Image recognition-based automatic analysis method and system for cell aging

Abstract

The application provides an image recognition-based automatic cell aging analysis method and system, and relates to the technical field of image analysis. The method comprises the following steps: obtaining a staining image and a spontaneous fluorescence image of target cells treated by a target compound; then performing pixel-level spectrum fusion on the two images to generate a fused image; then inputting the fused image into a feature learning network to perform feature learning and obtain an aging feature, wherein a generator in the network generates a pseudo-active vector based on the fused image, and a discriminator distinguishes the pseudo-active vector from a real active vector; then performing feature splicing on the aging feature and the real active vector to form a joint feature vector, analyzing the joint feature vector, mining the correlation between different feature dimensions, and generating a structured feature representation; finally, inputting the structured feature representation into a phenotype classifier for processing, and outputting an analysis atlas of cell aging. The application realizes multi-dimensional accurate identification and visual analysis of the cell aging state.

Description

Technical Field

[0001] This application relates to the field of image analysis technology, and in particular to an automatic cell senescence analysis method and system based on image recognition. Background Technology

[0002] Automated cell senescence analysis technology is an important research tool in the fields of drug screening and toxicity assessment. By analyzing changes in cell morphology and physiological characteristics after compound treatment, it can provide key data support for anti-aging drug development and environmental pollutant assessment.

[0003] Existing image recognition-based methods for analyzing cell senescence typically acquire stained images of cells after they have been treated with compounds. They then extract morphological or texture features from the images to distinguish senescent cells from normal cells. Some technical solutions combine multiple staining markers to improve recognition accuracy and present the analysis results in a visual form for scientific observation.

[0004] However, in drug or compound treatment scenarios, cellular senescence phenotypes exhibit significant heterogeneity and dynamic changes. Metabolic changes in some weakly senescent phenotype cells are difficult to fully reveal through conventional staining. Furthermore, the pharmacological effects of compounds on cells can be confused with cellular senescence characteristics in images, limiting the accuracy and subtype differentiation capabilities of existing analytical methods. Therefore, existing technologies suffer from insufficient accuracy in identifying cellular senescence phenotypes. Summary of the Invention

[0005] This application provides an automatic cell senescence analysis method and system based on image recognition, which solves the problem of insufficient multi-dimensional accurate identification and visualization analysis capabilities of cell senescence states in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an automated cell senescence analysis method based on image recognition, comprising:

[0007] Acquire staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites;

[0008] The stained image and the autofluorescence image are spectrally fused at the pixel level to generate a fused image;

[0009] The fused image is input into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound.

[0010] The aging features are concatenated with the real activity vector to form a joint feature vector. The joint feature vector is then analyzed using a relational network to uncover the correlation between different feature dimensions and generate a structured feature representation.

[0011] The structured feature representation is input into a phenotypic classifier, which processes the structured feature representation and outputs an analytical map of cell senescence.

[0012] Optionally, the step of analyzing the joint feature vector using a relational network to uncover the correlations between different feature dimensions and generate a structured feature representation includes:

[0013] The joint feature vector is input into a relational network, in which a virtual relational graph is constructed, and the nodes in the virtual relational graph correspond to different feature dimensions in the joint feature vector;

[0014] The virtual relation graph is encoded using a graph autoencoder to learn a low-dimensional latent representation of the nodes;

[0015] Based on the values ​​of each feature dimension in the joint feature vector, and combined with the low-dimensional latent representation, the association strength between any two nodes in the virtual relationship graph is calculated.

[0016] Based on the association strength, determine the edge connection weights between all nodes in the virtual relationship graph;

[0017] A graph attention network is used to adaptively perform weighted information propagation and aggregation operations on the feature dimension representations of nodes in the virtual relationship graph based on the edge connection weights, so as to update the feature dimension representation corresponding to each node;

[0018] The updated feature dimension representations of all nodes are combined to form a structured feature representation.

[0019] Optionally, the graph attention network is used to adaptively perform weighted information propagation and aggregation operations on the feature dimension representations of nodes in the virtual relationship graph based on the edge connection weights, in order to update the feature dimension representation corresponding to each node, including:

[0020] In the graph attention network, each node in the virtual relationship graph is sequentially taken as a target node, and the following operations are performed for each target node:

[0021] Based on the current feature representation of the target node and the current feature representation of each neighboring node, as well as the edge connection weights of the edges connecting the target node and the corresponding neighboring nodes, the attention weight distribution is calculated.

[0022] Based on the attention weight distribution, the current feature representations of all neighboring nodes of the target node are dynamically weighted;

[0023] The current feature representations of all adjacent nodes after weighting are aggregated to form the context-enhanced feature of the target node;

[0024] The current feature representation of the target node is fused with the context-enhanced features to generate the updated feature dimension representation of the target node.

[0025] Optionally, the step of inputting the fused image into a feature learning network for feature learning to obtain aging features includes: using a generator in the feature learning network to generate a pseudo-activity vector based on the fused image, and using a discriminator in the feature learning network to distinguish the pseudo-activity vector from the real activity vector, comprising:

[0026] The fused image is input into the generator, and the fused image is abstracted and transformed layer by layer through the multi-layer coding structure in the generator to output a pseudo-active vector.

[0027] The pseudo-activity vector and the real activity vector are input into the discriminator. In the discriminator, the feature distribution difference between the pseudo-activity vector and the real activity vector is measured to obtain the difference measurement result.

[0028] Based on the difference measurement results, feature feedback information is generated and transmitted back to the generator. The generator then iteratively adjusts the parameters of the multi-layer coding structure according to the feature feedback information.

[0029] When the preset number of iterations is reached, the fused image is used to extract features using the adjusted multi-layer coding structure to obtain aging features.

[0030] Secondly, this application provides an automated cell senescence analysis system based on image recognition, comprising:

[0031] The acquisition module is used to acquire staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites;

[0032] The fusion module is used to perform pixel-level spectral fusion of the stained image and the autofluorescence image to generate a fused image;

[0033] The input module is used to input the fused image into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound.

[0034] The splicing module is used to splice the aging features with the real activity vector to form a joint feature vector. The joint feature vector is then analyzed using a relational network to uncover the correlation between different feature dimensions and generate a structured feature representation.

[0035] The output module is used to input the structured feature representation into the phenotype classifier, which processes the structured feature representation and outputs an analysis map of cell senescence.

