图1 基于植物根超分辨成像的去噪增强、自动识别及量化分析流程图 A：以拟南芥幼苗为研究对象，在活体材料中对根细胞中内质网的结构和动态进行研究，方框所示区域为伸长区（Bar=2 μm）；B：激光共聚焦显微镜采集的内质网图像；C：使用SIM对内质网进行成像；D：SIM对内质网的成像结果；E：利用增强去噪框架处理后的成像结果；F：利用Swint-ResU-Net对采集图像进行内质网结构特征分割（绿色为整体流；红色为内质网体，深蓝色为管，浅蓝色为池）；G：对采集量化的9种特征参数（整体流面积占内质网总面积比例和管面积占内质网总面积比例作为举例）做统计；H：对内质网结构参数进行量化并分析各参数间相关性。

Fig. 1 Workflow for de-noising, enhancement, automatic identification and quantitative analysis based on the super-resolution imaging in the root zone of plant A: The structure and dynamics of ER（endoplasmic reticulum）in the root cells were studied in living material using Arabidopsis seedlings, with the region shown in the box as the elongation zone（Bar = 2 μm）. B: ER images acquired by confocal microscopy. C: Imaging of ER using SIM. D: Imaging results of ER using SIM. E: Imaging results processed by utilizing the enhancement denoising framework. F: Segmentation of ER structural features on the acquired images using Swint-ResU-Net（green markers are bulk flow; red markers are fusiform body, darks blue markers are the tubules, and light blue markers are cisternaes）. G: Statistics were done on the nine feature parameters quantified by the acquisition（the ratio of the areal bulk flow to the total area of the ER and the ratio of the areal tubule to the total area of the ER as examples）. H: Quantitative and correlation analysis of the ER structural parameters

图2 Blind2Unblind去噪框架的概述 A：训练步骤。全局掩码器Ω(·)向噪声图像y添加盲点来创建掩码体块。接着，全局感知掩蔽映射器对去噪体块盲点处采样得到h（fθ(Ωy)）。同时，去噪器fθ(·)将y作为输入，输出抑噪结果fθ(y)。重新可见损失以不可见项h（fθ(Ωy)）为梯度更新媒介，实现从盲到可见的过渡。此外，正则项被用来稳定训练。B：使用训练的去噪模型进行推理。去噪网络直接从噪声图像y中生成去噪图像，无需额外操作

Fig. 2 Overview of the Blind2Unblind de-noising framework A: Training process. The global masker Ω(·) introduces blind spots into the noisy image y, creating masked patches. Subsequently, the global perc; eptual masking mapper samples h（fθ(Ωy)）at the blind spots of the denoised patches. Simultaneously, the denoiser fθ(·) takes y as input and produces the denoised output fθ(y). The re-visible loss employs the imperceptible term h（fθ(Ωy)） as a gradient update medium, facilitating the transition from blind to visible. Additionally, regularization terms are utilized to stabilize the training. B: Inference using the trained denoising model. The denoising network directly generates denoised images from the noisy image y without the need for additional operations.

图3 时间序列中内质网的结构变化 A：拍摄时间点为1时的内质网结构成像结果；B：拍摄时间点为20时的内质网结构成像结果；C：A和B的重叠结果，右侧白色方框为星号区域放大后的展示（Bar=2 μm）；D：圆圈内的三叉点和多叉点在时间序列内的动态变化；E：箭头指示生长端在时间序列内的动态变化；F：圆圈内的管在时间序列内的动态变化；G：圆圈内的池在时间序列内的动态变化；H：3、4、5 d中内质网池和整体流面积占内质网比例在时间序列下的动态变化；I：3、4、5 d中内质网整体流面积占内质网比例在时间序列下的动态变化；J：3、4、5 d中管面积占内质网面积的比例在时间序列下的动态变化（n=21个细胞）

Fig. 3 Structural changes of ER in the time series A: Imaging results of ER structure when the time point of shooting is 0. B: Imaging results of ER structure when the time point of shooting is 20. C: Overlapping results of A and B. The white box on the right side is the demonstrated zoomed the asterisked area. Bar = 2 μm. D: Dynamic change of the three-way junction and multi-way junction inside the circle inside the time series. E: Arrows indicate the dynamics of the growth tip within the time series. F: Dynamic changes of tubules within circles within the time series. G: Dynamic changes of cisternae within circles within the time series. H: Dynamic changes of ER cisternae and bulk flow area under the time series in day 3, 4, and 5. I: Dynamic changes of ER bulk flow area as a proportion of endoplasmic reticulum under the time series of in day 3, 4, and 5. J: Dynamic changes of areal tubule as a proportion of ER area in day 3, 4, and 5 under the time series（n=21 cells）

图4 伸长区细胞在第4天以及RALF胁迫条件下的量化分析 A：伸长区细胞正常生长状态下第4天时成像的识别结果；B：在RALF胁迫下，伸长区细胞第4天的成像识别结果（箭头所指示从左往右依次为：多叉点，生长点和三叉点）；C：CK与RALF胁迫下，池和整体流面积的量化；D：CK与RALF胁迫下，整体流面积的量化；E. CK与RALF胁迫下，管面积占内质网面积比例的量化；F：CK与RALF胁迫下，节点数的量化；G：CK与RALF胁迫下，内质网体个数的量化。*: P< 0.05，**: P< 0.01, ***: P< 0.001（n=21个细胞）

Fig. 4 Elongation zone cells analyzed at day 4 and under RALF stress conditions A: Identification results of elongation zone cells imaged at day 4 under normal growth conditions. B: Identification results of elongation zone cells imaged at day 4 under RALF stress（From left to right, as indicated by the arrows, the sequence is as follows: multi-way junction, growth tip and three-way junction）. C: Quantification of areal cisternae and bulk flow under CK and RALF stress. D: Quantification of areal bulk flow under CK and RALF stress. E: Ratio of areal tubule to areal ER under CK and RALF stress quantification. F: Quantification of the number of nodes under CK and RALF stress. G: Quantification of the number of fusiform body under CK and RALF stress. *: P < 0.05, **:P < 0.01, ***: P< 0.001（n=21 cells）

图5 9种特征参数在第3天、4天、5天测量数据的相关性对比图 A：各结构参数在3、4、5 d量化数据的Pearson’s相关性对比图。B：各结构参数之间的相关性图，红线是正相关性，绿线表示负相关性；相关性关系数量与结构参数圆圈颜色呈正比。C：9种参数特征的相对变化率。*: P< 0.05，**: P< 0.01, ***: P< 0.001

Fig. 5 Comparative correlation analysis of nine feature parameters on the day 3, 4, and 5 of measurement A: Pearson's correlation analysis for quantified data of structural parameters on day 3, 4, and 5, respectively. B: The inter-parameter correlations with red lines denoting positive correlations and green lines representing negative correlations; the intensity of correlation is proportional to the circle color. C: The relative rate of change for the nine feature parameters. *: P< 0.05，**: P< 0.01, ***: P< 0.001