A method and device for constructing a structured light illumination microscope image reconstruction model
The structured light illumination microscope image reconstruction model built by Bayesian neural network solves the problems of resolution limitations and reconstruction reliability of traditional microscopes, and achieves clear super-resolution reconstruction and error assessment.
Patent Information
- Application Number
- CN202411458339.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-18
- Publication Date
- 2025-12-19
- Estimated Expiration
- 2044-10-18
AI Technical Summary
The resolution limit of traditional optical microscopes restricts the observation of microscopic life processes, and the reconstructed images from structured light illumination microscopes have low reliability and cannot quantify uncertainty.
A structured light illumination microscope image reconstruction model was constructed using a Bayesian neural network. By collecting paired datasets and dividing them into training and test sets, the neural network was trained to output the reconstruction mean and standard deviation, thus quantifying the uncertainty of the image.
It achieves clear and reliable super-resolution reconstruction under low signal-to-noise ratio conditions, quantifies the uncertainty of the reconstructed image, improves the reliability of the reconstruction, and avoids artifacts.
Smart Images
Figure CN119379908B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of image processing, and more particularly relates to a method and device for constructing a structured light illumination microscope image reconstruction model. BACKGROUND
[0002] Due to the diffraction effect of light and the limited aperture of the optical system, the traditional optical microscope has a resolution limit, the value of which is about λ / 2NA, wherein λ is the wavelength of the emitted light, and NA represents the numerical aperture of the objective lens. Generally, the resolution limit of the traditional optical microscope is about 250 nanometers, which severely limits the observation of biologists on the life processes of smaller scales.
[0003] In order to break this shackles, a series of super-resolution microscopic imaging methods have emerged. These methods obtain sub-diffraction information of the sample beyond the resolution limit through special light path design and other means, and realize super-resolution imaging. The emergence of super-resolution microscopic imaging methods has brought a series of major breakthroughs in the field of life sciences, and the inventors have therefore won the Nobel Prize in Chemistry in 2014. Among the many super-resolution imaging methods, structured light illumination microscopy is a technology that uses cosine structured light to modulate the sample optically to achieve a two-fold improvement in imaging resolution, has the advantages of low excitation power, high photon efficiency, and is very suitable for live cell imaging. However, as a computational imaging method, the imaging quality of structured light illumination microscopy depends heavily on the reconstruction algorithm in the backend. The reconstruction algorithm is responsible for extracting the modulated super-resolution information from multiple original images and reconstructing a super-resolution image that is isotropic.
[0004] In recent years, deep learning technology has been widely used in the reconstruction of structured light illumination microscopy images, but due to the black box nature of deep learning, the reconstruction credibility in practice cannot be guaranteed. At the same time, existing image reconstruction methods cannot quantify the uncertainty of the reconstructed image, and cannot provide additional reconstruction confidence information for users. SUMMARY
[0005] In view of the defects of the related art, the purpose of the present application is to provide a method and device for constructing a structured light illumination microscope image reconstruction model, which aims to solve the problems of low credibility of the reconstructed image and inability to quantify the uncertainty of the reconstructed image.
[0006] To achieve the above purpose, in a first aspect, the present application provides a method for constructing a structured light illumination microscope image reconstruction model, comprising:
[0007] S1, collecting a pair of data sets D={I raw ,I GT} of cells in a structured light illumination microscope, and dividing them into a training set D1 and a test set D2; the data set D includes original images I rawand ground truth image I GT ;
[0008] S2, construct a deep neural network model of the structured light illumination microscope reconstruction, and deploy it as a Bayesian neural network; the deep neural network model outputs the reconstruction mean μ and the reconstruction standard deviation σ of the reconstruction super-resolution image distribution;
[0009] S3, infer the posterior distribution of the parameters of the Bayesian neural network, train the parameters of the Bayesian neural network in S2 by using the training data D1 and the preset loss function, and obtain a plurality of different parameter sets;
[0010] S4, input the test data D2 into the trained Bayesian neural network for model inference, obtain a plurality of reconstruction mean μ and reconstruction standard deviation σ of the reconstruction super-resolution image distribution, calculate the mean of the plurality of reconstruction mean μ as the reconstructed super-resolution image, calculate the mean of the plurality of reconstruction standard deviation σ as the reconstruction data uncertainty, and calculate the standard deviation of the plurality of reconstruction mean μ as the reconstruction model uncertainty.
