An organoid microscopic image reconstruction method based on a multi-scale recurrent diffusion model

Abstract

The application discloses a kind of organoids microscopic image reconstruction methods based on multi-scale cyclic diffusion model, belong to computer vision, deep learning technical field.The application proposes multi-scale cyclic organoids microscopic diffusion reconstruction method to reconstruct high-quality organoids microscopic image, specifically, this method combines generated diffusion model with traditional iterative algorithm, effectively extracts features from the noise input of parameter Gaussian process.Subsequently, by adding cyclic hidden connection, the ability of model capture representation is enhanced, so as to speed up the convergence speed of training process.At the same time, the application uses multi-scale image as input, so that the model can obtain more comprehensive and richer image information, thereby enhancing the feature capture ability and robustness of model.

Description

Technical Field

[0001] This invention relates to a method for reconstructing organoid microscopic images based on a multi-scale cyclic diffusion model, which is used to reconstruct high-quality organoid images and belongs to the fields of computer vision and deep learning technology. Background Technology

[0002] Organoids, resembling tissues and organs, are three-dimensional cultures grown in vitro that can approximate source organs or tissues and create a model that simulates the in vivo environment. Organoids can simulate the structure and function of real organs to a great extent and are easier to manipulate than animal models. They can also be used to study the mechanisms of disease occurrence and development, serving as important clinical research models for translating basic cancer research findings into new treatment strategies and enabling personalized and precise drug screening for patients. However, due to various factors, medical imaging equipment often generates noise when acquiring image data. This noise may stem from limitations of the equipment itself, such as sensor instability and imaging system defects. Environmental factors such as electromagnetic waves and vibrations can also cause interference. Furthermore, the complexity of the human body itself is a significant cause of noise. This noise affects image visualization, reduces image quality, and may interfere with doctors' diagnoses, sometimes even misleading them to screen for the wrong drugs, leading to incorrect diagnoses or treatment plans. Therefore, medical image reconstruction has become an urgent problem to be solved.

[0003] Recently, deep learning algorithms have achieved great success in medical image processing, with a large amount of research focused on tasks such as biomarker prediction, disease classification, and cell type prediction. However, traditional organoid image reconstruction methods for bright-field images are ineffective in handling noise. Bright-field organoid images have low contrast, and organoids are highly adhesive, have high background noise, and blurred boundaries, all of which affect the reconstruction results. Summary of the Invention

[0004] To improve the image quality of organoid microscopic images, this invention provides a method for reconstructing organoid microscopic images based on a multi-scale cyclic diffusion model. The technical solution is as follows:

[0005] The first objective of this invention is to provide a method for reconstructing organoid microscopic images, comprising:

[0006] Step 1: Acquire images of organoid culture well plates;

[0007] Step 2: Based on the organoid culture plate images, collect bright-field images and acquire organoid cell images at intervals;

[0008] Step 3: Using the series of organoid cell images obtained in Step 2 as the initial state of the diffusion process, Gaussian noise is gradually added to the original images. The addition of Gaussian noise is a Markov chain process, in which the pixel values ​​of the image are transformed to values ​​closer to a Gaussian distribution at each step, ultimately obtaining a fully Gaussian noise image.

[0009] Step 4: Gradually restore the full Gaussian noise map from the noisy image to the noise-free image. This process is also a Markov chain process. At each step, the pixel value is adjusted to be closer to the value of the original noise-free image. In the denoising process, the recursive iterative convolutional network RICN is used to speed up the acquisition of image features.

[0010] Step 5: Use a multi-scale image feature extraction method to extract features from the image obtained in Step 4, and set the diffusion coefficient and denoising coefficient. Through multiple iterations and optimizations, ensure effective noise removal while preserving the details and features of the image, and obtain the reconstructed image.

[0011] Optionally, step 3 includes:

[0012] Given an original image x0, where x0~q(x0) represents the observed data x0 sampled according to a certain distribution q, a series of images x0, x1, ..., x2 are obtained by adding noise T times. T This process is based on variance. Gradually add noise to x0 to obtain x T Finally, we obtain the following equation:

[0013]

[0014] The image obtained at each time t is only related to the image at time t-1, as shown in the following formula:

[0015]

[0016] Where q(*) represents the conditional probability distribution, q(x) t |x t-1 () represents the prediction of the state x at time t, given the state at time t-1. t-1 The probability distribution; Let I represent a Gaussian distribution, and let I represent the identity matrix.

