Cell activity detection method based on fluorescence labeling

Abstract

The invention provides a cell activity detection method based on fluorescence labeling, and relates to the technical field of cell analysis. The cell activity detection method based on fluorescence labeling comprises the following steps: S1, co-incubating a cell sample to be detected with a mixed solution of propidium iodide, a JC-1 mitochondrial membrane potential dye, a DCFH-DA active oxygen probe, a Caspase-3 / 7 activated substrate and a Hoechst nuclear dye, completing synchronous multiple labeling, and generating a multi-labeled cell sample, S2, detecting the cell activity based on the multi-labeled cell sample, five-channel full-automatic scanning is carried out through a high-content imaging system, Z-stack images are collected, and a multispectral image data set is generated. The five probes are mixed and incubated at one time through a synchronous multiple labeling process, so that the operation time is remarkably shortened, and the risk of cell injury caused by multiple times of centrifugation is reduced; z-stack full-automatic scanning is combined with three-dimensional image acquisition, the capturing capability of weak fluorescence signals is improved through multi-focal plane fusion, and information loss caused by cell thickness or position offset in traditional two-dimensional imaging is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cell analysis, in particular to a cell activity detection method based on fluorescent labeling. BACKGROUND

[0002] The technical field of cell analysis focuses on the development and application of a series of methodologies and tools aimed at qualitatively and quantitatively studying the physical, chemical and biological properties of individual cells or cell populations. The core goal of this field is to gain a deep understanding of the structure, function, state of cells and their dynamic response to external stimuli (such as drugs, environmental factors, pathogens). Key technical means usually include but are not limited to microscopic imaging (such as fluorescence microscopy, confocal microscopy, high-content imaging system), flow cytometry, molecular probe labeling (such as antibodies, fluorescent dyes, genetically encoded indicators), biosensors, and increasingly dependent on bioinformatics and artificial intelligence algorithms. Among them, a cell activity detection method based on fluorescent labeling refers to the use of multiple specific fluorescent probes (such as propidium iodide, JC-1, DCFH-DA, Caspase-3 / 7 activation substrate, Hoechst 33342) to simultaneously label different activity-related indicators of cells (such as cell membrane integrity, mitochondrial membrane potential, reactive oxygen species level, apoptosis protease activity, cell nucleus), combined with high-content imaging technology to obtain multi-channel fluorescence images, and through advanced image processing, single-cell segmentation, multi-dimensional feature extraction, unsupervised clustering, machine learning classification and spatial distribution analysis, etc. Steps, a technical solution is realized to accurately, multi-parameter, high-throughput determine and quantitatively analyze the activity state (such as healthy, early apoptosis, high oxidative stress, etc.) of a large number of single cells.

[0003] Traditional multi-probe labeling requires step-by-step incubation and washing operations, and repeated centrifugation steps can easily induce mechanical stress damage, leading to unintended changes in cell activity state before detection; flow cytometry supports multi-parameter detection, but is limited by the liquid flow focusing mechanism and cannot preserve the original spatial position information of the cells, making it difficult to analyze the influence of the microenvironment on the cell state; conventional clustering methods rely on linear dimensionality reduction or pre-set threshold segmentation of subgroups, and have insufficient ability to analyze complex nonlinear relationships between high-dimensional features, which may mask key biological subtypes; artificial annotation classification rules require expert experience to set fixed thresholds, have limited generalization ability in the face of batch-to-batch experimental variations, and the annotation process is time-consuming and labor-intensive; population statistics only output percentage means, ignoring the state fluctuations at the single-cell level and the transition states between subgroups, reducing the sensitivity to weak treatment effects; two-dimensional imaging lacks Z-axis information, resulting in incomplete cell stereoscopic structure data, especially for thick samples or aggregated cells, which can easily cause cross interference of fluorescence signals; spatial heterogeneity analysis often uses simple partition counting methods, which cannot quantify local density gradient changes and are difficult to identify spatial aggregation patterns of micron-scale functional units. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a cell activity detection method based on fluorescent labeling, which solves the problems that traditional multi-probe labeling needs to be implemented in steps, incubation and cleaning operations, repeated centrifugation steps are easy to induce mechanical stress damage, and the cell activity state changes unexpectedly before detection; flow cytometry supports multi-parameter detection, but is limited by the liquid flow focusing mechanism and cannot retain the original spatial position information of the cells, making it difficult to analyze the influence of the microenvironment on the cell state; conventional clustering methods rely on linear dimension reduction or preset threshold to segment subgroups, and have insufficient ability to analyze complex nonlinear relationships between high-dimensional features, which may mask key biological subtypes.

[0005] To achieve the above object, the present application is implemented by the following technical scheme: a cell activity detection method based on fluorescent labeling, comprising the following steps:

[0006] S1: co-incubate the cell sample to be tested with a mixed solution of iodinated propyl, JC-1 mitochondrial membrane potential dye, DCFH-DA reactive oxygen probe, Caspase-3 / 7 activated substrate and Hoechst nuclear dye, complete synchronous multi-labeling, and generate a multi-labeled cell sample;

[0007] S2: based on the multi-labeled cell sample, five-channel full-automatic scanning is performed by a high-content imaging system, Z-stack images are collected, and a multi-spectral image data set is generated;

[0008] S3: based on the multi-spectral image data set, a deep learning segmentation model is used to process the Hoechst channel image, generate cell nucleus and cytoplasm contour masks, and generate single-cell spatial positioning data;

[0009] S4: based on the single-cell spatial positioning data and the multi-spectral image data set, the fluorescence intensity, morphological and textural features of each cell are calculated, and a single-cell multi-dimensional feature matrix is generated;

[0010] S5: based on the single-cell multi-dimensional feature matrix, through nonlinear dimension reduction and density clustering processing, the distribution pattern of cell subgroups is identified, and cell subgroup distribution data is generated;

[0011] S6: based on the single-cell multi-dimensional feature matrix and the cell subgroup distribution data, a graph neural network is used to train a semi-supervised classification model, learn multi-parameter association rules, and generate a multi-state classifier;

[0012] S7: based on the multi-state classifier, the single-cell multi-dimensional feature matrix is processed, and the state labels of health / early apoptosis / high oxidative stress are output, and single-cell state classification results are generated;

[0013] S8: Based on the single-cell state classification results, the proportion of cells in different states in each experimental group and the significant differences are calculated, and a population activity quantitative report is generated;

[0014] S9: Based on the single-cell state classification results and single-cell spatial positioning data, a two-dimensional state distribution heat map is generated by kernel density estimation, and an activity spatial heterogeneity report is generated.

