07

2024

-

08

ELight · Cover | Next Generation Fluorescent Physical Fidelity Deconvolution Technology

Author:


Recently, Peking University Prof. Xi PengThe team came up with a Deconvolution algorithm with regularization constraints for multiresolution analysis of fluorescence imagingThe fidelity problem of the deconvolution algorithm is solved and its noise robustness is further improved. Based on the multi-resolution analysis deconvolution algorithm, high-quality images can be obtained with a lower photon budget under the premise of fidelity, which improves the speed of fluorescence imaging and the imaging time history, and is expected to help the discovery of new biological phenomena.
 

The article is based on"Multi-resolution analysis enables fidelity-ensured deconvolution for fluorescence microscopyThe title is published ineLight. The research work has been strongly supported by the National Natural Science Foundation of China's regional joint key projects, general projects and other projects. Hou Yiwei, a doctoral student at Peking University, is the independent first author of the paper. Professor Xi Peng of Peking University and Dr. Meiqi Li are the co-corresponding authors of the paper. The main contributors to the paper also include Fu Yunzhe, a doctoral student at Peking University, and Wang Wenyi and Gai Xichuan, engineers at Beijing Ai Rui Precision Instrument Company. The work was completed by the Department of Biomedical Engineering, Future Institute of Technology, Peking University as the first unit.

Fluorescence microscopy is an important tool in live cell imaging. The proposal of super-resolution microscope further breaks the diffraction limit of fluorescence microscope, making it possible to observe more fine cell structure. However, random noise and optical blur in the imaging process affect the quality of fluorescence imaging. In live-cell imaging, high photon dose can have an impact on cell state and can lead to photobleaching. In ultrafast imaging, insufficient exposure time also restricts the acquisition of high signal-to-noise ratio images. In general, the key parameters of imaging speed, duration and resolution in fluorescence imaging are greatly restricted by each other.
 
Deconvolution algorithms (Deconvolution) have found great application in the field of fluorescence imaging because they have the potential to improve imaging duration and spatial and temporal resolution by recovering fluorescence signals from noise with a low photon budget. Although algorithmic compensation has shown great potential in enhancing these key aspects of fluorescence imaging, its fidelity, especially in terms of improving resolution, is still questioned. The development of a high-fidelity deconvolution algorithm that can work at a lower signal-to-noise ratio can further improve the imaging time and speed, and provide more possibilities for the discovery of new biological phenomena.
 
In order to solve the above problems, this paper reshapes the noise control model and resolution enhancement mechanism of the fluorescence imaging deconvolution algorithm, and proposes the multi-resolution analysis (MRA) deconvolution algorithm. This method effectively improves the minimum signal-to-noise ratio limit that the deconvolution algorithm can recover, and solves the defect that the traditional deconvolution algorithm is prone to produce artifacts when the resolution is improved.
 
The author first analyzes the defects of the traditional total variation (Total variation, TV) and Hessian (Hessian) regularization based on spatial variation, and thinks that although the continuous features extracted are partially consistent with the characteristics of fluorescence imaging, they will confuse the fluorescence signal with the edge of the background, and it is difficult to extract the signal from the extremely high noise. Based on the specific labeling of fluorescence imaging and the characteristics of biological structures, the authors propose for the first time two basic properties to distinguish fluorescence signal from noise:(1) high contrast across edges, and (2) high continuity along edges. In mathematical implementation, the authors use multi-scale Framelet and Curvelet transformation to extract these two key features in the fluorescence signal, called the MRA regularization method. Through the high and low signal-to-noise ratio of various imaging conditions to image restoration experiments, the authors verified that the MRA regularization method can further improve the signal-to-noise ratio by 2-10dB compared to the regularization method based on spatial variation.
 

