Cell morphology dynamic monitoring method and system fused with time sequence image analysis

Abstract

The invention discloses a cell morphology dynamic monitoring method and system fused with time sequence image analysis, and the method comprises the steps: firstly obtaining a time sequence image sequence of a cell at a plurality of continuous time points, and extracting an initial feature vector set containing the morphology, size, texture and the like of the cell; inputting the time sequence image into a time sequence image multi-focus fusion model to obtain a fusion feature vector, and forming a refined feature vector through dimension transformation and feature screening; and inputting an improved long-short-term memory network to learn a cell state evolution mode, determining the state of the cell at each time point, generating a change track and predicting a future state. The system comprises an image acquisition and initial feature extraction unit, a time sequence image multi-focus fusion unit, a dimension transformation and feature screening unit, an improved long-short-term memory network unit, a cell state determination unit and a cell state prediction unit, all the units are connected in sequence, image features at different time points are associated through a fusion model, the network is optimized in combination with cell parameters, and the cell state is predicted. The accurate monitoring on the dynamic change of the cell morphology is realized.

Description

Technical Field

[0001] The present invention relates to the field of dynamic monitoring of cell morphology, and in particular to a method and system for dynamic monitoring of cell morphology integrated with time-series image analysis. Background Art

[0002] In the interdisciplinary research of animal husbandry and biomedicine, the dynamic morphological changes of preadipocytes in the intramuscular muscles of sheep fetuses are the core observation objects to reveal the mechanism of intramuscular fat deposition. With the development of high-resolution time-series microscopy technology, multi-focus image sequences of cells at continuous time nodes can be obtained. These images contain key morphological features such as cell size, outline, and fat droplet distribution. However, due to the dynamic nature of cell growth and the focusing limitations of the imaging system, a single frame image cannot fully reflect the temporal correlation of cell morphology. Traditional static analysis methods have difficulty in analyzing the morphological evolution laws during cell proliferation and differentiation. Therefore, it is necessary to build a dynamic monitoring technology system that can integrate multi-focus time-series information and adapt to cell characteristics.

[0003] Existing technologies have significant deficiencies when processing this type of cell time-series images. First, the conventional multi-focus image fusion model does not design a time-series association mechanism for the morphological characteristics of sheep fetal intramuscular precursor adipocytes. It only performs fusion based on the clarity weight of a single-frame image, resulting in the severing of the correlation between cell morphological characteristics at different time points. It is unable to retain the details of the continuous evolution of fat droplets from tiny particles to fused states, as well as the dynamic expansion information of cell boundaries during the proliferation process. Second, when processing cell time-series data, the general long-short-term memory network does not incorporate cell-specific parameters such as fat synthesis rate and intracellular lipid content. The network gating mechanism cannot adapt to the morphological mutation characteristics of cells during the differentiation stage, resulting in deviations in the time-series prediction of cell states, making it difficult to accurately capture the key nodes of cell transformation from fibroblast-like morphology to adipocyte morphology. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a method and system for dynamic monitoring of cell morphology integrating time-series image analysis.

[0005] The technical solution adopted by the present invention is a method for dynamic monitoring of cell morphology integrated with time-series image analysis, comprising the following steps:

[0006] Step S1: obtaining a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, performing feature extraction on different frame images in the sequence, and obtaining a set of initial feature vectors including cell morphology, size, and texture;

[0007] Step S2: Input the initial feature vector set into a pre-built time-series image multi-focus fusion model. This model fuses the features at different moments based on the correlation characteristics of the cell features at different moments, outputs a fused feature vector, and enhances the representation of the dynamic changes of the cell features.

[0008] Step S3: performing dimensionality transformation and feature screening on the fused feature vector, converting it into a preset dimensional space through a preset mapping rule, and retaining the calibration features according to a preset screening criterion associated with the characteristics of sheep fetal intramuscular preadipocytes to form a refined feature vector;

[0009] Step S4: The refined feature vector is input into an improved long short-term memory network. The network improves the gating mechanism and weight update rules of the long short-term memory network based on the dynamic change characteristics of sheep fetal intramuscular preadipocytes. By learning the evolution pattern of cell states over time, the network outputs a cell state feature sequence.

[0010] Step S5: Based on the cell state feature sequence and in combination with pre-set state judgment rules related to the morphology, proliferation, and differentiation of sheep fetal intramuscular preadipocytes, the specific state of the cells at different time points is determined, and a cell state change trajectory is generated;

[0011] Step S6: Based on the cell state change trajectory, the state of the sheep fetus intramuscular preadipocytes in the future unit time is predicted, and the prediction result is output.

[0012] Furthermore, the time series image multi-focus fusion model is constructed as follows: Assume that the input time series image feature vector set is F = {f1, f2, ..., f T}, where T is the number of time points, f t is the eigenvector at time t, and defines the cell feature correlation matrix C, where the element C ij Indicates the degree of correlation between the cell characteristics at time i and time j, Among them, N is the dimension of the feature vector, is the mean of different dimensions of the feature vector at the i-th moment, and the feature vectors at different moments are weighted and fused through the correlation matrix C. The fused feature vector F fusion The m-th dimension element of Among them, the weight w t，m The cell feature importance parameter α is determined by C and the preset cell feature importance parameter α m Jointly determine,

[0013] Furthermore, in the improved long short-term memory network, the update formula of the memory unit is as follows: Assume that the input at time t is x t , the hidden state at the previous moment is h t-1 , the memory unit is ct-1 , input gate i t 、Forget Gate t , output gate o t The calculations are i t =σ(W ix x t +W ih h t-1 +b i ),f t =σ(W fx x t +W fh h t-1 +b f ),o t =σ(W ox x t +W oh h t-1 +b o ), where σ is the activation function, W ix 、W ih 、W fx 、W fh 、W ox 、W oh is the weight matrix, b i 、b f 、b o As the bias vector, a new cell state regulator β was added to the sheep fetal intramuscular preadipocytes. t , the memory cell is updated to c t =f t ⊙c t-1 +i t ⊙tanh(W cx x t +β t W ch h t-1 +b c ), where β t The fat accumulation parameter A of the current cell t and proliferation rate parameter P t Through Beta t =γ1A t +γ2P t Determine, γ1, γ2 are preset coefficients, and ⊙ is element-by-element multiplication.

[0014] Furthermore, in the process of dimension transformation and feature screening of the fused feature vector, the dimension transformation adopts a linear transformation based on the principal component analysis of cell morphology. Let the fused feature vector F fusion , calculate its covariance matrix S, perform eigenvalue decomposition on S to obtain the eigenvector matrix V, and select the first K eigenvectors corresponding to the preset eigenvalues ​​to form the transformation matrix V K, then the transformed eigenvector During feature screening, the screening thresholds θ1 and θ2 were set based on the fat droplet quantity characteristics and cell boundary clarity characteristics of sheep fetal intramuscular preadipocytes. If the element values ​​corresponding to these two features in the transformed feature vector were greater than θ1 and θ2, respectively, the dimensional feature was retained, otherwise it was discarded.

[0015] Furthermore, when determining the specific state of cells at different time points, the cell state judgment rule is set to be a combination of multiple state judgment sub-rules, each sub-rule corresponds to a characteristic of the cell, such as the cell proliferation state, and the number of cells at the current time point is set to N. t , the number of cells at the previous time point is N t-1 ,like The cells are judged to be in a state of rapid proliferation, δ1 is a threshold value set based on the proliferation characteristics of sheep fetal intramuscular preadipocytes; for the cell differentiation state, the expression level of the preset differentiation marker protein in the cell is set as E t , if E t >δ2, the cell is judged to be in a differentiated state. δ2 is a threshold determined based on the differentiation characteristics of the cell. The final state of the cell is determined by combining the judgment results of different sub-rules.

