system capable of stimulating and monitoring hundreds of dendritic spines in various conditions. For this purpose, we present a program capable of automatically identifying dendritic spines in live, fluorescent tissue. Our software relies on a machine learning approach to minimize any need for parameter tuning from the user. Custom thresholding and binarization functions serve to “clean” fluorescent images, and a neural network is trained using features based on the relative shape of the spine perimeter and its corresponding dendritic backbone. Our algorithm is rapid, flexible, has over 90% accuracy in spine detection, and bundled with our user-friendly, open-source, MATLAB-based software package for spine analysis.

Michael S. Smirnov, Tavita R. Garrett, Ryohei Yasuda. PLOS ONE 05 July 2018.