A method and system for extracting features from stem cell microscopic images

Abstract

The present application relates to the field of image processing technology and discloses a method and system for extracting features from stem cell microscopic images. The method comprises: segmenting a time-lapse 3D confocal microscopic image of a stem cell using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image; performing feature extraction processing on a marker fluorescence image based on the mask image to obtain a radial distribution feature vector and a pluripotency maintenance index; performing temporal quantization processing on a state change sequence based on the index to obtain differentiation trajectory feature data; performing spatial feature extraction processing on a subcellular organelle region using a topological map construction algorithm to obtain a spatial correlation network map; and performing feature fusion processing using a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result. The present application improves the accuracy of stem cell state recognition and the reliability of differentiation trajectory prediction.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method and system for extracting features from stem cell microscopic images. Background Art

[0002] Existing methods for extracting stem cell microscopic image features are primarily based on traditional image processing algorithms and machine learning techniques. They assess cell status by analyzing static image information such as stem cell morphology, fluorescence labeling intensity, and cell boundary outlines. These methods typically use image segmentation techniques such as threshold segmentation, edge detection, and region growing to separate stem cells from the background. They then extract geometric and statistical parameters such as cell area, perimeter, circularity, and texture features. Finally, classifiers such as support vector machines, random forests, or neural networks are used to identify cell types and assess their quality. Existing feature extraction methods have achieved some success in processing single-time-point stem cell images, enabling basic cell segmentation and morphological analysis.

[0003] However, existing technologies have significant deficiencies, mainly manifested in the lack of the ability to accurately identify differences in chromatin structure within stem cell nuclei. Traditional fixed threshold segmentation methods have difficulty accurately distinguishing subnuclear structures such as heterochromatin, euchromatin, and nucleoli, resulting in low accuracy in nuclear-cytoplasmic separation. At the same time, existing methods mainly focus on the overall morphological characteristics of cells, ignoring the spatial distribution patterns and temporal expression changes of stem cell differentiation markers, and are unable to effectively capture the dynamic evolution of the pluripotency state of stem cells. In addition, existing technologies lack in-depth analysis of the spatial arrangement characteristics of subcellular organelles, fail to establish a correlation between organelle distribution patterns and the functional state of stem cells, and have limited capabilities in multidimensional feature fusion and temporal analysis.

[0004] Based on the analysis of the above technical defects, the lack of an adaptive segmentation algorithm for chromatin texture features makes it impossible to accurately identify nuclear substructures; secondly, there is the problem of extracting spatial-temporal features of stem cell differentiation markers. Existing methods cannot simultaneously capture the radial distribution pattern and time evolution trajectory of the markers; this in turn leads to the problem of topological analysis of subcellular organelle spatial association networks, and there is a lack of systematic methods for quantifying the spatial arrangement features of organelles; ultimately, it leads to the problem of effective fusion of multi-dimensional heterogeneous features. Existing technologies find it difficult to organically integrate temporal dynamic features, spatial topological features, and biomarker expression features, which affects the accuracy of stem cell state identification and the reliability of differentiation trend prediction. Summary of the Invention

[0005] The present application provides a method and system for extracting features from stem cell microscopic images, which solves the problems of accurate segmentation of nuclear and cytoplasmic structures in stem cell microscopic images, coordinated extraction of temporal and spatial features of pluripotency markers, subcellular organelle topological association analysis, and multidimensional feature fusion classification, significantly improving the accuracy of stem cell state identification and the reliability of differentiation trajectory prediction.

[0006] In a first aspect, the present application provides a method for extracting features from a stem cell microscopic image, the method comprising:

[0007] The time-lapse 3D confocal microscopy images of stem cells were segmented using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image.

[0008] performing a multi-morphological feature collaborative extraction process on the stem cell differentiation marker fluorescence image according to the nuclear-cytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index;

[0009] Based on the pluripotency maintenance index, a time series feature quantification process is performed on the image sequence of stem cell state changes to obtain differentiation trajectory feature data;

[0010] The radial distribution feature vector is used to perform spatial feature extraction processing on the subcellular organelle image region through a topological map construction algorithm to obtain a subcellular organelle spatial association network map;

[0011] The differentiation trajectory feature data and the subcellular organelle spatial association network diagram are subjected to feature fusion processing by a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

[0012] Optionally, the segmentation process of the time-lapse 3D confocal microscopy image of the stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image includes:

[0013] performing 3×3 neighborhood gray-level co-occurrence matrix calculation on the time-lapse 3D confocal microscopy image of the stem cells to obtain contrast parameters, entropy parameters, and correlation parameters;

[0014] A first texture complexity threshold is set to 0.3 based on the contrast parameter, and a second texture complexity threshold is set to 2.8 based on the entropy parameter to obtain a chromatin region discrimination criterion;

[0015] Performing nucleolus candidate region marking processing on the image pixels according to the chromatin region discrimination standard to obtain an initial nucleolus segmentation region;

[0016] The initial nucleolus segmentation region is processed by morphological opening operation to remove the noise region with an area less than 50 pixels to obtain a purified nucleolus region;

[0017] Connected domain analysis is performed based on the purified nucleolus region to obtain the nucleocytoplasmic separation mask image.

[0018] Optionally, performing multi-morphological feature collaborative extraction processing on the stem cell differentiation marker fluorescence image according to the nuclear-cytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index includes:

[0019] Based on the nucleocytoplasmic separation mask image, the coordinates of the cell nucleus center of gravity are determined as the radial analysis origin, and the cell nucleus region is divided into 5 concentric rings at 20% of the equivalent radius to obtain the intranuclear radial partition structure;

[0020] Performing statistical analysis on the fluorescence intensity of the differentiation marker in each ring in the intranuclear radial partition structure to obtain a mean fluorescence intensity value, a standard deviation value, and a peak density value;

[0021] The cell cytoplasm region is divided into 36 sector regions at 10-degree intervals with the cell nucleus center of gravity as the pole, to obtain a cytoplasm polar coordinate partition structure;

[0022] The radial distribution feature vector is obtained by calculating the fluorescence intensity integral value, gradient change rate and spatial clustering coefficient of the marker expression in each sector region based on the cytoplasmic polar coordinate partition structure;

[0023] The pluripotency maintenance index was obtained by performing weighted sum calculation based on the expression intensities of the three differentiation markers at weight ratios of 0.4, 0.35, and 0.25.

[0024] Optionally, the step of performing time series feature quantification processing on the stem cell state change image sequence based on the pluripotency maintenance index to obtain differentiation trajectory feature data includes:

[0025] According to the numerical range of the pluripotency maintenance index, the stem cell state is divided into four biological states: pluripotency maintenance state, early differentiation initiation state, differentiation progress state and terminal differentiation state, to obtain a cell state classification standard;

[0026] Based on the cell state classification standard, the state transition frequency statistics of the pluripotency maintenance index changes at consecutive time points are processed to obtain a state transition frequency matrix;

[0027] The state transition frequency matrix is ​​subjected to normalized probability calculation processing according to a 2-hour time step to obtain state transition probability data;

[0028] The change rate of the pluripotency maintenance index between adjacent time points was calculated and analyzed to obtain the differentiation rate index and differentiation direction discrimination parameter;

[0029] A time series trajectory modeling process is performed based on the state transition probability data and the differentiation direction discrimination parameter to obtain the differentiation trajectory feature data.

[0030] Optionally, performing spatial feature extraction processing on the subcellular organelle image region using the radial distribution feature vector through a topological map construction algorithm to obtain a subcellular organelle spatial association network map includes:

[0031] Performing centroid coordinate extraction processing on the fluorescent labeled images of mitochondria, endoplasmic reticulum, and Golgi apparatus based on the radial distribution feature vector to obtain a set of spatial position coordinates of each subcellular organelle;

[0032] The spatial position coordinate set is used to establish an intracellular coordinate system with the center of gravity of the cell nucleus as the reference origin, and the Euclidean distance between each subcellular organelle is calculated to obtain an inter-subcellular organelle distance matrix;

[0033] According to the inter-subcellular organelle distance matrix, 5 microns is set as the adjacency determination threshold, and a connection edge relationship is established for the subcellular organelles whose distance is less than the adjacency determination threshold to obtain topological connection relationship data;

[0034] Based on the topological connection relationship data, an undirected graph structure is constructed with the subcellular organelle centroids as vertices and the adjacency relationships as edges to obtain an initial topological network graph;

[0035] The graph theory parameters of clustering coefficient, connectivity distribution and path length are calculated and processed on the initial topological network graph to obtain the subcellular organelle spatial association network graph.

