Detection Of Dendritic Spines In Microscope Images Based On Deep Learning And Image Segmentation

The study of brain neurons is a hot issue in the 21 st century,the arrival of artificial intelligence era is of epoch-making significance.A typical neuron consists of a cell body(soma),dendrites,and an axon.Neurons connect and transmit information through synapses with each other.Numerous spinous protrusions can be seen on the surface of dendrites,called dendritic spines,which are important sites for the formation of synapses between neurons.Medical researches show that dendritic spines have electrical properties and plasticity.These characteristics are related to the growth and development of nerve cells and the establishment and disappearance of synapses,and they play a key role in recognition,learning,memory,and neurological diseases.Therefore,the detection of dendritic spines is the basis for studying the structure of dendritic spines,including number,morphology,density,distribution and changes,and is of great significance to the life sciences.This subject comes from the National Major Research Fund 973 project.With the development of science and technology,modern fluorescence microscopy makes it possible to obtain high-resolution neuronal images,thus providing a powerful basis and tool for the study about the detection of dendritic spines.Previous methods about the detection of dendritic spines are based on the morphological characteristics of dendrites and dendritic spines in a tiny region of the neuron’s image,and require a high-resolution image.So the robustness of previous methods is not strong,and the methods can’t adapt to big data.At present,the development and application of deep learning in image processing is making rapid progress because of self-learning and adaptive capabilities.This paper proposes a new method about deep learning to achieve the segmentation of dendritic spines under big data.We use the two-dimensional fluorescence image of the rat’s brain as the source of the sample set of the neural network,artificially mark the dendritic spines in the fluorescence images,and then combine the residual neural network and atrous convolution to build and train a deep semantic segmentation network to achieve the segmentation of the dendritic spines.The initial segmentation is not enough,some dendritic spines will still stick together.We finally use the density peak clustering algorithm to achieve every single dendritic spine,and complete the task successfully.The experiment shows that this method gets a high accuracy and good robustness for the detection of dendritic spines.