Morphological Classification of Neurons Based Deep Residual Multiscale Convolutional Neural Network

Abstract

The study of neuron morphological classification has important application value to improve the accuracy and efficiency of three-dimensional reconstruction of neurons. However, due to the complex structure of neurons and the existence of global and local self-similarity in morphological distribution, it brings great difficulties to the classification of neuron morphology. Therefore, a new neuronal morphological classification model based on deep residual multiscale convolutional neural network is proposed. Firstly, the overall architecture of the model is based on the fast connection idea of ResNet, which can effectively prevent network model degradation. Secondly, by using the residual connection module, the input information is directly transferred to the output layer through a shortcut, so as to simplify the goal and difficulty of feature learning. Finally, the multi-scale convolution module is combined for feature extraction, and the dilated convolution with different dilation rates is adopted to increase the receiving field to expand the diversity of features, so as to improve the classification accuracy. To verify the effectiveness of the model, experiments are carried out on the neuron morphology classification dataset. The experimental results show that the accuracy, precision, sensitivity and specificity of our method reach 90.11%, 89.63%, 90.77% and 93.27%, respectively. Compared with other classification models (VGG, ResNet, RNN), the proposed model has better classification effect.