[0015] Preferably, the S1 generating a multi-label cell sample comprises the following steps:

[0016] S101: Take the cell suspension to be tested, remove the original culture medium by centrifugation, resuspend the cells with phosphate buffered saline, and obtain a pretreated cell suspension;

[0017] S102: Dissolve iodinated propyl, JC-1 dye, 2',7'-dichlorodihydrofluorescein diacetate, Caspase-3 / 7 green activation substrate and Hoechst 33342 in serum-free medium, mix evenly, and prepare five-channel fluorescence probe working solution;

[0018] S103: Mix the pretreated cell suspension with the five-channel fluorescence probe working solution at a volume ratio of 1:1, incubate at 37°C in the dark for 20 minutes, and form a probe labeling reaction system;

[0019] S104: Centrifuge the probe labeling reaction system, discard the supernatant to remove unbound fluorescent probes, resuspend the cells with fresh culture medium, and obtain a multi-label cell sample.

[0020] Preferably, the S2 generating a multi-spectral image data set comprises the following steps:

[0021] S201: Use an automatic dispenser to transfer the multi-label cell sample to a black transparent 96-well plate, inoculate 5000 cells per well, and form a standardized detection plate;

[0022] S202: Configure the optical parameters of the high-content imaging system, set the corresponding filters for different fluorescent probes: Hoechst 33342, iodinated propyl, JC-1 monomer / JC-1 polymer, DCFH-DA, Caspase-3 / 7 activation substrate, and generate a multi-spectral imaging parameter set;

[0023] S203: Use a 20x objective lens, select 9 fields of view for each well of the standardized detection plate according to the multi-spectral imaging parameter set, and collect 8 layers of images along the Z-axis direction with a step of to obtain the original three-dimensional image data set;

[0024] S204: Apply a maximum width projection algorithm to the Z-stack image sequence of each field of view in the original three-dimensional image dataset to generate a multi-spectral image dataset.

[0025] Preferably, the S3 generating single cell spatial positioning data comprises the following steps:

[0026] S301: Separate the two-dimensional projection image corresponding to the Hoechst 33342 channel from the multi-spectral image dataset as a cell nucleus reference image;

[0027] S302: Input the cell nucleus reference image into a pre-trained U-shaped network architecture semantic segmentation deep learning model to identify and segment the outline of each cell nucleus, generating a cell nucleus binary mask;

[0028] S303: Centered on the position of each cell nucleus in the cell nucleus binary mask, expand a fixed pixel radius outward to predict and generate the corresponding cytoplasmic region, obtaining a single cell region prediction map;

[0029] S304: Profile correction is performed on the single cell region prediction map to optimize the cell boundary through edge gradient detection and exclude the interference of adjacent cell overlapping areas, finally generating accurate single cell spatial positioning data.

[0030] Preferably, the S4 generating single cell multi-dimensional feature matrix comprises the following steps:

[0031] S401: Based on the single cell spatial positioning data, assign a unique ID to each segmented cell to generate a single cell identification index table;

[0032] S402: According to the single cell identification index table, calculate the mean, median and standard deviation of the fluorescence intensity in each cell region from the five fluorescence channels of the multi-spectral image dataset, respectively, to generate a single cell fluorescence intensity matrix;

[0033] S403: Based on the single cell spatial positioning data, calculate the morphological parameters of each cell, including area, perimeter, circularity and aspect ratio, to generate a single cell morphological feature matrix;

[0034] S404: Based on the single cell spatial positioning data and the Hoechst channel image in the multi-spectral image dataset, calculate the texture feature parameters of each cell region using the gray level co-occurrence matrix algorithm, including entropy and contrast, to generate a single cell texture feature matrix;

[0035] S405: Integrate the data in the single cell fluorescence intensity matrix, single cell morphological feature matrix and single cell texture feature matrix according to the single cell ID to form a single cell multi-dimensional feature matrix containing multi-dimensional parameters.

[0036] Preferably, the S5 generating cell subpopulation distribution data comprises the following steps:

[0037] S501: For each dimension parameter in the single-cell multi-dimensional feature matrix, apply the Z-score method for standardization processing to eliminate the dimension effect and generate a standardized feature matrix;

[0038] S502: Apply the t-distributed Stochastic Neighbor Embedding algorithm to reduce the high-dimensional features in the standardized feature matrix to a 2-dimensional space to generate a low-dimensional embedding coordinate set;

[0039] S503: On the low-dimensional embedding coordinate set, apply a density-based spatial clustering algorithm with a neighborhood radius , and a minimum number of points MinPts=20 to perform clustering analysis on the cells to generate a cell cluster label set;

[0040] S504: According to the cell cluster label set and the low-dimensional embedding coordinate set, calculate the distribution center of each cell subpopulation in the 2-dimensional space to generate cell subpopulation distribution data.

[0041] Preferably, the S6 generating multi-state classifier comprises the following steps:

[0042] S601: Based on the single-cell multi-dimensional feature matrix, construct a K-nearest neighbor graph connecting each cell to its 10 nearest neighbor cells with cosine similarity as the metric to generate a cell relationship topology graph;

[0043] S602: According to the cell subpopulation distribution data, label representative cell subpopulation samples with state labels by experts to generate a training sample label set;

[0044] S603: Construct a semi-supervised classification model based on a graph convolutional neural network. The model contains three layers of graph convolutional layers, each with 128 nodes;

[0045] S604: Use the cross-entropy loss function and the Adam optimizer to train and optimize the initial classification model with a learning rate of 0.001 and iterate for 50 rounds to finally obtain a trained multi-state classifier.

[0046] Preferably, the S7 generating single-cell state classification results comprises the following steps:

[0047] S701: Input the single-cell multi-dimensional feature matrix into the trained multi-state classifier to obtain the original probability distribution of each cell belonging to each state category through forward propagation calculation;

[0048] S702: For each cell, according to its original state probability distribution, adopt the maximum probability decision rule to select the state category with the highest probability as its preliminary state label to generate a preliminary state label set;

[0049] S703: Set a confidence threshold, remove low-confidence samples with probability values ​​lower than the threshold from the initial state label set, and generate a high-confidence state label set;

[0050] S704: Convert the numerical labels in the high-confidence state label set into corresponding text state identifiers to generate the final single-cell state classification result.

[0051] Preferably, the S8 step of generating a quantitative report on population activity includes the following steps:

[0052] S801: Based on the single-cell state classification results, the number of cells in each state in different experimental groups is counted, and a state distribution frequency table is generated.

[0053] S802: Based on the state distribution frequency table, calculate the percentage of cells in different states in each experimental group and generate a state percentage matrix;

[0054] S803: Apply the analysis of variance algorithm to the state proportion matrix to calculate the F-value and corresponding p-value of the state proportion difference between different experimental groups, and generate a significant difference index.

[0055] S804: Integrates the state percentage matrix and significant difference indicators to generate a structured report containing quantitative data on population activity.

[0056] Preferably, the step S9 of generating an active spatial heterogeneity report includes the following steps:

[0057] S901: Combining single-cell state classification results and single-cell spatial positioning data, establish the correspondence between the two-dimensional coordinates of each cell on the culture plate and its state label, and generate a spatial state mapping table.

[0058] S902: For each cell state in the spatial state mapping table, the kernel density estimation algorithm is applied independently, using a Gaussian kernel function to calculate its density distribution in two-dimensional space, generating a state density distribution field.

