Stem cell-oriented culture environment real-time monitoring method and system

Abstract

The invention discloses a stem cell-oriented culture environment real-time monitoring method and system, and particularly relates to the field of stem cell-oriented culture environment real-time monitoring. Comprising the following steps: performing sub-pixel-level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on acquired original cell image data, extracting internal texture feature information of cells, and constructing a texture orientation distribution tensor data set; and performing principal component feature extraction, kernel principal component analysis and manifold embedding dimensionality reduction on the texture orientation distribution tensor data set, constructing a low-dimensional feature space maintaining a local structure, and generating a cell texture density distribution model based on kernel density estimation. The early recognition of the cell fate heterogeneity and the dynamic intervention of the culture process under the unmarked condition are realized through time sequence modeling and classification discrimination based on cell image texture evolution characteristics in combination with density analysis and environmental feedback regulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of real-time monitoring of stem cell-oriented culture environment, and more particularly, to a real-time monitoring method and system for stem cell-oriented culture environment. BACKGROUND

[0002] In the process of stem cell culture, the early differentiation cell population usually presents highly consistent macro-morphological features, including cell size, contour roundness and surface smoothness, showing a so-called "pseudo-synchronous" state. Although the appearance tends to be uniform, the cells have already undergone potential differentiation at the level of gene expression and subcellular organelle reconstruction, leading to implicit divergence in cell fate.

[0003] The current mainstream image classification-based method mainly relies on macro-morphological features or gray texture differences of microscopic images for class discrimination, lacking the ability to sensitively capture the evolution process of the microstructure inside the cells. Especially under the condition of real-time monitoring of live cells without labeling and intervention, the traditional image classification model is limited by resolution and feature extraction dimension, and it is difficult to distinguish the cell population with consistent surface morphology but different functional states, resulting in a systematic blind area in the discrimination of early differentiation fate of stem cells.

[0004] Therefore, the prior art cannot effectively identify the cell fate heterogeneity in the pseudo-synchronous population at the early stage of stem cell differentiation through image classification, which becomes a technical bottleneck restricting the monitoring and regulation of stem cell culture. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a real-time monitoring method and system for stem cell-oriented culture environment, which realizes early identification of cell fate heterogeneity and dynamic intervention in the culture process under the condition of no labeling by means of time series modeling and classification discrimination based on cell image texture evolution characteristics, combined with density analysis and environmental feedback regulation, to solve the problems proposed in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a real-time monitoring method for stem cell-oriented culture environment, comprising:

[0007] S1, by performing sub-pixel level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on the collected original cell image data, extracting the internal texture feature information of the cells, and constructing a texture orientation distribution tensor data set;

[0008] S2, performing principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction on the texture orientation distribution tensor data set, constructing a low-dimensional feature space that preserves local structure, and generating a cell texture density distribution model based on kernel density estimation;

[0009] S3, constructing a cell texture evolution trajectory dataset by organizing the cell texture density distribution model according to a time sequence, extracting time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generating texture evolution feature encoding, and completing classification and identification of cell fate states through an image classification model;

[0010] S4, screening potential fate divergence cell groups by cross analysis of the texture density distribution model based on local density estimation and entropy calculation of the cell fate state classification results, completing heterogeneity classification correction based on an iterative discrimination network, and outputting corrected cell classification labels;

[0011] S5, generating a cell fate divergence monitoring atlas by updating the corrected cell classification labels to a preset cell culture monitoring platform in real time, adjusting the culture environment based on early warning signals and synchronously updating the texture evolution feature encoding, iteratively optimizing the image classification model, and constructing a stem cell fate monitoring and regulation system combined with environmental feedback.

[0012] In a preferred embodiment, the S1 further comprises: collecting original cell images in the stem cell culture process through a microscopic imaging device to form a cell image dataset, performing sub-pixel level interpolation reconstruction on the cell image dataset to obtain a cell image matrix;

[0013] Performing local contrast enhancement processing on the cell image matrix to generate a cell internal texture enhancement image, performing multi-scale wavelet transform on the cell internal texture enhancement image to extract a cell multi-scale texture feature atlas, and performing local gradient direction analysis on the multi-scale texture feature atlas to obtain a texture orientation distribution tensor dataset;

[0014] In the sub-pixel level interpolation reconstruction process, a local gray neighborhood is constructed for each pixel point in the cell image dataset, a local gray distribution function is defined to calculate a first-order gradient matrix and a second-order gradient matrix, and a Taylor expansion model of local gray variation is established using gradient information; based on the Taylor expansion model, a Lagrange interpolation function is used to interpolate the sub-pixel position between the original pixel grid points to generate a sub-pixel interpolation grid.

[0015] In a preferred embodiment, the S1 further comprises: in the sub-pixel interpolation grid, a directional interpolation weight matrix is constructed by a gradient direction vector, the local response of the Lagrange interpolation function is adjusted based on the directional interpolation weight matrix, and the contour diffusion bias introduced by isotropic interpolation is corrected; the cell image matrix is reconstructed with the continuous gray prediction value of the interpolation node to form a sub-pixel level cell image matrix with improved resolution and continuous local gray variation;

[0016] The local contrast enhancement processing includes but is not limited to: constructing a local gray level histogram of a fixed window by a cell image matrix, calculating a local contrast stretching function based on the local gray level histogram, adjusting the gray value distribution of each pixel by a local adaptive histogram equalization method, combining a local gradient amplitude constraint to suppress high-frequency noise enhancement, and generating a cell internal texture enhancement image;

[0017] In the multi-scale wavelet transform process, a multi-scale decomposition hierarchy is constructed by performing a discrete wavelet decomposition operation on the cell internal texture enhancement image, an approximate sub-band coefficient matrix and horizontal, vertical and diagonal detail sub-band coefficient matrices are generated in each decomposition layer, a set of sub-band coefficient matrices is recursively updated based on an iterative decomposition strategy with increasing scale, and a multi-scale texture feature spectrum containing different spatial frequency information is formed.

[0018] In the local gradient direction analysis process, based on each detail sub-band coefficient matrix in the multi-scale texture feature spectrum, a gradient vector field of each coefficient matrix is calculated using a local finite difference method, a direction angle is solved for the gradient vector of each pixel point in the gradient vector field, and a local gradient direction matrix is generated; the gradient direction matrices of each scale are modeled by high-order tensor stacking, and a texture orientation distribution tensor dataset describing the direction distribution of the cell internal texture is constructed.