[0016] Furthermore, when predicting the state of sheep fetal intramuscular preadipocytes in the future unit time, a prediction model based on a state transition probability matrix is ​​used, and the determined cell state set is S={s1, s2, ..., s M}, element P of the state transition probability matrix P ij Indicates that from state s i Transfer to state s j The probability of P is obtained by statistical analysis of the historical cell state change trajectory. The current cell state is s c , when predicting the state of the next L time steps, the state s of the lth step l Depend on Determine, where s0 = s c , and combined with the influence of the cell state on the external environmental factors E, the state transition probability matrix is ​​modified. ω is the adjustment coefficient, E ij is the parameter of the degree of influence of environmental factors on the transition from state i to state j.

[0017] Furthermore, in step S3, the dimension transformation and feature screening specifically include:

[0018] Step S31: Calculate the covariance matrix of the fused feature vector, and use the covariance matrix to reflect the correlation between features of different dimensions;

[0019] Step S32: Perform eigenvalue decomposition based on the covariance matrix to obtain eigenvectors, and sort the eigenvectors according to the size of the eigenvalues. The larger the eigenvalue, the greater the variance of the data in that direction, and the more information it contains.

[0020] Step S33: Select the first K eigenvectors that exceed the preset eigenvalue to construct a transformation matrix, and use the transformation matrix to perform linear transformation on the fused eigenvector to reduce the dimension and reorganize the features to highlight the abnormal cell features;

[0021] Step S34: According to the calibration characteristics of the fat droplet number and cell boundary clarity of the sheep fetal intramuscular preadipocytes, a screening threshold for different dimensions of the transformed feature vector is set, and the calibration feature dimensions that meet the cell characteristics are retained based on the threshold screening.

[0022] Furthermore, in step S4, the improved long short-term memory network learning cell state evolution pattern includes:

[0023] Step S41: The refined feature vectors are sequentially input into the improved long short-term memory network in chronological order. The input gate in the network controls the degree to which the current input information enters the memory unit based on the input features and the hidden state at the previous moment;

[0024] Step S42: The forget gate determines the proportion of the previous moment information to be retained in the memory unit based on the input and the previous moment state, so that the network can remember the long-term state change characteristics of the cell;

[0025] Step S43: By using the cell state regulating factor and combining the current cell fat accumulation and proliferation rate parameters, the degree of integration of new information during memory unit updating is adjusted to adapt the network to the characteristics of sheep fetal intramuscular preadipocytes;

[0026] Step S44: The output gate determines the hidden state of the output according to the memory cell state and the current input. The hidden state includes the state evolution information of the cell at the current moment and before.

[0027] Furthermore, in step S5, determining the cell state and generating the change trajectory includes:

[0028] Step S51: Matching the cell state feature sequence with pre-set different state judgment rules. Different judgment rules correspond to different cell characteristics, including proliferation, differentiation, and morphological changes. The specific state category of the cell at each time point is determined through rule matching.

[0029] Step S52: Recording the cell status information at each time point to form a cell status time series, and visually displaying the changes of cell status over time through the time series;

[0030] Step S53: Based on the cell state time series, a curve fitting or path planning algorithm is used to generate a cell state change trajectory, which presents the cell state transition process and trend.

[0031] A cell morphology dynamic monitoring system integrated with time-series image analysis, comprising:

[0032] The image acquisition and initial feature extraction unit is used to obtain a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, and perform feature extraction on different frame images. The output includes a set of initial feature vectors of cell morphology, size, and texture, which is connected to the time-series image multi-focus fusion unit to provide input data for it;

[0033] The time series image multi-focus fusion unit receives the feature vector set output by the initial feature extraction unit, fuses the features at different times using the constructed multi-focus fusion model based on the correlation characteristics of the cell features at different times, and outputs the fused feature vector, which is connected to the dimension transformation and feature screening unit to provide the fused feature data;

[0034] The dimensionality transformation and feature screening unit performs dimensionality transformation and feature screening on the fused feature vector, converts it to a preset dimensional space based on the cell characteristic association rules, retains the calibration features, and outputs the refined feature vector to the improved long short-term memory network unit;

[0035] An improved long short-term memory network unit receives the refined feature vector, learns the evolution pattern of the cell state over time through the improved long short-term memory network, and outputs a cell state feature sequence to the cell state determination unit;

[0036] The cell state determination unit determines the specific state of the cell at different time points based on the cell state feature sequence and the preset cell state judgment rules, generates a cell state change trajectory, and transmits the trajectory information to the cell state prediction unit;

[0037] The cell state prediction unit uses a prediction model to predict the state of sheep fetal intramuscular preadipocytes in the future unit time based on the cell state change trajectory and outputs the prediction result.

[0038] Beneficial Effects: The present invention proposes a method and system for dynamic monitoring of cell morphology by integrating time-series image analysis. This method addresses the problem that conventional multi-focus image fusion models are unable to correlate cell morphological features at different time points. The time-series image multi-focus fusion model designs a dynamic weighting mechanism based on the feature correlation characteristics of sheep fetal intramuscular preadipocytes at different times. This model can fuse the focused image features at each time point, fully preserving the details of the continuous evolution of fat droplets from tiny particles to fused states, as well as the dynamic expansion information of cell boundaries as they proliferate, thus avoiding the fragmentation of feature correlation. The improved long-short-term memory network addresses the deficiency that general long-short-term memory networks are difficult to adapt to the morphological mutation characteristics of cell differentiation stages. The improved long-short-term memory network incorporates cell-specific parameters such as fat synthesis rate and intracellular lipid content, and optimizes the gating mechanism, enabling it to accurately capture the key nodes of the cell transition from fibroblast-like morphology to adipocyte morphology, thereby improving the accuracy of cell state time-series prediction. At the same time, by transforming and screening the dimensionality of the fused features, and combining the state judgment rules with cell characteristics, comprehensive and accurate monitoring of the dynamic changes in the morphology of sheep fetal intramuscular preadipocytes is achieved, providing reliable technical support for analyzing the mechanism of intramuscular fat deposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method steps of the present invention;

[0040] Figure 2 It is a diagram of the system unit composition of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, a method for dynamic monitoring of cell morphology integrated with time-series image analysis comprises the following steps:

[0043] Step S1: obtaining a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, performing feature extraction on different frame images in the sequence, and obtaining a set of initial feature vectors including cell morphology, size, and texture;

[0044] Specifically, this step systematically acquires time-series images of fetal sheep intramuscular preadipocytes over continuous time dimensions and extracts initial features that comprehensively reflect the cell's characteristics. These initial features cover multiple key dimensions. Cell morphology includes contour irregularity (quantified by the deviation of the contour from a perfect circle) and cell extension area (the two-dimensional space occupied by the cell on the culture surface); cell size includes cell diameter range (measured by the maximum straight-line distance between two points on the cell's edge) and cell volume (calculated using three-dimensional modeling); and cell texture includes lipid droplet density (the number of lipid droplets per unit area) and intracellular granule roughness (quantified by the degree of surface roughness). The extraction of these features provides the raw, multi-dimensional foundational data for subsequent multi-focus fusion and deep learning analysis. This is crucial because only by fully capturing the raw information on the dynamic changes of cells at different time points can the subsequent fusion, analysis, and prediction processes have a reliable and comprehensive input source, avoiding biased final results due to missing or incomplete raw data.