[0036] Optionally, performing graph theory parameter calculation processing on the initial topological network graph, such as clustering coefficient, connectivity distribution, and path length, to obtain the subcellular organelle spatial association network graph includes:

[0037] Performing triangle connection relationship statistics processing on adjacent vertex pairs of each vertex in the initial topological network graph to obtain the number of triangles and the number of connection triplets;

[0038] Based on the number of triangles and the number of connection triplets, a clustering coefficient is calculated according to a ratio of three times the number of triangles to the number of connection triplets to obtain a network clustering coefficient parameter;

[0039] Performing statistical analysis on the number of connected edges of each vertex in the initial topological network graph to obtain the connectivity value and connectivity distribution frequency of each vertex;

[0040] Performing probability density calculation on the connectivity distribution frequency according to the degree interval to obtain connectivity distribution parameters;

[0041] Based on the graph theory shortest path algorithm, the shortest connection path length between any two vertices in the initial topological network graph is traversed and calculated to obtain an average path length parameter and the subcellular organelle spatial association network graph.

[0042] Optionally, the step of performing feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram using a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result includes:

[0043] Performing feature vector concatenation processing on the differentiation trajectory feature data and the graph theory parameters of the subcellular organelle spatial association network diagram to obtain a 32-dimensional input feature vector;

[0044] Based on the 32-dimensional input feature vector, a three-layer feedforward neural network structure is constructed, with 32 neurons in the input layer, 16 neurons in the hidden layer, and 4 neurons in the output layer, to obtain a multi-layer perceptron network architecture;

[0045] Performing nonlinear transformation processing on hidden layer neurons of the multilayer perceptron network architecture using a rectified linear unit activation function to obtain hidden layer feature representation;

[0046] The hidden layer feature representation is subjected to an iterative update process of network weights and bias parameters through a back-propagation algorithm to obtain trained and optimized network parameters;

[0047] Based on the network parameters optimized by the training, the output layer is subjected to classification calculation processing of the pluripotency maintenance degree, differentiation tendency score, proliferation activity index and cell cycle stage to obtain the stem cell image feature classification result.

[0048] In a second aspect, the present application provides a stem cell microscopic image feature extraction system, the stem cell microscopic image feature extraction system comprising:

[0049] A segmentation module is used to segment the time-lapse 3D confocal microscopy images of stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image;

[0050] an extraction module, configured to perform a multi-morphological feature collaborative extraction process on the stem cell differentiation marker fluorescence image based on the nucleocytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index;

[0051] a quantification module, configured to perform time series feature quantification processing on a sequence of stem cell state change images based on the pluripotency maintenance index to obtain differentiation trajectory feature data;

[0052] A topology module is used to perform spatial feature extraction processing on the subcellular organelle image region using the radial distribution feature vector through a topology map construction algorithm to obtain a subcellular organelle spatial association network map;

[0053] A fusion module is used to perform feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram through a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

[0054] In a third aspect, a stem cell microscopic image feature extraction device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the stem cell microscopic image feature extraction device executes the above-mentioned stem cell microscopic image feature extraction method.

[0055] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on the computer, the computer executes the above-mentioned stem cell microscopic image feature extraction method.

[0056] In the technical solution provided by this application, a chromatin texture-sensitive adaptive threshold segmentation algorithm is used to accurately segment the time-lapse 3D confocal microscopy images of stem cells, effectively solving the problem of insufficient segmentation accuracy of the traditional fixed threshold method when processing the complex structures of heterochromatin and euchromatin in the stem cell nucleus. The algorithm is based on the texture parameters of the 3×3 neighborhood grayscale co-occurrence matrix and can accurately identify the differentiated distribution characteristics of subnuclear structures such as nucleoli and heterochromatin, significantly improving the accuracy of nuclear cytoplasm separation. The collaborative extraction of multi-morphological features achieves accurate quantification of the spatial expression pattern of stem cell differentiation markers through the establishment of radial distribution patterns and polar coordinate partitioning structures, overcoming the limitations of existing technologies that only focus on overall fluorescence intensity while ignoring spatial distribution information. The calculation of the pluripotency maintenance index provides a standardized quantitative indicator for stem cell status assessment. The temporal feature quantification processing dynamically models the stem cell differentiation trajectory based on the Markov chain state transition model, breaking through the technical bottleneck of the traditional method based on single time point analysis. The differentiation trajectory feature data can predict the future state evolution trend of stem cells, providing a theoretical basis for the precise control of the stem cell differentiation process. By establishing a spatial association network of subcellular organelles, the topological graph construction algorithm achieved the first systematic quantitative analysis of the spatial arrangement patterns of key organelles, such as mitochondria, endoplasmic reticulum, and Golgi apparatus. Graph-theoretic parameters such as clustering coefficient, connectivity distribution, and average path length revealed the intrinsic correlation between organelle distribution and stem cell functional status. The multi-layer perceptron fusion algorithm organically integrated temporal dynamic features with spatial topological features, achieving nonlinear fusion of multidimensional heterogeneous data. The four classification indicators in the output layer provide a comprehensive evaluation system for stem cell quality control and differentiation status monitoring.

[0057] The texture complexity threshold used in the chromatin texture-sensitive algorithm directly corresponds to the low-complexity characteristics of heterochromatin. The weighting mechanism for the pluripotency maintenance index reflects the varying importance of Oct4, Nanog, and CD34 markers in stem cell biology. The weighting ratios are based on statistical analysis of extensive experimental data, ensuring the biological plausibility and clinical applicability of the index calculation. The application of a Markov chain state transition model fully accounts for the stochastic and probabilistic nature of stem cell differentiation. The model parameters are based on fundamental laws of stem cell biology. The 2-hour time step matches the timescale of changes in stem cell differentiation marker expression. The state transition probability matrix accurately reflects the transition patterns between different differentiation stages. The 5-micrometer adjacency threshold used in the topological graph construction algorithm is based on the distance between organelle membrane contacts and molecular exchange. This threshold ensures that the topological network accurately reflects the functional interactions between organelles. The calculation of graph-theoretic parameters provides a quantitative tool for understanding the internal spatial organization of stem cells. The design of the multi-layer perceptron network architecture takes into account the dimensional characteristics of stem cell feature data and the complexity of the classification task. The 32-16-4 network structure avoids overfitting while ensuring classification accuracy. The choice of the rectified linear unit activation function adapts to the distribution characteristics of stem cell feature data. The parameter setting of the backpropagation algorithm has been optimized through training on a large number of stem cell samples, ensuring the network's efficiency and accuracy in stem cell image feature classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 This is a schematic diagram of an embodiment of the method for extracting features from a stem cell microscopic image in an embodiment of the present application;

[0060] Figure 2 This is a schematic diagram of an embodiment of a stem cell microscopic image feature extraction system in an embodiment of the present application;

[0061] Figure 3 4 is a schematic block diagram of the structure of a stem cell microscopic image feature extraction device in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The embodiments of the present application provide a method and system for extracting features from stem cell microscopic images. The terms "first," "second," "third," "fourth," and so on (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0063] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, a method for extracting features from a stem cell microscopic image includes:

[0064] Step S101: Segmenting the time-lapse 3D confocal microscopy image of stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image;

[0065] Step S102: performing multi-morphological feature collaborative extraction processing on the stem cell differentiation marker fluorescence image according to the nuclear-cytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index;

[0066] Step S103: performing time series feature quantification processing on the stem cell state change image sequence based on the pluripotency maintenance index to obtain differentiation trajectory feature data;

[0067] Step S104: performing spatial feature extraction processing on the subcellular organelle image region using the radial distribution feature vector through a topological map construction algorithm to obtain a subcellular organelle spatial association network map;

[0068] Step S105 : performing feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram using a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

[0069] It is understandable that the execution subject of this application can be a stem cell microscopic image feature extraction system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking a server as the execution subject as an example.

[0070] Specifically, a 3×3 neighborhood gray-level co-occurrence matrix is ​​calculated for the input stem cell image. The gray-level co-occurrence matrix represents the spatial distribution of grayscale values ​​between adjacent pixels in the image and describes image texture features by counting the frequency of occurrence of pixel pairs with different grayscale levels. This matrix calculation generates three key texture descriptors: contrast, entropy, and correlation. The contrast parameter reflects the severity of local grayscale changes in the image, the entropy parameter measures the complexity of image information, and the correlation parameter describes the strength of linear relationships between adjacent pixels. The chromatin texture-sensitive algorithm sets a first texture complexity threshold of 0.3 and a second texture complexity threshold of 2.8 based on the distinct texture characteristics of heterochromatin and euchromatin in the stem cell nucleus. Pixels whose texture parameters meet these conditions are marked as candidate nucleoli. A morphological opening operation removes noisy regions smaller than 50 pixels by first eroding and then dilating. Connected domain analysis merges spatially adjacent pixels into the same region, ultimately generating an accurate nucleocytoplasmic separation mask image.