[0019] In a preferred embodiment, the S2 further comprises: obtaining a texture principal component feature vector set by principal component feature extraction of the texture orientation distribution tensor dataset, and performing kernel principal component analysis on the texture principal component feature vector set to form a high-dimensional texture feature space.

[0020] An adjacency graph is constructed based on the high-dimensional texture feature space, an adjacency relationship is defined based on the Euclidean distance between feature vectors, a weighted graph Laplacian matrix is generated, low-order feature vector groups are extracted by feature decomposition, and manifold embedding dimension reduction mapping from the high-dimensional feature space to the low-dimensional feature space is completed; in the low-dimensional feature space, the local density value of each sample point is calculated based on the kernel density estimation method, the sample point weight is reweighted combined with the density distribution function, and a cell texture density distribution model is constructed.

[0021] In the principal component feature extraction process, the covariance matrix decomposition of the texture orientation distribution tensor dataset is performed, the eigenvalues and corresponding eigenvectors are calculated, the first several principal component eigenvectors sorted by eigenvalue size are selected, and a texture principal component eigenvector set is constructed.

[0022] In the process of performing kernel principal component analysis, a kernel function mapping is constructed based on the texture principal component eigenvector set, an inner product matrix of the sample in the high-dimensional kernel feature space is calculated, eigenvalue decomposition is performed based on the kernel matrix after centering, the corresponding eigenvector group is extracted, and a high-dimensional texture feature space is formed.

[0023] In a preferred embodiment, the S3 further comprises: arranging the cell texture density distribution model in a time sequence, constructing a cell texture evolution trajectory dataset, performing time slicing processing on the cell texture evolution trajectory dataset, and generating a segmented texture trajectory feature set;

[0024] performing dynamic time warping on the segmented texture trajectory feature set to obtain a time-aligned texture trajectory vector set, inputting the time-aligned texture trajectory vector set into a deep attention feature encoding network to generate a texture evolution feature encoding, and inputting the texture evolution feature encoding into an image classification model to output a cell fate state classification result;

[0025] Wherein, when the time-aligned texture trajectory vector set is input into the deep attention feature encoding network, a three-dimensional time sequence feature tensor is constructed by the time-aligned texture trajectory vector set, the feature similarity of any two time step vectors in the tensor is defined as a weight matrix based on vector dot product, a multi-head mechanism is used to calculate the weight matrix in parallel to obtain a set of local self-attention weights, a weighted feature vector set is formed by performing weighted summation on each time step vector and the corresponding weight matrix, linear transformation and layer normalization operation are performed on the weighted feature vector set to construct a normalized feature representation tensor, an autoregressive structure across time steps is introduced based on the normalized feature representation tensor, an autoregressive parameter matrix is defined, and a linear combination of the current time step feature vector and the past time step feature vector is calculated to generate a cross-step autoregressive encoding representation;

[0026] Based on the normalized feature representation tensor, a multi-layer structure composed of a self-attention calculation layer and an autoregressive encoding layer is sequentially stacked, local feature correlation calculation and time sequence autoregressive mapping are performed on each layer input according to a fixed stacking order, feature representation is passed and updated layer by layer, and finally a texture evolution feature encoding tensor is generated;

[0027] When the texture evolution feature encoding tensor is input into the image classification model, an implicit space feature representation vector is obtained by linear mapping of the texture evolution feature encoding tensor; a multi-layer fully connected neural network is input based on the implicit space feature representation vector, a linear transformation matrix and a nonlinear activation function are defined for each layer, and layer-by-layer feature extraction and mapping are performed; based on the output feature vector of the last layer, a linear classification weight matrix is defined, and matrix multiplication operation is performed on the feature vector and the classification weight matrix to obtain a classification score vector; based on the element comparison operation of the classification score vector, the label corresponding to the score is selected, and the cell fate state classification result is output;

[0028] Wherein after the texture evolution feature coding tensor is generated, an image classification model is constructed based on the texture evolution feature coding tensor, the image classification model is formed by alternately arranging full connection mapping layers and nonlinear activation layers to form a feature transformation sequence, linear transformation of a feature vector is performed through the full connection mapping layer, and then nonlinear mapping is performed under the action of an activation function to extract a hidden space feature representation; layer-by-layer forward transmission is maintained in the feature transformation sequence until a final feature vector is output; a classification score vector is generated by performing matrix multiplication operation based on the final feature vector and a classification weight matrix, and a classification label of a cell fate state is determined by performing element comparison operation on the classification score vector.

[0029] In a preferred embodiment, the S4 further comprises: performing local density estimation on the cell fate state classification result, extracting a functional heterogeneity abnormal cell region, and generating a cell population heterogeneity entropy distribution map based on entropy value calculation of the functional heterogeneity abnormal cell region;

[0030] Cross-analyzing the cell population heterogeneity entropy distribution map and a texture density distribution model, screening a potential fate divergence cell population, inputting the potential fate divergence cell population into an iterative discriminant network, performing heterogeneity classification correction, and outputting a corrected cell classification label;

[0031] In the local density estimation, a density estimation function based on a kernel function is constructed in a low-dimensional feature space by using a feature vector set of the cell fate state classification result, each feature vector is selected as a center point, a local sample density value in the neighborhood of the center point is calculated by using a kernel density function, a cell local density distribution matrix is generated, and a functional heterogeneity abnormal cell region is determined according to a density value gradient change;

[0032] In the entropy value calculation, a probability distribution model is established by using each local density value in the functional heterogeneity abnormal cell region, a probability weight corresponding to each cell is calculated based on a density value normalization result, a discrete probability distribution function is constructed by using the probability weight, an entropy value accumulation operation is performed on the probability distribution function according to a discrete entropy formula, and a cell population heterogeneity entropy distribution map is generated.

[0033] In a preferred embodiment, the S4 further comprises: in the iterative discriminant network, a feature vector set of the potential fate divergence cell population is inputted, a class division is performed on the input feature vector based on an initial discriminant function, a classification error rate is calculated according to the division result, a discriminant function parameter is updated according to the classification error rate, the class division of the feature vector is performed again by using the updated discriminant function, the discriminant and parameter update process is repeated for a fixed number of iterations or according to an error convergence condition, and finally a cell classification label after heterogeneity classification correction is outputted.