[0036] Thirdly, this application provides an electronic device, comprising:

[0037] Memory, used to store computer programs;

[0038] A processor is configured to execute the computer program to implement the steps of the image recognition-based automated cell senescence analysis method as described in the first aspect above.

[0039] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the image recognition-based automatic cell senescence analysis method described in the first aspect above.

[0040] The technical solution provided in this application has the following beneficial effects:

[0041] This application simultaneously acquires image data from two different modalities to obtain enzyme activity staining information and metabolite autofluorescence information, providing multi-dimensional raw data for subsequent analysis. Then, pixel-level spectral fusion processing is performed on the two images to generate a fused image that simultaneously contains enzyme activity and metabolite distribution characteristics. Next, the fused image is fed into a feature learning network, where a generator produces pseudo-activity vectors, which are then distinguished from the real activity vectors by a discriminator, thereby extracting aging depth features related to the compound's mechanism of action. Subsequently, the extracted aging features are concatenated with the pre-constructed real activity vectors to form a joint feature vector, and a relational network is used to analyze this vector to uncover the intrinsic correlations between different feature dimensions, generating a structured feature representation. Finally, the structured feature representation is input into a phenotypic classifier for processing, outputting a cell aging analysis atlas containing aging probability and subtype information.

[0042] This process enables the effective integration of multimodal data and the extraction of deep features, improving the accuracy of cell senescence status determination and the refinement of subtype differentiation, and providing intuitive analytical results for drug screening research.

[0043] Furthermore, in the relational network analysis process, firstly, joint feature vectors are input to construct a virtual relational graph, where each node corresponds to a different feature dimension; then, a graph autoencoder is used to encode and learn the virtual relational graph to obtain the low-dimensional latent representation of the nodes, and the association strength and edge connection weights between nodes are calculated by combining feature values; next, a graph attention network is used to adaptively weighted propagate and aggregate the node features based on the calculated weights, thereby updating the feature dimension representation of each node; finally, all the updated node features are combined to form a structured feature representation.

[0044] This process enhances the discriminative power and biological interpretability of feature representations by modeling and optimizing the relationships between feature dimensions.

[0045] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating an automated cell senescence analysis method based on image recognition, provided for embodiments of this application;

[0048] Figure 2 A schematic diagram illustrating a specific implementation of an image recognition-based automated cell senescence analysis method provided in this application embodiment;

[0049] Figure 3 This is a schematic diagram of the structure of an automatic cell senescence analysis system based on image recognition, provided in an embodiment of this application. Detailed Implementation

[0050] To address the problems of existing technologies, this application proposes an automated cell senescence analysis method based on image recognition. First, it simultaneously acquires cell staining images and autofluorescence images processed with compounds, obtaining enzyme activity information and metabolite distribution information respectively. The two images are then fused at the pixel level to generate a fused image, thereby achieving complementary enhancement of multi-dimensional senescence features at the data source. Next, the fused image is fed into a feature learning network, where an adversarial mechanism between the generator and discriminator extracts deep senescence features related to the compound's mechanism of action. This feature is then concatenated with a real activity vector constructed based on the compound's activity spectrum to form a joint feature vector. Subsequently, a relational network is used to analyze the joint feature vector, uncovering the intrinsic correlations between different feature dimensions to generate a structured feature representation. Finally, a phenotypic classifier outputs an analytical map containing senescence probability and subtype information.

[0051] Therefore, this method enhances the identifiability of weakly aging phenotype cells by fusing multimodal data at the source and mining deep features, and achieves effective differentiation of different aging subtypes, fundamentally solving the problem of insufficient identification accuracy caused by phenotypic heterogeneity and pharmacological effects in existing technologies.

[0052] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] The core of this application is to provide an automated method for analyzing cell senescence based on image recognition, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0054] Step 101: Obtain staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites.

[0055] In step 101, target cells refer to the cell population to be analyzed after treatment with the target compound, and target compound refers to the analyte used to treat cells to induce or intervene in the cellular senescence process.

[0056] The staining image is image data acquired through bright-field imaging. This image data is used to show the distribution and intensity of the colored precipitates generated by the reaction of aging-related enzymes and staining substrates in cells after treatment with the target compound.

[0057] Autofluorescence images are image data acquired after excitation by a light source of a specific wavelength. This image data is used to present the intensity distribution of fluorescence signals spontaneously emitted by aging-related metabolites in cells after excitation.

[0058] In this embodiment, the cell sample treated with the target compound is first placed in a multispectral imaging system, and the staining image is acquired under bright field illumination. At the same time, the excitation light source is switched to a specific wavelength and a corresponding emission filter is used to acquire the autofluorescence image under the same field of view, thereby obtaining two types of image data reflecting enzyme activity information and metabolite distribution information of the same cell population under the same conditions.

[0059] Step 102: Perform pixel-level spectral fusion of the stained image and the autofluorescence image to generate a fused image.

[0060] The fused image is image data generated through pixel-level spectral fusion processing. The information of each pixel position in this image data is fused with the imaging features of the corresponding positions in the two original images.

[0061] In this embodiment, step 102 includes the following process:

[0062] Step 1021: Determine the correspondence between the pixel value matrix of the stained image and the pixel value matrix of the autofluorescence image.

[0063] In step 1021, the pixel value matrix of the stained image refers to a two-dimensional array composed of the pixel values ​​of all pixels in the stained image arranged according to their spatial positions; the pixel value matrix of the autofluorescence image refers to a two-dimensional array composed of the pixel values ​​of all pixels in the autofluorescence image arranged according to their spatial positions.

[0064] The correspondence refers to the one-to-one mapping relationship between two pixel value matrices in spatial location. This correspondence is established based on the same field of view and pixel resolution when the two images are acquired.

[0065] In this embodiment of the application, the stained image and the autofluorescence image are first read to obtain the pixel resolution and field of view information of the two images. By comparing the spatial coordinate parameters of the two images, the mapping relationship between the positions of the two images at the pixel level is determined.

[0066] Step 1022: Based on the correspondence, define a combination coefficient for the pixel value at each corresponding position.

[0067] In step 1022, the combination coefficient refers to the weighting parameter assigned to the pixel values ​​of the stained image and the pixel values ​​of the autofluorescent image in subsequent combination calculations.

