Neuronal Morphology Classification based on Improved Residual Network

To address the problem that traditional convolutional neural networks have difficulty in fully extracting important neuronal morphological features when classifying neuron types, resulting in low accuracy of neuron classification. We propose a neuron morphology classification method based on the optimization of the ResNet18 network model. First, ResNet18 is used as the base network and initialized using the weights trained by ImageNet. Second, at the back-end of the base network, we use a feature reconstruction module to suppress edge feature loss and the drawbacks associated with the padding strategy. Finally, experiments are conducted on the NeuroMorpho-rat dataset to verify the effectiveness of the proposed method. The experimental results showed that the accuracy of 2 classification on Img_raw, Img_resample and Img_XYalign datasets reached 94.32%, 85.37% and 86.74%, respectively, and the accuracy of 12 classification reached 94.15%, 85.47% and 85.81%, respectively, which effectively improved the neuron compared to the original residual network the accuracy of morphological classification was effectively improved compared to the original residual network.