[0017] Optionally, step 4 uses a U-net network to denoise the full Gaussian noise map, and the formula for this process is as follows:

[0018]

[0019] Where θ is a parameter of the U-net network. It is a Gaussian distribution, where I is the identity matrix, such asFigure 3 As shown, CR-ResBlocki represents the i-th layer in the U-net network;

[0020] Model p in the reverse diffusion stage θ (X 0：T Considering the joint probability distribution p(x) T And how to calculate x in the previous step at each time step t. t-1 conditional probability distribution p θ (x t-1 |x t ), where p θ (x t-1 |x t The distribution is derived from a Gaussian distribution, and its mean and variance are controlled by the parameter θ, μ. θ (x t ,t) represents the state at time step t, based on the current state x t The mean of the Gaussian distribution calculated with parameter θ, ∑ θ (x t ,t) represents the state x at time step t. t Under the given conditions, the covariance matrix of the Gaussian distribution is calculated based on the parameter θ.

[0021] Optionally, in step 4, the recursive iterative convolutional network (RICN) is combined with the residual blocks in the U-net network to concatenate the input and output of the previous round as the input of the next round. The node weights of the RICN are calculated as follows:

[0022]

[0023] Among them, W i and W k B represents the filter from the input layer to the hidden convolutional layer and the filter from the hidden layer to the hidden recurrent convolutional layer, respectively. i Let i represent the bias term, i represent the layer, and k represent the iteration step. This represents the node weight of layer i at the k-th iteration step, and... Figure 2 The settings are the same in the middle;

[0024] The latent features are processed by a tuned linear unit ReLU activation function. In each iteration, the recursive iterative convolutional network (RICN) updates the state of its internal nodes based on the given input. The RICN is defined as follows:

[0025] X rec =G N (G N-1 (…(G1(X u ))))

[0026] in, G represents the input of a certain iteration or the output of the previous iteration. N (*) represents the network function in each iteration of the optimization step, N represents the number of iterations, and X represents the network function in each iteration. u This represents an undersampled image sequence of length T, which is also the input to the network.

[0027] A second objective of this invention is to provide an organoid microscopic image reconstruction system, comprising:

[0028] The image acquisition module is used to acquire organoid cell images;

[0029] An organoid image reconstruction module is configured to perform image reconstruction based on the organoid cell image;

[0030] The organoid image reconstruction module includes a diffusion generation network, a recursive iterative convolutional network, and a multi-scale image feature extraction module.

[0031] The diffusion generation network includes a forward noise-adding module and a backward noise-denoising module. The forward noise-adding module uses the organoid cell image as the initial state of the diffusion process and gradually adds Gaussian noise to the original image. Adding Gaussian noise is a Markov chain process, and at each step, the pixel values ​​of the image are changed to values ​​closer to the Gaussian distribution, ultimately obtaining a full Gaussian noise image. The backward noise-denoising module gradually restores the full Gaussian noise image from the noisy image to a noise-free image. This process is also a Markov chain process, and at each step, the pixel values ​​are adjusted to values ​​closer to the original noise-free image.

[0032] The recursive iterative convolutional network is used to accelerate the acquisition of image features during the denoising process of the backward denoising module.

[0033] The multi-scale image feature extraction module is a pyramid denoising structure where each branch focuses on features at one scale, in order to extract global information while preserving local details.

[0034] Optionally, the processing procedure of the forward noise-adding module includes:

[0035] Given an original image x0, where x0~q(X0) represents the observed data x0 sampled according to a certain distribution q, a series of images x0, x1, ..., x2 are obtained by adding noise T times. T This process is based on variance. Gradually add noise to x0 to obtain x T Finally, we obtain the following equation:

[0036]

[0037] The image obtained at each time t is only related to the image at time t-1, as shown in the following formula:

[0038]

[0039] Where q(*) represents the conditional probability distribution, q(x) t |x t-1 () represents the prediction of the state x at time t, given the state at time t-1. t-1 The probability distribution; Let I represent a Gaussian distribution, and let I represent the identity matrix.

[0040] Optionally, the backward denoising module uses a U-net network to denoise the full Gaussian noise map, and the formula for this process is as follows:

[0041]

[0042] Where θ is a parameter of the U-net network. It is a Gaussian distribution, where I is the identity matrix, such as Figure 3 As shown, CR-ResBlocki represents the i-th layer in the U-net network;

[0043] Model p in the reverse diffusion stage θ (X 0：T Considering the joint probability distribution p(x) T And how to calculate x in the previous step at each time step t. t-1 conditional probability distribution p θ (x t-1 |x t ), where p θ (x t-1 |x t The distribution is derived from a Gaussian distribution, and its mean and variance are controlled by the parameter θ, μ. θ (x t ,t) represents the state at time step t, based on the current state x t The mean of the Gaussian distribution calculated with parameter θ, ∑ θ (x t ,t) represents the state x at time step t. t Under the given conditions, the covariance matrix of the Gaussian distribution is calculated based on the parameter θ.