[0011] Optionally, S1 specifically comprises:
[0012] S11, under the same field of view, change the direction and phase of the structured light, collect the data of the cell sample under the structured light illumination microscope, and obtain a group of original images I raw ;
[0013] S12, adjust the excitation power or exposure time of the structured light, obtain a plurality of groups of original images I raw with different signal-to-noise ratios in the same field of view, and obtain the original image data set in the same field of view;
[0014] S13, select a group of original images with the highest signal-to-noise ratio from the original image data set in the same field of view, and reconstruct to obtain the ground truth image I GT ;
[0015] S14, adjust the field of view of the structured light illumination microscope, repeat S11-S13, collect a plurality of groups of paired data, and form the data set D={I raw , I GT}.
[0016] Optionally, the preprocessing comprises background removal, normalization and data enhancement.
[0017] Optionally, the construction of the deep neural network model is based on a residual convolutional neural network architecture.
[0018] Optionally, S3 specifically comprises:
[0019] S31, inferring the posterior distribution of the parameters of the Bayesian neural network, training the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2 by using the training data D1 to obtain a plurality of different parameter sets;
[0020] In addition, other posterior inference algorithms can also be used, such as variational inference, MC-dropout or Deepensemble.
[0021] S32, freezing the trained network parameters in S31, training the standard deviation quantization layer parameters of the Bayesian neural network in the Bayesian model deployment module using the training data D1 to obtain a trained Bayesian neural network with a plurality of different parameter sets.
[0022] Optionally, the feature mapping layer includes a convolutional layer for mapping the input original image to a high-dimensional feature space.
[0023] Optionally, S4 specifically includes:
[0024] S41, inputting the test data D2 into the trained Bayesian neural network for a model inference to obtain K reconstruction means μ and reconstruction standard deviations σ;
[0025] S42, calculating the mean of the K reconstruction means μ as the reconstruction result of the reconstructed super-resolution image:
[0026]
[0027] Calculate the mean of the plurality of reconstruction standard deviations σ as the data uncertainty of the reconstruction:
[0028]
[0029] Calculate the standard deviation of the plurality of reconstruction means μ as the model uncertainty of the reconstruction:
[0030]
[0031] Secondly, the application also provides a device for constructing a structured light illumination microscope image reconstruction model, comprising:
[0032] A data acquisition module is configured to acquire a pair of data sets D = {I raw ,I GT} of cells under a structured light illumination microscope and divide the data sets D into a training set D1 and a test set D2; the data sets D include original images I raw with different signal-to-noise ratios and true value images I GT .
[0033] a model construction module, configured to construct a deep neural network model of a structured light illumination microscope reconstruction and deploy the deep neural network model as a Bayesian neural network; the deep neural network model outputs a reconstruction mean value mu and a reconstruction standard deviation sigma of a reconstructed super-resolution image distribution;
[0034] a Bayesian model training module, configured to infer a posterior distribution of parameters of the Bayesian neural network, train the parameters of the Bayesian neural network in the model construction module by using training data D1 and a preset loss function, and obtain a plurality of different parameter sets;
[0035] an image reconstruction model testing module, configured to input test data D2 into the trained Bayesian neural network for model inference, obtain a plurality of reconstruction mean values mu and a plurality of reconstruction standard deviations sigma of a reconstructed super-resolution image distribution, calculate a mean value of the plurality of reconstruction mean values mu as a reconstructed super-resolution image, calculate a mean value of the plurality of reconstruction standard deviations sigma as data uncertainty of reconstruction, and calculate a standard deviation of the plurality of reconstruction mean values mu as model uncertainty of reconstruction.