[0034] In a preferred embodiment, the S5 further comprises: updating the corrected cell classification label to the cell culture monitoring platform in real time, forming a cell fate divergence monitoring atlas, setting a divergence threshold based on the cell fate divergence monitoring atlas, and generating a culture process early warning signal in real time; wherein the cell culture monitoring platform includes but is not limited to: through the integration of cell image data acquisition module, feature extraction module and classification result visualization module, real-time receiving and processing cell classification label and its corresponding time sequence information, constructing cell fate divergence monitoring atlas, providing continuous dynamic display of cell differentiation state and divergence risk assessment results, supporting state monitoring and abnormal early warning in the culture process, thereby forming a cell culture monitoring platform; the cell fate divergence monitoring atlas includes the corrected cell classification label, cell spatial position information, time sequence information, texture feature encoding, local divergence density distribution and divergence event record;

[0035] The early warning signal is fed back to the culture environment regulation system to adjust the culture parameters, intervene the abnormal differentiation trend, and update the texture evolution feature encoding library to the culture environment adjustment result, iteratively optimize the image classification model, and continuously execute the texture evolution feature encoding update and image classification model optimization to form the stem cell fate monitoring and regulation;

[0036] The culture environment regulation system refers to the environment control platform preset in the cell culture process, which usually integrates dynamic regulation devices of temperature, humidity, gas composition, nutrient medium concentration and fluid shear force, etc. for continuously adjusting the culture microenvironment. When adjusting the culture parameters to intervene the abnormal differentiation trend, the divergence trend is analyzed by receiving the early warning signal, the change direction of the culture conditions corresponding to the abnormal area is determined, and the temperature controller, gas flow module or nutrient medium supply system is adjusted according to the feedback result to change the local or overall culture environment, thereby inhibiting the expansion of abnormal differentiated cell population or inducing cells to return to the target fate state.

[0037] After adjusting the culture environment, the environmental parameter changes recorded during the adjustment process are synchronously associated with the cell images collected at the corresponding time, the new texture evolution feature encoding is extracted, and the texture evolution feature encoding library is updated; the image classification model is retrained based on the updated feature encoding, the classification weight parameters are iteratively optimized, the model is adapted to the cell state feature distribution under the culture environment change, the classification accuracy is continuously improved, and finally the stem cell fate monitoring and regulation system combined with environmental feedback and feature evolution is formed.

[0038] A real-time monitoring system for stem cell culture environment, comprising a feature extraction module, a spatial construction module, a state discrimination module, a divergence correction module, a monitoring and regulation module;

[0039] The feature extraction module is used for performing sub-pixel level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on the collected original cell image data, extracting internal texture feature information of cells, and constructing a texture orientation distribution tensor data set;

[0040] The space construction module is used for performing principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction on the texture orientation distribution tensor data set, constructing a low-dimensional feature space that maintains local structure, and generating a cell texture density distribution model based on kernel density estimation;

[0041] The state discrimination module is used for organizing the cell texture density distribution model according to a time sequence, constructing a cell texture evolution trajectory data set, extracting time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generating texture evolution feature encoding, and completing classification and discrimination of cell fate states through an image classification model;

[0042] The divergence correction module is used for performing local density estimation and entropy calculation on the cell fate state classification results, screening potential fate divergence cell groups through cross analysis combined with the texture density distribution model, completing heterogeneity classification correction based on an iterative discrimination network, and outputting corrected cell classification labels;

[0043] The monitoring and regulation module is used for updating the corrected cell classification labels to a preset cell culture monitoring platform in real time, generating a cell fate divergence monitoring atlas, adjusting a culture environment based on an early warning signal and synchronously updating the texture evolution feature encoding, iteratively optimizing the image classification model, and constructing a stem cell fate monitoring and regulation system combined with environmental feedback.

[0044] Technical effects and advantages of the present application:

[0045] 1. Through the texture feature extraction method based on sub-pixel level interpolation reconstruction, multi-scale wavelet transform and local gradient direction analysis, the internal microscopic structure changes of cells can be sensitively captured under the condition of no labeling, early identification of stem cell fate heterogeneity under the background of consistent surface morphology is realized, and the resolution and feature dimension limitations of traditional classification methods are broken through;

[0046] 2. Through the combination of principal component feature extraction and kernel principal component analysis, further supplemented by manifold embedding dimension reduction to construct a low-dimensional feature space, the local structure relationship of cell populations is maintained, the modeling ability for potential nonlinear evolution rules among cell texture features is enhanced, and a stable feature expression basis is provided for classification;

[0047] 3. Through the time sequence organization and dynamic time warping processing of the cell texture density distribution model, the cell texture evolution trajectory is constructed, combined with the deep attention mechanism and autoregressive encoding, the dynamic features of cell state evolution can be captured, and continuous monitoring and time sequence correlation analysis of stem cell fate changes can be realized.

[0048] 4. By cross analysis based on local density estimation and entropy calculation, potential fate divergence cell population is screened, and iterative discriminant network is combined for classification correction, so that the recognition accuracy of heterogeneous cell population is effectively improved, and the misjudgment and missed detection of traditional classification method on weak heterogeneity divergence are reduced. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The overall timing diagram of the stem cell culture environment real-time monitoring method of the application.

[0050] Figure 2 The timing diagram of the stem cell image feature extraction and trajectory modeling method of the application.

[0051] Figure 3 The timing diagram of the stem cell fate state classification and potential divergence screening method of the application.

[0052] Figure 4 The timing diagram of the cell fate divergence classification correction and monitoring update method of the application.

[0053] Figure 5 The timing diagram of the cell fate divergence monitoring and culture environment regulation method of the application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0055] Referring to the drawings in the specification Figures 1-5 An embodiment of the application is a real-time monitoring method for a stem cell culture environment, which comprises the following steps.

[0056] S1, sub-pixel interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis are performed on the collected original cell image data to extract internal texture feature information of the cells and construct a texture orientation distribution tensor data set;

[0057] S2, principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction are performed on the texture orientation distribution tensor data set to construct a low-dimensional feature space that preserves local structure, and a cell texture density distribution model is generated based on kernel density estimation;

[0058] S3, constructing a cell texture evolution trajectory dataset by organizing the cell texture density distribution model in a time sequence, extracting time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generating texture evolution feature encoding, and completing classification and identification of cell fate states through an image classification model;

[0059] S4, screening potential fate divergence cell groups through cross analysis of the texture density distribution model based on local density estimation and entropy calculation of the cell fate state classification results, completing heterogeneity classification correction based on an iterative discrimination network, and outputting corrected cell classification labels;

[0060] S5, generating a cell fate divergence monitoring atlas by updating the corrected cell classification labels to a preset cell culture monitoring platform in real time, adjusting the culture environment based on early warning signals and synchronously updating the texture evolution feature encoding, iteratively optimizing the image classification model, and constructing a stem cell fate monitoring and regulation system combined with environmental feedback.