[0059] S903: Maps the state density distribution field values ​​of each state to 256 levels of pseudo-color to generate a heat map that intuitively displays the spatial distribution density of cells in different states.

[0060] S904: On the state-space heatmap, identify and mark spatial clustering regions with density values ​​exceeding twice the global mean, and generate an analysis report reflecting the spatial heterogeneity of cell activity.

[0061] This invention provides a method for detecting cell viability based on fluorescent labeling. It has the following beneficial effects:

[0062] The application completes the incubation of the five probes at one time through a synchronous multiple labeling process, significantly shortens the operation time and reduces the risk of cell damage caused by multiple centrifugation; Z-stack full-automatic scanning combined with three-dimensional image acquisition improves the capture ability of weak fluorescent signals through multi-focus plane fusion, and overcomes the information loss caused by cell thickness or position deviation in traditional two-dimensional imaging; the deep learning segmentation model directly generates a cytoplasm mask based on the nucleus characteristics, avoids the subjectivity of manual parameter adjustment, and enhances the robustness of contour recognition under complex cell morphology; a multi-dimensional feature matrix integrates intensity, morphology and texture parameters to construct a more comprehensive cell state representation system, providing high information density input for subsequent analysis; nonlinear dimensionality reduction and density clustering automatically identify the implicit subpopulation structure without presetting biological marker thresholds to reveal population heterogeneity; the graph neural network learns semi-supervised classification rules using the topological relationship between cells, effectively uses a small amount of labeled samples to realize high-precision state determination, and reduces the dependence on large-scale labeled data; the confidence filtering mechanism eliminates low-reliability prediction results to improve the biological interpretability of the final state label; the spatial kernel density estimation quantifies the local cell state aggregation characteristics, converts the microscopic distribution mode into a visual heat map, and provides spatial dimension evidence chain for microenvironment influence research. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 It is a schematic diagram of the main steps of the application;

[0064] Figure 2 It is a detailed schematic diagram of S1 of the application;

[0065] Figure 3 It is a detailed schematic diagram of S2 of the application;

[0066] Figure 4 It is a detailed schematic diagram of S3 of the application;

[0067] Figure 5 It is a detailed schematic diagram of S4 of the application;

[0068] Figure 6 It is a detailed schematic diagram of S5 of the application;

[0069] Figure 7 It is a detailed schematic diagram of S6 of the application;

[0070] Figure 8 It is a detailed schematic diagram of S7 of the application;

[0071] Figure 9 It is a detailed schematic diagram of S8 of the application;

[0072] Figure 10 It is a detailed schematic diagram of S9 of the application. DETAILED DESCRIPTION

[0073] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0074] Embodiments:

[0075] As shown in the drawings, the embodiments of the present application provide a cell activity detection method based on fluorescent labeling, comprising the following steps: Figures 1-10

[0076] S1: co-incubate the cell sample to be tested with a mixed solution of iodinated propidium, JC-1 mitochondrial membrane potential dye, DCFH-DA reactive oxygen probe, Caspase-3 / 7 activated substrate and Hoechst nuclear dye, complete synchronous multiple labeling, and generate a multi-labeled cell sample;

[0077] S2: based on the multi-labeled cell sample, perform five-channel full-automatic scanning through a high-content imaging system, collect Z-stack images, and generate a multi-spectral image data set;

[0078] S3: based on the multi-spectral image data set, process the Hoechst channel image by using a deep learning segmentation model, generate a cell nucleus and cytoplasm contour mask, and generate single-cell spatial positioning data;

[0079] S4: based on the single-cell spatial positioning data and the multi-spectral image data set, calculate the fluorescence intensity, morphological and textural features of each cell, and generate a single-cell multi-dimensional feature matrix;

[0080] S5: based on the single-cell multi-dimensional feature matrix, identify the cell subpopulation distribution pattern through nonlinear dimension reduction and density clustering processing, and generate cell subpopulation distribution data;

[0081] S6: based on the single-cell multi-dimensional feature matrix and the cell subpopulation distribution data, train a semi-supervised classification model by using a graph neural network, learn multi-parameter association rules, and generate a multi-state classifier;

[0082] S7: based on the multi-state classifier, process the single-cell multi-dimensional feature matrix, output state labels such as healthy / early apoptosis / high oxidative stress, and generate single-cell state classification results;

[0083] S8: based on the single-cell state classification results, calculate the proportion of cells in different states and the significant difference in each experimental group, and generate a population activity quantitative report;

[0084] ​S9: Based on the single-cell state classification result and the single-cell spatial positioning data, a two-dimensional state distribution heat map is generated by kernel density estimation, and an active spatial heterogeneity report is generated.

[0085] S1 generating a multi-labeled cell sample comprises the following steps:

[0086] S101: Take the cell suspension to be tested, remove the original culture medium by centrifugation, resuspend the cells with phosphate buffered saline to obtain a pretreated cell suspension;

[0087] Take the cell suspension to be tested and place it in a centrifuge tube. Remove the original culture medium by centrifugation (centrifugal force 300 x g, duration 5 minutes) and discard the supernatant. Add pre-cooled phosphate buffered saline (PBS) to the cell pellet, mix gently by blowing to resuspend the cells completely and remove any visible clumps to obtain a pretreated cell suspension. The cell density of the pretreated cell suspension can be adjusted to about .

[0088] S102: Dissolve iodinated propylene, JC-1 dye, 2',7'-dichlorodihydrofluorescein diacetate, Caspase-3 / 7 green activation substrate and Hoechst 33342 in serum-free medium, mix evenly to prepare a five-channel fluorescence probe working solution;

[0089] Perform the operation in the dark. Dissolve iodinated propylene (PI), JC-1 dye, 2',7'-dichlorodihydrofluorescein diacetate (DCFH-DA), Caspase-3 / 7 green activation substrate and Hoechst 33342 in serum-free medium, respectively, to prepare a stock solution of each probe or directly dilute to the required working concentration. Specifically, the working concentration of iodinated propylene is , the working concentration of JC-1 dye is , the working concentration of DCFH-DA probe is , the working concentration of Caspase-3 / 7 activation substrate is 5 , and the working concentration of Hoechst 33342 is . Mix the above five probe working solutions in equal proportions, vortex to ensure uniformity, and prepare a five-channel fluorescence probe working solution.

[0090] S103: Mix the pretreated cell suspension with the five-channel fluorescence probe working solution at a volume ratio of 1:1, incubate at 37°C in the dark for 20 minutes, and form a probe labeling reaction system;

[0091] Take the pretreated cell suspension and mix it with an equal volume of the five-channel fluorescence probe working solution in the same container. Place the mixed solution in a 37°C constant temperature environment and incubate it in the dark at a speed of 180 rpm for 20 minutes to form a probe labeling reaction system.

[0092] S104: Centrifuge the probe labeling reaction system, discard the supernatant to remove unbound fluorescent probes, resuspend the cells with fresh culture medium, and obtain a multi-labeled cell sample.