[0071] For collaborative polymorphic feature extraction based on the nuclear-cytoplasmic separation mask image, the algorithm determines the coordinates of the nucleus's centroid as the geometric center for radial analysis. The nucleus is then divided into five concentric rings at 20% of the equivalent radius, where the equivalent radius is the radius of a circle equal to the nucleus's area. Fluorescence intensities of differentiation markers such as Oct4, Nanog, and CD34 are calculated within each ring, and the mean fluorescence intensity, standard deviation, and peak density are calculated. The standard deviation reflects the dispersion of fluorescence intensity, and the peak density indicates the concentration of high-intensity pixels. The cytoplasm is partitioned using polar coordinates, with the nucleus's centroid as the pole. The cytoplasm is divided into 36 sectors at 10-degree angular intervals. For each sector, the fluorescence intensity integral, gradient change rate, and spatial clustering coefficient are calculated. The fluorescence intensity integral represents the cumulative sum of the fluorescence intensities of all pixels within the region, the gradient change rate describes the rate of spatial variation of fluorescence intensity, and the spatial clustering coefficient quantifies the degree of fluorescence signal aggregation. The radial distribution feature vector is composed of the characteristic parameters of all rings and fan-shaped regions. The pluripotency maintenance index is calculated by weighted summation of the expression intensities of the three markers at weight ratios of 0.4, 0.35, and 0.25.

[0072] Temporal feature quantification categorizes the pluripotency maintenance index into four biological states based on numerical ranges: the pluripotency maintenance state corresponds to an index greater than or equal to 0.7, the early differentiation initiation state corresponds to 0.4 to 0.7, the ongoing differentiation state corresponds to 0.2 to 0.4, and the terminal differentiation state corresponds to less than 0.2. The algorithm counts the frequency of state transitions between consecutive time points and constructs a 4×4 state transition frequency matrix, in which each element records the number of transitions from state i to state j. Normalization is performed using a 2-hour time step, and the state transition probability data is obtained by dividing the transition frequency by the total number of observations. The differentiation rate index is calculated by calculating the ratio of the difference in the pluripotency maintenance index between adjacent time points to the time interval. The differentiation direction discriminant parameter is calculated based on a weighted combination of the time derivatives of the expression intensities of three markers. The differentiation trajectory feature data contains temporal dynamic information such as state transition probabilities, differentiation rates, and direction discrimination. The topological map construction algorithm first extracts the centroid coordinates of the fluorescently labeled images of mitochondria, endoplasmic reticulum, and Golgi apparatus. The centroid coordinates are the weighted average position of all pixel coordinates within the region. An intracellular coordinate system was established with the center of mass of the cell nucleus as the reference origin, and the Euclidean distance between each subcellular organelle was calculated. The Euclidean distance represents the straight-line distance between two points. A threshold of 5 microns was set as the adjacency threshold. When the distance between two subcellular organelles was less than this threshold, a connection edge relationship was established. The topological connection relationship data recorded all subcellular organelle pairs that met the adjacency conditions, constructing an undirected graph structure with the subcellular organelle centroids as vertices and adjacency relationships as edges. The clustering coefficient was calculated by counting the proportion of triangles formed between adjacent vertices at each vertex. The connectivity distribution recorded the frequency distribution of the number of connecting edges to each vertex. The average path length was calculated using the shortest path algorithm to calculate the mean of the shortest connection distance between any two vertices.

[0073] The multi-layer perceptron fusion algorithm concatenates differentiation trajectory feature data and graph-theoretic parameters of the subcellular organelle spatial association network diagram into a 32-dimensional input feature vector. This feature vector contains parameters such as state transition probability, differentiation rate, clustering coefficient, and connectivity distribution. The three-layer feedforward neural network structure includes 32 neurons in the input layer, 16 neurons in the hidden layer, and 4 neurons in the output layer. Each neuron receives the output signals of all neurons in the previous layer. The hidden layer uses a rectified linear unit activation function for nonlinear transformation, which remains unchanged for positive values ​​and resets negative values ​​to zero. The backpropagation algorithm calculates the error between the predicted output and the true label, propagates the error signal forward layer by layer, and updates the network weights and bias parameters. The output layer generates four classification results: the degree of pluripotency maintenance, differentiation propensity score, proliferation activity index, and cell cycle stage, which correspond to different biological states and functional characteristics of stem cells.

[0074] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0075] The 3×3 neighborhood gray-level co-occurrence matrix of time-lapse 3D confocal microscopy images of stem cells was calculated to obtain contrast parameters, entropy parameters, and correlation parameters.

[0076] The first texture complexity threshold is set to 0.3 based on the contrast parameter, and the second texture complexity threshold is set to 2.8 based on the entropy parameter to obtain the chromatin region discrimination criteria;

[0077] According to the chromatin region discrimination criteria, the image pixels are marked with the candidate nucleolus regions to obtain the initial nucleolus segmentation region;

[0078] The initial nucleolus segmentation area is processed by morphological opening operation to remove the noise area with an area less than 50 pixels to obtain the purified nucleolus area;

[0079] Connected domain analysis is performed based on the purified nucleolus region to obtain a nucleocytoplasmic separation mask image.

[0080] Specifically, the 3×3 neighborhood gray-level co-occurrence matrix (GLCM) of time-lapse 3D confocal microscopy images of stem cells is calculated by scanning a 3×3 window region consisting of each pixel in the image and its eight neighboring pixels. The frequency of occurrence of pixel pairs with different gray levels at specific directions and distances is counted to establish a texture description matrix. The gray-level co-occurrence matrix (GLCM) is a square matrix whose rows and columns correspond to the grayscale levels of the image, and the matrix element values ​​represent the number of occurrences of pixel pairs with corresponding gray levels within a specified spatial relationship. The calculation process first determines the orientation angle of the pixel pairs, typically selecting four main directions: 0, 45, 90, and 135 degrees. The frequency of gray-level combinations of neighboring pixels in each direction is then counted. The contrast parameter is calculated by taking the weighted sum of the squares of the distances between each element in the GLCM and the diagonal line. It reflects the severity of local grayscale changes in the image; larger values ​​indicate stronger texture contrast. The entropy parameter calculates the information entropy based on the probability distribution of the elements in the GLCM, measuring the complexity and randomness of the texture. Higher entropy values ​​indicate complex and irregular texture structures. The correlation parameter describes the strength of the linear relationship between pixel grayscale values ​​in the row and column directions of the gray-level co-occurrence matrix. Values ​​close to 1 indicate strong positive correlation, values ​​close to -1 indicate strong negative correlation, and values ​​close to 0 indicate no linear relationship. The chromatin region discrimination criteria are based on the different textural properties of heterochromatin and euchromatin in stem cell nuclei. Heterochromatin typically exhibits high electron density and a relatively simple texture pattern, while euchromatin exhibits a more complex texture structure. The first texture complexity threshold of 0.3 is set based on the contrast parameter. When the contrast parameter of a pixel region is below this threshold, it indicates that the region has a relatively uniform grayscale distribution, consistent with the texture characteristics of heterochromatin. The second texture complexity threshold of 2.8 is determined based on the entropy parameter. When the entropy parameter is below this threshold, the region's texture information is relatively simple and orderly, also consistent with the characteristics of heterochromatin regions. The chromatin region discrimination criteria require that pixels simultaneously meet the dual conditions of a contrast parameter less than 0.3 and an entropy parameter less than 2.8. Only pixels that meet these two constraints are labeled as candidate nucleolar regions, thus eliminating confounding factors such as cytoplasm and intercellular spaces.

[0081] Nucleolus candidate region labeling involves scanning the entire stem cell image pixel by pixel. Texture parameters of a 3×3 neighborhood are calculated for each pixel and compared to the chromatin region discrimination criteria. Binarization is used in the labeling process, with pixels that meet the criteria assigned a value of 1 to represent nucleolus candidates and pixels that do not meet the criteria assigned a value of 0 to represent background pixels. The initial nucleolus segmentation consists of all pixels labeled 1, but the segmentation result contains a large number of isolated noisy pixels and discontinuous small regions. This noise primarily arises from optical interference during image acquisition, uneven staining, and other intracellular structures with similar texture features. The spatial distribution of the initial segmented regions is fragmented, requiring further morphological processing to remove noise and connect discontinuous regions.

[0082] Morphological opening is a composite morphological operation consisting of erosion followed by dilation. Erosion reduces the foreground area by performing a logical AND operation on the structuring element and the image, while dilation expands the foreground area by performing a logical OR operation on the structuring element and the image. Erosion eliminates noise points and fine connections smaller than the structuring element, while dilation restores the shape of over-eroded regions. Noise regions smaller than 50 pixels are completely eliminated by erosion, as these small regions are insufficient to withstand the erosion effects of the structuring element. After opening, the purified nucleus region exhibits a more regular shape, with internal voids filled and boundaries smoothed and continuous, while preserving the nucleus's primary morphological features and spatial location.