[0045] During the implementation process, an environmental system suitable for the culture and imaging of fetal ovine intramuscular preadipocytes was first established. High-resolution microscopy equipment with a resolution of 1024×1024 pixels, autofocus, and time-lapse photography was used. The culture environment was maintained at a constant temperature of 37°C and a 5% CO₂ atmosphere to ensure optimal physiological activity of the cells. Shooting intervals were set at 2 hours, and continuous recording lasted for 48 hours, resulting in a 24-frame time-series image sequence, each clearly capturing cellular morphological details. During the image feature extraction phase, an algorithm based on edge detection and region segmentation was used to accurately identify cells within each frame, eliminating interference from background and impurities. Specific parameters were calculated for each identified cell: contour irregularity was determined by the ratio of contour perimeter to the perimeter of a circle of equal area, with a value range of 0.3-0.8, which effectively distinguishes morphological differences between cells under different physiological conditions. Extended area was calculated by pixel count, with a value range of 500-1200 μm. 2 This interval covers the typical area range of sheep fetal intramuscular preadipocytes during proliferation and differentiation; the fat droplet distribution density is calculated by counting the number of cells per 100 μm 2The number of fat droplets within a region is determined, ranging from 1 to 5, reflecting the initial state of intracellular fat synthesis. The roughness of intracellular particles is calculated using the gray-level co-occurrence matrix, with a value between 15 and 40 grayscale values, effectively characterizing the fineness of intracellular structure. This process ultimately yields a set of initial feature vectors encompassing these parameters for all cells at 24 time points. Each vector corresponds to a comprehensive cell feature at a specific time point.

[0046] Step S2: Input the initial feature vector set into a pre-built time-series image multi-focus fusion model. This model fuses the features at different moments based on the correlation characteristics of the cell features at different moments, outputs a fused feature vector, and enhances the representation of the dynamic changes of the cell features.

[0047] Specifically, in this step, the initial feature vector set extracted in step S1 is input into a specially constructed multi-focus fusion model of time-series images, and the features at different time points are deeply fused through the model. The core capability of this model is to assign differentiated weights to the features at each moment and fuse them based on the intrinsic correlation characteristics between the features of cells at different moments (such as the gradual relationship between the cell sizes at adjacent time points, the continuous change trend of fat droplet distribution, etc.). The fused feature vector finally outputted can integrate the effective information of multiple time points and significantly enhance the overall characterization capability of the dynamic change characteristics of cells. Its significance lies in that it effectively solves the problem of incomplete information caused by focus differences in single-frame images (such as different clarity in different areas). By integrating the features of multiple time points, it can make up for the limitations of single-time point images in information capture, improve the comprehensiveness and accuracy of features, and provide higher-quality input data for subsequent dimensionality transformation, network learning and other steps, thereby improving the reliability of the entire monitoring system.

[0048] In specific implementation, the initial feature vector set of 24 image frames obtained in step S1 is first input into the time-series image multi-focus fusion model in chronological order. The model first activates the feature association analysis module, which quantifies the strength of the association between features at each moment by calculating the correlation coefficient. For example, for cell size features at adjacent moments, when the size change rate is between 5% and 15%, a strong positive correlation is determined, and the correlation strength is set to 0.7-0.9; when the change rate exceeds this range, the correlation strength is reduced to 0.3-0.6. Simultaneously, the consistency of the changing trends of fat droplet distribution density is analyzed. When the distribution density changes in the same direction (increasing or decreasing) at three consecutive time points, the fat droplet features at these time points are determined to be strongly correlated, and the correlation strength is set to 0.6-0.8. Based on these correlation analysis results, the model's dynamic weight assignment module assigns corresponding weights to the features at each moment. The weights are dynamically adjusted between 0.05 and 0.2 based on the feature correlation strength, with higher correlation strengths corresponding to larger weights. During the weighted fusion process, the model uses a feature-by-feature dimension fusion approach, performing a fusion calculation for each feature dimension (such as cell expansion area and fat droplet distribution density). The feature values ​​at 24 time points are summed according to their corresponding weights to obtain the fusion value for that dimension. In this way, the resulting fused feature vector contains comprehensive information on all key features related to cellular dynamics across the 24 image frames. The fusion value for each feature dimension is the result of weighted integration of information from multiple time points, providing a more comprehensive picture of the dynamic changes in cells over a 48-hour period.

[0049] Step S3: performing dimensionality transformation and feature screening on the fused feature vector, converting it into a preset dimensional space using a preset mapping rule, and retaining the calibration features based on a preset screening criterion associated with the characteristics of sheep fetal intramuscular preadipocytes to form a refined feature vector;

[0050] Specifically, this step mainly performs two core operations on the fusion feature vector output by step S2: dimensionality transformation and feature screening. Dimension transformation converts the high-dimensional fusion feature vector to a lower-dimensional and more suitable space through specific linear or nonlinear mapping rules. Its purpose is to reduce the redundancy of the feature space and reduce the complexity of the data. Feature screening is based on screening criteria closely related to the biological characteristics of sheep fetal intramuscular preadipocytes, retaining those features that play a key role in cell state judgment, proliferation and differentiation monitoring, etc., and eliminating those features with low information contribution or interference, and finally forming a refined feature vector. The significance of this step is that reducing the data dimension through dimensional transformation can significantly reduce the computational burden of the subsequent improved long-short term memory network and improve the network's operating efficiency; and retaining key features through feature screening can improve the pertinence and effectiveness of the features, avoid the interference of irrelevant information on the network learning process, thereby improving the accuracy of the network's learning of cell state evolution patterns, and laying the foundation for the efficient and accurate operation of the entire monitoring system.

[0051] During the specific implementation process, the dimensionality transformation stage adopts a linear mapping rule based on principal component analysis, which can achieve dimensionality compression while retaining the original feature information to the maximum extent. First, the covariance matrix of the fused feature vector is calculated to analyze the correlation between the feature dimensions, and then the principal component with the highest contribution is selected based on the size of the eigenvalue. The original 80-dimensional fused feature vector is converted to a 30-dimensional space through this mapping rule. During the conversion process, the relative relationship between the features is strictly maintained to ensure that the converted features can still reflect the inherent structure of the original features. In the feature screening stage, a multi-dimensional screening criterion related to the characteristics of sheep fetal intramuscular preadipocytes is first established: in terms of cell morphology, features with a contour irregularity change rate greater than 8% are retained. This is because a change rate exceeding 8% usually means that the cells may be in a critical stage of morphological remodeling. In terms of cell size, features with a diameter change range of more than 10μm are retained. Changes in this range can effectively reflect the volume growth of cells during proliferation or volume changes during differentiation. In terms of cell texture, fat droplet distribution density fluctuations greater than 2 / 100μm are retained. 2 The amplitude of this fluctuation is typically correlated with the level of intracellular fat anabolism activity. During the screening process, the transformed 30-dimensional feature vectors are individually examined, retaining those that meet the aforementioned criteria and removing those that do not. Ultimately, 15-20 key features are selected from the 30-dimensional transformed features to form refined feature vectors. These refined feature vectors incorporate key information across multiple dimensions, such as cell morphology, size, and texture, while removing redundant and interfering content. This allows for more efficient learning of the subsequently improved long-short-term memory network.