[0093] The completed incubation probe labeling reaction system is centrifuged (centrifugal force 300 x g, duration 5 minutes), and the supernatant is carefully discarded to remove unbound free fluorescent probes. Fresh complete medium (such as RPMI-1640) pre-warmed to 37°C is added to the cell pellet, and the cells are gently resuspended to a uniform dispersed state, and finally a multi-labeled cell sample is obtained. After resuspension, the cell density can be adjusted to about , and the cell survival rate should be no less than 95%.

[0094] S2 generates a multi-spectral image data set, including the following steps:

[0095] S201: Use an automatic dispenser to transfer the multi-labeled cell sample to a black bottom 96-well plate, inoculate 5000 cells per well to form a standardized detection plate;

[0096] Based on the multi-labeled cell sample, start the automated liquid handling system, set the dispensing volume to 100 μl per well (corresponding to a cell density ), use an 8-channel pipette head to evenly distribute the sample to each well of the 96-well plate, and after dispensing, let stand for 5 minutes to allow the cells to settle. Microscopic inspection confirms that the actual number of cells in wells A1, D5, and H12 is 4987 ± 35, 5032 ± 41, and 4995 ± 29 (mean ± standard deviation), respectively, which meets the quality control range of 5000 ± 5%, and the preparation of the standardized detection plate is completed.

[0097] S202: Configure the optical parameters of the high-content imaging system, set the corresponding filters for different fluorescent probes: Hoechst 33342, propidium iodide, JC-1 monomer / JC-1 polymer, DCFH-DA, Caspase-3 / 7 activated substrate, and generate a multi-spectral imaging parameter set;

[0098] Create a new experiment protocol in the imaging control software, select the Hoechst 33342 channel, set the excitation filter center wavelength 350 nm (bandwidth 40 nm), and the emission filter center wavelength 460 nm (bandwidth 50 nm); set the propidium iodide channel to excitation 535 nm (bandwidth 30 nm) / emission 617 nm (bandwidth 36 nm); set the JC-1 monomer channel to excitation 514 nm (bandwidth 20 nm) / emission 529 nm (bandwidth 25 nm), and the JC-1 polymer channel to reuse the same excitation filter, and change the emission to 590 nm (bandwidth 35 nm); set the DCFH-DA channel to excitation 502 nm (bandwidth 15 nm) / emission 530 nm (bandwidth 20 nm); set the Caspase-3 / 7 channel to excitation 502 nm (bandwidth 18 nm) / emission 530 nm (bandwidth 25 nm), and optimize the exposure time of each channel through pre-experiment: Hoechst 50 ms, PI 100 ms, JC-1 monomer 80 ms, JC-1 polymer 70 ms, DCFH-DA 120 ms, Caspase-3 / 7 90 ms, and save the parameter set as "five-probe imaging_V1".

[0099] S203: Using a 20x objective, select 9 fields of view for each well of the standardized detection plate according to the multi-spectral imaging parameter set, and step along the Z-axis direction to Collect 8 layers of images as steps to obtain the original three-dimensional image data set;

[0100] Load the standardized detection plate onto the stage, select a 20x air objective (NA=0.45), focus to the bottom reference surface of the well plate, set the Z-axis scanning range from -10 μm below the reference surface to +25 μm above the reference surface (covering the cell layer thickness), step interval 5 μm, single well scanning time calculation: 8 layers x 5 channels x 0.5 seconds per layer = 20 seconds, total time for 96 wells of the whole plate is 32 minutes, after executing the scan, obtain the original data file (including 96 wells x 9 fields of view x 8 layers x 5 channels = 34560 16-bit TIFF images), and store it as a three-dimensional image data set.

[0101] S204: Apply the maximum width projection algorithm to the Z-stack image sequence of each field of view in the original three-dimensional image data set to generate a multi-spectral image data set.

[0102] Call the MIP function of the image processing library to compare the intensity values of the 8-layer Z-stack image of a single field of view pixel by pixel, take the maximum width value of each pixel position to generate a two-dimensional projection, for example: the intensity value of the (512, 512) pixel point of a field of view in the Hoechst channel is , take the maximum value 1560 as the output value, repeat the process until all pixels are processed, five channels independently execute the algorithm, and finally generate a single view five-channel fusion image. A total of 4320 two-dimensional images (96 wells x 9 views x 5 channels) are generated to form a multispectral image dataset.

[0103] S3 generates single-cell spatial positioning data, including the following steps:

[0104] S301: Separate the two-dimensional projection image corresponding to the Hoechst33342 channel from the multispectral image dataset as the cell nucleus reference image;

[0105] Based on the multispectral image dataset, call the channel separation function of the image processing software, input the parameter to specify the channel index number as 0 (corresponding to the Hoechst33342 channel), and execute to extract the single-channel two-dimensional image matrix. The matrix size is 512x512 pixels, the bit depth is 16 bits, and it is stored as an independent image file with the naming rule "Plate1_WellA1_Site1_Hoechst.tif". After extracting all 8640 images, they are summarized into a cell nucleus reference image dataset.

[0106] S302: Input the cell nucleus reference image into the pre-trained U-shaped network architecture semantic segmentation deep learning model to identify and segment the outline of each cell nucleus and generate a cell nucleus binary mask;

[0107] Load the pre-trained U-Net model weight file (the model structure includes 4 layers of encoder and 4 layers of decoder, and the training data is derived from 5000 labeled cell nucleus images). After normalizing the cell nucleus reference image to the range of 0-1, input the model, and the output probability map is binarized by 0.5 threshold value. For example: if the original output value of a pixel point is 0.78>0.5, it is determined as a cell nucleus region, and connected component analysis is performed to mark independent cell nuclei, generating a binary mask file. The average number of identified cell nuclei in a single view is 356±28, and the mask file is stored as a "_mask" suffix file with the same name as the reference image.

[0108] S303: Take the position of each cell nucleus in the cell nucleus binary mask as the center and expand a fixed pixel radius outward to predict and generate the corresponding cytoplasmic region to obtain a single-cell region prediction map;

[0109] Read the cell nucleus binary mask file, traverse each connected domain marked area, calculate its centroid coordinates (such as the centroid X=125.3, Y=87.6 of area R), make a circular expansion with a radius of 15 pixels (corresponding to the actual size of 9.75 μm) centered on the centroid, generate the cytoplasm prediction area, and trigger the overlap processing when the overlap rate of the two expanded areas is >40%: calculate the bisector of the overlapping area, shrink the boundary of the overlapping area along the bisector to an overlap rate <5%, and finally output the single cell area prediction map containing the nucleus and cytoplasm area, stored as GeoJSON format spatial data.

[0110] S304: Profile correction is performed on the single cell area prediction map, the cell boundary is optimized by edge gradient detection, and the interference of adjacent cell overlapping area is excluded, and finally the accurate single cell spatial positioning data is generated.

