基于U-Net卷积神经网络的目标分子人机交互观测方法

An Observation Method of Human-Computer Interaction for Identification of Target Molecules Based on U-Net Convolutional Neural Network

对生命活动中发挥重要作用的分子进行观测是发现生命活动内在机理的重要手段。现有的生物医学影像处理方法大多集中在对特定物质的检测和识别,难以适应不断变化的科研需要。为此,本研究提出了一种基于U-Net卷积神经网络的人机交互方法以识别生物医学影像中所有同类分子,如:细胞、蛋白质等。首先利用U-Net卷积网络将待观察的分子影像转换为深度特征图,然后使用目标分子的特征在整个特征图上进行匹配,以检测出所有感兴趣的同类分子。之后利用通道和空间可靠的判别式相关滤波器构建多目标跟踪器以实现对目标分子的持续追踪。结果表明,该方法可以通过简单的人机交互快速检测出感兴趣的同类分子,获取目标分子的数量、分布以及相互作用等重要信息,Attention-basedU-Net和U-Net在从细胞核、人类蛋白质图谱、细菌和血红细胞数据集中随机抽取的200张静态测试影像上的各项指标表现稳定,平均精度的平均值分别为0.9125和0.8981,同时对小鼠干细胞动态影像中的目标跟踪准确且保持稳定,证明了方法的有效性,可满足生命科学研究中对微观生命过程观测的需要。

The observation of molecules that play an important role in life activities is an important way to discover intrinsic mechanisms of the life activities.Most of existing biomedical image processing methods focus on the detection and identification of specific substances,however,it is difficult to adapt to changing demands of scientific research.To this end,this paper proposed a human-computer interaction method based on U-Net convolutional neural network to identify all the same molecules in biomedical images,such as nucleic cell,proteins,etc.First,the U-Net convolutional network was used to convert molecular images to deep feature maps,and then the features of the target molecules were used to match on the entire feature map to detect all the same molecules of interest.Then,the CSR-DCF(discriminative correlation filter with channel and spatial reliability)algorithm was used to build a multi-target tracker to achieve continuous tracking of the target molecules.Experimental results showed that the proposed method was able to quickly detect similar molecules of interest through simple human-computer interaction,and obtain important information on the number,distribution and interactions of target molecules.Attention-based U-Net and U-Net performed consistently on 200 static test images randomly selected from Nucleus,Human Protein Atlas,Bacteria and Blood Red Cell datasets,with average precision mean values of 0.9125 and 0.8981,respectively.At the same time,the tracking of targets in the dynamic images of mouse stem cells was accurate and stable,proving the effectiveness of the method to meet the needs of microscopic life process observation in life science research.