[0036] The above technical scheme conceived by the present application can achieve the following beneficial effects compared with the prior art:
[0037] 1. A structured light illumination microscope image reconstruction model construction method, the neural network reconstruction model in the present application is deployed as a Bayesian neural network and adopts a block training method to respectively train a feature mapping layer outputting a reconstruction mean value, a deep feature extraction layer, a mean value reconstruction layer, and a standard deviation quantization layer outputting a reconstruction standard deviation. Due to the unique modeling method, training method, and posterior distribution sampling average, the present application has excellent reconstruction performance, and can resolve clear, reliable, and artifact-free super-resolution structures even under low signal-to-noise ratio imaging conditions. Model uncertainty can reflect the generalization performance of the trained model, and data uncertainty reflects the fuzziness of input to output; by outputting model uncertainty and data uncertainty, the uncertainty of the reconstructed super-resolution image can be quantified.
[0038] 2. A structured light illumination microscope image reconstruction model construction method, the reconstructed data uncertainty and the reconstructed model uncertainty obtained by the present application can help users effectively identify unreliable results and improve the reliability of the structured light illumination microscope. The reconstruction uncertainty result can be used to evaluate the reconstruction error under the condition of no reference image.
[0039] 3. A structured light illumination microscope image reconstruction model construction method, in view of the fact that calculating an imaging model may easily lead to untrustworthy error structures or reconstruction artifacts in many cases, the present application identifies unreliable model results by quantifying the uncertainty in the image reconstruction process, and ultimately enables deep learning-based computational imaging to avoid such problems. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1 is a flowchart of a method for constructing a structured light illumination microscope image reconstruction model provided by an embodiment of the present application.
[0041] Figure 2 is a result diagram of a method for constructing a structured light illumination microscope image reconstruction model provided by an embodiment of the present application. DETAILED DESCRIPTION
[0042] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.
[0043] The content involved in the above embodiments will be described below in combination with a preferred embodiment.
[0044] Embodiment One
[0045] The present application provides a method for constructing a structured light illumination microscope image reconstruction model, comprising:
[0046] S1, collecting a pair of data sets D={I raw ,I GT} of cells under a structured light illumination microscope, and dividing them into a training set D1 and a test set D2; the data set D includes original images I raw with different signal-to-noise ratios and true value images I GT ;
[0047] S2, constructing a deep neural network model for structured light illumination microscope reconstruction, and deploying it as a Bayesian neural network; the deep neural network model outputs the reconstruction mean μ and the reconstruction standard deviation σ of the reconstructed super-resolution image distribution;
[0048] S3, inferring the posterior distribution of the parameters of the Bayesian neural network, training the parameters of the Bayesian neural network in S2 by using the training data D1 and a preset loss function, and obtaining a plurality of different parameter sets;
[0049] S4, inputting the test data D2 into the trained Bayesian neural network for model inference, obtaining a plurality of reconstruction means μ and reconstruction standard deviations σ of the reconstructed super-resolution image distribution, calculating the mean of the plurality of reconstruction means μ as the reconstructed super-resolution image, calculating the mean of the plurality of reconstruction standard deviations σ as the reconstructed data uncertainty, and calculating the standard deviation of the plurality of reconstruction means μ as the reconstructed model uncertainty.
[0050] S1 specifically comprises:
[0051] S11, under the same field of view, changing the direction and phase of the structured light, collecting the data of the cell sample under the structured light illumination microscope, and obtaining a group of original images I raw ;
[0052] S12, adjusting the excitation power or exposure time of the structured light, obtaining a group of original images I raw with different signal-to-noise ratios in the same field of view, obtaining the original image dataset in the same field of view;
[0053] S13, selecting a group of original images with the highest signal-to-noise ratio from the original image dataset in the same field of view, and reconstructing to obtain the true value image I GT ;
[0054] S14, adjusting the field of view of the structured light illumination microscope, repeating S11-S13, collecting a plurality of pairs of data, and forming a dataset D={I raw ,I GT}.