[0052] Step S4: The refined feature vector is input into an improved long short-term memory network. The network improves the gating mechanism and weight update rules of the long short-term memory network based on the dynamic change characteristics of sheep fetal intramuscular preadipocytes. By learning the evolution pattern of cell states over time, the network outputs a cell state feature sequence.

[0053] Specifically, this step inputs the refined feature vector obtained in step S3 into the improved long short-term memory network, and the evolution pattern of the state of the pre-adipocytes in the muscle of sheep fetus is learned through the network. The improved long short-term memory network is directed to the unique dynamic change characteristics of the pre-adipocytes in the muscle of sheep fetus (such as the rapid morphological changes in the proliferation stage, the large accumulation of fat droplets in the differentiation stage, etc.), and the gating mechanism (input gate, forgetting gate, output gate) and weight update rule of the traditional long short-term memory network are targeted and optimized. The improved gating mechanism can more accurately control the input, forgetting and output of information, and the weight update rule can better adapt to the characteristic change intensity of cells in different physiological stages. Its significance is that, through this targeted improvement, the network can more effectively capture the subtle changes and long-term dependencies of cells at continuous time points, accurately learn the entire evolution pattern of cells from the initial state to the proliferation state and then to the differentiation state, and provide accurate and reliable time series feature basis for the determination of the state of cells at each time point in the subsequent steps, thereby improving the analytical ability of the entire monitoring system to the dynamic changes of cells.

[0054] In practice, refined feature vectors are first fed into a modified long-short-term memory network in chronological order. The number of nodes in the network's input layer matches the dimensionality of the refined feature vectors (15-20 dimensions). The network's input gate controls the proportion of the refined feature vector at the current time point that enters the memory cell. This proportion is dynamically adjusted based on the correlation between the current feature and the cell's historical state, and is set between 0.3 and 0.7. When differentiation-related parameters such as fat droplet density within the current feature change significantly, the input proportion is increased to 0.5-0.7 to ensure that key information enters the memory cell. The forget gate regulates the proportion of information retained from the previous time point in the memory cell. This proportion is dynamically adjusted based on the cell's proliferation rate, and is adjusted within a range of 0.2-0.6. When the cell proliferation rate increases by 10%-20% every two hours (a period of rapid proliferation), the forget gate's retention proportion is increased to 0.4-0.6 to retain more historical information related to proliferation. When the proliferation rate drops below 10%, the retention proportion is reduced to 0.2-0.4 to reduce the retention of redundant information. The output gate determines the hidden state information to be output at the next moment based on the current state of the memory unit and the current input features, ensuring that the output information can reflect the current core state of the cell. In terms of weight update rules, the network uses the amount of cellular fat accumulation as a key regulatory parameter. When the amount of fat accumulation increases by 5%-15% every 2 hours (in the active period of fat synthesis), the weight value of features related to fat metabolism increases by 10%-20% to enhance the network's sensitivity to differentiation-related features. Through the synergistic effect of the above-mentioned gating mechanism and weight update rules, after the feature input and processing of 24 time points, the network outputs a cell state feature sequence containing comprehensive state information such as the cell's morphology, size, texture, etc. at each time point. This sequence can clearly reflect the dynamic evolution trajectory of the cell within 48 hours, providing detailed and accurate time series data support for subsequent state determination.

[0055] Step S5: Based on the cell state feature sequence and in combination with pre-set state judgment rules related to the morphology, proliferation, and differentiation of sheep fetal intramuscular preadipocytes, the specific state of the cells at different time points is determined, and a cell state change trajectory is generated;

[0056] Specifically, this step is based on the cell state feature sequence output by step S4, combined with the pre-set multi-dimensional state judgment rules related to the morphological changes, proliferation activities, differentiation processes, etc. of the sheep fetal intramuscular preadipocytes, to judge the specific state of the cells at each time point, and connect these states in chronological order to generate cell state change trajectories. Among them, the state judgment rules are formulated based on the biological characteristics of the sheep fetal intramuscular preadipocytes and a large amount of experimental data, and can accurately distinguish the different physiological states of the cells (such as static state, proliferation state, differentiation state, etc.). The significance of this step is that the abstract, high-dimensional cell state feature sequence is converted into specific, understandable cell state information. Through the generated state change trajectory, the dynamic change process of the cell in a continuous time can be presented intuitively and clearly, including the time nodes of the state transition, the duration of each state, and other key information. It provides a clear and clear context for researchers to understand the physiological activity laws of cells and grasp the growth and development process of cells. It is a key link in connecting feature analysis with actual biological interpretation.

[0057] During the specific implementation process, the pre-stored state judgment rule library is first called, which contains multiple sets of judgment conditions for different cell states. For the judgment of proliferation state, the rule is set as follows: when the increase rate of cell extension area in the cell state feature sequence exceeds 15% every 2 hours, and the reduction rate of contour irregularity every 2 hours exceeds 8%, the cell at that time point is judged to be in a proliferation state. This is because the cell will increase in volume and tend to have regular morphology during the proliferation process. For the judgment of differentiation state, the rule is set as follows: when the increase in the number of fat droplet distribution density exceeds 2 / 100μm every 2 hours, the cell will be judged to be in a proliferation state. 2 , and when the degree of intracellular particle roughness decreases by more than 10 grayscale values ​​every 2 hours, the cell at that time point is judged to be in a differentiated state, which is consistent with the characteristics of large-scale formation of fat droplets and delicate intracellular structure during cell differentiation. The judgment of the static state is the default state when the above-mentioned proliferation and differentiation state judgment conditions are not met. During the judgment process, the above rules are applied one by one to the cell state feature sequence of 24 time points, and each time point obtains a unique state judgment result (proliferation, differentiation or static). Subsequently, the state results of these time points are arranged in chronological order, and the states of adjacent time points are connected by linear interpolation to generate a continuous cell state change trajectory. The trajectory is visualized with time as the horizontal axis and cell state as the vertical axis, clearly marking the specific time periods from the starting time point to the 48-hour time point when the cell is in a proliferation state, differentiation state or static state, as well as the precise time nodes of the transition between different states, providing a structured historical state data foundation for subsequent state prediction.

[0058] Step S6: Based on the cell state change trajectory, the state of the sheep fetus intramuscular preadipocytes in the future unit time is predicted, and the prediction result is output.

[0059] Specifically, this step predicts the state of sheep fetal intramuscular preadipocytes in the future based on the cell state change trajectory generated in step S5. The prediction process is based on the historical laws and intrinsic trends of cell state changes, combined with the current physiological parameters of the cells, to infer the possible states at multiple time points in the future. Its significance lies in that it can provide researchers with forward-looking information on cell development in advance, so that they can predict the proliferation and differentiation trends of cells and promptly discover possible abnormal states or key conversion nodes. This has important guiding value for related biological research (such as exploring the mechanism of fat deposition) and experimental intervention (such as adjusting culture conditions to regulate the direction of cell differentiation), and helps researchers formulate response strategies in advance, optimize experimental plans, improve research efficiency and success rate, and avoid experimental deviations or waste of resources caused by passive responses to cell state changes.