[0111] Sobel edge gradient detection is performed on the single cell area prediction map, the boundary pixel gradient amplitude is calculated, the gradient threshold is set to 30 (8-bit grayscale range 0-255), the boundary points with gradient value <30 are determined as fuzzy boundary points, and the adjacent position with gradient >50 is moved along the normal direction. For adjacent cell overlapping area (the distance between two cell boundaries <3 pixels), forced separation is performed: translate the boundary along the direction of the two cell nucleus connecting line to a distance ≥5 pixels, output the corrected polygon vertex coordinate set, and store it as a spatial positioning database containing cell ID and vertex coordinate list.

[0112] S4 generates a single cell multi-dimensional feature matrix including the following steps:

[0113] S401: Based on the single cell spatial positioning data, a unique ID is assigned to each segmented cell, and a single cell identification index table is generated;

[0114] Based on the single cell spatial positioning data file (format GeoJSON), all polygon areas are traversed, and a unique ID is assigned according to the "plate number_well number_view number_sequence number" rule, for example: the third cell in the second view of well 2 in plate 1A1 is marked as P1_WA1_S2_C003, an index table containing the mapping relationship between ID and corresponding polygon vertex coordinates is established, stored as a CSV file, and the table header is Cell_ID, Vertices_List. Complete the single cell identification index table generation.

[0115] S402: According to the single cell identification index table, the mean, median and standard deviation of the fluorescence intensity in each cell area are calculated from the five fluorescence channels of the multi-spectral image data set, respectively, to generate a single cell fluorescence intensity matrix;

[0116] Read single cell ID index table, extract corresponding polygon region by ID, perform region pixel value statistics on each channel (Hoechst / PI / JC-1 monomer / JC-1 polymer / DCFH-DA) of multispectral image dataset: calculate the mean value of the sum of all pixel intensity values in the region divided by the total number of pixels, take the median value after sorting the pixel values, and the standard deviation formula is the pixel value, μ is the mean value, and n is the number of pixels), for example: the PI channel pixel value of a certain cell [102, 98, 105, 110], the mean value is (102+98+105+110) / 4=103.75, the median value is the third value 105 after sorting [98, 102, 105, 110], and the standard deviation , the output matrix row title is cell ID, and the column title is channel_statistic (such as PI_Mean, PI_Median, PI_Std), and a single cell fluorescence intensity matrix is generated.

[0117] S403: Based on the single cell spatial positioning data, calculate the morphological parameters of each cell, including area, perimeter, circularity and aspect ratio, and generate a single cell morphological feature matrix;

[0118] Based on the polygon vertex coordinates in the single cell spatial positioning data, calculate the morphological parameters: the area is calculated by the shoelace formula to calculate the polygon region (unit pixel²), the perimeter is the cumulative Euclidean distance between adjacent vertices, the circularity (A is the area and P is the perimeter), and the aspect ratio is the ratio of the long side to the short side of the minimum enclosing rectangle, for example: the vertex coordinates of a certain cell [(10, 15), (12, 17), (14, 16)], , , , the aspect ratio of the enclosing rectangle with length 12 and width 8 is 1.5, the output matrix column title is Area, Perimeter, Circularity, Aspect_Ratio, and a single cell morphological feature matrix is generated.

[0119] S404: Based on the single cell spatial positioning data and the Hoechst channel image in the multispectral image dataset, use the gray level co-occurrence matrix algorithm to calculate the texture feature parameters of each cell region, including entropy and contrast, and generate a single cell texture feature matrix;

[0120] For the Hoechst channel image, extract the gray level image (16-bit to 8-bit) for each cell polygon region, set the gray level co-occurrence matrix parameters: distance d=1 pixel, angle 0°, gray level 16, calculate the contrast (i,j is the gray value, and p is the joint probability), and the entropy value ​, example: a certain area gray matrix [[50, 52], [53, 55]], co-occurrence matrix , output matrix column title , generate single cell texture feature matrix.

[0121] S405: integrate the data in the single cell fluorescence intensity matrix, single cell morphological feature matrix and single cell texture feature matrix according to the single cell ID to form a single cell multi-dimensional feature matrix containing multi-dimensional parameters.

[0122] Take the ID column of the single cell identification index table as the reference, merge the three feature matrices in ID order: fluorescence intensity matrix (15 columns), morphological feature matrix (4 columns), and texture feature matrix (2 columns) to form a feature matrix with a total of 21 columns, example: , stored as a two-dimensional array with a row number equal to the total number of cells (such as 356) and a column number of 21, output the single cell multi-dimensional feature matrix.

[0123] S5 generates cell subpopulation distribution data, including the following steps:

[0124] S501: apply Z-score method to standardize each dimension parameter in the single cell multi-dimensional feature matrix to eliminate the influence of dimension and generate a standardized feature matrix;

[0125] Based on the single cell multi-dimensional feature matrix (dimension 356x21), calculate the arithmetic mean μ and standard deviation σ of each feature parameter column by column, and perform standardization conversion on the feature vector of each cell: new value=(original value-μ) / σ, example: a certain cell The original value is 123.45, the column μ=85.3, σ=12.7, and the converted value is (123.45-85.3) / 12.7≈3.00, complete the standardization of 21-dimensional features of 356 cells, output matrix feature value distribution range [-3.2, 3.8], conform to the standard normal distribution characteristics, generate a standardized feature matrix.

[0126] S502: apply t-distributed stochastic neighbor embedding algorithm to reduce high-dimensional features in the standardized feature matrix to 2-dimensional space to generate a low-dimensional embedding coordinate set;

[0127] Set the t-SNE algorithm parameters: perplexity 30, learning rate 200, iteration number 1000, and use PCA dimension reduction result for initialization. Perform dimension reduction on the standardized feature matrix: calculate the high-dimensional space cell similarity probability , low-dimensional space similarity , the y_i coordinates are iteratively optimized by KL divergence minimization, example: cell A and B high-dimensional distance 0.5, low-dimensional initial distance 1.2, after 50 iterations, the distance converges to 0.6, the final output of 356 cell two-dimensional coordinate set, coordinate range X [-42.3, 38.7], Y [-35.6, 41.2], generate low-dimensional embedding coordinate set.

[0128] S503: On the low-dimensional embedding coordinate set, apply the density-based spatial clustering algorithm, set the neighborhood radius , the minimum number of points MinPts = 20, and the cell clustering analysis is generated.

[0129] In the two-dimensional coordinate space, set the DBSCAN parameters: neighborhood radius ε = 0.5 (determined according to the 75th percentile of the sample distance 0.48), minimum number of points MinPts = 20 (about 5.6% of the total number of cells), traverse all points, if the number of points P ε neighborhood is ≥ 20, mark it as a core point, and the density connected core points are classified into the same cluster, non-core points are classified into adjacent clusters, and boundary points are marked separately, example: the ε neighborhood of point (12.3, -5.7) has 25 points, which is marked as a core point, and the neighborhood of point (13.1, -6.2) has 18 adjacent points but is not a core point, which is still classified into the same cluster, finally generating a clustering label set, containing 3 main clusters (cluster 0: 142 cells, cluster 1: 98 cells, cluster 2: 78 cells) and 38 noise points.