[0055] Reference Figure 1 , Figure 1 is the overall flowchart of the scheme.
[0056] Imaging the cell sample using sinusoidal interference fringe illumination, and the imaging model is:
[0057]
[0058] Where θ is the azimuth angle of the structured light, p θ is the light wave vector of the structured light, c θ is the modulation depth of the structured light, is the initial phase of the structured light, f(x) represents the fluorescence distribution of the sample, h(x) is the point spread function of the imaging system, represents convolution operation.
[0059] For the same field of view, the structured light is repeated three times in the spatial direction with an interval angle of 120°, and the initial phase is also repeated three times, i.e. m=0,±1, and nine times of imaging are performed with structured light in different directions and phases, to obtain a group of nine frames of diffraction-limited original images I raw . By adjusting the excitation power or exposure time, a plurality of groups of original images with different signal-to-noise ratios can be collected. At the same time, a group of original images with extremely high signal-to-noise ratio are collected and a super-resolution image I GT with a resolution of about 100 nanometers is reconstructed using the traditional reconstruction method. The above operation is repeated for different fields of view of the cell sample, a plurality of pairs of data are collected, and a dataset D={I raw ,I GT} is formed.
[0060] The collected data set D is preprocessed, which includes background removal, normalization and data enhancement. First, I GT The camera background is subtracted from the image; then the background fluorescence is removed by using a background removal algorithm; then I raw And I GT Normalization is performed to scale the image dynamic range to the range [0, 1] for subsequent neural network processing; then, random cropping and random flipping data augmentation techniques are used to increase data diversity to avoid overfitting of the subsequent neural network training; finally, the preprocessed data set is randomly divided into training set D1 and test set D2 in the ratio of 4:1, which are used for subsequent model training and testing.
[0061] The structured light illumination microscope reconstruction deep neural network model is constructed, the input is set as a group of diffraction-limited original images (a group of nine original images), and the output target is the distribution of the reconstructed super-resolution image, and the mapping dimension is
[0062] The distribution of the reconstructed super-resolution image is a spatially heterogeneous Gaussian distribution. The reconstructed super-resolution image distribution can be parameterized as a spatially varying mean and standard deviation, so the actual output of the neural network model is the mean and standard deviation of the reconstructed super-resolution image distribution.
[0063] Optionally, S3 specifically includes:
[0064] S31, the SGLD algorithm is used to infer the posterior distribution of the parameters of the Bayesian neural network, and the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2 are trained by using the training data D1 to obtain a plurality of different parameter sets;
[0065] S32, the trained network parameters in S31 are frozen, the standard deviation quantization layer parameters of the Bayesian neural network in the Bayesian model deployment module are trained using the training data D1, and the trained Bayesian neural network with a plurality of different parameter sets is obtained.
[0066] The constructed neural network model is based on a residual convolutional neural network architecture, which takes a convolutional layer, a ReLU activation function, an attention mechanism, and a skip connection as a basic unit, and is composed of four parts: a feature mapping layer, a deep feature extraction layer, a mean reconstruction layer, and a standard deviation quantization layer. The feature mapping layer is a convolutional layer that maps nine original images to a high-dimensional feature space. The deep feature extraction layer is composed of a plurality of residual groups and a skip connection. The deep feature extraction layer includes a plurality of residual groups, each of which is composed of a plurality of residual blocks. The mean reconstruction layer includes an upsampling layer and a convolutional layer. The standard deviation quantization layer includes a convolutional layer and an activation function. The feature mapping layer maps the input original image to a high-dimensional feature space. The deep feature extraction layer extracts modulated super-resolution information in the high-dimensional feature space. The mean reconstruction layer and the standard deviation quantization layer are used to output the mean and standard deviation of the super-resolution image, respectively. The data type of the trainable parameters of the neural network is single-precision floating point (32-bit floating point number). Further, the deep neural network model in the present scheme is not limited to the residual convolutional neural network architecture.