[0044] Optionally, the recursive iterative convolutional network (RICN) is combined with the residual blocks in the U-net network to concatenate the input and output of the previous round as the input of the next round. The node weights of the RICN are calculated as follows:

[0045]

[0046] Among them, W i and W k B represents the filter from the input layer to the hidden convolutional layer and the filter from the hidden layer to the hidden recurrent convolutional layer, respectively. i Let i represent the bias term, i represent the layer, and k represent the iteration step. This represents the node weight of layer i at the k-th iteration step, and... Figure 2 The middle nodes are set to be the same;

[0047] The latent features are processed by a tuned linear unit ReLU activation function. In each iteration, the recursive iterative convolutional network (RICN) updates the state of its internal nodes based on the given input. The RICN is defined as follows:

[0048] X rec =G N (G N-1 (…(G1(X u ))))

[0049] in, G represents the input of a certain iteration or the output of the previous iteration. N (*) represents the network function in each iteration of the optimization step, N represents the number of iterations, and X represents the network function in each iteration. u This represents an undersampled image sequence of length T, which is also the input to the network.

[0050] A third objective of this invention is to provide an electronic device, including a memory and a processor;

[0051] The memory is used to store computer programs;

[0052] The processor is configured to implement the organoid microscopic image reconstruction method as described above when executing the computer program.

[0053] A fourth objective of the present invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the organoid microscopic image reconstruction method as described in any of the preceding claims.

[0054] The beneficial effects of this invention are:

[0055] This invention proposes a multi-scale recurrent organoid microscopic diffusion reconstruction method to reconstruct high-quality organoid microscopic images. Specifically, this invention combines a generative diffusion model with a traditional iterative algorithm to effectively extract features from noisy inputs of a parametric Gaussian process. Subsequently, by adding recurrent hidden connections, the model's ability to capture representations is enhanced, thereby accelerating the convergence speed of the training process. Simultaneously, this invention uses multi-scale images as input, enabling the model to acquire more comprehensive and richer image information, thus enhancing feature capture capability and model robustness.

[0056] This invention utilizes a diffusion generation network, a recursive iterative convolutional network, and a multi-scale image feature extraction module to reconstruct organoid images, effectively improving the quality of organoid microscopic images and solving the problem of unclear organoid microscopic images caused by factors such as imaging equipment or individual differences in human anatomy. The intelligent prediction method provided by this invention can achieve the acquisition of more comprehensive, richer, and higher-quality organoid microscopic images. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the basic process of the organoid microscopic image reconstruction method of the present invention.

[0059] Figure 2 This is a structural diagram of the organoid microscopic image reconstruction model based on the multi-scale cyclic diffusion model in Embodiment 2 of the present invention.

[0060] Figure 3 This is a schematic diagram of the recursive iterative convolution module structure in Example 2.

[0061] Figure 4 This is a schematic diagram of the multi-scale feature extraction module in Embodiment 2 of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0063] Example 1:

[0064] This embodiment provides a method for reconstructing organoid microscopic images, including:

[0065] Step 1: Acquire images of organoid culture well plates;

[0066] Step 2: Based on the organoid culture plate images, collect bright-field images and acquire organoid cell images at intervals;

[0067] Step 3: Using the series of organoid cell images obtained in Step 2 as the initial state of the diffusion process, Gaussian noise is gradually added to the original image. Adding Gaussian noise is a Markov chain process. Each step changes the pixel values ​​of the image to values ​​that are closer to the Gaussian distribution, and finally a full Gaussian noise image is obtained.

[0068] Step 4: Gradually restore the Gaussian noise map from the noisy image to the noise-free image. This process is also a Markov chain process. At each step, the pixel values ​​are adjusted to be closer to the values ​​of the original noise-free image. In the denoising process, the recursive iterative convolutional network RICN is used to speed up the acquisition of image features.

[0069] Step 5: Use a multi-scale image feature extraction method to extract features from the image obtained in Step 4, and set the diffusion coefficient and denoising coefficient. Through multiple iterations and optimizations, ensure effective noise removal while preserving the details and features of the image.

