An immune cell state analysis system based on image processing technology

Abstract

The application relates to the fields of biomedical image processing and intelligent control technology, in particular to an immune cell state analysis system based on an image processing technology; the system comprises feature extraction, state evaluation, decision generation and self-adaptive correction modules; the system extracts features by using a space-time graph neural network, the core of which is to calculate visual semantic entropy based on classification probability and feature response field, and to solve decision confidence weight by combining a cell motion diffusion index; accordingly, an AI regulation instruction and a conservative instruction based on a kinetic tolerance boundary are weighted and fused to generate a final instruction and to self-adaptively calibrate boundary parameters according to an observation error; by quantifying visual uncertainty and analyzing motion characteristics, the application effectively overcomes image blurring and non-biological interference, and significantly improves the recognition precision of active cells and the system robustness in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of biomedical image processing and intelligent control technology, specifically to an immune cell state analysis system based on image processing technology. Background Technology

[0002] With the development of computer vision and pattern recognition technologies, intelligent analysis of cell state based on microscopic image sequences has become an important research direction in the field of biomedical engineering. Existing mainstream solutions typically employ convolutional neural networks or deep learning models such as vision to extract features, segment, and classify acquired cell images in order to monitor cell growth status.

[0003] However, existing image processing and analysis technologies face significant challenges in scenarios involving the in vitro expansion and closed-loop regulation of immune cells:

[0004] First, feature extraction and representation of low signal-to-noise ratio images are challenging. Because immune cells are extremely sensitive to light, imaging equipment must operate under extremely low illumination conditions to maintain cell viability, and exposure time is often limited to capture rapid dynamic changes. This results in acquired image sequences generally exhibiting low contrast, high photon shot noise, and complex background textures. Traditional gradient-based edge detection or purely data-driven deep learning models are highly susceptible to noise interference when processing such blurred images, leading to disordered intermediate layer feature response fields and difficulty in extracting robust, discriminative features, thus resulting in false positives or false negatives.

[0005] Secondly, pattern recognition confusion arises from non-biological mimicry. In complex culture environments, numerous non-biological impurities or apoptotic cell debris exist. These interfering elements are highly similar to normal, active immune cells in static visual features, exhibiting severe "inter-class visual homogeneity." Existing image classification algorithms mostly rely on single-frame static features, lacking the ability to deeply mine and jointly analyze the motion patterns of targets over time, making it difficult to effectively eliminate these non-biological mimicry interferences, leading to decreased recognition accuracy.

[0006] Finally, there is the lack of quantification of uncertainty in visual models and the associated decision-making risks. Existing end-to-end image analysis models typically output deterministic classification results, lacking an effective mechanism for measuring the reliability of their own predictions. When the input image has poor imaging quality or is located in an out-of-distribution sample region, the model often gives a high-confidence erroneous prediction. In closed-loop systems involving automated control, this "black box" prediction, lacking physical constraints and confidence assessment, is highly susceptible to generating erroneous control commands, failing to meet the stringent requirements for system robustness and safety boundaries in high-risk biomanufacturing processes.

[0007] Therefore, there is an urgent need to develop an image analysis system that can deeply integrate spatiotemporal visual features and physical dynamic constraints, focusing on solving the problems of feature extraction robustness in low signal-to-noise ratio environments, anti-interference recognition of dynamic patterns, and uncertainty quantification in visual prediction. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides an immune cell state analysis system based on image processing technology. Specifically, the technical solution of this invention includes:

[0009] The feature extraction module is used to acquire immune cell image sequences and extract classification probability distribution vectors and intermediate layer feature response fields using a spatiotemporal graph neural network.

[0010] The state assessment module is used to calculate the visual semantic entropy based on the classification probability distribution vector and the intermediate layer feature response field, and to map the classification probability distribution vector into a cell normalized activity index, and to establish the dynamic tolerance boundary based on the cell normalized activity index.

[0011] The decision generation module is used to calculate the diffusion index based on the immune cell image sequence, combine the visual semantic entropy and the diffusion index to solve the decision confidence weight, obtain the AI ​​regulation command predicted based on the spatiotemporal graph neural network, generate the conservative command based on the dynamic tolerance boundary, and perform weighted fusion of the AI ​​regulation command and the conservative command based on the decision confidence weight to generate the final regulation command.

[0012] The adaptive correction module is used to calibrate the attenuation coefficient in the dynamic tolerance boundary based on the observation error after executing the final control command.

