基于Group-Depth U-Net的电子显微图像中神经元结构分割

Segmentation of neural structure in electron microscopy image based on Group-Depth U-Net

针对电子显微(EM)成像存在边界有损、模糊不均匀以及神经元结构本身轮廓纹理复杂难以定位的问题,提出一种深层卷积神经网络模型Group-Depth U-Net,以实现EM图像中神经元结构的自动分割。该模型采用更加深层的U-Net架构作为骨架网络,以获取更加丰富的图像特征信息;同时采用分组卷积网络结构,使模型更加高效、防止过拟合,从而提高分割的准确性与效率。公开的数据集实验表明该模型相比U-Net达到了更好的分割准确率。

Aiming at the problems of electron microscopy(EM)imaging,such as boundary damage,fuzzy inhomogeneity and difficulty of localization due to the complex contour texture of neural structure itself,a deep convolutional neural network model,Group-Depth U-Net,is proposed to realize automatic segmentation of neural structure in EMimage.In the proposed model,a deeper U-Net architecture is used as the backbone network to obtain more abundant image feature information.Meanwhile,group convolutional network structure is adopted to make the model more efficient and prevent over-fitting,thus improving the accuracy and efficiency of segmentation.The experiments on the open data set show that the proposed model achieves a higher segmentation accuracy than U-Net.