[0130] S504: According to the cell clustering label set and the low-dimensional embedding coordinate set, calculate the distribution center of each cell subpopulation in the two-dimensional space, and generate cell subpopulation distribution data.

[0131] Grouped by clustering label, calculate the distribution center of each cluster in two-dimensional space: for all cell coordinates (x, y) in cluster k Calculate the arithmetic mean, the center of gravity , example: cluster 0 contains 142 cells, X coordinate sum Σ = 1853.6, center of gravity X = 1853.6 / 142 ≈ 13.05, Y coordinate sum Σ = -624.3, center of gravity Y ≈ -4.40, calculate the center of gravity coordinates of other clusters in the same way, output data format is the mapping table of cluster label and center of gravity coordinates: Cluster0: (13.05, -4.40), Cluster1: (-15.20, 22.17), Cluster2: (28.33, -18.65), generate cell subpopulation distribution data.

[0132] S6 generates a multi-state classifier including the following steps:

[0133] S601: Based on the single-cell multi-dimensional feature matrix, the cosine similarity is taken as the measurement to construct the K-nearest neighbor graph of each cell and its 10 nearest neighbor cells, and a cell relationship topology graph is generated;

[0134] Based on the single-cell multi-dimensional feature matrix (356 cells x 21 features), the cosine similarity between cells is calculated: for cell A feature vector [a1, a2,..., a21] and cell B vector [b1, b2,..., b21], the similarity , for example: cell A vector [1.2, 0.8, 2.1], cell B vector [0.9, 1.1, 1.8], numerator = 1.2 x 0.9 + 0.8 x 1.1 + 2.1 x 1.8 = 5.58, denominator , similarity ≈ 0.928, select the top 10 neighbors of each cell with the highest similarity to construct edges and store them in the adjacency list format: Cell003: [Cell045, Cell128,..., Cell301], and generate a cell relationship topology graph.

[0135] S602: According to the cell subpopulation distribution data, the representative cell subpopulation samples are labeled by experts, and a training sample label set is generated;

[0136] According to the cell subpopulation distribution data (cluster 0-2 gravity coordinates), select the core area cells (within ±2 standard deviations from the gravity center) in the low-dimensional embedding coordinate set, and label the state label according to the morphology and fluorescence characteristics: cluster 0 core area cells with complete nuclei, PI negative, and high mitochondrial membrane potential, labeled as healthy; cluster 1 cells with Caspase-3 / 7 positive and cell shrinkage, labeled as early apoptosis; cluster 2 cells with DCFH-DA intensity > 20000 and nuclear fragmentation, labeled as high oxidative stress, a total of 142 cells (healthy 50, early apoptosis 45, high oxidative stress 47) are labeled, stored as a label file: CellID, Label (such as P1_WA1_S2_C045: Healthy), and a training sample label set is generated.

[0137] S603: Construct a semi-supervised classification model based on graph convolutional neural network. The model contains three layers of graph convolutional layers, each with 128 nodes;

[0138] Construct a graph convolutional neural network: the number of input layer nodes is 21 (corresponding to the feature dimension), the weight matrix of the first graph convolutional layer is , the adjacency matrix A is the degree matrix , the output is , the second layer H2 = ReLU (D^{-1 / 2}AD^{-1 / 2}H1W2) , the third layer H3 = Softmax (D^{-1 / 2}AD^{-1 / 2}H2W3) , Example: cell node feature vector input [0.82, -1.2, …, 0.33], output probability distribution [0.91, 0.06, 0.03] after three-layer propagation, generate initial classification model.

[0139] S604: Train and optimize the initial classification model using the cross-entropy loss function and Adam optimizer, with learning rate set to 0.001, iterate for 50 rounds, and finally obtain the trained multi-state classifier.

[0140] Set the loss function True label one-hot encoding Predicted probability), Adam optimizer parameters β1=0.9, β2=0.999, initial learning rate 0.001, each round of training: forward propagation to calculate the predicted value, backward propagation to update the weight, example: cell true label [1, 0, 0], predicted value [0.85, 0.10, 0.05], loss L=-1×log0.85≈0.16, after 50 iterations, the loss decreases to 0.02, the validation set accuracy increases to 93.7%, output the final model weight file.

[0141] S7 generates single-cell state classification results, including the following steps:

[0142] S701: Input the single-cell multi-dimensional feature matrix into the trained multi-state classifier, calculate the original probability distribution of each cell belonging to each state category through forward propagation;

[0143] Input the single-cell multi-dimensional feature matrix (356 cells × 21 features) into the trained graph convolutional neural network model, perform forward propagation calculation: first layer output = ReLU (standardized adjacency matrix × feature matrix × weight matrix W1), second layer output = ReLU (standardized adjacency matrix × first layer output × W2), third layer output = Softmax (standardized adjacency matrix × second layer output × W3), example: cell 's feature vector input, after three-layer calculation, output the original probability distribution [0.88, 0.09, 0.03], indicating the probability of healthy state 88%, early apoptosis 9%, high oxidative stress 3%, complete all cell calculations to generate the original probability distribution matrix.

[0144] S702: For each cell, according to its original state probability distribution, use the maximum probability decision rule to select the state category with the highest probability as its preliminary state label, and generate the preliminary state label set.

[0145] ​maximizing the probability vector [p1, p2, p3] of each cell: if p1=0.88>p2=0.09 and p1>p3=0.03, then the preliminary state label = 0 (corresponding to the healthy state), example: a cell probability [0.45, 0.50, 0.05], the maximum value 0.50 is located in the second dimension, label = 1 (early apoptosis), the output preliminary state label set format is CellID: Label (such as ).

[0146] S703: Set a confidence threshold, remove low-confidence samples in the preliminary state label set whose probability value is lower than the threshold, and generate a high-confidence state label set;

[0147] Set the confidence threshold θ=0.85 (based on the validation set recall rate of 92%), check the highest probability value of each cell: if the cell The highest probability 0.88>0.85, the label is retained, if a cell probability [0.40, 0.35, 0.25] has the highest value 0.40<0.85, it is marked as a low-confidence sample and removed, after screening, 278 cells are retained (accounting for 78% of the total number), and a high-confidence state label set is generated.

[0148] S704: Convert the numerical labels in the high-confidence state label set into corresponding text state identifiers to generate the final single-cell state classification results.

[0149] Establish a numerical label-text mapping relationship: 0→"Healthy", 1→"EarlyApoptosis", 2→"HighOxidativeStress", traverse the high-confidence state label set, and convert the numerical label into a text identifier, example: Convert to , output the final state classification result file.

[0150] S8 generates a population activity quantitative report including the following steps:

[0151] S801: Based on the single-cell state classification results, count the number of cells in different experimental groups, and generate a state distribution frequency table;

[0152] Based on the single-cell state classification result file (format: CellID, Status), group the number of cells in each state according to the experimental group label (such as CTRL group, Dose1 group, Dose2 group): traverse all cell records, if the cell ID contains "CTRL" and Status="Healthy", then the CTRL group healthy cell count +1, example: CTRL group has 120 cells, statistics of healthy number 92, early apoptosis 18, high oxidative stress 10, generate frequency table format: experimental group, Healthy, EarlyApoptosis, HighOxidativeStress (such as CTRL, 92, 18, 10).