[0068] In this embodiment of the application, based on the correspondence between each pixel position already determined in the two images, an initial combination coefficient is assigned to each pair of corresponding pixel positions. This initial combination coefficient is used to adjust the relative contribution ratio of the stained image pixel value and the autofluorescent image pixel value in the subsequent combination calculation.

[0069] Step 1023: Adjust the combination coefficient based on the correlation characteristics between the color information of enzyme activity and the fluorescence information of metabolites in the cell senescence process.

[0070] In step 1023, the color information of enzyme activity refers to the color intensity changes in the staining image that reflect the strength of aging-related enzyme activity, and the fluorescence information of metabolites refers to the fluorescence intensity changes in the autofluorescence image that reflect the distribution and content of aging-related metabolites.

[0071] Correlation characteristics refer to the biological correlation between changes in enzyme activity and accumulation of metabolites during cell senescence.

[0072] In this embodiment of the application, prior knowledge about the biological correlation between enzyme activity color depth and metabolite fluorescence intensity in the field of cell senescence research is introduced. Based on this prior knowledge, the initial combination coefficient set in step 1022 is dynamically adjusted so that in pixel regions where enzyme activity information and metabolite fluorescence information have a strong biological correlation, the corresponding combination coefficient can enhance the synergistic expression of the two during fusion.

[0073] Step 1024: Using the adjusted combination coefficients, perform a combination calculation on the pixel values ​​at corresponding positions in the pixel value matrix of the stained image and the pixel value matrix of the autofluorescent image to obtain a fused pixel value matrix.

[0074] In step 1024, the fused pixel value matrix refers to a new two-dimensional array of pixel values ​​generated after combined calculations. Each pixel value in this array integrates information from the corresponding position in the original stained image and the autofluorescence image.

[0075] In the embodiments of this application, for each pair of corresponding pixel positions, the adjusted combination coefficients are weighted by the pixel values ​​of the stained image and the autofluorescence image, and the results of the weighting operation are summarized to form a fused pixel value matrix covering the entire image field of view.

[0076] Step 1025: Generate a fused image based on the fused pixel value matrix.

[0077] In this embodiment of the application, the fusion pixel value matrix calculated in step 1024 is image-converted according to its two-dimensional spatial structure, and each fusion pixel value in the matrix is ​​mapped to an image display unit at the corresponding position, thereby generating a fusion image that can simultaneously present enzyme activity distribution characteristics and metabolite distribution characteristics.

[0078] This application establishes a spatial correspondence between stained images and autofluorescent images, and dynamically adjusts the combination coefficients based on the biological correlation characteristics between enzyme activity and metabolites during cell senescence, thereby achieving effective fusion of the two modal images at the pixel level. This enables the generated fused image to more comprehensively reflect the multidimensional features of cell senescence.

[0079] Step 103: Input the fused image into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound.

[0080] The feature learning network employs a generative adversarial network (GAN) architecture. The generator uses an encoder structure consisting of five convolutional layers. The first convolutional layer uses 32 3×3 kernels with a stride of 2 and padding of 1, followed by a batch normalization layer and a ReLU activation function. The second convolutional layer uses 64 3×3 kernels with a stride of 2 and padding of 1, followed by a batch normalization layer and a ReLU activation function. The third convolutional layer uses 128 3×3 kernels. The first convolutional layer has a stride of 2 and padding of 1, followed by a batch normalization layer and a ReLU activation function. The second convolutional layer uses 256 3×3 convolutional kernels with a stride of 2 and padding of 1, followed by a batch normalization layer and a ReLU activation function. The third convolutional layer uses 512 3×3 convolutional kernels with a stride of 2 and padding of 1, followed by a batch normalization layer and a ReLU activation function. Finally, a global average pooling layer and a fully connected layer are connected to map the features into a 128-dimensional pseudo-activity vector.

[0081] The discriminator adopts a multilayer perceptron structure, consisting of three fully connected layers. The first fully connected layer maps the input 128-dimensional vector to 64-dimensional vectors, followed by a LeakyReLU activation function and a random deactivation layer with a dropout rate of 0.3. The second fully connected layer maps the 64-dimensional vector to 32-dimensional vectors, followed by a LeakyReLU activation function and a random deactivation layer with a dropout rate of 0.3. The third fully connected layer maps the 32-dimensional vector to 1-dimensional vectors, followed by a Sigmoid activation function to output the difference measurement result.

[0082] The training process of the feature learning network includes: first, preparing a training dataset containing a large number of fused images and their corresponding real active vectors; then, inputting the fused images into a generator to obtain pseudo-active vectors; inputting the pseudo-active vectors and real active vectors into a discriminator to obtain a difference measurement result; and calculating the loss function of the generator and the loss function of the discriminator. The generator's loss function aims to make the pseudo-active vectors judged as real by the discriminator, while the discriminator's loss function aims to correctly distinguish between the pseudo-active vectors and the real active vectors.

[0083] The Adam optimizer is used to alternately update the parameters of the generator and discriminator. The learning rate of the generator is set to 0.0002, the learning rate of the discriminator is set to 0.0001, the batch size is set to 32, and the training epochs are set to 100. In each training epoch, the generator parameters are fixed and the discriminator parameters are updated first, and then the discriminator parameters are fixed and the generator parameters are updated. Through repeated iterative training, the generator can extract pseudo-active vectors from the fused image that are close to the distribution of the real active vectors. When the training reaches the preset number of epochs, the encoder part of the generator is extracted separately as a feature extraction network for subsequent processing.

[0084] It should be noted that the above structure is exemplary, and this application does not impose specific limitations on the internal structure design of the feature learning network. It can be set according to the actual situation.

[0085] The pseudo-activity vector refers to the vector data output by the generator after processing the fused image. This vector data simulates the form of the compound's activity vector. The true activity vector refers to the vector data pre-constructed based on the known molecular pathway activity data of the target compound. This vector data is used to characterize the interaction features between the target compound and aging-related pathways.

[0086] Senescence features refer to the abstract image features extracted from fused images after processing by a feature learning network. These features contain cellular senescence phenotype information associated with the mechanism of action of the target compound.

[0087] In this embodiment, step 103 includes the following process:

[0088] Step 1031: Input the fused image into the generator, and perform layer-by-layer abstraction and transformation on the fused image through the multi-layer coding structure in the generator to output a pseudo-active vector.

[0089] In step 1031, the multi-layer coding structure refers to the encoder part inside the generator, which consists of multiple sequentially connected neural network layers. This encoder part is used to perform layer-by-layer feature extraction and dimension transformation on the input image.