[0067] The input data is processed by the feature mapping layer and the deep feature extraction layer to obtain deep features:
[0068] F = Conv 1→64 (I raw )+ RG 64→64 ((… RG 64→64 (Conv 1→64 (I raw )))
[0069] where Conv represents a convolutional layer, RG represents a residual group, and the subscript represents the change in the number of feature channels.
[0070] Each residual group is composed of a plurality of residual blocks (Residual Block):
[0071] RG 64→64 (x) = x + RB 64→64 (… RB 64→64 (x))
[0072] Each residual block is specifically:
[0073] RB 64→64 (x) = x + CA(Conv 64→64 (ReLU(Conv 64→64 (x))))
[0074] where ReLU is an activation function and CA is a channel attention mechanism.
[0075] The mean reconstruction layer uses an upsampling module and a convolutional layer to reconstruct the extracted deep features into a super-resolution image:
[0076] μ = Conv 64→1 (PS(F))
[0077] where PS denotes PixelShuffle up-sampling.
[0078] The standard deviation quantization layer is also based on the deep feature, using the convolution layer and the activation function to predict the standard deviation of the reconstructed image distribution:
[0079] σ = ReLU(Conv 64→64 (ReLU(Conv 64→64 (F))))
[0080] The neural network model built is modeled as a Bayesian neural network, that is, the network parameters are no longer a set of fixed values, but are subject to a certain probability distribution. In Bayesian inference, the parameter distribution is called the posterior distribution, that is, φ ~ p(φ|D1), and any sample φ ~ p(φ|D1) in the posterior distribution can well fit the training set data D1.
[0081] The inference of the posterior distribution will be implemented using the SGLD algorithm in the subsequent training steps. Specifically, the SGLD algorithm is used to infer the posterior distribution of the parameters of the Bayesian neural network. The inference process uses the training data D1 and the dual-domain loss function to train the feature mapping layer parameters, the deep feature extraction layer parameters and the mean reconstruction layer parameters of the Bayesian neural network in S2, and obtains multiple different parameter sets. After each iteration of the network parameters, the SGLD algorithm samples a Gaussian noise and applies it to the network parameters to approximate the true posterior distribution.
[0082] where the random Gaussian noise: η t ~ N(0, v t ), the subscript t represents the tth iteration, v t represents the network learning rate of the tth iteration
[0083] where the loss function used in training is the output reconstructed mean μ and the training target I GT The dual-domain error in the spatial and frequency domains, the dual-domain loss function is: where, denotes the Fourier transform.
[0084] The parameters obtained by training are the network parameters related to feature extraction and reconstructed mean in the Bayesian neural network, and the trained parameters are frozen.
[0085] The training data D1 divided is used to train the standard deviation quantization layer in the Bayesian neural network, and the loss function used in training is the Gaussian heteroscedastic loss:
[0086] The test data D2 and each set of parameters are input into the trained Bayesian neural network to perform reconstruction. The different sets of parameters obtained from the posterior distribution by the SGLD algorithm can be used to infer the distribution of multiple reconstructed super-resolution images. The mean of the means of multiple distributions is calculated as the final reconstruction result. The mean of the standard deviations of multiple distributions is calculated as the final reconstruction standard deviation (also known as data uncertainty). The standard deviation of the means of multiple distributions is calculated as the model uncertainty.
[0087] Optionally, S4 specifically includes:
[0088] S41. Input the test data D2 into the trained Bayesian neural network to perform a model inference to obtain K reconstruction mean μ and reconstruction standard deviation σ.