[0060] In the specific implementation process, the cell state change trajectory generated in step S5 is first deeply analyzed to extract the underlying cell state transition patterns. By statistically analyzing the state data at 24 time points, the preceding characteristics of transitions between different states are identified. For example, before transitioning from the proliferation state to the differentiation state, the cell's fat accumulation increases by more than 12% over two consecutive time points (4 hours). Simultaneously, the cell's proliferation rate shows a decreasing trend, from a 15% increase every 2 hours to less than 8% every 2 hours. These preceding characteristics form the basis for determining state transitions. The prediction timeframe is set to the next eight time points (16 hours), consistent with the time interval of historical data (2 hours / time point). During the prediction process, the cell's current physiological parameters are incorporated as initial conditions into the calculation, including key parameters such as current fat accumulation (10-30 μg per cell), proliferation rate (15% every 2 hours), and contour irregularity (0.5-0.6). By matching these current parameters with historical state transition patterns, a prediction algorithm based on state transition probabilities is used to calculate the probability of being in different states at each future time point. For example, if the current amount of fat accumulated in a cell is 20 μg per cell and has increased by 12% for one consecutive time point, the probability of being in a differentiated state at the second time point (four hours later) is predicted to be 60%-80%, and the probability of being in a proliferative state is 20%-40%. The final prediction output includes the state probability distribution at each of the next eight time points, clearly showing the future development trend of the cell state and providing researchers with specific, quantitative, forward-looking reference data.

[0061] Preferably, the time series image multi-focus fusion model is constructed as follows: suppose the input time series image feature vector set is F = {f1, f2, ..., f T}, where T is the number of time points, f t is the eigenvector at time t, and defines the cell feature correlation matrix C, where the element C ij Indicates the degree of correlation between the cell characteristics at time i and time j, Among them, N is the dimension of the feature vector, is the mean of different dimensions of the feature vector at the i-th moment, and the feature vectors at different moments are weighted and fused through the correlation matrix C. The fused feature vector F fusion The m-th dimension element of Among them, the weight w t，m The cell feature importance parameter α is determined by C and the preset cell feature importance parameter α m Jointly determine,

[0062] Specifically, the model achieves more accurate feature fusion by quantifying the degree of correlation between cellular features. The model first determines the number of time points and the feature vectors at each moment in the input time-series image feature vector set. It then defines a cell feature correlation matrix to quantify the correlation between features at different moments. The element values ​​in the correlation matrix reflect the degree of correlation between features at different moments. This correlation is calculated based on the mean value of the feature vector at each moment and the eigenvalues ​​of each dimension. After quantification using a specific method, it provides a basis for subsequent weighted fusion. In implementation, the dimension of the feature vector and the mean value of the feature vector at each moment are first determined. The values ​​of each element in the correlation matrix are then calculated. The weights of the feature vectors at each moment in fusion are determined based on the correlation matrix and a preset cell feature importance parameter. The weights are then weighted and accumulated according to the feature dimension to obtain the element values ​​of each dimension in the fused feature vector. This process fully utilizes the correlation between cellular features at different moments, ensuring that the fused feature vector more comprehensively captures information about cellular dynamics. This overcomes the problem of incomplete information in single-frame images, provides higher-quality fused features for subsequent analysis, and enhances the overall monitoring method's ability to capture cellular dynamics.

[0063] Preferably, in the improved long short-term memory network, the update formula of the memory unit is as follows: Let the input at time t be x t , the hidden state at the previous moment is h t-1 , the memory unit is c t-1 , input gate i t 、Forget Gate t , output gate o t The calculations are i t =σ(W ix x t +Wih h t-1 +b i ),f t =σ(W fx x t +W fh h t-1 +b f ),o t =σ(W ox x t +W oh h t-1 +b o ), where σ is the activation function, W ix 、W ih 、W fx 、W fh 、W ox 、W oh is the weight matrix, b i 、b f 、b o As the bias vector, a new cell state regulator β was added to the sheep fetal intramuscular preadipocytes. t , the memory cell is updated to c t =f t ⊙c t-1 +i t ⊙tanh(W cx x t +β t W ch h t-1 +b c ), where β t The fat accumulation parameter A of the current cell t and proliferation rate parameter P t Through Beta t =γ1A t +γ2P t Determine, γ1, γ2 are preset coefficients, and ⊙ is element-by-element multiplication.

[0064] Specifically, a cell state regulator factor is introduced to adapt to the characteristics of ovine fetal intramuscular preadipocytes. The network retains the basic structure of the input, forget, and output gates of the traditional long-short-term memory network. The values ​​of each gate are calculated through activation functions, weight matrices, and bias vectors to control the input, forgetting, and output of information. Furthermore, the newly added cell state regulator factor is composed of the cell's fat accumulation parameter and proliferation rate parameter combined with preset coefficients to reflect the cell's current physiological state. During implementation, the values ​​of the input, forgetting, and output gates are calculated based on the input refined feature vector, the previous hidden state, the associated weight matrix, and the bias vector. The cell state regulator factor is then calculated based on the current cell's fat accumulation parameter and proliferation rate parameter, combined with preset coefficients. The regulator factor then adjusts the memory unit update process, allowing the memory unit to not only consider the previous memory information and current input information, but also the cell's physiological state parameters. This allows the memory unit update to better adapt to the characteristic changes of ovine fetal intramuscular preadipocytes at different physiological stages, enhancing the network's ability to learn the cell state evolution pattern and improving the accuracy of the cell state feature sequence output.

[0065] Preferably, in the process of dimension transformation and feature screening of the fused feature vector, the dimension transformation adopts a linear transformation based on the principal component analysis of cell morphology, and the fused feature vector F is set to fusion , calculate its covariance matrix S, perform eigenvalue decomposition on S to obtain the eigenvector matrix V, and select the first K eigenvectors corresponding to the preset eigenvalues ​​to form the transformation matrix V K , then the transformed eigenvector During feature screening, the screening thresholds θ1 and θ2 were set based on the fat droplet quantity characteristics and cell boundary clarity characteristics of sheep fetal intramuscular preadipocytes. If the element values ​​corresponding to these two features in the transformed feature vector were greater than θ1 and θ2, respectively, the dimensional feature was retained, otherwise it was discarded.

[0066] Specifically, principal component analysis (PCA) is used to perform dimensionality transformation and feature selection based on cell characteristics. During the dimensionality transformation phase, the covariance matrix of the fused feature vector is first calculated. The covariance matrix reflects the correlation between features across dimensions. The eigenvalue decomposition of the covariance matrix is ​​then performed to obtain an eigenvector matrix. The first several eigenvectors with the largest eigenvalues ​​are selected to form a transformation matrix, which is then used to perform a linear transformation on the fused feature vector to achieve dimensionality reduction. During the feature selection phase, thresholds are set based on the number of fat droplets and cell boundary clarity in ovine fetal intramuscular preadipocytes. The values ​​of each dimension of the transformed feature vector are then evaluated, and dimensional features with values ​​exceeding the threshold are retained. During implementation, the covariance matrix is ​​first calculated and eigenvalue decomposition is performed to determine the appropriate transformation matrix dimensions and construct the transformation matrix. The fused feature vector is then transformed to obtain a reduced-dimensional feature vector. Specific thresholds for fat droplet number and cell boundary clarity are then set. The values ​​of each dimension of the transformed feature vector are then examined one by one, retaining those that meet the threshold requirements and removing those that do not. This process reduces redundant information in features, retains features that are crucial for judging cell status, and improves the efficiency and accuracy of subsequent network learning.