[0090] In this embodiment of the application, the fused image generated in step 102 is first fed into the generator part of the feature learning network as input data. The multi-layer coding structure inside the generator processes the input fused image layer by layer. Each layer performs convolution operation and non-linear activation on the image data, and gradually extracts local texture features, morphological structure features and global distribution features in the image. After multi-layer processing, a fixed-dimensional vector data is finally output, which is the pseudo-active vector.

[0091] In practical applications, a multi-layer coding structure consisting of 5 convolutional layers is set up. A fused image with a size of 256×256 pixels is input into this coding structure. The first convolutional layer uses 32 3×3 convolutional kernels to extract the edge and texture features of the image and outputs 32 feature maps. The second convolutional layer uses 64 3×3 convolutional kernels to further extract local structural features. Subsequent layers increase the number of feature maps and gradually reduce the spatial resolution. Finally, the extracted features are mapped into a 128-dimensional vector data through a fully connected layer. This 128-dimensional vector data is the pseudo-active vector.

[0092] Step 1032: Input the pseudo-activity vector and the real activity vector into the discriminator. In the discriminator, perform feature distribution difference measurement on the pseudo-activity vector and the real activity vector to obtain the difference measurement result.

[0093] In step 1032, the difference measurement result refers to the numerical result output after the feature distribution difference measurement, which reflects the degree of closeness between the pseudo-activity vector and the true activity vector.

[0094] In this embodiment of the application, the pseudo-active vector generated in step 1031 and the pre-constructed real active vector are simultaneously input into the discriminator part of the feature learning network. The discriminator processes the two input vectors through a multi-layer fully connected network and calculates the distribution difference of the two vectors in the high-dimensional feature space. The distribution difference includes the degree of deviation of the values ​​of each dimension of the vector and the similarity of the overall distribution shape of the vector. Finally, a difference measurement result is output.

[0095] In practical applications, a 128-dimensional pseudo-active vector and a 128-dimensional real active vector are input into a discriminator consisting of three fully connected layers. The first fully connected layer maps the two input vectors to a 64-dimensional latent space, the second fully connected layer further maps them to 32 dimensions, and the third fully connected layer outputs a scalar value as a difference measure. This scalar value ranges from 0 to 1. The closer the value is to 1, the smaller the difference in the distribution of the two vectors, and the closer the value is to 0, the greater the difference in the distribution of the two vectors.

[0096] Step 1033: Based on the difference measurement result, generate feature feedback information and pass the feature feedback information back to the generator. The generator iteratively adjusts the parameters of the multi-layer coding structure according to the feature feedback information.

[0097] In step 1033, the feature feedback information refers to the signal data generated based on the difference measurement results to guide the adjustment of generator parameters.

[0098] In this embodiment of the application, feature feedback information is generated based on the difference measurement result obtained in step 1032. The feature feedback information includes the discriminator's evaluation of the current generator output quality. The feature feedback information is passed back to the generator. The generator calculates the gradient direction of parameter adjustment based on the feedback information and updates the weights and bias parameters of each convolution kernel in its multi-layer coding structure. Through multiple iterations, the pseudo-active vector output by the generator gradually approaches the real active vector in feature distribution.

[0099] In practical applications, the number of iterations is set to 500. During each iteration, the discriminator outputs a difference measurement result. The generator calculates the gradient of the loss function with respect to the parameters of each layer based on the result, and updates the parameters using an optimization algorithm with a learning rate of 0.001. After 500 iterations, the difference measurement result of the pseudo-active vector and the real active vector output by the generator in the feature distribution reaches 0.92.

[0100] Step 1034: When the preset number of iterations is reached, the fused image is used to extract features using the adjusted multi-layer coding structure to obtain aging features.

[0101] In step 1034, the preset iteration count refers to the number of loops for adjusting the generator parameters as preset.

[0102] The embodiments of this application do not specifically limit the value of the preset number of iterations; it can be set according to the actual situation.

[0103] In this embodiment of the application, after the generator parameters have been adjusted to the preset number of iterations, the multi-layer coding structure after multiple iterations in step 1033 is fixed, the new fused image is input into the multi-layer coding structure, the fused image is encoded through the structure, and feature data containing cell senescence phenotype information related to the mechanism of action of the target compound is extracted from the image. This feature data is the senescence feature.

[0104] In practical applications, after reaching the preset 500 iterations, the adjusted 5-layer convolutional coding structure is used for feature extraction. A new fused image is input, and after processing by the coding structure, a 128-dimensional vector data is output. This 128-dimensional vector data is the aging feature.

[0105] This application enables a feature learning network to extract aging features that are deeply related to the mechanism of action of the target compound from the fused image through the interactive operation of the generator and discriminator, thus achieving an effective correlation between image phenotypic information and compound activity information.

[0106] Step 104: The aging features are concatenated with the real activity vector to form a joint feature vector. The joint feature vector is analyzed using a relational network to explore the correlation between different feature dimensions and generate a structured feature representation.

[0107] The joint feature vector refers to the vector data generated after feature concatenation, where each feature dimension in the vector data corresponds to a specific value in the aging feature or the true activity vector.

[0108] A relational network is a network structure used to mine the relationships between various feature dimensions in a vector. This network structure achieves feature dimension relationship mining by constructing a graph model and analyzing the relationships between nodes in the graph.

[0109] Structured feature representation refers to new feature data generated after processing by relational networks. This feature data not only contains the numerical information of the original feature dimensions, but also the correlation information between each feature dimension obtained through mining.

[0110] This application does not impose specific limitations on the type of relational network, the structural design of its internal structure, the parameter design, the training process, etc., and can set them accordingly based on the actual situation.

[0111] In this embodiment, such as Figure 2 As shown, step 104 includes the following process:

[0112] Step 1041: Input the joint feature vector into the relation network, and construct a virtual relation graph in the relation network. The nodes in the virtual relation graph correspond to different feature dimensions in the joint feature vector.

[0113] In step 1041, the virtual relationship graph refers to graph structure data constructed with feature dimensions as nodes. Each node in the graph structure represents a specific feature dimension in the joint feature vector. A node is the basic unit in the virtual relationship graph, and each node corresponds to a feature dimension and carries the numerical information of that feature dimension.