In terms of resolution enhancement mechanisms, the authors first demonstrate through simulation experiments that Richardson-Lucy iteration can spontaneously produce artifacts due to its statistical estimation properties. Although the iterative inverse filtering based on the physical model can restore the blurred image perfectly without noise, its high-frequency detail recovery requires a large number of iterations, is easily overwhelmed by noise and is poorly compatible with the noise control regularization based on spatial variation. The MRA regularization method proposed by the author provides a more accurate fluorescence noise model, which can effectively control the noise in the inverse filtering iteration and retain the high-frequency details of its inference. The author further proposes a strategy to accelerate FISTA iteration, which can complete thousands of constrained inverse filtering iterations in a short calculation time to recover high-frequency details.

Figure 1:MRA deconvolution for better noise control and fidelity resolution improvement. a. Verification of MRA noise control using SIM high and low SNR. B. Comparison of MRA and Hessian regularization in continuous frame image processing. c. Spacious image of U2OS cell mitochondria, Huygens, sparse and MRA deconvolution results and SR-SIM results as truth reference. d. Linear nonlinear SIM image pairs obtained from the open source BioSR dataset. The non-linear SIM is used as the comparison truth value of the deconvolution result of the linear SIM by Huygens, Sparse and MRA. e. Intensity distribution along the line indicated by the two white arrows in d.

Judging the authenticity of the resolution improvement of the deconvolution algorithm requires strict verification using a physical super-resolution microscope. Using wide-field-SIM, linear-nonlinear SIM,Confocal-SoRa, and Confocal-STED multiple low-high resolution data pairs, the authors verified that MRA deconvolution has significantly superior fidelity compared to the deconvolution algorithm with Richardson-Lucy as the resolution improvement mechanism, and the resolved high-frequency detail structure is almost consistent with the corresponding high-resolution reference image and even super-resolution microscope.
 
The authors further propose a layer cut multi-resolution analysis deconvolution algorithm (SecMRA), which introduces a new design of biased sparse regularization, which can effectively improve the performance in strong background image processing. The wide-field image after SecMRA deconvolution can achieve a layer-cutting effect similar to that of confocal microscopy, which can further reduce the phototoxicity of live cell imaging while providing high-quality images. The effectiveness of the proposed method is verified in various imaging modalities such as widefield, Lightsheet, Confocal, SoRa, SIM, STED and Expansion microscopy.
 
The authors show an example of using the MRA deconvolution algorithm to support low-light, low-toxicity fluorescence imaging to study cellular activity. Using MRA deconvolution, the authors achieved long-term super-resolution imaging of microtubules and lysosomes, and observed the close relationship between lysosomal movement and microtubule structure. The authors also used wide-field microscopy to achieve more than half an hour around the nucleus of the mitochondria and endoplasmic reticulum co-labeling imaging, observed the rapid extension of mitochondria after contraction and finally bound to the endoplasmic reticulum phenomenon.
 
In order to generalize this technique, the authors also propose a scheme to solve the hyperparameter adjustment in the regularized deconvolution algorithm. By using the curve wave coefficient sparsity to estimate the noise intensity, the automatic determination of parameters is realized. At the same time, in order to facilitate peer researchers, the author open source all the source code and fluorescence image raw data, and write easy to interact with the GUI software, in order to facilitate users and developers to use. The authors hope that MRA can become a widely used next-generation high-fidelity deconvolution tool for biologists and microscopists.
 

At the same time, for the booming model computing super-resolution and deep learning super-scoring fields, the author hopes that the physical truth value determination proposed in this work can become a widely used authenticity judgment tool, and through open source related low-high resolution physical imaging datasets, it can provide objective evaluation criteria for further improving the performance of different algorithms.

Figure 2:SecMRA deconvolution powers long-term live-cell imaging.A- B. Two-color imaging of microtubule lysosomes. SecMRA deconvolution effectively removes noise, out-of-focus background in the image and improves resolution. of CG. SecMRA deconvolution for endoplasmic reticulum and mitochondrial interaction imaging.

Chongqing Chongfan Technology Co., Ltd. Optical platform film thickness meter OCT optical coherence tomography LabVIEW control