[0089] S42. Calculate the mean value μ of the K groups of reconstructions as the reconstruction result of the super-resolution image:
[0090]
[0091] The mean of multiple reconstruction standard deviations σ is calculated as the data uncertainty of the reconstruction:
[0092]
[0093] The standard deviation of the multiple reconstruction means μ is calculated as the model uncertainty of the reconstruction:
[0094]
[0095] like Figure 2 As shown, compared to existing methods, the reconstruction results of this invention have higher resolution and better spatial detail resolution. Furthermore, the quantified data uncertainty of this invention indicates the standard deviation of the super-resolution image distribution, which, combined with subsequent statistical analysis, allows for the determination of the reliability of each region in the reconstructed image. The quantified model uncertainty of this invention describes the consistency of reconstruction results obtained by sampling the posterior distribution of model parameters differently; the magnitude of the model uncertainty can be used to determine the degree of matching between the neural network model and the input data.
[0096] The embodiment of the present application trains the parameters to obtain multiple sets of parameter sets, so that multiple results can be obtained when the model is inferred, by deploying the constructed neural network model as a Bayesian neural network to make it conform to an unknown posterior distribution. Meanwhile, the distribution of the reconstructed super-resolution image is designed as a spatial heteroscedastic Gaussian distribution, so that the actual output of the neural network model is the mean and standard deviation of the distribution of the reconstructed super-resolution image, so that the mean of multiple means and the standard deviation of multiple standard deviations can be calculated to obtain the mean of multiple distribution means as the final reconstruction result, the mean of multiple distribution standard deviations as the final reconstruction standard deviation representing data uncertainty, and the standard deviation of multiple distribution means as model uncertainty. The technical problems of low feasibility of the reconstructed image and the inability to quantify the uncertainty of the reconstructed image are solved, accurate image reconstruction is achieved, and the uncertainty of the reconstructed super-resolution image can be quantified, and the reconstructed data uncertainty and the reconstructed model uncertainty can help users effectively identify unreliable results.
[0097] Embodiment two
[0098] The present application also provides a structured light illumination microscope image reconstruction model construction device for executing the method in embodiment one, comprising:
[0099] A data acquisition module is configured to acquire a pair of data sets D={I raw ,I GT} of cells under a structured light illumination microscope and divide them into a training set D1 and a test set D2; the data set D includes original images I raw with different signal-to-noise ratios and true value images I GT .
[0100] A model construction module is configured to construct a deep neural network model for structured light illumination microscope reconstruction and deploy it as a Bayesian neural network; the deep neural network model outputs a reconstruction mean μ and a reconstruction standard deviation σ of the distribution of the reconstructed super-resolution image.
[0101] A Bayesian model training module is configured to infer the posterior distribution of the parameters of the Bayesian neural network, train the parameters of the Bayesian neural network in the model construction module using the training data D1 and a preset loss function, and obtain multiple different sets of parameters.
[0102] An image reconstruction model test module is configured to input test data D2 into the trained Bayesian neural network to perform model inference and obtain multiple sets of reconstruction mean μ and reconstruction standard deviation σ of the distribution of the reconstructed super-resolution image; calculate the mean of multiple reconstruction means μ as the reconstructed super-resolution image, calculate the mean of multiple reconstruction standard deviations σ as the reconstructed data uncertainty, and calculate the standard deviation of multiple reconstruction means μ as the reconstructed model uncertainty.
[0103] The device for constructing a structured light illumination microscope image reconstruction model provided by the embodiment of the present application is used for executing the method for constructing a structured light illumination microscope image reconstruction model provided by any embodiment of the present application, and has corresponding function modules and beneficial effects.
[0104] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.
Claims
1. A method for constructing an image reconstruction model using structured light illumination microscopy, characterized in that, include: S1. Collect paired datasets of cells under structured light illumination microscopy. And divided into training sets and test set The dataset D includes original images with different signal-to-noise ratios. and truth image ; S2. Construct a deep neural network model for reconstruction using structured light illumination microscopy and deploy it as a Bayesian neural network; the deep neural network model outputs the reconstruction mean of the super-resolution image distribution. and reconstruction standard deviation ; S3. Infer the posterior distribution of the parameters of the Bayesian neural network by using training data. The parameters of the Bayesian neural network in S2 are trained using a preset loss function to obtain multiple different parameter sets; S4. Test data The trained Bayesian neural network is used for model inference to obtain the reconstruction mean of multiple sets of reconstructed super-resolution image distributions. and reconstruction standard deviation ; Calculate the mean of multiple reconstructions The mean of the values is used as the reconstructed super-resolution image, and multiple reconstruction standard deviations are calculated. The mean of the reconstruction data is used as the uncertainty of the reconstruction data, and multiple reconstruction means are calculated. The standard deviation is used as the uncertainty of the reconstructed model.