[0083] Connected domain analysis uses a region growing or flood fill algorithm to identify and label spatially adjacent sets of pixels. Starting with a seed pixel in the purified nucleolus region, the algorithm expands the search for adjacent foreground pixels along four or eight neighborhoods. The core of connected domain analysis is to establish adjacency relationships between pixels, typically using 4-connectivity or 8-connectivity criteria. 4-connectivity considers adjacent pixels in the four directions of up, down, left, and right, while 8-connectivity also includes adjacent pixels in the four diagonal directions. During the analysis, each connected region is assigned a unique label value; pixels with the same label belong to the same connected region. The nucleocytoplasm separation mask image is generated by merging the connected regions, where the foreground region corresponds to the nucleus and the background region corresponds to the cytoplasm. The pixel value in the mask image directly indicates whether a location belongs to the nucleus. Taking a typical human umbilical cord mesenchymal stem cell as an example, after grayscale co-occurrence matrix calculation of its 3D confocal image, the contrast parameters of the nucleolus region are concentrated between 0.15 and 0.25, the entropy parameters are between 1.8 and 2.5, and the correlation parameters are maintained above 0.7. These values ​​are significantly lower than the corresponding parameters of the cytoplasm region, where the contrast parameters are usually greater than 0.4, the entropy parameters are greater than 3.2, and the correlation parameters are less than 0.5. The nucleolus region is successfully separated from the complex intracellular environment by the set double threshold standard. The morphological opening operation removes approximately 15 noise spots with an area of ​​less than 50 pixels. The connected domain analysis identifies three main nucleolar connected regions and generates a complete nuclear-cytoplasmic separation mask image.

[0084] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0085] Based on the nuclear-cytoplasmic separation mask image, the coordinates of the cell nucleus center of gravity were determined as the radial analysis origin. The cell nucleus region was divided into five concentric rings at 20% of the equivalent radius to obtain the radial partition structure within the nucleus.

[0086] The fluorescence intensity of the differentiation markers in each ring in the radial partition structure of the nucleus was statistically analyzed to obtain the mean fluorescence intensity value, standard deviation value and peak density value;

[0087] The cell cytoplasm region is divided into 36 sectors at 10-degree intervals with the center of gravity of the cell nucleus as the pole, and the cytoplasm polar coordinate partition structure is obtained;

[0088] Based on the cytoplasmic polar coordinate partition structure, the integrated value of the fluorescence intensity, gradient change rate and spatial clustering coefficient of the marker expression in each sector area were calculated and processed to obtain the radial distribution feature vector;

[0089] The pluripotency maintenance index was obtained by weighted summation calculation based on the expression intensities of the three differentiation markers at weight ratios of 0.4, 0.35, and 0.25.

[0090] Specifically, the coordinates of the center of gravity of the nucleus are determined based on the spatial distribution of all foreground pixels in the nucleocytoplasmic separation mask image. The center of gravity position is obtained by calculating the arithmetic mean of the x and y coordinates of all pixels within the nucleus. The center of gravity calculation process identifies all foreground pixels with a pixel value of 1 in the mask image. The x- and y-coordinates of these pixels are then summed to obtain the sum in the x direction and the y direction, respectively. Finally, the x and y coordinates of the center of gravity are divided by the total number of foreground pixels. The equivalent radius is the radius of a circular region equal to the actual area of ​​the nucleus. It is calculated by dividing the nucleus area by pi and taking the square root. This equivalence eliminates the influence of irregular nucleus shape on radial analysis. Concentric zoning divides the equivalent radius into five annular regions at a 20% ratio. The first zoning covers the area from the center of gravity to 20% of the equivalent radius, the second zoning covers 20% to 40%, and so on until the fifth zoning covers 80% to 100% of the area. The radial partition structure within the kernel is established through polar coordinate transformation. The position of each pixel is uniquely determined by the radial distance from the center of gravity and the angle relative to the horizontal axis. The pixels are assigned to the corresponding annulus according to the radial distance.

[0091] Statistical analysis of fluorescence intensity for differentiation markers quantitatively assessed the expression distribution of three key proteins, Oct4, Nanog, and CD34, within each annulus. The mean fluorescence intensity (MFI) was calculated by summing the fluorescence intensities of all pixels within the annulus and dividing the sum by the total number of pixels. This reflects the overall level of marker expression within that annulus. The standard deviation (SD) measures the dispersion of fluorescence intensity within the annulus. The calculation process first calculates the square of the difference between the fluorescence intensity of each pixel and the mean, then averages and takes the square root of all squared differences. A large standard deviation indicates a nonuniform distribution of fluorescence intensity, while a small standard deviation indicates a relatively uniform distribution. The peak density (PDI) was calculated by calculating the ratio of the number of pixels within the annulus whose fluorescence intensity exceeded a specific threshold to the total number of pixels. The threshold was set at 1.5 times the mean fluorescence intensity for that annulus. The PDI reflects the distribution of highly expressed regions within the annulus. These three statistical parameters were calculated for each of the three markers, generating a total of 45 values ​​that constitute the set of intranuclear characteristic parameters.

[0092] The cytoplasmic polar coordinate partitioning structure uses the center of mass of the cell nucleus as the polar coordinate origin and divides the cytoplasm into 36 equal sectors along the angular direction. Each sector covers an angular range of 10 degrees, and the radial range of the sector extends from the nuclear boundary to the cell boundary. Angle calculation uses the inverse tangent function to determine the angle of each cytoplasmic pixel relative to the positive horizontal axis. The angle range is 0 to 360 degrees, and the angular space is divided into 36 intervals of 10 degrees. The boundaries of the sector are bounded by two rays originating from the pole and the cell outline segment. The number of pixels contained in the sector varies depending on the cell shape and the cytoplasmic thickness in that direction. The polar coordinate partitioning method can capture the directional distribution characteristics of cytoplasmic marker expression and is particularly suitable for analyzing the spatial patterns of cytoskeletal proteins and membrane-bound proteins with polar distribution.

[0093] Calculation of the fluorescence intensity integral involves summing the fluorescence intensity values ​​of all pixels within a sector. The magnitude of this integral reflects both the total amount of marker expression in that direction and the thickness of the cytoplasm in that direction, with directions of greater thickness typically exhibiting higher integral values. The gradient change rate describes the rate of change in fluorescence intensity in the radial direction. This calculation is performed by dividing the sector into equal segments along the radial direction, calculating the ratio of the fluorescence intensity difference to the distance between adjacent segments, and then averaging the gradient values ​​of all segments to obtain the gradient change rate for that sector. The spatial clustering coefficient quantifies the degree of marker clustering within a sector. It is obtained by normalizing the ratio of the variance to the mean of the fluorescence intensity. A high clustering coefficient indicates a clustered distribution of the marker, while a low clustering coefficient indicates a relatively uniform distribution. The radial distribution feature vector, composed of the fluorescence intensity integral values, gradient change rates, and spatial clustering coefficients for all sectors, encompasses a total of 108 characteristic parameters and comprehensively describes the spatial distribution pattern of the marker within the cytoplasm.

[0094] The pluripotency maintenance index (PLMI) is calculated based on the combined expression levels of three differentiation markers: Oct4, Nanog, and CD34, within the cell. Weights of 0.4, 0.35, and 0.25 correspond to the importance of these markers in maintaining stem cell pluripotency, respectively. Oct4, as a core pluripotency transcription factor, receives the highest weight of 0.4; Nanog, as a key regulator of self-renewal, receives a weight of 0.35; and CD34, as a stem cell surface marker, receives a weight of 0.25. The weighted summation calculation begins by normalizing the mean fluorescence intensity of each marker within the cell by subtracting the minimum value and dividing by the range to ensure that the fluorescence intensities of the different markers are within the same numerical range. These normalized intensities are then multiplied by the corresponding weighting coefficients. Finally, the three weighted values ​​are summed to yield the PLMI. The index ranges from 0 to 1, with values ​​closer to 1 indicating that the cell is closer to pluripotency and values ​​closer to 0 indicating that the cell is more likely to be differentiated. Taking a typical induced pluripotent stem cell as an example, after its nucleus was divided into five concentric rings, the average fluorescence intensity of Oct4 in the inner ring was significantly higher than that in the outer ring. The Oct4 intensity of the first ring was 1.8 times that of the inner ring. The standard deviation was smaller in the inner ring, indicating relatively uniform expression. The peak density reached the highest value in the core area. After the cytoplasmic area was divided into 36 sector-shaped areas, CD34 showed a higher integral value and aggregation coefficient in the sector-shaped area facing the culture medium, and the gradient change rate increased significantly in the cell edge area. The pluripotency maintenance index obtained after weighted calculation of the three markers was 0.82, indicating that the cells were in a highly pluripotent state.

