A Deep Learning Approach For Generalized Neuron Segmentation In Two-Photon Ca²⁺ Imaging

Ca²⁺indicators and two-photon microscopy techniques have enabled the large-scale recording of neural activity in live animals,contributing to our understanding of brain function underlying intelligent behavior.However,automated,stable,and accurate neuron segmentation is crucial for processing these data.This paper bases on the understanding of the previously proposed neuron segmentation method NeuroSeg,aims at solving the challenges of neuron morphological differences and background interference in two-photon Ca²⁺imaging data and focuses on the generalized segmentation of neurons in two-photon Ca²⁺imaging data based on deep learning.We proposed a neuron segmentation technique called NeuroSeg-Ⅱ.Its network structure was optimized by three methods based on Mask region-based convolutional neural network（Mask R-CNN）,including image pre-processing,data enhancement,and network architecture optimization,which improved the performance of neuron segmentation in two-photon Ca²⁺imaging.Our method achieved well generalized neuron segmentation in two-photon Ca²⁺imaging data with different Ca²⁺indicators,depths,brain regions,imaging scales,and activity levels.In the aspect of image preprocessing,we used mean projection,maximum projection,and correlation map to extract the image information and temporal information of active neurons in 2D images.The method can supplement the neural activity information lost in the original 2D images.We also used a Generative adversarial network（GAN）to increase the image resolution of the dataset,and to enhance the network’s feature extraction capabilities.In terms of network optimization,we used multi-scale image feature extraction to improve the CNN structure’s ability to extract neuron features.Our method was applied to two-photon Ca²⁺imaging data recorded under different conditions,it had a certain degree of generality,accurately segments neurons with different morphologies,and could segment neurons with small target features in large field-of-view quickly and transfer new neuron images for learning.We also incorporated an attention mechanism into the network model,allowing the network to deepen the weight of the neuron region and focus on learning the features of that region,improving the performance of neuron segmentation.Our method achieves the best results on the Neurofinder dataset compared to other state-of-the-art neuron segmentation methods.The results show that NeuroSeg-Ⅱ can accurately and stably segment neurons from two-photon Ca²⁺imaging data,providing a new multiscale brain imaging data processing tool for neuroscience research.