[0153] S802: According to the state distribution frequency table, calculate the percentage of different state cells in each experimental group, and generate a state proportion matrix;

[0154] Calculate the percentage of each row data in the frequency table: healthy proportion = healthy cell number / total cell number in the group x 100%, example: CTRL group healthy proportion , similarly calculate the early apoptosis proportion , high oxidative stress proportion = 8.33%, generate percentage matrix: experimental group, Healthy(%), EarlyApoptosis(%), HighOxidativeStress(%)(such as CTRL, 76.67, 15.00, 8.33).

[0155] S803: Apply variance analysis algorithm to the state proportion matrix to calculate the F value and corresponding p value of the difference in state proportion between different experimental groups, and generate a significant difference index;

[0156] Perform one-way ANOVA on each state column of the percentage matrix independently: set k=3 experimental groups , the number of samples in the group n=3 (the number of repeated experiments), calculate the between-group variance , within-group variance , F value = MSB / MSW, example: healthy state column data: CTRL group [76.67, 75.23, 77.12], Dose1 group [68.45, 70.11, 67.89], Dose2 group [52.33, 50.78, 54.17], total mean , group mean , , , , , F=628.44 / 5.27≈119.25, p<0.001 by looking up F distribution table, generate difference index: state, F value, p value (such as Healthy, 119.25, <0.001).

[0157] S804: Integrate the state proportion matrix and the significance difference indicator to generate a structured report containing quantitative data of population activity.

[0158] Integrate the percentage matrix and the difference indicator: Construct a structured table with the first column as the experimental group and the subsequent columns as the percentage of each state (retain two decimal places). Add statistical rows at the bottom to mark the F value and p value. Example report fragment:

[0159] Experimental group Healthy (%) Early Apoptosis (%) High Oxidative Stress (%) CTRL 76.67 15.00 8.33 Dose 1 68.82 25.18 6.00 Dose 2 52.43 38.57 9.00 F value 119.25 94.31 0.87 p value <0.001 <0.001 0.43

[0160] Generate the final report file (CSV / PDF format).

[0161] S9 Generate activity space heterogeneity report including the following steps:

[0162] S901: Combine single-cell state classification results and single-cell spatial positioning data to establish a correspondence between the two-dimensional coordinates of each cell on the culture plate and its state label, generating a spatial state mapping table;

[0163] Based on the single-cell state classification result file (CellID: Status) and the spatial positioning database (CellID: Vertices), establish a mapping relationship: parse the cell ID to obtain the plate number, well number, and field number (e.g. ), extract the mean of the cell vertex coordinates as the center coordinates (e.g. X=125.3, Y=87.6), associate the state label (e.g. Healthy), and generate the table: CellID, X, Y, Status (e.g. ,125.3,87.6,Healthy), covering all 278 high-confidence cells.

[0164] S902: For each cell state in the spatial state mapping table, independently apply the kernel density estimation algorithm, using a Gaussian kernel function, to calculate its density distribution in two-dimensional space, generating a state density distribution field;

[0165] Perform kernel density estimation for each state independently: set the Gaussian kernel function bandwidth h=40 pixels (calculated according to the Silverman rule), grid resolution 5 pixels / grid, and calculate the density at grid points (d is the Euclidean distance from G to the cell center), example: the density of healthy state cells at point G(100,100) is calculated , the contribution value , and the total contribution of all healthy cells is f(g)=0.037, generating 256x256 grid density field data.

[0166] S903: Map the density field value of each state to 256-level pseudo-color, generate a heat map to intuitively show the spatial distribution density of different states of cells.

[0167] Linearly map the density field value to [0, 255]: Let the minimum density , the maximum density , the density value of the grid point g , the color value , example: When C = round(255 x 0.036 / 0.151) ≈ 61, the corresponding pseudo-color encoding table 61→RGB(120, 180, 80), generate a three-state heat map (PNG format, resolution 1280x1024).

[0168] S904: On the state space heat map, identify and mark the spatial aggregation area with density value exceeding 2 times the global mean value, and generate an analysis report reflecting the spatial heterogeneity of cell activity.

[0169] Calculate the global density mean μ = 0.042, and the standard deviation σ = 0.018, set the aggregation threshold , scan the heat map data, mark the continuous area with density > 0.078 (adjacent grid connected domain ≥ 5 grids), example: the coordinate range X [80-120], Y [60-90] area density 0.081-0.095, marked as aggregation area 1, output report contains: aggregation area coordinate range, maximum density value, involved cell number (such as aggregation area 1: X80-120 / Y60-90, max = 0.095, cells = 32).

[0170] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting cell viability based on fluorescent labeling, characterized in that, Includes the following steps: S1: The cell sample to be tested is co-incubated with a mixed solution of propidium iodide, JC-1 mitochondrial membrane potential dye, DCFH-DA reactive oxygen species probe, Caspase-3 / 7 activating substrate and Hoechst nuclear dye to complete simultaneous multiple labeling and generate multi-labeled cell samples. S2: Based on the multi-labeled cell sample, a five-channel fully automated scan is performed using a high-content imaging system to acquire Z-stack images and generate a multispectral image dataset; S3: Based on the multispectral image dataset, a deep learning segmentation model is used to process the Hoechst channel image to generate cell nucleus and cytoplasm contour masks, and generate single-cell spatial localization data. S4: Based on the single-cell spatial localization data and multispectral image dataset, calculate the fluorescence intensity, morphological and texture features of each cell to generate a single-cell multidimensional feature matrix; S5: Based on the single-cell multidimensional feature matrix, nonlinear dimensionality reduction and density clustering are used to identify cell subpopulation distribution patterns and generate cell subpopulation distribution data. S6: Based on the single-cell multidimensional feature matrix and cell subpopulation distribution data, a graph neural network is used to train a semi-supervised classification model, learn multi-parameter association rules, and generate a multi-state classifier. S7: Based on the multi-state classifier, the multi-dimensional feature matrix of a single cell is processed to output state labels such as healthy / early apoptosis / high oxidative stress, and generate single-cell state classification results; S8: Based on the single-cell state classification results, calculate the proportion and significant differences of cells in different states within each experimental group, and generate a quantitative report on population activity; S9: Based on the single-cell state classification results and single-cell spatial location data, a two-dimensional state distribution heatmap is generated through kernel density estimation, and an activity space heterogeneity report is generated.

2. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S1 method for generating multi-labeled cell samples includes the following steps: S101: Take the cell suspension to be tested, remove the original culture medium by centrifugation, and resuspend the cells with phosphate buffer to obtain a pretreated cell suspension; S102: Propidium iodide, JC-1 dye, 2',7'-dichlorodihydrofluorescein diacetate, Caspase-3 / 7 green activating substrate and Hoechst33342 were dissolved in serum-free culture medium and mixed evenly to prepare a five-channel fluorescent probe working solution. S103: Mix the pretreated cell suspension with the five-channel fluorescent probe working solution at a volume ratio of 1:1, and incubate at 37°C in the dark with shaking for 20 minutes to form a probe labeling reaction system; S104: Centrifuge the probe-labeled reaction system, discard the supernatant to remove unbound fluorescent probes, resuspend the cells in fresh culture medium to obtain multi-labeled cell samples.

3. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S2 method for generating a multispectral image dataset includes the following steps: S201: Use an automated dispensing device to transfer multi-labeled cell samples into black transparent 96-well plates, seeding 5000 cells per well to form a standardized detection plate; S202: Configure the optical parameters of the high-content imaging system and set corresponding filters for different fluorescent probes: Hoechst33342, propidium iodide, JC-1 monomer / JC-1 polymer, DCFH-DA, Caspase-3 / 7 activating substrate, and generate a multispectral imaging parameter set. S203: Using a 20x objective lens, based on the multispectral imaging parameter set, select 9 fields of view for each well of the standardized detection plate, and acquire 8 layers of images along the Z-axis in 5μm increments to obtain the original 3D image dataset. S204: Apply the maximum width projection algorithm to the Z-stack image sequence of each field of view in the original 3D image dataset for 3D fusion processing to generate a multispectral image dataset.

4. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S3 process for generating single-cell spatial localization data includes the following steps: S301: Separate the two-dimensional projection image corresponding to the Hoechst33342 channel from the multispectral image dataset and use it as the reference image of the cell nucleus; S302: Input the cell nucleus baseline image into the pre-trained U-shaped network architecture semantic segmentation deep learning model, identify and segment the contour of each cell nucleus, and generate a cell nucleus binary mask; S303: Using the position of each cell nucleus in the binary mask of the cell nucleus as the center, expand outward with a fixed pixel radius to predict and generate the corresponding cytoplasmic region, and obtain a single cell region prediction map; S304: Contour correction is performed on the single-cell region prediction map, cell boundaries are optimized through edge gradient detection, and interference from overlapping areas of adjacent cells is eliminated, ultimately generating accurate single-cell spatial localization data.

5. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S4 method for generating a single-cell multidimensional feature matrix includes the following steps: S401: Based on single-cell spatial localization data, assign a unique ID to each segmented cell and generate a single-cell identifier index table; S402: Based on the single-cell identifier index table, calculate the mean, median and standard deviation of fluorescence intensity in each cell region from the five fluorescence channels of the multispectral image dataset to generate a single-cell fluorescence intensity matrix. S403: Based on single-cell spatial localization data, calculate the morphological parameters of each cell, including area, perimeter, roundness and aspect ratio, and generate a single-cell morphological feature matrix; S404: Based on the Hoechst channel images in the single-cell spatial localization data and multispectral image dataset, the gray-level co-occurrence matrix algorithm is used to calculate the texture feature parameters of each cell region, including entropy and contrast, to generate a single-cell texture feature matrix. S405: Integrate the data from the single-cell fluorescence intensity matrix, single-cell morphology feature matrix, and single-cell texture feature matrix according to the single-cell ID to form a single-cell multidimensional feature matrix containing multidimensional parameters.

6. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S5-generated cell subpopulation distribution data includes the following steps: S501: Apply the Z-score method to standardize the parameters of each dimension in the multidimensional feature matrix of a single cell to eliminate the influence of dimensions and generate a standardized feature matrix. S502: Apply the t-distribution random neighborhood embedding algorithm to reduce the high-dimensional features in the standardized feature matrix to a 2-dimensional space and generate a low-dimensional embedding coordinate set. S503: Apply density-based spatial clustering on a low-dimensional embedded coordinate set and set the neighborhood radius. With a minimum number of points MinPts=20, perform cluster analysis on the cells to generate a set of cell cluster labels; S504: Based on the cell clustering label set and the low-dimensional embedding coordinate set, calculate the centroid of the distribution of each cell subpopulation in two-dimensional space, and generate cell subpopulation distribution data.

7. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S6 step of generating a multi-state classifier includes the following steps: S601: Based on the multidimensional feature matrix of a single cell, using cosine similarity as a metric, construct a K-nearest neighbor graph connecting each cell with its 10 nearest neighbor cells to generate a cell relationship topology graph. S602: Based on cell subpopulation distribution data, experts label representative cell subpopulation samples with status tags to generate a training sample label set; S603: Construct a semi-supervised classification model based on a graph convolutional neural network. This model contains three graph convolutional layers, each with 128 nodes; S604: The initial classification model is trained and optimized using the cross-entropy loss function and the Adam optimizer. The learning rate is set to 0.001, and the training is iterated for 50 rounds to finally obtain a trained multi-state classifier.

8. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S7 method for generating single-cell state classification results includes the following steps: S701: Input the multidimensional feature matrix of a single cell into the trained multi-state classifier, and obtain the original probability distribution of each cell belonging to each state category through forward propagation calculation; S702: For each cell, based on its original state probability distribution, the maximum probability decision rule is used to select the state category with the highest probability as its initial state label, and a preliminary state label set is generated; S703: Set a confidence threshold, remove low-confidence samples with probability values ​​lower than the threshold from the initial state label set, and generate a high-confidence state label set; S704: Convert the numerical labels in the high-confidence state label set into corresponding text state identifiers to generate the final single-cell state classification result.

9. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S8 method for generating a quantitative report on population activity includes the following steps: S801: Based on the single-cell state classification results, the number of cells in each state in different experimental groups is counted, and a state distribution frequency table is generated. S802: Based on the state distribution frequency table, calculate the percentage of cells in different states in each experimental group and generate a state percentage matrix; S803: Apply the analysis of variance algorithm to the state proportion matrix to calculate the F-value and corresponding p-value of the state proportion difference between different experimental groups, and generate a significant difference index. S804: Integrates the state percentage matrix and significant difference indicators to generate a structured report containing quantitative data on population activity.

10. The method for detecting cell viability based on fluorescent labeling according to claim 1, characterized in that: The S9 method for generating an active spatial heterogeneity report includes the following steps: S901: Combining single-cell state classification results and single-cell spatial positioning data, establish the correspondence between the two-dimensional coordinates of each cell on the culture plate and its state label, and generate a spatial state mapping table. S902: For each cell state in the spatial state mapping table, the kernel density estimation algorithm is applied independently, using a Gaussian kernel function to calculate its density distribution in two-dimensional space, generating a state density distribution field. S903: Maps the state density distribution field values ​​of each state to 256 levels of pseudo-color to generate a heat map that intuitively displays the spatial distribution density of cells in different states. S904: On the state-space heatmap, identify and mark spatial clustering regions with density values ​​exceeding twice the global mean, and generate an analysis report reflecting the spatial heterogeneity of cell activity.

Images