2. The method as described in claim 1, characterized in that, S1 specifically includes: S11. Under the same field of view, change the direction and phase of the structured light, collect data of cell samples under a structured light illumination microscope, and obtain a set of raw images. ; S12. Adjust the excitation power or exposure time of the structured light to acquire multiple sets of original images with different signal-to-noise ratios for the same field of view. This yields the original image dataset with the same field of view; S13. Select the set of original images with the highest signal-to-noise ratio from the original image dataset of the same field of view, and reconstruct the ground truth image. ; S14. Adjust the field of view of the structured light illumination microscope, repeat S11-S13, and collect multiple sets of paired data to form a dataset. .
3. The construction method as described in claim 1, characterized in that, The images in the dataset D are preprocessed images, and the preprocessing includes background removal, normalization, and data augmentation.
4. The construction method as described in claim 1, characterized in that, The deep neural network model is constructed based on a residual convolutional neural network architecture.
5. The construction method as described in claim 1, characterized in that, S3 specifically includes: S31. Infer the posterior distribution of the parameters of the Bayesian neural network by using training data. The parameters of the feature mapping layer, the deep feature extraction layer, and the mean reconstruction layer of the Bayesian neural network in S2 were trained to obtain multiple different parameter sets. S32. Freeze the network parameters trained in S31 and use the training data. The standard deviation quantization layer parameters of the Bayesian neural network in the training Bayesian model deployment module are obtained to obtain a trained Bayesian neural network with multiple sets of different parameters.
6. The construction method as described in claim 5, characterized in that, The feature mapping layer includes a convolutional layer that maps the input raw image to a high-dimensional feature space.
7. The method as described in claim 1, characterized in that, S4 specifically includes: S41, Test data Input the trained Bayesian neural network into a model inference process to obtain... K Reconstruction mean and reconstruction standard deviation ; S42, Calculation K Group Reconstruction Mean The mean value is used as the reconstruction result of the super-resolution image: Calculate multiple reconstruction standard deviations The mean is used as the uncertainty of the reconstructed data: Calculate multiple reconstruction means The standard deviation is used as the uncertainty of the reconstructed model: 。 8. A device for constructing an image reconstruction model using structured light illumination microscopy, characterized in that, include: The data acquisition module is used to acquire paired datasets of cells under structured light illumination microscopy. And divided into training sets and test set The dataset D includes original images with different signal-to-noise ratios. and truth image ; The model building module is used to construct a deep neural network model for reconstruction using structured light illumination microscopy and deploy it as a Bayesian neural network; the deep neural network model outputs the reconstruction mean of the super-resolution image distribution. and reconstruction standard deviation ; The Bayesian model training module is used to infer the posterior distribution of the parameters of a Bayesian neural network by using training data. The parameters of the Bayesian neural network in the model building module are trained with a preset loss function to obtain multiple different parameter sets; The image reconstruction model testing module is used to process test data. The trained Bayesian neural network is used for model inference to obtain the reconstruction mean of multiple sets of reconstructed super-resolution image distributions. and reconstruction standard deviation ; Calculate the mean of multiple reconstructions The mean of the values is used as the reconstructed super-resolution image, and multiple reconstruction standard deviations are calculated. The mean of the reconstruction data is used as the uncertainty of the reconstruction data, and multiple reconstruction means are calculated. The standard deviation is used as the uncertainty of the reconstructed model.