[0095] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0096] According to the numerical range of the pluripotency maintenance index, the stem cell state is divided into four biological states: pluripotency maintenance state, early differentiation initiation state, differentiation ongoing state and terminal differentiation state, thus obtaining the cell state classification standard;

[0097] Based on the cell state classification standard, the state transition frequency matrix was obtained by statistically processing the changes of the pluripotency maintenance index at consecutive time points.

[0098] The state transition frequency matrix is ​​normalized and processed with a 2-hour time step to obtain the state transition probability data;

[0099] The change rate of the pluripotency maintenance index between adjacent time points was calculated and analyzed to obtain the differentiation rate index and differentiation direction discrimination parameter;

[0100] Based on the state transition probability data and differentiation direction discrimination parameters, time series trajectory modeling is performed to obtain differentiation trajectory feature data.

[0101] Specifically, the four-category classification standard for stem cell states is based on the numerical range of the pluripotency maintenance index. The pluripotency maintenance state corresponds to an index value between 0.7 and 1.0, indicating that cells express high levels of pluripotency markers and have the ability to self-renew. The early differentiation initiation state corresponds to an index value between 0.4 and 0.7, indicating that cells are beginning to lose some pluripotency characteristics but have not yet committed to a specific differentiation direction. The ongoing differentiation state corresponds to an index value between 0.2 and 0.4, indicating that cells are differentiating toward a specific cell lineage and have a significant decrease in pluripotency marker expression. The terminal differentiation state corresponds to an index value between 0 and 0.2, indicating that cells have lost pluripotency and acquired functional characteristics of mature cells. The cell state classification standard is automated through threshold judgment. The algorithm reads the pluripotency maintenance index value at each time point and automatically assigns a corresponding state label based on the index's numerical range. This classification process does not rely on manual intervention, ensuring objectivity and consistency of state classification. State transition frequency statistics are based on the state change records of consecutive time points in the time-lapse image sequence. The algorithm traverses the state sequence of each cell at all observed time points and counts the number of transitions from state i to state j. The state transition frequency matrix is ​​a 4×4 square matrix with row indices representing the starting state and column indices representing the target state. The matrix element values ​​record the number of observations of the corresponding state transition. The statistical process uses a sliding window approach, with the state at the current time point as the starting state and the state at the next time point as the target state. All pairs of adjacent time points are examined one by one, and the corresponding transition counts are accumulated. The diagonal elements of the state transition frequency matrix represent the number of times the state remains unchanged, while the off-diagonal elements represent the number of times the state transitions occur. The sum of the elements in each row of the matrix equals the total number of occurrences of the starting state during the observation period.

[0102] Normalized probability calculation converts the state transition frequency into a probability value, ensuring that the sum of the elements in each row is equal to 1. The calculation method is to divide each element of the frequency matrix by the sum of the elements in its row. The 2-hour time step is used as the standard time unit for probability calculation. All state transition observations are normalized to this time scale to eliminate the impact of different observation intervals on probability estimation. State transition probability data reflects the probability of stem cells transitioning from one state to another within 2 hours. The larger the probability value, the more likely the transition is to occur. The probability value of 0 means that this type of transition did not occur during the observation period. The normalized probability matrix has the Markov property. The current state completely determines the state distribution at the next moment, without considering historical state information. This property simplifies the mathematical modeling of the stem cell differentiation process.

[0103] The differentiation rate index is calculated by calculating the ratio of the difference in the pluripotency maintenance index between adjacent time points to the time interval. A positive value indicates increased pluripotency, a negative value indicates decreased pluripotency, and the absolute value reflects the severity of the change. The differentiation direction discriminant parameter is calculated based on the time derivative of the expression intensity of three markers: Oct4, Nanog, and CD34. Decreased expression of Oct4 and Nanog, combined with the pattern of changes in CD34 expression, can indicate the differentiation direction of the stem cell. Mesenchymal differentiation is typically accompanied by a sustained decrease in CD34 expression and an upregulation of specific mesenchymal markers, while neuroectodermal differentiation is characterized by a rapid decrease in CD34 expression and the onset of neural marker expression. The differentiation direction discriminant parameter is calculated by weighting the rate of change of the three markers. The weight coefficient is determined based on the role of each marker in indicative of each differentiation direction. The sign and magnitude of the parameter value together indicate the direction and intensity of differentiation.

[0104] Time-series trajectory modeling integrates state transition probability data and differentiation direction discriminant parameters to construct a dynamic evolution model of stem cell differentiation. The model can predict the state distribution and differentiation trend of cells at future time points. Trajectory modeling uses a probabilistic graphical model framework to represent the stem cell differentiation process as a random walk in state space. The state transition probability determines the direction of the walk, and the differentiation direction parameter provides additional constraint information. The differentiation trajectory feature data includes the state transition probability matrix, the differentiation rate time series, the differentiation direction parameter series, and the predicted future state probability distribution. These data comprehensively describe the spatiotemporal dynamics of the stem cell differentiation process. Taking a 72-hour time-lapse observation of a group of 100 mesenchymal stem cells as an example, all cells were in the pluripotency maintenance state at the initial moment. After 24 hours, 15 cells transferred to the early differentiation initiation state. After 48 hours, 8 of these 15 cells entered the differentiation process state, 2 returned to the pluripotency maintenance state, and 5 remained in the early differentiation initiation state. After 72 hours, 3 cells finally reached the terminal differentiation state. The state transition frequency matrix showed that the probability of transition from the pluripotency maintenance state to the early differentiation initiation state was 0.15, and the probability of returning to the pluripotency maintenance state from the early differentiation initiation state was 0.13. The differentiation rate index gradually decreased after reaching a peak in the differentiation initiation stage. The differentiation direction discrimination parameter showed that most cells tended to differentiate in the bone marrow mesenchymal direction.

[0105] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0106] Based on the radial distribution feature vector, the centroid coordinates of the fluorescent labeled images of mitochondria, endoplasmic reticulum and Golgi apparatus were extracted and processed to obtain the spatial position coordinate set of each subcellular organelle;

[0107] The spatial position coordinate set is used to establish an intracellular coordinate system with the center of gravity of the cell nucleus as the reference origin, and the Euclidean distance between each subcellular organelle is calculated to obtain the subcellular organelle distance matrix;

[0108] According to the distance matrix between subcellular organelles, 5 microns is set as the adjacency determination threshold, and the subcellular organelles whose distance is less than the adjacency determination threshold are connected to obtain topological connection relationship data;

[0109] Based on the topological connection relationship data, an undirected graph structure is constructed with the subcellular organelle centroid as the vertex and the adjacency relationship as the edge to obtain the initial topological network diagram;

[0110] The graph theory parameters of clustering coefficient, connectivity distribution and path length were calculated and processed on the initial topological network graph to obtain the subcellular organelle spatial association network graph.

[0111] Specifically, radial distribution feature vectors guide the extraction of subcellular organelle centroid coordinates based on the specific fluorescent marker signals of three organelles: mitochondria, labeled with MitoTracker Red dye, exhibit red fluorescence; the endoplasmic reticulum, labeled with ER-Tracker Green dye, exhibits green fluorescence; and the Golgi apparatus, immunolabeled with GM130 antibody, exhibits blue fluorescence. The centroid coordinate extraction process first performs binary segmentation for each fluorescence channel, marking regions with fluorescence intensity above background noise as foreground pixels. Connected domain analysis is then used to identify individual organelle structures. Each connected domain represents an individual organelle. The centroid coordinates are calculated by calculating the weighted average of the coordinates of all pixels within the connected domain, with the weight being the corresponding pixel's fluorescence intensity. This weighted calculation method ensures that the centroid position more accurately reflects the organelle's geometric center and the center of gravity of the fluorescence distribution. The spatial clustering coefficient parameter in the radial distribution feature vector is used to screen candidate regions for centroid extraction. Regions with high clustering coefficients are prioritized for centroid calculation to avoid misidentification of fluorescence noise as organelle structures. The spatial position coordinate set contains the x- and y-coordinates of the centroids of all identified organelles, along with the corresponding organelle type identifier and fluorescence intensity information.

[0112] An intracellular coordinate system was established with the nuclear center of mass as the origin, with the x-axis defined along the long axis of the cell and the y-axis perpendicular to it. The orientation of the coordinate system was determined by principal component analysis of the cell outline. A coordinate transformation was performed to convert the absolute pixel coordinates of all subcellular organelles into relative coordinates relative to the nuclear center of mass. This transformation involved two steps: translation, which shifted the origin to the nuclear center of mass, and rotation, which aligned the coordinate axes with the principal axis of the cell. Euclidean distance was calculated using a standard two-dimensional distance formula. For two organelles with coordinates xi, yi and xj, yj, the Euclidean distance between them was equal to the square root of the squared sum of the squared y-coordinate difference. The inter-subcellular organelle distance matrix is ​​a symmetric matrix whose rows and columns correspond to individual organelles. The matrix elements represent the Euclidean distances between corresponding pairs of organelles, and zero diagonal elements represent the distances between organelles themselves. The construction of this distance matrix lays the foundation for subsequent adjacency determination and topological analysis.