[0070] Step 6: After steps 3-5, the reconstructed image is obtained.

[0071] Example 2

[0072] This embodiment provides a method for organoid microscopic image reconstruction based on a multi-scale cyclic diffusion model. See [link to relevant documentation]. Figures 1-4 The method includes:

[0073] Step 1: Acquire images of organoid culture well plates;

[0074] Step 2: Based on the organoid culture plate images, collect bright-field images and acquire organoid cell images at intervals;

[0075] Step 3: Use the series of organoid cell images obtained in Step 2 as the initial state for the diffusion process; gradually add Gaussian noise to the initial state images. This step can be viewed as a Markov chain process, where each step transforms the pixel values ​​of the image into values ​​closer to a Gaussian distribution. As the steps proceed, the noise gradually increases, and the image quality gradually decreases, approaching a fully Gaussian noise image.

[0076] The process of adding noise is considered a forward process. In this forward process, given the organoid cell images x0~q(x0) in the initial state, a series of images x0,x1,…,x are obtained by adding noise T times cumulatively. T This process is based on variance. Gradually add noise to x0 to obtain x TFinally, we obtain the following equation:

[0077]

[0078] The image obtained at each time t is only related to the image at time t-1, as shown in the following formula:

[0079]

[0080] Before the image is input into the network, it is processed by a multi-scale image feature extraction module ε( Figure 4 (As shown) Images are divided into different scales, and these images are input into a reconstruction network for reconstruction. Figure 2 The multi-scale feature fusion module D in the model represents the fusion of image features at different scales to output a single image, which can be seen as the inverse operation of ε.

[0081] Step 4: After obtaining the image from Step 3, begin to gradually restore the image from the noisy image to the noise-free image. This process is also a Markov chain process, in which the pixel values ​​are adjusted to be closer to the values ​​of the original noise-free image at each step.

[0082] This process, used for noise removal, is called the reverse process. In this process, U-net is used as the denoising network to parameterize the process, progressively learning the noise distribution during the forward diffusion process and predicting the reverse distribution p. θ This process restores the original image distribution x0 from a perfectly standard Gaussian distribution, as shown in the following formula:

[0083]

[0084] Where θ is a parameter of the denoising U-net. It follows a Gaussian distribution, where I is the identity matrix. The model p for the backdiffusion stage... θ (X 0:T Considering the joint probability distribution p(x) T And how to calculate x in the previous step at each time step t. t-1 conditional probability distribution p θ (x t-1 |x t ), where p θ (x t-1 |x t The distribution is derived from a Gaussian distribution, and its mean and variance are controlled by the parameter θ.

[0085] In the above denoising process, a recursive iterative convolutional network (RICN) is added to accelerate the acquisition of texture and shape features of the image.

[0086] By combining a recursive iterative convolutional network with the residual block CR-Resblock in the U-net model, the input and output of the previous round are concatenated and used as the input of the next round.

[0087] Furthermore, this embodiment proposes a new method for calculating node weights, which can store the weights of nodes at each layer. The node weights are calculated as follows:

[0088]

[0089] In the recursive iterative convolutional network of this embodiment, W i and W k These represent the filters from the input layer to the hidden convolutional layer and the filters from the hidden layer to the hidden recurrent convolutional layer, respectively. B i This is the bias term. The latent features are processed by an Integrating Linear Unit (ReLU) activation function, i.e., σ(x) = max(0,x). In each iteration, the RICN updates the state of its internal nodes based on the given input. The neural model is defined as follows:

[0090] X rec =G N (G N-1 (···(G1(X u ))))

[0091] in, Represents the undersampled image at the i-th iteration. The progressive reconstruction results show that the weights of each node in this module consist of two parts. Therefore, the hidden-to-hidden iterative connections in each unit can pass the feature information collected in the previous iteration to subsequent iterations. This allows each iteration to optimize not only based on the output image but also based on the hidden features from the previous iteration. The hidden connection convolution "remembers" valuable features, avoiding redundant computation. Furthermore, as the number of iterations increases, the information of each node is reset, resulting in more convincing features to support feature extraction in the next part. Because the weight parameters are shared throughout the iterations, the number of parameters is reduced, leading to better generalization properties.

[0092] Step 5: Utilize multi-scale image feature extraction methods, i.e., through... Figure 4 Pooling operations and the U-net module ( Figure 4 The PU module shown performs further feature extraction on the image features obtained in step 4, and selects appropriate diffusion coefficients and denoising coefficients. Through multiple iterations and optimizations, it ensures effective noise removal while preserving the details and features of the image. The reconstructed image is obtained through the diffusion-based reverse denoising process.