[0114] In this embodiment, the aging features obtained in step 1034 are first concatenated with the real activity vector used in step 1032. The two vectors are merged in dimensional order to form a joint feature vector with the number of dimensions equal to the sum of the two. Then, the joint feature vector is input into the relational network. The relational network treats each feature dimension in the joint feature vector as an independent node and assigns the corresponding value of the dimension to each node as the initial feature of the node, thereby constructing a virtual relational graph with the number of nodes equal to the dimension of the joint feature vector.

[0115] In practical applications, the 128-dimensional aging features are concatenated with the 128-dimensional real activity vector to obtain a 256-dimensional joint feature vector. The relational network uses the 256 feature dimensions of this 256-dimensional joint feature vector as 256 nodes, and the initial feature value of each node is the value corresponding to that dimension, thereby constructing a virtual relational graph containing 256 nodes.

[0116] Step 1042: Encode the virtual relation graph using a graph autoencoder to learn the low-dimensional latent representation of the nodes.

[0117] In step 1042, the graph autoencoder refers to a network model used for encoding learning of graph structure data. This model compresses the feature information of nodes in the graph into a low-dimensional space through the encoder, and then reconstructs the original graph structure through the decoder. In this application, the graph autoencoder is a structure independent of the relational network.

[0118] Low-dimensional latent representation refers to the node feature vector obtained after encoding by a graph autoencoder. This feature vector has a lower dimension while preserving the relationship information between nodes.

[0119] This application does not impose specific limitations on the type of graph autoencoder, the structural design of its internal structure, the parameter design, the training process, etc., and corresponding settings can be made according to the actual situation.

[0120] In this embodiment of the application, the virtual relationship graph constructed in step 1041 is input into the graph autoencoder. The encoder part of the graph autoencoder encodes the initial features of each node in the graph and the potential relationships between the nodes, maps the original features of each node to a lower-dimensional latent space, and generates a low-dimensional latent representation corresponding to each node. This low-dimensional latent representation condenses the important information of the original feature dimension and initially reflects the association possibility between nodes.

[0121] In practical applications, a virtual relationship graph containing 256 nodes is input into a graph autoencoder. The encoder of the graph autoencoder adopts a two-layer graph convolutional network structure. The first layer of the graph convolutional network maps the 256-dimensional initial features of each node to 128-dimensional latent features. The second layer of the graph convolutional network further maps the 128-dimensional latent features to a 64-dimensional low-dimensional latent representation. After processing by the encoder, a 64-dimensional low-dimensional latent representation of 256 nodes is obtained.

[0122] Step 1043: Based on the values ​​of each feature dimension in the joint feature vector and combined with the low-dimensional latent representation, calculate the association strength between any two nodes in the virtual relationship graph.

[0123] In step 1043, the association strength refers to a value used to measure the degree of association between any two nodes in the virtual relationship graph. The larger the value, the stronger the association between the feature dimensions corresponding to the two nodes.

[0124] In this embodiment of the application, for each pair of nodes in the virtual relationship graph, the association strength between the two nodes is calculated based on the numerical values ​​of the original feature dimensions of the two nodes and the low-dimensional latent representations of the two nodes obtained in step 1042, through a preset similarity calculation function. This calculation process takes into account both the similarity of the original feature values ​​and the similarity of the latent representations, so that the determination of the association strength depends on both the numerical characteristics of the features themselves and the learned latent relationships.

[0125] In practical applications, for nodes i and j, let their original eigenvalues ​​be respectively... and Their 64-dimensional low-dimensional latent representations are respectively and Using formula Calculate the association strength, where This indicates the strength of the association between node i and node j. This represents the balance coefficient, with a value of [value missing]. , This represents the original feature value of node i. This represents the original feature value of node j. This represents the 64-dimensional low-dimensional latent representation of node i. The 64-dimensional low-dimensional latent representation of node j is used to calculate the association strength between node i and node j using this formula. It is 0.76.

[0126] Step 1044: Based on the association strength, determine the edge connection weights between all nodes in the virtual relationship graph.

[0127] In step 1044, the edge connection weight refers to the weight value of the edge connecting two nodes in the virtual relation graph. This weight value is used to control the intensity of information transmission during the information propagation process of the graph attention network.

[0128] In this embodiment of the application, based on the association strength between each pair of nodes calculated in step 1043, positive edge connection weights are assigned to node pairs with association strength exceeding the set threshold through preset threshold filtering or normalization processing, and zero or small weights are assigned to node pairs with association strength below the set threshold, thereby determining the edge connection weight matrix between all nodes in the virtual relationship graph. This edge connection weight matrix will serve as the basis for subsequent graph attention network processing.

[0129] In practical applications, the association strength threshold is set to 0.5. The association strength between node i and node j calculated in step 1043 is 0.76, which is greater than 0.5. Therefore, it is determined that there is an edge connection between node i and node j, and the edge connection weight is directly adopted using the association strength value of 0.76. For node pairs with an association strength lower than 0.5, the edge connection weight is set to 0. After processing all node pairs, a 256×256 edge connection weight matrix is ​​obtained.

[0130] Step 1045: Using a graph attention network, based on the edge connection weights, adaptively perform weighted information propagation and aggregation operations on the feature dimension representations of nodes in the virtual relationship graph to update the feature dimension representation corresponding to each node.

[0131] This application does not impose specific limitations on the type of graph attention network, the structural design of its internal structure, the parameter design, the training process, etc., and can set them accordingly based on the actual situation.

[0132] Step 1045 may specifically include the following steps:

[0133] A1: In the graph attention network, each node in the virtual relationship graph is sequentially taken as a target node, and the following operations are performed for each target node:

[0134] In step A1, the target node refers to the node selected as the central node in the current processing step, which will aggregate information from its neighboring nodes to update its own feature representation.

[0135] In this embodiment, the graph attention network traverses 256 nodes in the virtual relationship graph, and sequentially uses each node as the target node for subsequent attention weight calculation and information aggregation operations.

[0136] A2: Based on the current feature representation of the target node and the current feature representation of each adjacent node, as well as the edge connection weights of the edges connecting the target node and the corresponding adjacent nodes, the attention weight distribution is calculated.

[0137] In step A2, neighboring nodes refer to other nodes in the virtual relationship graph that are connected to the target node by edges; attention weight distribution refers to the set of attention weight values ​​of the target node to all its neighboring nodes. This weight distribution reflects the degree of importance the target node attaches to different neighboring nodes when aggregating information.