[0113] The 5-micrometer threshold for determining adjacency is based on the biological characteristics of organelle interactions. Organelles within this distance range typically interact through membrane contacts, molecular exchange, or functional coordination. The determination process traverses all elements of the distance matrix. When the matrix element value is less than 5 micrometers, a connecting edge relationship is established between the corresponding organelle pairs. This connecting edge indicates that the two organelles are spatially adjacent and may have a functional connection. Topological connectivity data is stored in the form of an adjacency list or adjacency matrix, which records a list of all neighboring organelles for each organelle. Connectivity relationships are symmetrical: if organelle A is adjacent to organelle B, then organelle B is also adjacent to organelle A. The process of establishing connectivity relationships also takes into account organelle type information. Connections between mitochondria and the endoplasmic reticulum, between the endoplasmic reticulum and the Golgi apparatus, and between organelles of the same type are recorded separately, providing topological information for subdivided categories for subsequent functional network analysis.

[0114] The construction of an undirected graph structure converts the topological connectivity data into a standard graph theory representation. The vertex set of the graph contains all identified organelle individuals, and each vertex carries attribute information such as organelle type, coordinate position, and fluorescence intensity. The edge set of the graph is composed of organelle pairs that satisfy the adjacency relationship. The weight of the edge can be set to the inverse of the distance between organelles. The closer the distance, the stronger the connection strength of the organelle. The construction process of the undirected graph creates a vertex index map, mapping the actual identifiers of the organelles to the vertex numbers in the graph, which facilitates subsequent graph theory algorithm processing. The initial topological network graph has the characteristics of a small-world network. Most organelles are connected to a few neighbors, and a few key organelles have a high degree of connectivity. This topological structure reflects the spatial organization pattern of organelle distribution within the cell.

[0115] The calculation of graph theory parameters includes the quantitative analysis of three core indicators. The clustering coefficient measures the local connectivity density of the network. It is calculated by counting the ratio of the actual number of edges between the neighboring vertices of each vertex to the theoretical maximum number of edges, and then averaging the clustering coefficients of all vertices. The connectivity distribution describes the statistical distribution of the number of vertex connections in the network. The distribution histogram is constructed by counting the number of vertices with the same connectivity. The shape of the connectivity distribution reflects the topological structure type of the network. The path length calculation uses the breadth-first search algorithm to find the shortest path between any two vertices. The average path length is the arithmetic mean of the shortest path lengths between all vertex pairs, reflecting the overall connectivity and information transmission efficiency of the network. The subcellular organelle spatial association network diagram integrates the topological structure and graph theory parameters to form a complete mathematical description of the spatial organization pattern of organelles. Taking a typical embryonic stem cell as an example, 85 mitochondria, 12 endoplasmic reticulum branches and 3 Golgi apparatus fragments were identified in the cell. After establishing the intracellular coordinate system, it was found that mitochondria were mainly distributed in a ring structure around the cell nucleus, endoplasmic reticulum branches extended from the nuclear membrane to the cell edge, and the Golgi apparatus was located on one side of the cell nucleus with a polarized distribution. Distance matrix calculation showed that the average distance between adjacent mitochondria was 2.3 microns, and the average distance between mitochondria and endoplasmic reticulum was 1.8 microns. After applying a 5-micron threshold, the topological network established contained 312 connecting edges. The clustering coefficient reached 0.68, indicating that the organelles were clustered. The connectivity distribution showed a power-law decay characteristic. The average path length was 3.2, indicating that any two organelles could be reached through an average of 3.2 steps.

[0116] In a specific embodiment, the step of calculating graph theory parameters such as clustering coefficient, connectivity distribution, and path length on the initial topological network graph may specifically include the following steps:

[0117] Perform triangle connection statistics on the adjacent vertex pairs of each vertex in the initial topological network graph to obtain the number of triangles and the number of connection triplets;

[0118] Based on the number of triangles and the number of connected triplets, the clustering coefficient is calculated by taking the ratio of 3 times the number of triangles to the number of connected triplets to obtain the network clustering coefficient parameter.

[0119] Perform statistical analysis on the number of connected edges of each vertex in the initial topological network graph to obtain the connectivity value and connectivity distribution frequency of each vertex;

[0120] The connectivity distribution frequency is processed by probability density calculation according to the degree interval to obtain the connectivity distribution parameters;

[0121] Based on the graph theory shortest path algorithm, the shortest connection path length between any two vertices in the initial topological network graph is traversed and calculated to obtain the average path length parameter and the subcellular organelle spatial association network graph.

[0122] Specifically, the triangle connectivity statistics analyze the local subgraph formed by each vertex and its neighboring vertices in the initial topological network graph. A triangle refers to the minimum closed loop structure composed of three vertices and three edges. The statistical process traverses each vertex in the network and checks whether all neighbor vertex pairs of the vertex are also connected to each other. If vertex A is adjacent to vertex B, vertex A is adjacent to vertex C, and vertex B is adjacent to vertex C, then the three vertices form a triangle. The triangle number statistics use a combined counting method to count the number of triangles with each vertex as an endpoint. The final total number of network triangles is the sum of the number of triangles of all vertices divided by 3. The division by 3 is because each triangle is counted repeatedly by three vertices. The number of connected triplets counts all possible three-vertex combinations, including three-vertex combinations that form triangles and three-vertex combinations that do not form triangles but have at least two connecting edges. Connected triplets reflect the potential connection density between vertices in the network.

[0123] The network clustering coefficient is calculated based on the ratio of three triangles to the number of connected triplets. This ratio measures the probability that a combination of three vertices in the network forms a fully connected triangle. The coefficient of three in the calculation formula comes from the fact that each triangle contains three edges, while each connected triplet contains at most three edges. This ratio reflects the local connectivity density of the network. The clustering coefficient ranges from 0 to 1. A coefficient close to 1 indicates a high degree of local clustering in the network, with most adjacent vertices connected to each other. A coefficient close to 0 indicates a relatively sparse network with a lack of direct connections between adjacent vertices. The network clustering coefficient parameter is expressed as a global clustering coefficient and an average local clustering coefficient. The global clustering coefficient is calculated based on the triangle and triple statistics of the entire network, while the average local clustering coefficient is the arithmetic mean of the local clustering coefficients of all vertices.

[0124] The statistical analysis of the number of connected edges traverses each vertex in the initial topological network graph and calculates the number of edges directly connected to the vertex. This number is called the vertex's degree or connectivity. The statistical process establishes a mapping relationship between vertices and connectivity, records the identifier of each vertex and the corresponding connectivity value, and forms a degree sequence data structure. The frequency statistics of the connectivity distribution are achieved by constructing a frequency histogram of the degree values. The number of vertices with the same connectivity is counted to obtain the frequency of occurrence of each degree value. The degree distribution frequency data reflects the distribution pattern of vertex connectivity in the network. A high frequency degree value indicates that there are more vertices with this connectivity in the network, and a low frequency degree value indicates that there are fewer vertices with the corresponding connectivity.

[0125] The connectivity distribution parameters are obtained by converting the connectivity distribution frequency into a probability density function. The probability density calculation divides the frequency of each degree value by the total number of vertices, ensuring that the sum of the probabilities of all degree values ​​is equal to 1. Degree interval partitioning divides the connectivity range into several equal intervals, each corresponding to a probability density value. The granularity of the interval partitioning is determined by the network size and the degree distribution dispersion. The shape characteristics of the connectivity distribution parameters reflect the topological structure type of the network. The normal distribution represents the characteristics of a random network, the power law distribution represents the characteristics of a scale-free network, and the Poisson distribution represents the characteristics of a uniform network. The distribution parameters also include statistical quantities of the degree distribution, such as the average degree, the variance of the degree distribution, and descriptive parameters such as skewness and kurtosis.

[0126] The shortest path algorithm uses a breadth-first search (BFS) method to calculate the shortest path length between any two vertices. Starting from the source vertex, the algorithm expands the search range layer by layer until it reaches the target vertex or traverses all reachable vertices. The traversal calculation process uses each vertex in the network as the source point, calculates the shortest path length from that vertex to all other vertices, and constructs a global shortest path distance matrix. The shortest path length is defined as the minimum number of edges required to connect two vertices. If two vertices are disconnected, the path length is set to infinity. The average path length parameter is calculated by taking the arithmetic mean of the shortest path lengths of all vertex pairs. This calculation excludes vertex pairs with infinite path lengths and only considers connected vertex pairs. The average path length reflects the overall connectivity and information transfer efficiency of the network. A short path length indicates that the network has small-world characteristics, allowing information to propagate quickly between vertices.