[0061] The S1 further comprises: collecting original cell images in the stem cell culture process through a microscopic imaging device to form a cell image dataset, and performing sub-pixel level interpolation reconstruction on the cell image dataset to obtain a cell image matrix;

[0062] Performing local contrast enhancement processing on the cell image matrix to generate a cell internal texture enhancement image, performing multi-scale wavelet transform on the cell internal texture enhancement image to extract a cell multi-scale texture feature atlas, and performing local gradient direction analysis on the multi-scale texture feature atlas to obtain a texture orientation distribution tensor dataset;

[0063] In the sub-pixel level interpolation reconstruction process, a local gray neighborhood is constructed for each pixel point in the cell image dataset, a local gray distribution function is defined to calculate a first-order gradient matrix and a second-order gradient matrix, and a Taylor expansion model of local gray variation is established using gradient information; based on the Taylor expansion model, a Lagrange interpolation function is used to interpolate the sub-pixel position between the original pixel points to generate a sub-pixel interpolation grid.

[0064] The S1 further comprises: in the sub-pixel interpolation grid, a directional interpolation weight matrix is constructed by a gradient direction vector, the local response of the Lagrange interpolation function is adjusted based on the directional interpolation weight matrix, and the contour diffusion bias introduced by isotropic interpolation is corrected; the cell image matrix is reconstructed using the continuous gray prediction value of the interpolation node to form a sub-pixel level cell image matrix with improved resolution and continuous local gray variation;

[0065] The local contrast enhancement processing includes but is not limited to: constructing a local gray level histogram of a fixed window by a cell image matrix, calculating a local contrast stretching function based on the local gray level histogram, adjusting the gray value distribution of each pixel by a local adaptive histogram equalization method, combining a local gradient amplitude constraint to suppress high-frequency noise enhancement, and generating a cell internal texture enhancement image;

[0066] In the multi-scale wavelet transform process, a discrete wavelet decomposition operation is performed on the cell internal texture enhancement image to construct a multi-scale decomposition hierarchy, an approximate sub-band coefficient matrix and horizontal, vertical and diagonal detail sub-band coefficient matrices are generated in each decomposition layer, a sub-band coefficient matrix set is recursively updated based on an iterative decomposition strategy with increasing scale, and a multi-scale texture feature spectrum containing different spatial frequency information is formed.

[0067] In the local gradient direction analysis process, a gradient vector field of each coefficient matrix is calculated using a local finite difference method based on the detail sub-band coefficient matrices in the multi-scale texture feature spectrum, a direction angle is solved for the gradient vector of each pixel point in the gradient vector field, and a local gradient direction matrix is generated; and the gradient direction matrices of different scales are modeled by high-order tensor stacking to construct a texture orientation distribution tensor dataset describing the direction distribution of the cell internal texture.

[0068] The S2 further includes: performing principal component feature extraction on the texture orientation distribution tensor dataset to obtain a texture principal component feature vector set, and performing kernel principal component analysis on the texture principal component feature vector set to form a high-dimensional texture feature space.

[0069] An adjacency graph is constructed based on the high-dimensional texture feature space, an adjacency relationship is defined based on the Euclidean distance between feature vectors, a weighted graph Laplacian matrix is generated, a feature decomposition is performed to extract a low-order feature vector group, and a manifold embedding dimension reduction mapping from the high-dimensional feature space to the low-dimensional feature space is completed; in the low-dimensional feature space, the local density value of each sample point is calculated based on the kernel density estimation method, the sample point weight is reweighted combined with the density distribution function, and a cell texture density distribution model is constructed.

[0070] In the principal component feature extraction process, covariance matrix decomposition is performed on the texture orientation distribution tensor dataset, the eigenvalues and corresponding eigenvectors are calculated, the first several principal component eigenvectors sorted by eigenvalue size are selected, and a texture principal component eigenvector set is constructed.

[0071] In the kernel principal component analysis process, a kernel function mapping is constructed based on the texture principal component eigenvector set, an inner product matrix of the samples in the high-dimensional kernel feature space is calculated, eigenvalue decomposition is performed on the kernel matrix after centering, the corresponding eigenvector group is extracted, and a high-dimensional texture feature space is formed.

[0072] The S3 further comprises: arranging the cell texture density distribution model in a time sequence, constructing a cell texture evolution trajectory dataset, performing time slicing processing on the cell texture evolution trajectory dataset, and generating a segmented texture trajectory feature set;

[0073] The segmented texture trajectory feature set is subjected to dynamic time warping to obtain a time-aligned texture trajectory vector set, the time-aligned texture trajectory vector set is input into a deep attention feature encoding network to generate texture evolution feature encoding, the texture evolution feature encoding is input into an image classification model, and a cell fate state classification result is output;

[0074] When the time-aligned texture trajectory vector set is input into the deep attention feature encoding network, a three-dimensional time sequence feature tensor is constructed from the time-aligned texture trajectory vector set, the feature similarity of any two time step vectors in the tensor is defined as a weight matrix based on vector dot product, a multi-head mechanism is used to calculate the weight matrix in parallel to obtain a set of local self-attention weights, a weighted feature vector set is formed by performing weighted summation on each time step vector and the corresponding weight matrix, linear transformation and layer normalization operations are performed on the weighted feature vector set to construct a normalized feature representation tensor, an inter-time step self-recurrent structure is introduced based on the normalized feature representation tensor, a self-recurrent parameter matrix is defined, and a linear combination of the current time step feature vector and the past time step feature vector is calculated to generate a cross-step self-recurrent encoding representation;

[0075] Based on the normalized feature representation tensor, a multi-layer structure composed of a self-attention calculation layer and a self-recurrent encoding layer is sequentially stacked, local feature correlation calculation and time sequence self-recurrent mapping are performed on each layer input according to a fixed stacking order, feature representation is passed and updated layer by layer, and finally a texture evolution feature encoding tensor is generated;

[0076] When the texture evolution feature encoding tensor is input into the image classification model, a hidden space feature representation vector is obtained by performing linear mapping on the texture evolution feature encoding tensor; a multi-layer fully connected neural network is input based on the hidden space feature representation vector, a linear transformation matrix and a nonlinear activation function are defined for each layer, and layer-by-layer feature extraction and mapping are performed; based on the output feature vector of the last layer, a linear classification weight matrix is defined, and matrix multiplication operation is performed on the feature vector and the classification weight matrix to obtain a classification score vector; based on the element comparison operation of the classification score vector, the label corresponding to the score is selected, and the cell fate state classification result is output;

[0077] Wherein after the texture evolution feature coding tensor is generated, an image classification model is constructed based on the texture evolution feature coding tensor, the image classification model is formed by alternately arranging full connection mapping layers and nonlinear activation layers to form a feature transformation sequence, linear transformation of a feature vector is performed through the full connection mapping layer, and then nonlinear mapping is performed under the action of an activation function to extract a hidden space feature representation; layer-by-layer forward transmission is maintained in the feature transformation sequence until a final feature vector is output; a classification score vector is generated by performing matrix multiplication operation based on the final feature vector and a classification weight matrix, and a classification label of a cell fate state is determined by performing element comparison operation on the classification score vector.