[0067] Preferably, when determining the specific state of cells at different time points, the cell state judgment rule is set to be a combination of multiple state judgment sub-rules, each sub-rule corresponds to a characteristic of the cell, such as the cell proliferation state, the number of cells at the current time point is set to N t , the number of cells at the previous time point is N t-1 ,like The cells are judged to be in a state of rapid proliferation, δ1 is a threshold value set based on the proliferation characteristics of sheep fetal intramuscular preadipocytes; for the cell differentiation state, the expression level of the preset differentiation marker protein in the cell is set as E t , if E t >δ2, the cell is judged to be in a differentiated state. δ2 is a threshold determined based on the differentiation characteristics of the cell. The final state of the cell is determined by combining the judgment results of different sub-rules.

[0068] Specifically, a cell's specific state is determined by combining multiple state-determination sub-rules. Each sub-rule corresponds to a cell characteristic, such as proliferation and differentiation states, each with its own determination criteria. For proliferation, the cell count at the current time point is compared with the previous time point, the percentage change in count is calculated, and then compared with a threshold value set based on the cell's proliferation characteristics. For differentiation, the cell's state is determined by comparing the expression level of specific differentiation marker proteins within the cell with a preset threshold. During implementation, the sub-rules required for cell state determination and the corresponding cell characteristic parameters, such as cell count and differentiation marker protein expression, are first identified. Specific thresholds are then set for each sub-rule. These thresholds are determined based on the biological characteristics of fetal sheep intramuscular preadipocytes. During the determination process, the cell's parameter values ​​at the current time point are obtained and compared with the relevant parameters at the previous time point to calculate the percentage change, or directly compared with the threshold value. The results of each sub-rule are then combined to determine the final cell state at that time point. This multi-rule combination of determinations can more comprehensively reflect the cell's physiological state, avoid the limitations of single-rule determinations, and improve the accuracy of cell state determination.

[0069] Preferably, when predicting the state of sheep fetal intramuscular preadipocytes in the future unit time, a prediction model based on a state transition probability matrix is ​​adopted, and the determined cell state set is S={s1, s2, ..., s M}, element P of the state transition probability matrix P ij Indicates that from state s i Transfer to state s j The probability of P is obtained by statistical analysis of the historical cell state change trajectory. The current cell state is s c , when predicting the state of the next L time steps, the state s of the lth step l Depend on Determine, where s0 = s c , and combined with the influence of the cell state on the external environmental factors E, the state transition probability matrix is ​​modified. ω is the adjustment coefficient, E ij is the parameter of the degree of influence of environmental factors on the transition from state i to state j.

[0070] Specifically, the cell state prediction model and correction process based on the state transition probability matrix are used. This model first determines the set of cell states and the state transition probability matrix. The matrix elements represent the probability of transitioning from one state to another, which is obtained through statistical analysis of historical cell state change trajectories. When predicting future states, starting with the current cell state, the state transition probability matrix is ​​used to determine the state at each future time point. The influence of external environmental factors on state transitions is also considered, and the state transition probability matrix is ​​corrected using adjustment coefficients and environmental factor influence parameters. In implementation, the historical cell state change trajectories are first compiled, the number of transitions between states is counted, the transition probabilities are calculated, and the state transition probability matrix is ​​constructed. The predicted time step and the current cell state are determined. During the prediction process, the future state is initially determined based on the uncorrected matrix. The influence parameters of external environmental factors on each state transition are then obtained. The matrix is ​​corrected using the adjustment coefficients, and the state at each future time point is recalculated using the corrected matrix. This process fully utilizes historical state information and considers the influence of environmental factors, making the prediction results more accurate and improving the accuracy of future cell state predictions.

[0071] Preferably, in step S3, the dimension transformation and feature screening specifically include:

[0072] Step S31: Calculate the covariance matrix of the fused feature vector, which reflects the correlation between features of different dimensions and prepares for subsequent principal component analysis;

[0073] Step S32: Perform eigenvalue decomposition based on the covariance matrix to obtain eigenvectors, and sort the eigenvectors according to the size of the eigenvalues. The larger the eigenvalue, the greater the variance of the data in that direction, and the more information it contains.

[0074] Step S33: Select the first K eigenvectors that exceed the preset eigenvalue to construct a transformation matrix, and use the transformation matrix to perform linear transformation on the fused eigenvector to reduce the dimension and reorganize the features to highlight the abnormal cell features;

[0075] Step S34: According to the calibration characteristics of the fat droplet number and cell boundary clarity of the sheep fetal intramuscular preadipocytes, a screening threshold for different dimensions of the transformed feature vector is set, and the calibration feature dimensions that meet the cell characteristics are retained based on the threshold screening.

[0076] Specifically, the dimensionality transformation and feature screening process involves a step-by-step process. The first step involves calculating the covariance matrix of the fused eigenvectors. The magnitude of the covariance matrix reflects the correlation between the features in each dimension, providing the basis for subsequent principal component analysis and ensuring accurate extraction of key features. The second step involves performing eigenvalue decomposition based on the covariance matrix, obtaining eigenvectors and sorting them by eigenvalue. Larger eigenvalues ​​indicate greater data variance in the direction represented by the corresponding eigenvector and richer information. This step provides a basis for the subsequent selection of key eigenvectors. The third step involves selecting the first few eigenvectors with larger eigenvalues ​​to construct a transformation matrix. This matrix is ​​then used to perform a linear transformation on the fused eigenvectors, achieving dimensionality reduction and feature reorganization. This allows the transformed features to highlight key information and reduce data complexity. The fourth step involves setting a screening threshold based on key characteristics of sheep fetal intramuscular preadipocytes, such as the number of fat droplets and cell boundary clarity. Using this threshold, the transformed eigenvectors are then screened for each dimension, retaining key feature dimensions that reflect cellular characteristics and eliminating redundant dimensions. These steps are interconnected, effectively improving the quality of the feature data through dimensionality reduction and feature screening, laying a solid foundation for subsequent network learning.

[0077] Preferably, in step S4, the improved long short-term memory network learning cell state evolution pattern includes:

[0078] Step S41: The refined feature vectors are sequentially input into the improved long short-term memory network in chronological order. The input gate in the network controls the degree to which the current input information enters the memory unit based on the input features and the hidden state at the previous moment;

[0079] Step S42: The forget gate determines the proportion of the previous moment information to be retained in the memory unit based on the input and the previous moment state, so that the network can remember the long-term state change characteristics of the cell;

[0080] Step S43: By using the cell state regulating factor and combining the current cell fat accumulation and proliferation rate parameters, the degree of integration of new information during memory unit updating is adjusted to make the network better adapted to the characteristics of sheep fetal intramuscular preadipocytes;

[0081] Step S44: The output gate determines the hidden state of the output based on the memory cell state and the current input. The hidden state includes the state evolution information of the cell at the current moment and before, which is used for subsequent cell state judgment and prediction.

[0082] Specifically, the improved long-short-term memory network learns cell state evolution in a step-by-step process. First, the refined feature vector is fed into the network in chronological order. The input gate determines the extent to which the current input information enters the memory cell based on the input features and the previous hidden state, ensuring that important current feature information is effectively received and providing timely input data for network learning. Second, the forget gate determines the proportion of memory cell information retained from the previous moment based on the input and the previous state. This allows the network to appropriately forget irrelevant information while retaining historical information valuable for understanding long-term cell state changes, thus avoiding excessive memory burden. Third, the cell state regulator, combined with parameters such as the current cell's fat accumulation and proliferation rate, adjusts the degree of incorporation of new information during memory cell updates. This ensures that memory cell updates adapt to changes in the cell's physiological state, allowing the network to better capture cell characteristics at different stages. Fourth, the output gate determines the output hidden state based on the memory cell state and the current input. This hidden state integrates current and historical cell state information, providing output data with temporal characteristics for subsequent cell state judgment and prediction. These steps work together to enable the improved network to more accurately learn cell state evolution patterns over time.