[0138] In this embodiment of the application, for the currently selected target node, the current feature representation of the target node and the current feature representation of all adjacent nodes that are connected to the target node by edges are obtained. Combined with the edge connection weights that connect the target node and each adjacent node determined in step 1044, the attention weight of the target node to each adjacent node is calculated by the attention mechanism calculation formula. The attention weights of all adjacent nodes are normalized to obtain the attention weight distribution.

[0139] In practical applications, taking node i as the target node, its current features are represented as follows: The set of nodes adjacent to it is For adjacent node j, Its current feature is represented as The edge connection weight is The attention coefficient calculation formula is used. Calculate attention weights , where W is a learnable parameter matrix, and || denotes vector concatenation operation. Represents the parameter matrix transpose, The attention weight distribution is obtained by calculating the modified linear unit activation function with leakage over all neighboring nodes and then performing softmax normalization. .

[0140] A3: Based on the attention weight distribution, dynamically weight the current feature representations of all neighboring nodes of the target node.

[0141] In this embodiment of the application, based on the attention weight distribution calculated in step A2, the current feature representations of all neighboring nodes of the target node are weighted, and the feature representation of each neighboring node is multiplied by the corresponding attention weight to obtain the weighted feature representation of the neighboring nodes.

[0142] In practical applications, for a neighboring node j of node i, the current feature representation of node j is... The weighted feature representation is obtained. .

[0143] A4: Aggregate the current feature representations of all adjacent nodes after weighting to form the context-enhanced feature of the target node.

[0144] In step A4, the context-enhanced feature refers to the feature vector obtained by aggregating neighbor node information, which reflects the context information provided by the neighbor nodes around the target node.

[0145] In this embodiment of the application, the weighted feature representations of all neighboring nodes obtained in step A3 are summed or averaged to aggregate the information of all neighboring nodes into a single context-enhanced feature vector. This context-enhanced feature vector condenses the information carried by the neighboring nodes around the target node.

[0146] A5: Fuse the current feature representation of the target node with the context enhancement features to generate the updated feature dimension representation of the target node.

[0147] In this embodiment of the application, the current feature representation of the target node is fused with the context-enhanced features generated in step A4. This can be done by adding the features and then applying a non-linear activation function, or by concatenating the features and then applying a linear transformation, to generate an updated feature dimension representation of the target node. This updated feature dimension representation includes both the node's original information and the context information of its neighboring nodes.

[0148] Step 1046: Combine the updated feature dimension representations of all nodes to form a structured feature representation.

[0149] In this embodiment of the application, the updated feature dimension representations of all 256 nodes generated in step A5 are arranged and combined according to the original order of the nodes in the virtual relationship graph to form a new feature vector with the same or transformed dimensions as the original joint feature vector. This new feature vector is the structured feature representation. Each feature dimension in the structured feature representation has been enhanced with neighbor node information and the relationship network has been recalibrated.

[0150] This application constructs a virtual relationship graph and sequentially applies a graph autoencoder for low-dimensional representation learning, association strength calculation, edge connection weight determination, and a graph attention network for weighted information propagation and aggregation. This enables in-depth mining and structured representation of the complex relationships between the various feature dimensions in the joint feature vector, allowing the final generated structured feature representation to more accurately reflect the intrinsic relationship pattern between aging features and the real activity vector.

[0151] Step 105: Input the structured feature representation into the phenotype classifier, which processes the structured feature representation and outputs an analysis map of cell senescence.

[0152] Among them, the phenotypic classifier refers to the classification network structure used to analyze and process the structured feature representation and output information on the cellular senescence state; the cellular senescence analysis map refers to the analysis results presented in the form of an image. This map contains the senescence probability value and senescence subtype category information of each analyzed cell, and is distinguished and displayed by visual markers such as different colors or symbols.

[0153] This application does not impose specific limitations on the type of phenotypic classifier, the structural design of its internal structure, the parameter design, the training process, etc., and can set them accordingly based on the actual situation.

[0154] In this embodiment, step 105 includes the following process:

[0155] Step 1051: The structured feature representation is nonlinearly transformed through at least one fully connected layer of the phenotypic classifier, and the transformed features are processed using a clustering algorithm to identify cell populations with similar features, so as to generate corresponding cell group labels.

[0156] In step 1051, the fully connected layer refers to a network layer in the phenotypic classifier that consists of multiple neurons with linear transformation and non-linear activation; the cell group label refers to the identification information of the group to which each cell belongs after clustering.

[0157] In this embodiment, the structured feature representation generated in step 1046 is first input into the phenotypic classifier. The structured feature representation is then transformed nonlinearly through at least one fully connected layer to map the original feature space to a feature space more suitable for classification. Then, a clustering algorithm is applied to the feature data obtained after the nonlinear transformation. The clustering algorithm automatically identifies cell groups with similar features based on the spatial distance between feature vectors and assigns a corresponding cell group label to each analyzed cell. The cell group label reflects the group category to which the cell belongs.

[0158] Step 1052: Using the probability generation module of the phenotypic classifier, a distribution vector containing the probabilities of different aging subtypes is generated for each cell based on the cell grouping label.

[0159] In step 1052, the probability generation module refers to the network part of the phenotypic classifier used to calculate the probability of each category; aging subtype refers to the different phenotypic categories presented during cell aging, such as drug-sensitive aging or stress-induced aging; the distribution vector refers to the vector data containing the probability value corresponding to each aging subtype.

[0160] In this embodiment of the application, the feature data after nonlinear transformation in step 1051 and the corresponding cell group labels are input to the probability generation module of the phenotype classifier. Based on the transformed features of each cell and combined with the overall distribution characteristics of its cell population, the probability generation module calculates the probability value of the cell belonging to each preset aging subtype category through the classification network. These probability values ​​are arranged in order of category to form a distribution vector, which is the distribution vector of the cell containing the probabilities of different aging subtypes.

[0161] Step 1053: Through the decision output layer of the phenotypic classifier, determine the senescence subtype category corresponding to each cell based on the distribution vector, calculate the confidence level of the senescence subtype category, and use the confidence level as the cell senescence probability.

[0162] In step 1053, the output layer refers to the network layer in the phenotypic classifier that ultimately outputs the classification result; the aging subtype category refers to the most likely aging subtype determined based on the distribution vector; the confidence level refers to the reliability of the cell belonging to the determined aging subtype category, usually taking the maximum probability value in the distribution vector as the confidence level; the cell aging probability refers to the confidence level of the cell being in an aging state, which is expressed in this application as the confidence level of the determined aging subtype category.

