基于双编码路径融合和双向ConvLSTM的神经元图像分割  

Neuron Image Segmentation Based on Dual Coding Path Fusion and Bidirectional ConvLSTM

作　　者：钱有为 何富运[1,2] 韦燕 冯慧玲[1] 胡聪 QIAN Youwei;HE Fuyun;WEI Yan;FENG Huiling;HU Cong(School of Electronic and Information Engineering,Guangxi Normal University,Guilin Guangxi 541004,China;Guangxi Key Laboratory of Automatic Detecting Technology and Instruments(Guilin University of Electronic Technology),Guilin Guangxi 541004,China)

机构地区：[1]广西师范大学电子与信息工程学院,广西桂林541004 [2]广西自动检测技术与仪器重点实验室(桂林电子科技大学),广西桂林541004

出　　处：《广西师范大学学报（自然科学版）》2023年第3期67-79,共13页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金(62062014);广西自然科学基金(2018GXNSFAA050024);广西师范大学重点科学研究计划(2018ZD007)。

摘　　要：目标分割是神经元图像分析中必不可少的步骤之一,分割的准确性会直接影响到神经元图像分析和重建的质量。在面对结构边界模糊、存在弱噪声或弱纤维信号的神经元图像时,已有的分割方法依然存在误差较大、识别信号不准等问题。为了解决这些问题,基于神经元的特征,本文提出一种基于双编码路径融合和双向ConvLSTM的深度学习网络(DFC-Net)用于神经元图像分割。首先,网络在编码器阶段采用双编码路径提取特征,其中第一路编码路径采用基于空洞卷积的密集连接网络作为固定特征提取器,第二路编码器采用深度残差网络作为特征提取网络;接着,使用密集连接ASPP网络作为桥梁连接编码器和解码器;最后,在跳跃连接中使用双向ConvLSTM结合编码器和解码器,在解码器阶段引入融合网络以融合2个编码器提取的特征,从而增强空间信息的传播。多组对比实验结果显示,本文提出的网络有效地提高了电子显微镜神经元图像的分割精度,在ISBI-2012和SNEMI3D数据集上的Sen、Dice分别达到0.952 7、0.958 9和0.941 6、0.912 7,平均准确率相比于其他U-Net变体网络提高2.93%。Target segmentation is one of the essential steps in neuronal image analysis.The accuracy of segmentation directly affects the quality of neuronal image analysis and reconstruction.In the face of fuzzy structure boundary,weak noise or weak fiber signal,the existing segmentation methods still have some problems,such as large error and inaccurate recognition signal.To solve these problems,based on the characteristics of neurons,a deep learning network based on dual coding path fusion and bidirectional ConvLSTM is proposed for neuronal image segmentation.Firstly,the network adopts dual coding paths to extract features in the encoder stage,in which the first encoding path adopts dense connection network based on dilated convolution as the fixed feature extractor,and the second encoder adopts deep residual network as the feature extraction network.Then,the densely connected ASPP network is used as a bridge to connect the encoder and decoder.Finally,bidirectional ConvLSTM is used to combine encoder and decoder in skip connection,and a fusion network is introduced in the decoder stage to fuse the features extracted by the two encoders,so as to enhance the spatial information propagation.The results of multiple comparative experiments show that the proposed network can effectively improve the segmentation accuracy of electron microscope neuron images.Sen and Dice on ISBI-2012 and SNEMI3D datasets reaches 0.9527,0.9589 and 0.9416,0.9127,respectively,and the average accuracy is higher than that of U-net and other models.

关 键 词：图像分割 神经元 双编码路径 D-ASPP 双向ConvLSTM 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]