[0078] The S4 further comprises: performing local density estimation on the cell fate state classification result, extracting a functional heterogeneity abnormal cell region, and generating a cell population heterogeneity entropy distribution map based on entropy calculation of the functional heterogeneity abnormal cell region;

[0079] The cell population heterogeneity entropy distribution map is cross-analyzed with a texture density distribution model to screen a potential fate divergence cell population, the potential fate divergence cell population is input into an iterative discriminant network for heterogeneity classification correction, and a corrected cell classification label is output;

[0080] In the local density estimation, a density estimation function based on a kernel function is constructed in a low-dimensional feature space by using a feature vector set of the cell fate state classification result, each feature vector is selected as a center point, a local sample density value in the neighborhood of the center point is calculated by using a kernel density function, a cell local density distribution matrix is generated, and a functional heterogeneity abnormal cell region is determined according to a density value gradient change;

[0081] In the entropy calculation, a probability distribution model is established by using each local density value in the functional heterogeneity abnormal cell region, a probability weight corresponding to each cell is calculated based on a density value normalization result, a discrete probability distribution function is constructed by using the probability weight, an entropy accumulation operation is performed on the probability distribution function according to a discrete entropy formula, and a cell population heterogeneity entropy distribution map is generated.

[0082] The S4 further comprises: in the iterative discriminant network, a feature vector set of the potential fate divergence cell population is input, a class division is performed on the input feature vector based on an initial discriminant function, a classification error rate is calculated according to the division result, a discriminant function parameter is updated according to the classification error rate, the class division of the feature vector is performed again by using the updated discriminant function, the discriminant and parameter update process is repeated for a fixed number of iterations or according to an error convergence condition, and finally a cell classification label after heterogeneity classification correction is output.

[0083] The S5 further comprises: updating the corrected cell classification label to a cell culture monitoring platform in real time to form a cell fate divergence monitoring map, setting a divergence threshold based on the cell fate divergence monitoring map, and generating a culture process early warning signal in real time; wherein the cell culture monitoring platform includes but is not limited to: through the integration of a cell image data acquisition module, a feature extraction module and a classification result visualization module, real-time reception and processing of cell classification labels and their corresponding time sequence information, construction of a cell fate divergence monitoring map, provision of continuous dynamic display of cell differentiation state and divergence risk assessment results, support for state monitoring and abnormal early warning during the culture process, thereby forming a cell culture monitoring platform; the cell fate divergence monitoring map includes the corrected cell classification label, cell spatial position information, time sequence information, texture feature encoding, local divergence density distribution and divergence event record;

[0084] The early warning signal is fed back to the culture environment regulation system to adjust the culture parameters, intervene the abnormal differentiation trend, and update the texture evolution feature encoding library synchronously to the adjustment result of the culture environment, and iteratively optimize the image classification model, through continuous execution of texture evolution feature encoding update and image classification model optimization, to form stem cell fate monitoring and regulation;

[0085] The culture environment regulation system refers to an environment control platform preset in the cell culture process, which is usually integrated with dynamic regulation devices of temperature, humidity, gas composition, nutrient base concentration and fluid shear force, etc. for continuously adjusting the culture microenvironment; when adjusting the culture parameters to intervene the abnormal differentiation trend, the divergence trend is analyzed by receiving the early warning signal, the change direction of the culture conditions corresponding to the abnormal area is determined, and the temperature controller, gas flow module or nutrient base supply system is adjusted according to the feedback result to change the local or overall culture environment, thereby inhibiting the expansion of abnormal differentiated cell population or inducing cells to return to the target fate state;

[0086] After adjusting the culture environment, the environmental parameter changes recorded during the adjustment process are synchronously associated with the cell images collected at the corresponding time, new texture evolution feature encodings are extracted, and the texture evolution feature encoding library is updated; the image classification model is retrained based on the updated feature encodings, the classification weight parameters are iteratively optimized, the model is adapted to the cell state feature distribution under the culture environment changes, the classification accuracy is continuously improved, and finally a stem cell fate monitoring and regulation system combined with environmental feedback and feature evolution is formed.

[0087] A real-time monitoring system for stem cell culture environment, comprising a feature extraction module, a spatial construction module, a state discrimination module, a divergence correction module, a monitoring and regulation module;

[0088] The feature extraction module is used for performing sub-pixel level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on the collected original cell image data, extracting cell internal texture feature information, and constructing a texture orientation distribution tensor data set;

[0089] The space construction module is used for performing principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction on the texture orientation distribution tensor data set, constructing a low-dimensional feature space that preserves local structure, and generating a cell texture density distribution model based on kernel density estimation;

[0090] The state discrimination module is used for organizing the cell texture density distribution model according to time sequence, constructing a cell texture evolution trajectory data set, extracting time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generating texture evolution feature encoding, and completing classification and discrimination of cell fate state through an image classification model;

[0091] The divergence correction module is used for performing local density estimation and entropy calculation on the cell fate state classification results, combining cross analysis of the texture density distribution model to screen potential fate divergence cell groups, completing heterogeneity classification correction based on an iterative discrimination network, and outputting corrected cell classification labels;

[0092] The monitoring and regulation module is used for updating the corrected cell classification labels to a preset cell culture monitoring platform in real time, generating a cell fate divergence monitoring atlas, adjusting the culture environment based on early warning signals and synchronously updating the texture evolution feature encoding, iteratively optimizing the image classification model, and constructing a stem cell fate monitoring and regulation system combined with environmental feedback.