[0093] This embodiment constructs an organoid microscopic image reconstruction model based on a multi-scale cyclic diffusion model. By inputting the organoid cell image obtained in step 2 into the trained organoid image reconstruction model, a clearer organoid cell image than the original noisy image can be obtained, while preserving the details and features of the original image.

[0094] This embodiment trains the organoid reconstruction model using the following training loss functions: mean squared error (MSE) loss function, structural similarity (SSIM) loss function, and cross-entropy (CE) loss function. The overall calculation formula is as follows:

[0095]

[0096] in The calculation formula is:

[0097]

[0098] Where y i Represents the true value. This represents the predicted value, where n is the number of samples.

[0099] in The calculation formula is:

[0100]

[0101] Where n is the number of samples, y i It is the actual value. This is a predicted value.

[0102] in, The formula is as follows:

[0103]

[0104] Where x and y represent the actual value and the predicted value, respectively, and μ x and μ y Let σ be the average of x and y, respectively. x and σ y Let σ be the standard deviation of x and y, respectively. xy Let C1 and C2 be the covariance of x and y, and C1 and C2 be two constants used to avoid the case where the denominator is 0.

[0105] Example 3:

[0106] This embodiment provides an organoid microscopic image reconstruction system for implementing the organoid microscopic image reconstruction method described in Embodiment 2, including:

[0107] The image acquisition module is used to acquire organoid cell images;

[0108] An organoid image reconstruction module is configured to perform image reconstruction based on the organoid cell image;

[0109] The organoid image reconstruction module includes a diffusion generation network, a recursive iterative convolutional network, and a multi-scale image feature extraction module.

[0110] The diffusion generation network includes a forward noise-adding module and a backward noise-denoising module. The forward noise-adding module uses the organoid cell image as the initial state of the diffusion process and gradually adds Gaussian noise to the original image. Adding Gaussian noise is a Markov chain process, and at each step, the pixel values ​​of the image are changed to values ​​closer to the Gaussian distribution, ultimately obtaining a full Gaussian noise image. The backward noise-denoising module gradually restores the full Gaussian noise image from the noisy image to a noise-free image. This process is also a Markov chain process, and at each step, the pixel values ​​are adjusted to values ​​closer to the original noise-free image.

[0111] The recursive iterative convolutional network is used to accelerate the acquisition of image features during the denoising process of the backward denoising module.

[0112] The multi-scale image feature extraction module is a pyramid denoising structure where each branch focuses on features at one scale, in order to extract global information while preserving local details.

[0113] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing organoid microscopic images, characterized in that, The method includes: Step 1: Acquire images of organoid culture well plates; Step 2: Based on the organoid culture plate images, collect bright-field images and acquire organoid cell images at intervals; Step 3: Using the series of organoid cell images obtained in Step 2 as the initial state of the diffusion process, Gaussian noise is gradually added to the original image. Adding Gaussian noise is a Markov chain process. Each step changes the pixel values ​​of the image to values ​​that are closer to the Gaussian distribution, and finally a full Gaussian noise image is obtained. Step 4: Gradually restore the full Gaussian noise map from the noisy image to the noise-free image. This process is also a Markov chain process. At each step, the pixel value is adjusted to be closer to the value of the original noise-free image. In the denoising process, the recursive iterative convolutional network RICN is used to speed up the acquisition of image features. Step 5: Use a multi-scale image feature extraction method to extract features from the denoised image obtained in Step 4, and set the diffusion coefficient and denoising coefficient. Through multiple iterations and optimizations, ensure effective noise removal while preserving the details and features of the image, and obtain the reconstructed image. Step 3 includes: Given the original image ,in Representing observation data Sampled according to a certain distribution q, and added cumulatively. Noise was used to obtain a series of images. This process is based on variance. Gradually towards Adding noise to obtain Finally, we obtain the following equation: Among them, each moment The obtained image is only related to time. The image is related to the following formula: in, ) represents the conditional probability distribution. Indicates the time. t In the state of -1, predict at time... t status The probability distribution; Indicates a Gaussian distribution. Represents the identity matrix; Step 4 uses the U-net network to denoise the full Gaussian noise map. The formula for this process is as follows: in, These are the parameters of the U-net network, and the model for the backdiffusion phase. Considering the joint probability distribution and each time step How to calculate the previous step conditional probability distribution ,in, Derived from the Gaussian distribution, its mean and variance are given by the parameters. control, This indicates that at time step t Next, based on the current state and parameters The calculated mean of the Gaussian distribution, Indicates at time step t and state Under the given conditions, the covariance matrix of the Gaussian distribution is calculated based on the parameter θ.