[0083] Preferably, in step S5, determining the cell state and generating the change trajectory includes:

[0084] Step S51: Matching the cell state feature sequence with pre-set different state judgment rules. Different judgment rules correspond to different cell characteristics, including proliferation, differentiation, and morphological changes. The specific state category of the cell at each time point is determined through rule matching.

[0085] Step S52: Recording the cell status information at each time point to form a cell status time series, and visually displaying the changes of cell status over time through the time series;

[0086] Step S53: Based on the cell state time series, a curve fitting or path planning algorithm is used to generate a cell state change trajectory. This trajectory can clearly present the cell state transition process and trend, providing a basis for subsequent state prediction.

[0087] Specifically, the sub-steps of step S5 are as follows: the first step is to match the cell state feature sequence with the preset state judgment rules. Each rule corresponds to different characteristics of the cell, such as proliferation, differentiation, morphological changes, etc. The matching process preliminarily determines the state category that the cell may belong to at each time point, providing direction for subsequent accurate judgment. The second step is to record the cell state information at each time point to form a cell state time series. The sequence is arranged in chronological order, clearly showing the state of the cell at different moments, and is the basic data for analyzing cell state changes. The third step is to generate a cell state change trajectory based on the cell state time series using curve fitting or path planning algorithms. The trajectory can intuitively present the transition process and overall trend of the cell state from one time point to another. It can not only reflect the time node of the state transition, but also display information such as the duration of each state. It provides structured historical state data support for subsequent state prediction, so that the prediction process can be based on a clear state evolution trajectory, thereby improving the reliability of the prediction.

[0088] like Figure 2 As shown, a cell morphology dynamic monitoring system integrating time-series image analysis includes:

[0089] The image acquisition and initial feature extraction unit is used to obtain a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, and perform feature extraction on different frame images. The output includes a set of initial feature vectors of cell morphology, size, and texture, which is connected to the subsequent time-series image multi-focus fusion unit to provide input data for it;

[0090] The time series image multi-focus fusion unit receives the feature vector set output by the initial feature extraction unit, fuses the features at different times using the constructed multi-focus fusion model based on the correlation characteristics of the cell features at different times, and outputs the fused feature vector, which is connected to the dimension transformation and feature screening unit to provide the fused feature data;

[0091] The dimensionality transformation and feature screening unit performs dimensionality transformation and feature screening on the fused feature vector, converts it to a preset dimensional space based on the cell characteristic association rules, retains the calibration features, and outputs the refined feature vector to the improved long short-term memory network unit;

[0092] An improved long short-term memory network unit receives the refined feature vector, learns the evolution pattern of the cell state over time through the improved long short-term memory network, and outputs a cell state feature sequence to the cell state determination unit;

[0093] The cell state determination unit determines the specific state of the cell at different time points based on the cell state feature sequence and the preset cell state judgment rules, generates a cell state change trajectory, and transmits the trajectory information to the cell state prediction unit;

[0094] The cell state prediction unit uses a prediction model to predict the state of sheep fetal intramuscular preadipocytes in the future unit time based on the cell state change trajectory and outputs the prediction result.

[0095] A method and system for dynamic monitoring of cell morphology, fused with time-series image analysis, effectively overcomes the inability of conventional image fusion to correlate cell morphological features at different time points through a multi-focus fusion model for time-series images. This model constructs a dynamic weighting mechanism based on the feature correlation characteristics of sheep fetal intramuscular preadipocytes at consecutive time points. This model accurately integrates focused image features at each moment, fully preserving details of the continuous evolution of fat droplets from tiny particles to fused states, as well as the dynamic expansion of cell boundaries during proliferation. This avoids feature fragmentation caused by focus differences, allowing the fused features to more comprehensively reflect the temporal changes in cell morphology.

[0096] The improved long-short-term memory network addresses the difficulty of general models in adapting to the morphological mutation characteristics of cell differentiation. By incorporating parameters such as the fat synthesis rate and intracellular lipid content unique to sheep fetal intramuscular preadipocytes, and adjusting the network gating mechanism and weight update rules, the network's ability to capture nonlinear patterns in cell state evolution is enhanced. When processing the key node where cells transition from a fibroblast-like morphology to an adipocyte morphology, the network effectively balances the weights of historical information and current features, avoiding the forgetting of historical information or the overfitting of current features. This significantly improves the accuracy of cell state time series predictions and provides reliable time series modeling support for accurately tracking dynamic changes in cell morphology.

[0097] In addition, the synergy of dimensional transformation and feature screening in the steps, as well as the targeted design of cell state judgment rules and prediction models for cell characteristics, further enhances the practicality of the system. Dimension transformation and feature screening retain core features based on key cell characteristics to reduce redundant information interference; state judgment and prediction are combined with cell biological characteristics to set rules, making the results more consistent with actual physiological processes. These designs together achieve comprehensive and accurate monitoring of the dynamic changes in the morphology of sheep fetal intramuscular preadipocytes, providing more complete technical support for analyzing the mechanism of intramuscular fat deposition, and effectively making up for the shortcomings of existing technologies in handling time-series monitoring of such cells.

[0098] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0099] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic monitoring of cell morphology by integrating time-series image analysis, characterized in that: The following steps are involved: Step S1: obtaining a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, performing feature extraction on different frame images in the sequence, and obtaining a set of initial feature vectors including cell morphology, size, and texture; Step S2: Input the initial feature vector set into a pre-built time-series image multi-focus fusion model. This model fuses the features at different moments based on the correlation characteristics of the cell features at different moments, outputs a fused feature vector, and enhances the representation of the dynamic changes of the cell features. Step S3: performing dimensionality transformation and feature screening on the fused feature vector, converting it into a preset dimensional space through a preset mapping rule, and retaining the calibration features according to a preset screening criterion associated with the characteristics of sheep fetal intramuscular preadipocytes to form a refined feature vector; Step S4: The refined feature vector is input into an improved long short-term memory network. The network improves the gating mechanism and weight update rules of the long short-term memory network based on the dynamic change characteristics of sheep fetal intramuscular preadipocytes. By learning the evolution pattern of cell states over time, the network outputs a cell state feature sequence. Step S5: Based on the cell state feature sequence and in combination with pre-set state judgment rules related to the morphology, proliferation, and differentiation of sheep fetal intramuscular preadipocytes, the specific state of the cells at different time points is determined, and a cell state change trajectory is generated; Step S6: Based on the cell state change trajectory, the state of the sheep fetus intramuscular preadipocytes in the future unit time is predicted, and the prediction result is output.

2. The method according to claim 1, characterized in that The construction of the time series image multi-focus fusion model is as follows: Assume that the input time series image feature vector set is F = {f1, f2, ..., f T }, where T is the number of time points, f t is the eigenvector at time t, and defines the cell feature correlation matrix C, where the element C ij Indicates the degree of correlation between the cell characteristics at time i and time j, Among them, N is the dimension of the feature vector, is the mean of different dimensions of the feature vector at the i-th moment, and the feature vectors at different moments are weighted and fused through the correlation matrix C. The fused feature vector F fusion The m-th dimension element of Among them, the weight w t，m It is determined by C and the preset cell feature importance parameter αm.