[0093] Overall, the present scheme focuses on the real-time monitoring of the dynamic evolution and potential differentiation divergence of cell fate state during stem cell culture, combines microscopic imaging and image processing technology, time sequence feature encoding technology, classification and discrimination method, and culture environment feedback regulation logic, and constructs a whole process flow from data acquisition, feature extraction, feature encoding, fate state classification to environment adaptation regulation, aiming to realize continuous monitoring of cell fate change, high-sensitivity identification of potential divergence state and intelligent intervention of culture environment;

[0094] In the cell culture process, the cell state evolution usually presents as subtle changes in phenotype, which is difficult to be accurately identified by naked eye or macroscopic morphology under traditional scales; To solve this problem, the method first acquires original cell image data in the culture process periodically based on a microscopic imaging device; Considering the characteristics of weak local structure changes, low contrast, unclear texture details and the like in the stem cell differentiation process, a sub-pixel level interpolation reconstruction method is adopted to improve the image resolution, specifically by constructing a local gray neighborhood, defining a gray distribution function, combining a first-order and second-order gradient matrix, establishing a local Taylor expansion model, and using a Lagrange interpolation function to predict the gray scale at a sub-pixel scale, effectively making up for the information loss caused by traditional pixel-level interpolation in the contour transition zone; In order to further enhance the texture details, a local contrast enhancement method is adopted, a contrast stretching function is constructed based on a local gray histogram, a local adaptive histogram equalization technique is combined to adjust the pixel gray scale distribution, and high-frequency noise is suppressed by combining the gradient amplitude constraint to form a texture enhanced image inside the cell;

[0095] In order to capture the spatial frequency information of the internal structure of the cell at different scales, the texture enhanced image is subjected to multi-scale wavelet transform; By constructing a decomposition level structure, the approximate sub-band and detail sub-band coefficient matrices are extracted at each scale, and the texture features in the horizontal, vertical and diagonal directions are decomposed; On this basis, the local finite difference method is applied to calculate the gradient vector field of the detail sub-band coefficient matrix, and the gradient direction of each pixel point is solved, and then the texture orientation distribution tensor dataset is constructed by tensor stacking; In this way, the texture features extracted can effectively capture the change rule of the internal microstructure of the cell at different scales and directions, and have good rotation invariance and scale sensitivity;

[0096] For the texture orientation distribution tensor dataset, in order to reduce the feature dimension and highlight the main variation information, first, principal component feature extraction is performed, based on covariance matrix decomposition, the principal component feature vectors in the front of the characteristic value are selected to construct a texture principal component feature vector set; Further, the kernel principal component analysis method is adopted, the high-order nonlinear feature separation ability is improved by kernel function mapping, and a high-dimensional texture feature space is constructed; In order to avoid the decline of discriminant ability caused by sample sparseness in high-dimensional space, the adjacency relationship is defined based on the Euclidean distance between feature vectors, a weighted graph Laplacian matrix is constructed, feature decomposition is performed, low-order feature vector groups are extracted, and manifold embedding dimension reduction mapping from high-dimensional to low-dimensional is realized; By keeping the local structure relationship of the neighborhood in the dimension reduction way, the internal geometric structure between the texture features of the cell is effectively preserved, which adapts to the subsequent time series modeling requirements;

[0097] In the low-dimensional feature space, the local density value of each sample point is calculated based on the kernel density estimation method, and a cell texture density distribution model is constructed. In the density modeling process, by selecting a reasonable kernel function and bandwidth parameter, the smoothness and resolution of local density estimation are controlled to ensure that the density distribution can reflect the local clustering characteristics of the cell population and also consider the overall distribution continuity. In this way, the texture density distribution model established can serve as an important basis for subsequent trajectory modeling and dynamic evolution analysis.

[0098] Around the dynamic evolution characteristics of cell fate state, the cell texture evolution trajectory dataset is constructed based on the time series organization cell texture density distribution model. By time slicing, the long time series is divided into several time segments, and the dynamic time warping method is used to align different trajectories in the time scale and eliminate the influence of the difference in cell state change rate. Then, a three-dimensional time series feature tensor is constructed, the vector dot product similarity between time steps is defined, the multi-head self-attention mechanism is used to calculate the local weight set in parallel, and the layer normalization and cross-step autoregressive modeling are combined to generate texture evolution feature encoding that captures long-term dependencies. In order to ensure that the feature extraction has nonlinear expression ability and high-order semantic modeling ability, the texture evolution feature encoding is input into an image classification model composed of multiple layers of fully connected mapping and nonlinear activation alternately, and the hidden space features are extracted layer by layer. Finally, the fate state classification label is generated based on the operation result of the classification weight matrix and the feature vector.

[0099] Considering that there may be sub-population heterogeneity differences within the stem cell population, local density estimation and entropy calculation methods are introduced based on fate state classification. By constructing a kernel function density estimation function based on the feature vector set, the local sample density is calculated, and the probability weight distribution is formed after normalization. According to the discrete entropy formula, the entropy value of the probability distribution is accumulated to generate a cell population heterogeneity entropy distribution map. Cross analysis is performed on the entropy distribution map and the texture density distribution model to screen potential fate divergence cell populations. For the screened cell populations, an iterative discriminant network is used for heterogeneity classification correction. Specifically, the initial discriminant function is used for class division, the classification error rate is calculated, the discriminant function parameters are updated based on the error rate, and the discriminant and update process is repeated until the set iteration number or error convergence condition is reached, and the corrected cell classification label is output.

[0100] The corrected cell classification label is updated to the cell culture monitoring platform in real time, and the platform generates a cell fate divergence monitoring atlas by integrating cell image data acquisition, feature extraction, and classification result visualization functions; the monitoring atlas includes cell classification labels, spatial location information, time sequence information, texture feature encoding, local divergence density distribution, and divergence event records; based on the monitoring atlas, a divergence threshold is set to determine whether to generate an early warning signal in real time; if an early warning signal is generated, the signal is fed back to the culture environment regulation system; the environment regulation system dynamically analyzes the cell divergence trend based on the early warning signal, determines the adjustment direction of the culture conditions corresponding to the abnormal area, and adjusts the temperature, humidity, gas composition, and nutrient base concentration to precisely intervene in the abnormal differentiation trend, inhibit the expansion of abnormal cells, and induce the cell fate state to return to the intended target.

[0101] After adjusting the culture environment, the system synchronously records the changes in environmental parameters during the adjustment process, synchronously associates the changes in environmental parameters with cell images, and updates the texture evolution feature encoding library; based on the updated feature encoding, the image classification model is retrained, the classification weight parameters are updated, and the dynamic adaptability of the model is improved; by continuously performing feature encoding updates and classification model optimization, a closed-loop feedback system for cell fate monitoring and culture environment regulation is formed.