[0163] In this embodiment of the application, the distribution vector generated in step 1052 is input to the decision output layer of the phenotype classifier. The decision output layer identifies the aging subtype category corresponding to the highest probability value from the distribution vector, determines the aging subtype category as the aging subtype category of the cell, and uses the highest probability value as the confidence level of the aging subtype category. At the same time, the confidence level is used as the cell aging probability output of the cell.

[0164] Step 1054: The cell senescence probability and the senescence subtype category, combined with the corresponding cell position coordinate information in the original image, are input into the atlas synthesis module of the phenotype classifier. The atlas synthesis module maps cells with different senescence probabilities and different senescence subtype categories to a new image corresponding to the original image space according to their corresponding position coordinates, so as to generate an analysis atlas of cell senescence.

[0165] In step 1054, the position coordinate information in the original image refers to the spatial position coordinates of each cell in the initially acquired stained image or autofluorescence image; the map synthesis module refers to the module in the phenotypic classifier used to convert the analysis results into a visual image; the new image refers to a blank image with the same spatial size as the original image, used to display the analysis results.

[0166] In this embodiment, the cell senescence probability and senescence subtype category of each cell determined in step 1053, together with the position coordinate information of the cell recorded in the original image, are input into the atlas synthesis module of the phenotypic classifier. The atlas synthesis module first creates a blank image with the same spatial size as the original image as a new image, and then iterates through each cell, using different colors or symbols to label the senescence subtype category of the cell at the corresponding position in the new image according to its position coordinate information, and uses the intensity of the color or the size of the symbol to represent the cell senescence probability value of the cell. After all cells are labeled, the cell senescence analysis atlas is generated.

[0167] In this embodiment, after step 105, the method further includes the following steps:

[0168] B1: Using 3D reconstruction technology, based on the aging probability and aging subtype of different cells in the analysis map, and combined with multiple batches of cell images collected at continuous time points, a dynamic 3D model of the evolution of cell aging state in the time dimension is generated.

[0169] In step B1, the multiple batches of cell images acquired at consecutive time points refer to cell image data acquired multiple times at fixed time intervals under the same field of view; the dynamic three-dimensional model refers to a cell senescence state model presented in three-dimensional form that contains information on the evolution of time. Each time slice in the model corresponds to the cell senescence distribution state at a time point.

[0170] In this embodiment of the application, after generating the cell senescence analysis map, multiple batches of cell images are acquired at consecutive time points under the same field of view. Steps 101 to 105 are performed on the image at each time point to obtain the cell senescence analysis map at the corresponding time point. Then, using three-dimensional reconstruction technology, the analysis maps at each time point are used as different time slices in the three-dimensional model. These slices are stacked and interpolated in the time dimension according to the time order to generate a dynamic three-dimensional model that can reflect the evolution of cell senescence state over time.

[0171] B2: Input the dynamic three-dimensional model into the time series prediction model, use the time series prediction model to extrapolate the progression trend of cell senescence in the dynamic three-dimensional model, predict the cell senescence state at future time points, and convert the prediction results into a displayable prediction analysis map.

[0172] In step B2, the time series prediction model refers to a machine learning model used to analyze time series data and predict future trends. This application does not impose specific limitations on the model type, internal structure design, parameter design, training process, etc. of the time series prediction model, and corresponding settings can be made according to the actual situation.

[0173] Future time points refer to time points after the current time point where actual data collection has not yet taken place, while predictive analysis maps are images generated based on prediction results that show the cellular aging state at future time points.

[0174] In this embodiment, the dynamic three-dimensional model generated in step B1 is input into a pre-trained time series prediction model. Based on the existing time slice data in the dynamic three-dimensional model, the time series prediction model analyzes the patterns of cell senescence probability and senescence subtype categories over time, and deduces the distribution of cell senescence states that may occur at a specified future time point. Finally, the predicted cell senescence state data at the future time point is converted into image form in the same way as in step 1054 to generate a predictive analysis map. This predictive analysis map can be used to predict the senescence change trend of the cell population at future time points.

[0175] This application achieves precise mapping from features to cell senescence probability and senescence subtype categories by processing structured feature representations layer by layer through a phenotypic classifier, and generates a visualized analysis map of cell senescence by combining location coordinate information; furthermore, by combining three-dimensional reconstruction technology with time series prediction models, it achieves dynamic modeling and future trend prediction of the cell senescence process in the time dimension, providing a more comprehensive analysis tool for drug screening and aging research.

[0176] Figure 3 A schematic diagram of the structure of an image recognition-based automated cell senescence analysis system provided in this application embodiment is shown below. Figure 3 As shown, the system includes:

[0177] The acquisition module 31 is used to acquire staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites.

[0178] The fusion module 32 is used to perform pixel-level spectral fusion of the stained image and the autofluorescence image to generate a fused image.

[0179] Input module 33 is used to input the fused image into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound.

[0180] The splicing module 34 is used to splice the aging features with the real activity vector to form a joint feature vector, and to analyze the joint feature vector using a relational network to explore the correlation between different feature dimensions and generate a structured feature representation.

[0181] The output module 35 is used to input the structured feature representation into the phenotype classifier, which processes the structured feature representation and outputs an analysis map of cell senescence.

[0182] The image recognition-based automatic cell senescence analysis system of this application is used to implement the aforementioned image recognition-based automatic cell senescence analysis method. Therefore, the specific implementation of the image recognition-based automatic cell senescence analysis system can be found in the embodiment section of the image recognition-based automatic cell senescence analysis method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0183] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described image recognition-based automatic cell senescence analysis methods.

[0184] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described image recognition-based automatic cell senescence analysis methods.

[0185] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0186] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the image recognition-based automatic cell senescence analysis method.

[0187] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0188] The above provides a detailed description of an automatic cell senescence analysis method and system based on image recognition provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. An automated method for analyzing cellular senescence based on image recognition, characterized in that, include: Acquire staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites; The stained image and the autofluorescence image are spectrally fused at the pixel level to generate a fused image; The fused image is input into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound. The aging features are concatenated with the real activity vector to form a joint feature vector. The joint feature vector is then analyzed using a relational network to uncover the correlation between different feature dimensions and generate a structured feature representation. The structured feature representation is input into a phenotypic classifier, which processes the structured feature representation and outputs an analytical map of cell senescence.