3. The method according to claim 2, characterized in that In the improved long short-term memory network, the update formula of the memory unit is as follows: Let the input at time t be x t , the hidden state at the previous moment is h t-1 , the memory unit is c t-1 , input gate i t 、Forget Gate t , output gate o t The calculations are i t =σ(W ix x t +W ih h t-1 +b i ),f t =σ(W fx x t +W fh h t-1 +b f ),o t =σ(W ox x t +W oh h t-1 +b o ), where σ is the activation function, W ix 、W ih 、W fx 、W fh 、W ox 、W oh is the weight matrix, b i 、b f 、b o As the bias vector, a new cell state regulator β was added to the sheep fetal intramuscular preadipocytes. t , the memory cell is updated to c t =f t ⊙c t-1 +i t ⊙tanh(W cx x t +β t W ch h t-1 +b c ), where β t The fat accumulation parameter A of the current cell t and proliferation rate parameter P t Through Beta t =γ1A t +γ2P t Determine, γ1, γ2 are preset coefficients, and ⊙ is element-by-element multiplication.

4. The method according to claim 1, wherein In the process of dimension transformation and feature screening of the fusion feature vector, the dimension transformation adopts the linear transformation based on the principal component analysis of cell morphology. Let the fusion feature vector F fusion , calculate its covariance matrix S, perform eigenvalue decomposition on S to obtain the eigenvector matrix V, and select the first K eigenvectors corresponding to the preset eigenvalues ​​to form the transformation matrix V K , then the transformed eigenvector During feature screening, the screening thresholds θ1 and θ2 were set based on the fat droplet quantity characteristics and cell boundary clarity characteristics of sheep fetal intramuscular preadipocytes. If the element values ​​corresponding to these two features in the transformed feature vector were greater than θ1 and θ2, respectively, the dimensional feature was retained, otherwise it was discarded.

5. The method according to claim 1, wherein When determining the specific state of cells at different time points, the cell state judgment rule is set to be a combination of multiple state judgment sub-rules, each sub-rule corresponds to a characteristic of the cell, such as the cell proliferation state, the number of cells at the current time point is set to N t , the number of cells at the previous time point is N t-1 ,like The cells are judged to be in a state of rapid proliferation, δ1 is a threshold value set based on the proliferation characteristics of sheep fetal intramuscular preadipocytes; for the cell differentiation state, the expression level of the preset differentiation marker protein in the cell is set as E t , if E t >δ2, the cell is judged to be in a differentiated state. δ2 is a threshold determined based on the differentiation characteristics of the cell. The final state of the cell is determined by combining the judgment results of different sub-rules.

6. The method according to claim 1, wherein When predicting the state of sheep fetal intramuscular preadipocytes in the future unit time, a prediction model based on a state transition probability matrix is ​​used, and the determined cell state set is S={s1, s2, ..., s M }, element P of the state transition probability matrix P ij Indicates that from state s i Transfer to state s j The probability of P is obtained by statistical analysis of the historical cell state change trajectory. The current cell state is s c , when predicting the state of the next L time steps, the state s of the lth step l Depend on Determine, where s0 = s c , and combined with the influence of the cell state on the external environmental factors E, the state transition probability matrix is ​​modified. ω is the adjustment coefficient, E ij is the parameter of the degree of influence of environmental factors on the transition from state i to state j.

7. The method according to claim 1, characterized in that In step S3, the dimension transformation and feature screening specifically include: Step S31: Calculate the covariance matrix of the fused feature vector, and use the covariance matrix to reflect the correlation between features of different dimensions; Step S32: Perform eigenvalue decomposition based on the covariance matrix to obtain eigenvectors, and sort the eigenvectors according to the size of the eigenvalues. The larger the eigenvalue, the greater the variance of the data in that direction, and the more information it contains. Step S33: Select the first K eigenvectors that exceed the preset eigenvalue to construct a transformation matrix, and use the transformation matrix to perform linear transformation on the fused eigenvector to reduce the dimension and reorganize the features to highlight the abnormal cell features; Step S34: According to the calibration characteristics of the fat droplet number and cell boundary clarity of the sheep fetal intramuscular preadipocytes, a screening threshold for different dimensions of the transformed feature vector is set, and the calibration feature dimensions that meet the cell characteristics are retained based on the threshold screening.

8. The method according to claim 1, characterized in that In step S4, the improved long short-term memory network learning cell state evolution pattern includes: Step S41: The refined feature vectors are sequentially input into the improved long short-term memory network in chronological order. The input gate in the network controls the degree to which the current input information enters the memory unit based on the input features and the hidden state at the previous moment; Step S42: The forget gate determines the proportion of the previous moment information to be retained in the memory unit based on the input and the previous moment state, so that the network can remember the long-term state change characteristics of the cell; Step S43: By using the cell state regulating factor and combining the current cell fat accumulation and proliferation rate parameters, the degree of integration of new information during memory unit updating is adjusted to adapt the network to the characteristics of sheep fetal intramuscular preadipocytes; Step S44: The output gate determines the hidden state of the output according to the memory cell state and the current input. The hidden state includes the state evolution information of the cell at the current moment and before.

9. The method according to claim 1, characterized in that In step S5, determining the cell state and generating the change trajectory includes: Step S51: Matching the cell state feature sequence with pre-set different state judgment rules. Different judgment rules correspond to different cell characteristics, including proliferation, differentiation, and morphological changes. The specific state category of the cell at each time point is determined through rule matching. Step S52: Recording the cell status information at each time point to form a cell status time series, and visually displaying the changes of cell status over time through the time series; Step S53: Based on the cell state time series, a curve fitting or path planning algorithm is used to generate a cell state change trajectory, which presents the cell state transition process and trend.

10. A cell morphology dynamic monitoring system integrated with time-series image analysis, characterized in that: include: The image acquisition and initial feature extraction unit is used to obtain a time-series image sequence of sheep fetal intramuscular preadipocytes at different consecutive time points, and perform feature extraction on different frame images. The output includes a set of initial feature vectors of cell morphology, size, and texture, which is connected to the time-series image multi-focus fusion unit to provide input data for it; The time series image multi-focus fusion unit receives the feature vector set output by the initial feature extraction unit, fuses the features at different times using the constructed multi-focus fusion model based on the correlation characteristics of the cell features at different times, and outputs the fused feature vector, which is connected to the dimension transformation and feature screening unit to provide the fused feature data; The dimensionality transformation and feature screening unit performs dimensionality transformation and feature screening on the fused feature vector, converts it to a preset dimensional space based on the cell characteristic association rules, retains the calibration features, and outputs the refined feature vector to the improved long short-term memory network unit; An improved long short-term memory network unit receives the refined feature vector, learns the evolution pattern of the cell state over time through the improved long short-term memory network, and outputs a cell state feature sequence to the cell state determination unit; The cell state determination unit determines the specific state of the cell at different time points based on the cell state feature sequence and the preset cell state judgment rules, generates a cell state change trajectory, and transmits the trajectory information to the cell state prediction unit; The cell state prediction unit uses a prediction model to predict the state of sheep fetal intramuscular preadipocytes in the future unit time based on the cell state change trajectory and outputs the prediction result.

Images