[0127] The subcellular organelle spatial association network graph integrates topological information and graph theory parameters to form a complete network description. The network graph contains information at three levels: vertex attributes, edge attributes, and global attributes. Vertex attributes record the type, coordinates, connectivity, and local clustering coefficient of each organelle. Edge attributes record the connected organelle pairs, connection distances, and connection weights. Global attributes include the network clustering coefficient, degree distribution parameter, and average path length. For example, a mitochondrial network of mesenchymal stem cells contains 127 mitochondrial vertices and 263 connecting edges. Triangle statistics reveal 89 triangles and 445 connected triplets within the network. The calculated global clustering coefficient is 0.60, indicating a distinct clustered distribution pattern among mitochondria. Connectivity statistics reveal that most mitochondria have a connectivity between 2 and 5, while a few hub mitochondria have a connectivity exceeding 10. The degree distribution exhibits a power-law decay. Shortest path calculations reveal an average path length of 3.8 between any two mitochondria, with a maximum path length of 9, indicating good connectivity and short information transmission distances.

[0128] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0129] The differentiation trajectory feature data and the graph theory parameters of the subcellular organelle spatial association network diagram are processed by feature vector splicing to obtain a 32-dimensional input feature vector;

[0130] A three-layer feedforward neural network structure is constructed based on the 32-dimensional input feature vector, with 32 neurons in the input layer, 16 neurons in the hidden layer, and 4 neurons in the output layer, resulting in a multi-layer perceptron network architecture.

[0131] The hidden layer neurons of the multi-layer perceptron network architecture are subjected to nonlinear transformation using the rectified linear unit activation function to obtain the hidden layer feature representation;

[0132] The hidden layer feature representation is processed through the back propagation algorithm to iteratively update the network weights and bias parameters to obtain the trained and optimized network parameters;

[0133] Based on the trained and optimized network parameters, the output layer is processed for classification and calculation of pluripotency maintenance degree, differentiation tendency score, proliferation activity index and cell cycle stage to obtain the stem cell image feature classification results.

[0134] Specifically, the feature vector concatenation process combines the differentiation trajectory feature data and the graph-theoretic parameters of the subcellular organelle spatial association network diagram in a fixed order into a unified numerical vector. The differentiation trajectory feature data includes 18 numerical values, including the 16 elements of the state transition probability matrix, the differentiation rate index, and the differentiation direction discriminant parameter. The graph-theoretic parameters of the subcellular organelle spatial association network diagram include 14 numerical values, including the clustering coefficient, average connectivity, degree distribution variance, and average path length. The concatenation process uses an array concatenation operation to sequentially arrange the numerical elements of the two types of feature data to form a one-dimensional feature vector of length 32. The first 18 positions of the vector store the differentiation trajectory features, and the last 14 positions store the network topology features. Before concatenating the feature vectors, numerical normalization is required to unify the feature parameters of different dimensions and numerical ranges to the range of 0 to 1. Normalization uses the minimum-maximum normalization method: the minimum value of each feature is subtracted from the feature value and then divided by the feature range. The 32-dimensional input feature vector comprehensively describes the temporal dynamic characteristics and spatial structural characteristics of stem cells.

[0135] The three-layer feedforward neural network structure adopts a fully connected architecture. The 32 neurons in the input layer correspond to the dimensions of the 32-dimensional feature vector, and each neuron receives the numerical input corresponding to the position in the feature vector. The hidden layer contains 16 neurons, with the number of neurons chosen to be half the input dimension to ensure the network's expressiveness while avoiding overfitting. The output layer has four neurons corresponding to four classification targets: pluripotency maintenance, differentiation propensity score, proliferation activity index, and cell cycle stage. Each output neuron generates a numerical value between 0 and 1 to represent the predicted value of the corresponding attribute. The multilayer perceptron network architecture defines the connections between neurons through weight matrices and bias vectors. The weight matrices from the input layer to the hidden layer are 32×16 in dimension, and from the hidden layer to the output layer are 16×4 in dimension. The bias vectors are 16 and 4 in dimension, respectively. The network architecture uses a forward propagation computational model, with information flowing unidirectionally from the input layer to the output layer, without any feedback or skip connections in between.

[0136] The rectified linear unit activation function (ReLU) performs a nonlinear transformation on the weighted inputs of hidden layer neurons. The function is defined as outputting the original value when the input value is greater than zero, and outputting a zero value when the input value is less than or equal to zero. The nonlinear transformation process first calculates the weighted input of each neuron in the hidden layer. The weighted input is equal to the sum of the product of the output of the input layer neuron and the corresponding weight plus the bias value. The ReLU activation function is then applied to the weighted input to obtain the output value of the hidden layer neuron. The advantages of the ReLU function are simple calculation, stable gradient calculation, and the ability to alleviate the gradient vanishing problem. The linear interval of the function maintains the gradient at 1, and the nonlinear interval sets negative values ​​to zero to introduce sparsity. The hidden layer feature representation consists of the output values ​​of 16 neurons. These values ​​reflect the expression of the input features in the hidden layer feature space. The hidden layer features have a higher level of abstraction and stronger discriminative ability than the original input features.

[0137] The backpropagation algorithm guides parameter updates by calculating the gradient of a loss function with respect to network parameters. The loss function, in the form of mean squared error, is calculated as the sum of the squares of the differences between the network output and the true label. Gradient calculation begins with the output layer and propagates backward to the input layer. The output layer gradient is directly based on the loss function, while the hidden layer gradient is derived from the gradient of the subsequent layer using the chain rule. Weight parameters are updated using gradient descent, with the new weight equal to the current weight minus the product of the learning rate and the gradient. The learning rate is set to 0.001 to control the step size of the parameter update. Bias parameters are updated using the same gradient descent rule, with the bias gradient equal to the error signal of the corresponding neuron. The iterative update process repeats the three steps of forward propagation to calculate the loss, backpropagation to calculate the gradient, and parameter update until the loss function converges or the preset number of iterations is reached. The optimized network parameters during training include the optimal values ​​of all weight matrices and bias vectors. These parameters encode the complex mapping between input features and output classifications.

[0138] The output layer classification calculation predicts the test sample based on the trained and optimized network parameters. The calculation process inputs the test sample's 32-dimensional feature vector into the network, sequentially calculating the hidden layer feature representation and the output layer response. The degree of pluripotency maintenance is represented by the value of the first output neuron: values ​​closer to 1 indicate that the cell is closer to pluripotency, while values ​​closer to 0 indicate that the cell is further away from pluripotency. The differentiation propensity score is generated by the second output neuron: a high score indicates a strong differentiation tendency, while a low score indicates that the cell tends to maintain its current state. The proliferation activity index corresponds to the third output neuron: a high index indicates active cell proliferation, while a low index indicates weak cell proliferation. The cell cycle stage is predicted by the fourth output neuron, with output values ​​in different ranges corresponding to different cell cycle stages: 0 to 0.25 corresponds to the G0 / G1 phase, 0.25 to 0.5 corresponds to the S phase, 0.5 to 0.75 corresponds to the G2 phase, and 0.75 to 1 corresponds to the M phase. The stem cell image feature classification results integrate these four output indicators into a comprehensive assessment report, providing quantitative analysis for stem cell quality control and differentiation status monitoring. Taking a specific embryonic stem cell sample as an example, after network processing, its 32-dimensional input feature vector outputs a pluripotency maintenance degree of 0.89, a differentiation tendency score of 0.12, a proliferation activity index of 0.76, and a cell cycle stage of 0.31. It is comprehensively judged that the cell is in the S phase with high pluripotency, low differentiation tendency, and high proliferation activity, which meets the typical characteristics of high-quality embryonic stem cells.

[0139] The above describes the stem cell microscopic image feature extraction method in the embodiment of the present application. The following describes the stem cell microscopic image feature extraction system in the embodiment of the present application. Figure 2 In one embodiment of the present application, a stem cell microscopic image feature extraction system includes:

[0140] A segmentation module is used to segment the time-lapse 3D confocal microscopy images of stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image;

[0141] an extraction module, configured to perform a multi-morphological feature collaborative extraction process on the stem cell differentiation marker fluorescence image based on the nucleocytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index;

[0142] a quantification module, configured to perform time series feature quantification processing on a sequence of stem cell state change images based on the pluripotency maintenance index to obtain differentiation trajectory feature data;

[0143] A topology module is used to perform spatial feature extraction processing on the subcellular organelle image region using the radial distribution feature vector through a topology map construction algorithm to obtain a subcellular organelle spatial association network map;

[0144] A fusion module is used to perform feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram through a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

[0145] above Figure 2 The stem cell microscopic image feature extraction system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The stem cell microscopic image feature extraction device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0146] Reference Figure 3 In an embodiment of the present invention, a stem cell microscopic image feature extraction device is also provided. The stem cell microscopic image feature extraction device can be a server, and its internal structure can be as follows: Figure 3 As shown. The stem cell microscopic image feature extraction device includes a processor, a memory, a display screen, an input device, a network interface and a database connected via a system bus. The computer-designed processor is used to provide computing and control capabilities. The memory of the stem cell microscopic image feature extraction device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the stem cell microscopic image feature extraction device is used to store the corresponding data in this embodiment. The network interface of the stem cell microscopic image feature extraction device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0147] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the stem cell microscopic image feature extraction device to which the solution of the present invention is applied.