2. The method according to claim 1, characterized in that, The step of analyzing the joint feature vector using a relational network to uncover the relationships between different feature dimensions and generate a structured feature representation includes: The joint feature vector is input into a relational network, in which a virtual relational graph is constructed, and the nodes in the virtual relational graph correspond to different feature dimensions in the joint feature vector; The virtual relation graph is encoded using a graph autoencoder to learn a low-dimensional latent representation of the nodes; Based on the values ​​of each feature dimension in the joint feature vector, and combined with the low-dimensional latent representation, the association strength between any two nodes in the virtual relationship graph is calculated. Based on the association strength, determine the edge connection weights between all nodes in the virtual relationship graph; A graph attention network is used to adaptively perform weighted information propagation and aggregation operations on the feature dimension representations of nodes in the virtual relationship graph based on the edge connection weights, so as to update the feature dimension representation corresponding to each node; The updated feature dimension representations of all nodes are combined to form a structured feature representation.

3. The method according to claim 2, characterized in that, The method employs a graph attention network, which adaptively performs weighted information propagation and aggregation operations on the feature dimension representations of nodes in the virtual relationship graph based on the edge connection weights, to update the feature dimension representation corresponding to each node, including: In the graph attention network, each node in the virtual relationship graph is sequentially taken as a target node, and the following operations are performed for each target node: Based on the current feature representation of the target node and the current feature representation of each neighboring node, as well as the edge connection weights of the edges connecting the target node and the corresponding neighboring nodes, the attention weight distribution is calculated. Based on the attention weight distribution, the current feature representations of all neighboring nodes of the target node are dynamically weighted; The current feature representations of all adjacent nodes after weighting are aggregated to form the context-enhanced feature of the target node; The current feature representation of the target node is fused with the context-enhanced features to generate the updated feature dimension representation of the target node.

4. The method according to claim 1, characterized in that, The step of inputting the fused image into a feature learning network for feature learning to obtain aging features includes: using a generator in the feature learning network to generate a pseudo-activity vector based on the fused image, and using a discriminator in the feature learning network to distinguish the pseudo-activity vector from the real activity vector, comprising: The fused image is input into the generator, and the fused image is abstracted and transformed layer by layer through the multi-layer coding structure in the generator to output a pseudo-active vector. The pseudo-activity vector and the real activity vector are input into the discriminator. In the discriminator, the feature distribution difference between the pseudo-activity vector and the real activity vector is measured to obtain the difference measurement result. Based on the difference measurement results, feature feedback information is generated and transmitted back to the generator. The generator then iteratively adjusts the parameters of the multi-layer coding structure according to the feature feedback information. When the preset number of iterations is reached, the fused image is used to extract features using the adjusted multi-layer coding structure to obtain aging features.

5. The method according to claim 1, characterized in that, The process of processing the structured feature representation by the phenotypic classifier to output an analytical map of cell senescence includes: The structured feature representation is nonlinearly transformed by at least one fully connected layer of the phenotypic classifier, and the transformed features are processed by a clustering algorithm to identify cell populations with similar features and generate corresponding cell group labels. The probability generation module of the phenotypic classifier generates a distribution vector containing the probabilities of different aging subtypes for each cell based on the cell grouping labels. The senescence subtype category corresponding to each cell is determined by the decision output layer of the phenotypic classifier based on the distribution vector, and the confidence level of the senescence subtype category is calculated. The confidence level is then used as the cell senescence probability. The cell senescence probability and the senescence subtype category, along with the corresponding cell position coordinates in the original image, are input into the atlas synthesis module of the phenotype classifier. The atlas synthesis module maps cells with different senescence probabilities and different senescence subtype categories to a new image corresponding to the original image space according to their corresponding position coordinates, so as to generate an analysis atlas of cell senescence.

6. The method according to claim 1, characterized in that, The step of performing pixel-level spectral fusion of the stained image and the autofluorescence image to generate a fused image includes: Determine the correspondence between the pixel value matrix of the stained image and the pixel value matrix of the autofluorescence image; Based on the correspondence, a combination coefficient is defined for the pixel value at each corresponding position; The combination coefficients are adjusted based on the correlation between color information of enzyme activity and fluorescence information of metabolites during cell senescence. Using the adjusted combination coefficients, the pixel values ​​at corresponding positions in the pixel value matrix of the stained image and the pixel value matrix of the autofluorescent image are combined and calculated to obtain a fused pixel value matrix. A fused image is generated based on the fused pixel value matrix.

7. The method according to claim 1, characterized in that, After outputting the analytical atlas of cellular senescence, it also includes: Using 3D reconstruction technology, based on the senescence probability and senescence subtype of different cells in the analysis map, and combined with multiple batches of cell images collected at continuous time points, a dynamic 3D model of the evolution of cell senescence state in the time dimension is generated. The dynamic three-dimensional model is input into the time series prediction model. The time series prediction model is used to extrapolate the progression trend of cell senescence in the dynamic three-dimensional model, predict the cell senescence state at future time points, and convert the prediction results into a displayable predictive analysis map.

8. An automated cell senescence analysis system based on image recognition, characterized in that, include: The acquisition module is used to acquire staining images and autofluorescence images of target cells, wherein the target cells are cells treated with the target compound, the staining images reflect aging-related enzyme activity, and the autofluorescence images reflect the distribution of aging-related metabolites; The fusion module is used to perform pixel-level spectral fusion of the stained image and the autofluorescence image to generate a fused image; The input module is used to input the fused image into a feature learning network for feature learning to obtain aging features. The generator in the feature learning network generates a pseudo-activity vector based on the fused image, and the discriminator in the feature learning network distinguishes the pseudo-activity vector from the real activity vector. The real activity vector is a vector pre-constructed based on the activity spectrum data of the target compound. The splicing module is used to splice the aging features with the real activity vector to form a joint feature vector. The joint feature vector is then analyzed using a relational network to uncover the correlation between different feature dimensions and generate a structured feature representation. The output module is used to input the structured feature representation into the phenotype classifier, which processes the structured feature representation and outputs an analysis map of cell senescence.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the image recognition-based automated cell senescence analysis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the image recognition-based automatic cell senescence analysis method as described in any one of claims 1 to 7.

Images