Classification Research Of The Multiple Synapse Morphology Based On Deep Learning

The synapse is a structure formed by the contact of two neurons,and a large number of the neurons form the neural circuits through the synapses.To some extent,the number of synapses can reflect the complexity of the neural circuit connections,which is also an aspect of the biological brain learning representation.The study of the synaptic plasticity by observing the changes in the number and morphology of the synapses under the large-volume electron microscopy has been a research hotspot for the neuroscientists in recent years.The formation of new synaptic connection or structural remodeling of the synapse is thought to be associated with underlying learning and memory.Subsequent biological studies have shown that behavioral learning or environmental stimulation can lead to changes of the synaptic morphology in vertebrate brains,resulting in the formation of the multiple synapses.By studying the morphological types and distribution of the multiple synapses,we can explore the relationship between multiple synapses and learning behaviors in the process of memory formation,storage and consolidation.According to the domestic and foreign literatures about the multiple synapses,the researches on the multiple synapses only focus on its formation and changes related to certain learning,stimulation and pathology.These studies have never involved the morphology classification of the multiple synapses.In view of the fact that deep neural network has surpassed the human level in many complex problems of the medical image processing,it can automatically extract the features of various biological structures from a large number of the medical images.At the same time,the development of the electron microscopy technology makes it possible to observe more detailed biological structures.Based on the deep learning,we carry out automatic classification research on the morphological types of the multiple synapse under electron microscopy images,and process the image data of the multiple synapses through the deep learning network,so this can provide a basis for the automatic classification method.In the large number of the neural electron microscopy images,how to efficiently and accurately classify the morphological types of the multiple synapses in order to greatly reduce the human labor is crucial for the study of large-volume multiple synapses.Starting from the composition of the multiple synapses,and based on the deep learning,this paper proposes and combines the Mask R-CNN network and Fusion Net network in the convolutional neural network to detect and segment the the micro-structure of the multiple synapses.Then,based on the segmentation results,a method based on morphology fitting of the synaptic cleft and vesicle band segmentation results is proposed to classify the morphology of the multiple synapses.The main steps can be summarized into three parts: the synaptic cleft detection and segmentation,the vesicle band segmentation and the multiple synapses classification based on the synaptic cleft and the vesicle band structure.Finally,the experiments on four groups of the electron microscope image data show that the average accuracy of the proposed method is about 97%,the experimental result on public dataset also shows that the accuracy of the proposed method can reach 96.5%.In addition,compared with the advanced deep learning network for the direct classification of the multiple synapses,this method has a higher classification accuracy.The classification results provide a basis for the quantitative statistics of the subsequent studies,and this automatic classification method can reduce the time of manual statistics,so that the researchers can pay more attention to the analysis of statistical results.This work aims to solve the classification problem of the multiple synapses under the large-scale data and provides more real and effective realistic basis for the relevant scientific conclusions.