基于机器学习的低信噪比细胞图像分割  

Cell Segmentation in Low Signal-to-Noise Ratio Microscopy Images Based on Machine Learning

作　　者：阎彧萱 卢金旺 宋奕泓 旷小宇 田原 傅静雁 YAN Yu-Xuan;LU Jin-Wang;SONG Yi-Hong;KUANG Xiao-Yu;TIAN Yuan;FU Jing-Yan(State Key Laboratory of Agrobiotechnology,College of Biological Sciences,China Agricultural University,Beijing 100193,China)

机构地区：[1]中国农业大学生物学院农业生物技术国家重点实验室,北京100193

出　　处：《农业生物技术学报》2021年第1期198-206,共9页Journal of Agricultural Biotechnology

基　　金：国家自然科学基金项目(31722032;31771494;31970662);国家级大学生创新创业训练计划项目(201910019040;202010019014);中国农业大学生命科学实验教学示范中心探索项目(20190102)。

摘　　要：荧光显微图像在生命科学研究中具有重要作用。利用计算机和机器学习可以有效处理大量图像数据,从而得到具有统计意义的结论。本研究针对低信噪比的细胞荧光图像,提出一种基于U-Net的细胞分割模型,并构建了单通道低信噪比细胞图像数据集用于模型训练和测试。本研究提出的模型使用不同尺度的卷积核提取特征,利用残差模块加深网络深度,并使用权重损失机制使机器学习过程更加关注于细胞边缘。相较于其他方法,在低信噪比荧光显微图像的细胞分割上,可以有效解决细胞与背景对比度低、胞内信号亮度分布不均的问题,其像素准确率、IoU(intersection-over-union)可分别达到87.6%和72.0%。本研究为细胞形态学研究、借助图像的高通量细胞筛选提供技术支持。Fluorescence microscopy image plays an important role in the research of life sciences. A large number of image data can be processed effectively using computer and machine learning, so that we could obtain the conclusion with statistical significance. In this study, a cell segmentation model was proposed based on U-Net for low signal-to-noise ratio(SNR) fluorescence images, and a cell image dataset for training and validation was constructed. In this method, convolution kernels were adopted to extract features, residual modules to deepen network, and weighted loss function to make machine learning process pay more attention to cell edges. Compared to several other methods, this algorithm shows better performance in cell segmentation for low signal-to-noise ratio fluorescence images with 87.6% pixel accuracy and intersectionover-union(IOU) 72.0%. This study provides technical support for cell morphology research and image-based high-throughput cell screening.

关 键 词：细胞分割 机器学习 低信噪比 U-Net 

分 类 号：S-03[农业科学]