2. The method according to claim 1, characterized in that, In step 4, the recursive iterative convolutional network (RICN) is combined with the residual blocks in the U-net network, concatenating the input and output of the previous round as the input of the next round. The node weights of the RICN are calculated as follows: in, and These represent filters from the input layer to the hidden convolutional layer and filters from the hidden layer to the hidden recurrent convolutional layer, respectively. Indicates the bias term. i Indicates the number of floors. k Indicates the iteration steps, Indicates the first i The layer in the first k Node weights during each iteration step; The latent features are processed by a tuned linear unit ReLU activation function. In each iteration, the recursive iterative convolutional network (RICN) updates the state of its internal nodes based on the given input. The RICN is defined as follows: in, This represents the input of a current iteration or the output of the previous iteration. This represents the network function used in each iteration of the optimization step. N Indicates the number of iterations. This represents an undersampled image sequence, which is also the input to the network.

3. An organoid microscopic image reconstruction system, characterized in that, The system includes: The image acquisition module is used to acquire organoid cell images; An organoid image reconstruction module is configured to perform image reconstruction based on the organoid cell image; The organoid image reconstruction module includes a diffusion generation network, a recursive iterative convolutional network, and a multi-scale image feature extraction module. The diffusion generation network includes a forward noise-adding module and a backward noise-denoising module. The forward noise-adding module uses the organoid cell image as the initial state of the diffusion process and gradually adds Gaussian noise to the original image. Adding Gaussian noise is a Markov chain process, and at each step, the pixel values ​​of the image are changed to values ​​closer to the Gaussian distribution, ultimately obtaining a full Gaussian noise image. The backward noise-denoising module gradually restores the full Gaussian noise image from the noisy image to a noise-free image. This process is also a Markov chain process, and at each step, the pixel values ​​are adjusted to values ​​closer to the original noise-free image. The recursive iterative convolutional network is used to accelerate the acquisition of image features during the denoising process of the backward denoising module. The multi-scale image feature extraction module is for the pyramid denoising structure, where each branch focuses on features at one scale, in order to extract global information while preserving local details. The processing procedure of the forward noise addition module includes: Given the original image ,in Representing observation data Sampled according to a certain distribution q, and added cumulatively. Noise was used to obtain a series of images. This process is based on variance. Gradually towards Adding noise to obtain Finally, we obtain the following equation: Among them, each moment The obtained image is only related to time. The image is related to the following formula: in, ) represents the conditional probability distribution. This represents the state at time t, given the state at time t-1. The probability distribution; Indicates a Gaussian distribution. Represents the identity matrix; The backward denoising module uses a U-net network to denoise the full Gaussian noise map, and the formula for this process is as follows: in, These are the parameters of the U-net network, and the model for the backdiffusion phase. Considering the joint probability distribution and each time step How to calculate the previous step conditional probability distribution ,in, Derived from the Gaussian distribution, its mean and variance are given by the parameters. control, This indicates that at time step t, based on the current state... and parameters The calculated mean of the Gaussian distribution, Indicates the state at time step t. Under the given conditions, the covariance matrix of the Gaussian distribution is calculated based on the parameter θ.

4. The organoid microscopic image reconstruction system according to claim 3, characterized in that, The Recursive Iterative Convolutional Network (RICN) is combined with the residual blocks in the U-net network to concatenate the input and output of the previous round as the input of the next round. The node weights of the RICN are calculated as follows: in, and These represent filters from the input layer to the hidden convolutional layer and filters from the hidden layer to the hidden recurrent convolutional layer, respectively. Indicates the bias term. i Presentation layer, k Indicates the iteration steps, Indicates the first i The layer in the first k Node weights during each iteration step; The latent features are processed by a tuned linear unit ReLU activation function. In each iteration, the recursive iterative convolutional network (RICN) updates the state of its internal nodes based on the given input. The RICN is defined as follows: in, This represents the input of a current iteration or the output of the previous iteration. This represents the network function used in each iteration of the optimization step. N Indicates the number of iterations. This represents an undersampled image sequence, which is also the input to the network.

5. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the organoid microscopic image reconstruction method as described in any one of claims 1 to 2 when executing the computer program.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the organoid microscopic image reconstruction method as described in any one of claims 1 to 2.

Images