光学学报, 2020, 40 (24): 2410001, 网络出版: 2020-11-23
基于深度学习的细胞骨架图像超分辨重建 下载: 1520次
Super-Resolution Reconstruction of Cytoskeleton Image Based on Deep Learning
图像处理 深度学习 图像超分辨重建 随机光学重构显微术 细胞骨架 image processing deep learning image super-resolution reconstruction stochastic optical reconstruction microscopy cytoskeleton
摘要
21世纪初诞生的超分辨光学成像技术在生命科学研究中发挥着巨大作用,极大地增强了人们探索微纳尺度亚细胞结构的能力,然而这些成像技术往往耗时长,成本高。如今,许多研究者致力于基于深度学习的图像超分辨重建算法的研究中。利用自主搭建的随机光学重构超分辨显微镜获得细胞微管骨架超分辨图像,然后采用双线性插值降采样法处理得到低分辨率输入图集,再分别使用传统的三次样条插值法和增强型深度超分辨率神经网络进行学习训练,实现低分辨率图像的超分辨重建。结果表明:通过深度学习所重建的各种降采样的图像效果均优于采用传统插值法得到的图像效果,尤其是二倍降采样重建图像在主观和客观评价指标上可比拟实验获得的微管骨架超分辨图像。基于增强型深度超分辨率神经网络的细胞骨架图像超分辨重建有望提供一种简捷、有效和高性价比的成像方法,可应用于对细胞骨架超微结构的快速预测。
Abstract
Super-resolution microscopy techniques invented at the beginning of the 21 st century provide unprecedented access to life science researches owing to its impressive ability of studying subcellular structures at the micrometer and nanometer scales. However, these techniques often require high cost of time and money. Recently, many researchers work on super-resolution image reconstruction algorithms based on deep learning. Herein, we obtained the super-resolution images of cell microtubule cytoskeletons by the self-built stochastic optical reconstruction microscopy (STORM), and then the bilinear interpolation down-sampling method was used to obtain the low-resolution input atlas. The traditional cubic spline interpolation method and the enhanced depth super-resolution neural network were used for learning and training to realize the super-resolution reconstruction of the low-resolution image. Results show that the effects of all kinds of down-sampling images reconstructed by deep learning are better than those obtained by traditional interpolation method; the super-resolution images of microtubule skeletons obtained by double down-sampling and experiments are comparable in subjective and objective evaluation indexes. Based on the enhanced depth super-resolution neural network, the super-resolution reconstruction of cytoskeleton images is expected to provide a simple, effective, and cost-effective imaging method, which can be applied to the rapid prediction of cytoskeleton super-microstructures.
胡芬, 林洋, 侯梦迪, 胡浩丰, 潘雷霆, 刘铁根, 许京军. 基于深度学习的细胞骨架图像超分辨重建[J]. 光学学报, 2020, 40(24): 2410001. Fen Hu, Yang Lin, Mengdi Hou, Haofeng Hu, Leiting Pan, Tiegen Liu, Jingjun Xu. Super-Resolution Reconstruction of Cytoskeleton Image Based on Deep Learning[J]. Acta Optica Sinica, 2020, 40(24): 2410001.