[0102] Overall, the present scheme effectively realizes early identification and intervention of cell fate divergence through deep coupling of texture feature extraction, manifold dimension reduction, time series modeling, classification discrimination, and environmental feedback regulation, breaks through the limitations of traditional monitoring based on macroscopic morphological indicators, establishes a real-time monitoring and regulation mechanism for stem cell culture environment, and has good application prospects and promotional value.

[0103] The above only describes preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for real-time monitoring of a stem cell oriented culture environment, characterized in that, The method comprises the following steps: S1, by performing sub-pixel level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on the collected original cell image data, extracting the internal texture feature information of the cells, and constructing a texture orientation distribution tensor data set; S2, performing principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction on the texture orientation distribution tensor data set, constructing a low-dimensional feature space that preserves local structure, and generating a cell texture density distribution model based on kernel density estimation; S3, by organizing the cell texture density distribution model according to the time sequence, constructing a cell texture evolution trajectory data set, extracting time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generating texture evolution feature encoding, and completing the classification and identification of cell fate state through an image classification model; S4, by performing local density estimation and entropy calculation on the cell fate state classification results, combining cross analysis of the texture density distribution model to screen potential fate divergence cell groups, and based on an iterative discriminant network, completing heterogeneity classification correction and outputting the corrected cell classification label; S5, by updating the corrected cell classification label to the preset cell culture monitoring platform in real time, generating a cell fate divergence monitoring atlas, adjusting the culture environment based on the early warning signal and synchronously updating the texture evolution feature encoding, iteratively optimizing the image classification model, and constructing a stem cell fate monitoring and regulation system combined with environmental feedback.

2. The real-time monitoring method of the culture environment for stem cells according to claim 1, characterized in that: S1 further comprises: collecting original cell images in the stem cell culture process through a microscopic imaging device to form a cell image data set, and performing sub-pixel level interpolation reconstruction on the cell image data set to obtain a cell image matrix; performing local contrast enhancement processing on the cell image matrix to generate a cell internal texture enhancement image, performing multi-scale wavelet transform on the cell internal texture enhancement image to extract a multi-scale texture feature atlas, and performing local gradient direction analysis on the multi-scale texture feature atlas to obtain a texture orientation distribution tensor data set; in the sub-pixel level interpolation reconstruction process, a local gray neighborhood is constructed for each pixel point in the cell image data set, a local gray distribution function is defined to calculate a first-order gradient matrix and a second-order gradient matrix, and a Taylor expansion model of local gray variation is established using gradient information; based on the Taylor expansion model, a Lagrange interpolation function is used to perform sub-pixel position interpolation between original pixel points to generate a sub-pixel interpolation grid.

3. The real-time monitoring method of the culture environment for stem cells according to claim 2, characterized in that: S1 further comprises: in the sub-pixel interpolation grid, a directional interpolation weight matrix is constructed by a gradient direction vector, the local response of the Lagrange interpolation function is adjusted combined with the directional interpolation weight matrix, and the contour diffusion bias introduced by isotropic interpolation is corrected; the cell image matrix is reconstructed by the continuous gray prediction value of the interpolation node to form the cell image matrix. In the multi-scale wavelet transform process, the discrete wavelet decomposition operation is performed on the texture-enhanced image inside the cell to construct a multi-scale decomposition hierarchy, to generate an approximate sub-band coefficient matrix and horizontal, vertical and diagonal detail sub-band coefficient matrices in each decomposition layer, and to recursively update the sub-band coefficient matrix set based on an iterative decomposition strategy to form a multi-scale texture feature spectrum containing different spatial frequency information; In the local gradient direction analysis process, the gradient vector field of each coefficient matrix is calculated using the local finite difference method based on the detail sub-band coefficient matrices in the multi-scale texture feature spectrum, the direction angle of the gradient vector of each pixel point in the gradient vector field is solved, and a local gradient direction matrix is generated; and the gradient direction matrices of different scales are stacked in a tensor stacking manner to model high-order tensors, and a texture orientation distribution tensor dataset describing the texture direction distribution inside the cell is constructed.

4. The real-time monitoring method of the stem cell-oriented culture environment according to claim 3, characterized in that: S2 further comprises: performing principal component feature extraction on the texture orientation distribution tensor dataset to obtain a texture principal component feature vector set, and performing kernel principal component analysis on the texture principal component feature vector set to form a high-dimensional texture feature space; An adjacency graph is constructed based on the high-dimensional texture feature space, an adjacency relationship is defined based on the Euclidean distance between feature vectors, a weighted graph Laplacian matrix is generated, feature decomposition is performed to extract a low-order feature vector set, and manifold embedding dimension reduction mapping from the high-dimensional feature space to the low-dimensional feature space is completed; in the low-dimensional feature space, the local density value of each sample point is calculated based on the kernel density estimation method, the sample point weight is reweighted by combining the density distribution function, and a cell texture density distribution model is constructed; In the principal component feature extraction process, covariance matrix decomposition is performed on the texture orientation distribution tensor dataset, the eigenvalues and corresponding eigenvectors are calculated, the top several principal component eigenvectors sorted by eigenvalue size are selected, and a texture principal component feature vector set is constructed; In the kernel principal component analysis process, a kernel function mapping is constructed based on the texture principal component feature vector set, the inner product matrix of the samples in the high-dimensional kernel feature space is calculated, the eigenvalue decomposition is performed based on the centralized kernel matrix, the corresponding feature vector set is extracted, and the high-dimensional texture feature space is formed.