[0148] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the stem cell microscopic image feature extraction method.

[0149] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a stem cell microscopic image feature extraction device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0151] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for extracting features from stem cell microscopic images, characterized in that: The method comprises: The time-lapse 3D confocal microscopy images of stem cells were segmented using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image. performing a multi-morphological feature collaborative extraction process on the stem cell differentiation marker fluorescence image according to the nuclear-cytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index; Based on the pluripotency maintenance index, a time series feature quantification process is performed on the image sequence of stem cell state changes to obtain differentiation trajectory feature data; The radial distribution feature vector is used to extract spatial features of the subcellular organelle image area through a topological map construction algorithm to obtain a subcellular organelle spatial association network map, including: extracting centroid coordinates of fluorescent labeled images of mitochondria, endoplasmic reticulum, and Golgi apparatus based on the radial distribution feature vector to obtain a set of spatial position coordinates of each subcellular organelle; establishing an intracellular coordinate system with the center of gravity of the cell nucleus as a reference origin for the spatial position coordinate set, calculating the Euclidean distance between each subcellular organelle to obtain an inter-subcellular organelle distance matrix; setting 5 microns as an adjacency determination threshold according to the inter-subcellular organelle distance matrix, establishing connection edge relationship processing for subcellular organelles with a distance less than the adjacency determination threshold, and obtaining topological connection relationship data; constructing an undirected graph structure with subcellular organelle centroids as vertices and adjacency relationships as edges based on the topological connection relationship data to obtain an initial topological network map; calculating graph theory parameters such as clustering coefficient, connectivity distribution, and path length on the initial topological network map to obtain the subcellular organelle spatial association network map; The differentiation trajectory feature data and the subcellular organelle spatial association network diagram are subjected to feature fusion processing by a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

2. The stem cell microscopic image feature extraction method according to claim 1, characterized in that: The method of segmenting the time-lapse 3D confocal microscopy image of stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image includes: performing 3×3 neighborhood gray-level co-occurrence matrix calculation on the time-lapse 3D confocal microscopy image of the stem cells to obtain contrast parameters, entropy parameters, and correlation parameters; A first texture complexity threshold is set to 0.3 based on the contrast parameter, and a second texture complexity threshold is set to 2.8 based on the entropy parameter to obtain a chromatin region discrimination criterion; Performing nucleolus candidate region marking processing on the image pixels according to the chromatin region discrimination standard to obtain an initial nucleolus segmentation region; The initial nucleolus segmentation region is processed by morphological opening operation to remove the noise region with an area less than 50 pixels to obtain a purified nucleolus region; Connected domain analysis is performed based on the purified nucleolus region to obtain the nucleocytoplasmic separation mask image.

3. The stem cell microscopic image feature extraction method according to claim 1, characterized in that: The step of performing multi-morphological feature collaborative extraction processing on the stem cell differentiation marker fluorescence image according to the nuclear-cytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index includes: Based on the nucleocytoplasmic separation mask image, the coordinates of the cell nucleus center of gravity are determined as the radial analysis origin, and the cell nucleus region is divided into 5 concentric rings at 20% of the equivalent radius to obtain the intranuclear radial partition structure; Performing statistical analysis on the fluorescence intensity of the differentiation marker in each ring in the intranuclear radial partition structure to obtain a mean fluorescence intensity value, a standard deviation value, and a peak density value; The cell cytoplasm region is divided into 36 sector regions at 10-degree intervals with the cell nucleus center of gravity as the pole, to obtain a cytoplasm polar coordinate partition structure; The radial distribution feature vector is obtained by calculating the fluorescence intensity integral value, gradient change rate and spatial clustering coefficient of the marker expression in each sector region based on the cytoplasmic polar coordinate partition structure; The pluripotency maintenance index was obtained by performing weighted sum calculation based on the expression intensities of the three differentiation markers at weight ratios of 0.4, 0.35, and 0.

25.

4. The stem cell microscopic image feature extraction method according to claim 1, characterized in that: The step of performing time series feature quantification processing on the stem cell state change image sequence based on the pluripotency maintenance index to obtain differentiation trajectory feature data includes: According to the numerical range of the pluripotency maintenance index, the stem cell state is divided into four biological states: pluripotency maintenance state, early differentiation initiation state, differentiation progress state and terminal differentiation state, to obtain a cell state classification standard; Based on the cell state classification standard, the state transition frequency statistics of the pluripotency maintenance index changes at consecutive time points are processed to obtain a state transition frequency matrix; The state transition frequency matrix is ​​subjected to normalized probability calculation processing according to a 2-hour time step to obtain state transition probability data; The change rate of the pluripotency maintenance index between adjacent time points was calculated and analyzed to obtain the differentiation rate index and differentiation direction discrimination parameter; A time series trajectory modeling process is performed based on the state transition probability data and the differentiation direction discrimination parameter to obtain the differentiation trajectory feature data.

5. The stem cell microscopic image feature extraction method according to claim 1, characterized in that: The step of performing graph theory parameter calculation processing on the initial topological network graph, such as clustering coefficient, connectivity distribution, and path length, to obtain the subcellular organelle spatial association network graph comprises: Performing triangle connection relationship statistics processing on adjacent vertex pairs of each vertex in the initial topological network graph to obtain the number of triangles and the number of connection triplets; Based on the number of triangles and the number of connection triplets, a clustering coefficient is calculated according to a ratio of three times the number of triangles to the number of connection triplets to obtain a network clustering coefficient parameter; Performing statistical analysis on the number of connected edges of each vertex in the initial topological network graph to obtain the connectivity value and connectivity distribution frequency of each vertex; Performing probability density calculation on the connectivity distribution frequency according to the degree interval to obtain connectivity distribution parameters; Based on the graph theory shortest path algorithm, the shortest connection path length between any two vertices in the initial topological network graph is traversed and calculated to obtain an average path length parameter and the subcellular organelle spatial association network graph.

6. The stem cell microscopic image feature extraction method according to claim 1, characterized in that: The step of performing feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram using a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result includes: Performing feature vector concatenation processing on the differentiation trajectory feature data and the graph theory parameters of the subcellular organelle spatial association network diagram to obtain a 32-dimensional input feature vector; Based on the 32-dimensional input feature vector, a three-layer feedforward neural network structure is constructed, with 32 neurons in the input layer, 16 neurons in the hidden layer, and 4 neurons in the output layer, to obtain a multi-layer perceptron network architecture; Performing nonlinear transformation processing on hidden layer neurons of the multilayer perceptron network architecture using a rectified linear unit activation function to obtain hidden layer feature representation; The hidden layer feature representation is subjected to an iterative update process of network weights and bias parameters through a back-propagation algorithm to obtain trained and optimized network parameters; Based on the network parameters optimized by the training, the output layer is subjected to classification calculation processing of the pluripotency maintenance degree, differentiation tendency score, proliferation activity index and cell cycle stage to obtain the stem cell image feature classification result.

7. A stem cell microscopic image feature extraction system, characterized in that: For implementing the stem cell microscopic image feature extraction method according to any one of claims 1 to 6, the stem cell microscopic image feature extraction system comprises: A segmentation module is used to segment the time-lapse 3D confocal microscopy images of stem cells using a chromatin texture-sensitive adaptive threshold segmentation algorithm to obtain a nuclear-cytoplasmic separation mask image; an extraction module, configured to perform a multi-morphological feature collaborative extraction process on the stem cell differentiation marker fluorescence image based on the nucleocytoplasmic separation mask image to obtain a radial distribution feature vector and a pluripotency maintenance index; a quantification module, configured to perform time series feature quantification processing on a sequence of stem cell state change images based on the pluripotency maintenance index to obtain differentiation trajectory feature data; A topology module is used to perform spatial feature extraction processing on the subcellular organelle image region using the radial distribution feature vector through a topology map construction algorithm to obtain a subcellular organelle spatial association network map; A fusion module is used to perform feature fusion processing on the differentiation trajectory feature data and the subcellular organelle spatial association network diagram through a multi-layer perceptron fusion algorithm to obtain a stem cell image feature classification result.

8. A stem cell microscopic image feature extraction device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method for extracting features from a stem cell microscopic image according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to perform the stem cell microscopic image feature extraction method according to any one of claims 1 to 6.

Images