5. The real-time monitoring method of the stem cell-oriented culture environment according to claim 4, characterized in that: S3 further comprises: arranging the cell texture density distribution model in a time sequence to construct a cell texture evolution trajectory dataset, and performing time slicing on the cell texture evolution trajectory dataset to generate a segmented texture trajectory feature set; performing dynamic time warping on the segmented texture trajectory feature set to obtain a time-aligned texture trajectory vector set, inputting the time-aligned texture trajectory vector set into a deep attention feature encoding network to generate a texture evolution feature encoding, and inputting the texture evolution feature encoding into an image classification model to output a cell fate state classification result; In the input of the time-aligned texture trajectory vector set into the deep attention feature encoding network, the time-aligned texture trajectory vector set is constructed into a three-dimensional time sequence feature tensor, the feature similarity of any two time step vectors in the tensor is defined as a weight matrix based on vector dot product, a multi-head mechanism is used to calculate the weight matrix in parallel, and a set of local self-attention weight sets is obtained; by performing weighted summation on each time step vector and the corresponding weight matrix, a weighted feature vector set is formed, a linear transformation and layer normalization operation are performed on the weighted feature vector set, a normalized feature representation tensor is constructed, an autoregressive structure across time steps is introduced based on the normalized feature representation tensor, an autoregressive parameter matrix is defined, and a linear combination of the current time step feature vector and the past time step feature vector is calculated to generate an autoregressive encoding representation across step lengths; Based on the normalized feature representation tensor, a multi-layer structure composed of a self-attention calculation layer and an autoregressive encoding layer is sequentially stacked, local feature correlation calculation and time sequence autoregressive mapping are performed on each layer input according to a fixed stacking order, feature representation is passed and updated layer by layer, and finally a texture evolution feature encoding tensor is generated; In the input of the texture evolution feature encoding tensor into the image classification model, a hidden space feature representation vector is obtained by performing linear mapping on the texture evolution feature encoding tensor; based on the hidden space feature representation vector, a multi-layer fully connected neural network is input, a linear transformation matrix and a nonlinear activation function are defined for each layer, and layer-by-layer feature extraction and mapping are performed; based on the output feature vector of the last layer, a linear classification weight matrix is defined, and a matrix multiplication operation is performed on the feature vector and the classification weight matrix to obtain a classification score vector; based on the element comparison operation of the classification score vector, the label corresponding to the score is selected, and the cell fate state classification result is output. Wherein after the generation of the texture evolution feature encoding tensor, an image classification model is constructed based on the texture evolution feature encoding tensor, the image classification model is formed by alternately arranging fully connected mapping layers and nonlinear activation layers to form a feature transformation sequence, linear transformation of the feature vector is performed through the fully connected mapping layer, and then nonlinear mapping is performed under the action of the activation function to extract the hidden space feature representation; maintain layer-by-layer forward transmission in the feature transformation sequence until the final feature vector is output; based on the final feature vector and the classification weight matrix, a matrix multiplication operation is performed to generate a classification score vector, and by performing an element comparison operation on the classification score vector, the classification label of the cell fate state is determined.

6. The real-time monitoring method of the stem cell-oriented culture environment according to claim 5, characterized in that: The S4 further comprises: performing local density estimation on the cell fate state classification result, extracting a functional heterogeneity abnormal cell region, and generating a cell population heterogeneity entropy distribution map based on entropy calculation of the functional heterogeneity abnormal cell region; cross-analyzing the cell population heterogeneity entropy distribution map and the texture density distribution model to screen a potential fate divergence cell population, inputting the potential fate divergence cell population into an iterative discriminant network for heterogeneity classification correction, and outputting a corrected cell classification label. In local density estimation, a kernel-based density estimation function is constructed in a low-dimensional feature space by classifying the feature vector set of the cell fate state classification result, each feature vector is selected as a center point, and the local sample density value in its neighborhood is calculated using the kernel density function to generate a cell local density distribution matrix, and the abnormal cell region of functional heterogeneity is determined by the gradient change of the density value; In entropy value calculation, the probability distribution of each local density value in the functional heterogeneity abnormal cell region is modeled, the probability weight corresponding to each cell is calculated based on the density value normalization result, the discrete probability distribution function is constructed using the probability weight, and the entropy value accumulation operation is performed on the probability distribution function according to the discrete entropy formula to generate the cell population heterogeneity entropy distribution graph.

7. The real-time monitoring method of the stem cell-oriented culture environment according to claim 6, characterized in that: The S4 further comprises: in the iterative discriminant network, the feature vector set of the latent fate divergence cell population is taken as input, the input feature vector is classified based on the initial discriminant function, then the classification error rate is calculated according to the classification result, the discriminant function parameters are updated according to the classification error rate, the classification of the feature vector is re-executed using the updated discriminant function, and the discriminant and parameter updating process is repeated according to the fixed iteration number or error convergence condition, and finally the cell classification label after heterogeneity classification correction is output.

8. The real-time monitoring method of the stem cell-oriented culture environment according to claim 7, characterized in that: The S5 further comprises: the corrected cell classification label is updated to the cell culture monitoring platform in real time to form a cell fate divergence monitoring graph, a divergence threshold is set based on the cell fate divergence monitoring graph, and a culture process warning signal is generated in real time; the cell fate divergence monitoring graph includes the corrected cell classification label, cell spatial position information, time sequence information, texture feature encoding, local divergence density distribution and divergence event record; The warning signal is fed back to the culture environment regulation system to adjust the culture parameters, intervene the abnormal differentiation trend, and update the texture evolution feature encoding library synchronously to the culture environment adjustment result, iteratively optimize the image classification model, and form the stem cell fate monitoring and regulation by continuously executing the texture evolution feature encoding update and the image classification model optimization.

9. A real-time monitoring system for a stem cell-oriented culture environment, comprising a feature extraction module, a space construction module, a state discriminant module, a divergence correction module, and a monitoring and regulation module, characterized in that: The feature extraction module is used to perform sub-pixel level interpolation reconstruction, local contrast enhancement processing, multi-scale wavelet transform and local gradient direction analysis on the collected original cell image data, extract the texture feature information inside the cell, and construct a texture orientation distribution tensor data set; The space construction module is used to perform principal component feature extraction, kernel principal component analysis and manifold embedding dimension reduction on the texture orientation distribution tensor data set, construct a low-dimensional feature space that preserves local structure, and generate a cell texture density distribution model based on kernel density estimation; The state discrimination module is used to organize the cell texture density distribution model according to a time sequence, construct a cell texture evolution trajectory data set, extract time-aligned texture trajectory features based on time slicing and dynamic time warping processing, generate texture evolution feature encoding, and complete the classification and discrimination of cell fate states through an image classification model. The divergence correction module is used to perform local density estimation and entropy calculation on the cell fate state classification results, cross-analyze and screen potential fate divergence cell groups in combination with the texture density distribution model, complete heterogeneity classification correction based on an iterative discrimination network, and output corrected cell classification labels. The monitoring and regulation module is used to update the corrected cell classification labels to a preset cell culture monitoring platform in real time, generate a cell fate divergence monitoring atlas, adjust the culture environment based on early warning signals and update the texture evolution feature encoding synchronously, iteratively optimize the image classification model, and construct a stem cell fate monitoring and regulation system combined with environmental feedback.

Images