[0013] Preferably, the state assessment module calculates visual semantic entropy, including:

[0014] Call the classification probability distribution vector and the intermediate layer feature response field;

[0015] Calculate the Shannon information entropy based on the classification probability distribution vector;

[0016] Calculate the regularization term based on the gradient of the intermediate layer feature response field;

[0017] Adding the Shannon information entropy to the weighted regularization term yields the visual semantic entropy.

[0018] Preferably, the state assessment module maps the classification probability distribution vector to a normalized cell activity index, including:

[0019] Extract the activation state probability and exhaustion state probability from the classification probability distribution vector;

[0020] Use the probability of the active state as the base value;

[0021] Alternatively, calculate the difference between the probability of the active state and the weighted probability of the exhausted state;

[0022] The baseline value or difference was determined as the normalized cell activity index.

[0023] Preferably, the state assessment module establishes the dynamic tolerance boundary, including:

[0024] Call upon normalized cell activity indices, maximum theoretical activation rate, Michaelis constant, and decay coefficient;

[0025] Based on the Michaelis-Menten kinetic equation, a differential constraint relationship was constructed for the change of normalized cell activity index over time.

[0026] The upper limit of the rate of change determined by the differential constraint relationship is defined as the dynamic tolerance boundary.

[0027] Preferably, the decision generation module calculates the diffusion index, including:

[0028] Calculate the mean square displacement of targets in immune cell image sequences under different time lag windows;

[0029] Establish the functional relationship between the logarithm of the mean square displacement and the logarithm of the time lag window;

[0030] Calculate the derivative of the function to obtain the diffusion index.

[0031] Preferably, the decision generation module calculates the decision confidence weights, including:

[0032] Invoke visual semantic entropy, entropy threshold, diffusion index, and preset sensitivity coefficient;

[0033] Calculate the first Sigmoid function term;

[0034] The first Sigmoid function term is used to characterize the decreasing credibility trend when the visual semantic entropy exceeds the entropy critical threshold.

[0035] Calculate the second Sigmoid function term;

[0036] The second Sigmoid function term is used to characterize the decreasing credibility trend when the diffusion index is less than the non-biological mimicry threshold.

[0037] Calculate the product of the first Sigmoid function term and the second Sigmoid function term;

[0038] The product is determined as the decision confidence weight.

[0039] Preferably, the decision generation module generates the final control instruction, including:

[0040] Invoke decision confidence weights, AI control instructions, and conservative instructions;

[0041] The product of the decision confidence weight and the AI ​​control command is calculated to obtain the aggressive policy component;

[0042] Calculate the difference between 1 and the decision confidence weight;

[0043] Calculate the product of the difference and the conservative instruction to obtain the conservative strategy component;

[0044] The radical strategy component and the conservative strategy component are added together to generate the final control instruction.

[0045] Preferably, the adaptive correction module calibrates the attenuation coefficient in the dynamic tolerance boundary, including:

[0046] Obtain the decay coefficient, adaptive learning rate, theoretical activity value for the next time step, and actual observed activity value for the next time step at the current time step;

[0047] Calculate the deviation between the theoretical activity value and the actual observed activity value;

[0048] The correction amount is obtained by multiplying the bias by the adaptive learning rate;

[0049] Add the current attenuation coefficient to the correction amount to generate the updated attenuation coefficient.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. This system effectively overcomes the challenges of image blurring and non-biological mimicry interference caused by cell phototoxicity limitations by calculating visual semantic entropy and diffusion index. Visual semantic entropy combines the disorder of classification probability with the gradient information of feature response field, which can accurately quantify the uncertainty of visual model and identify adversarial noise. At the same time, by using diffusion index to analyze the motion characteristics of the target in time series, non-biological impurities can be effectively eliminated through differences in physical motion patterns. This not only improves the system's recognition accuracy of active cells, but also enhances its robustness against interference in complex imaging environments.

[0052] 2. This system introduces a tolerance boundary based on the Michaelis-Menten kinetic equation and constructs a dual-strategy fusion mechanism based on decision confidence. The system can generate not only AI-driven regulatory instructions that pursue efficiency, but also conservative instructions that maintain minimum metabolic requirements. By evaluating the credibility of visual perception in real time, the system can prioritize the execution of efficient instructions when the image is clear, and smoothly transition to conservative strategies when the image is blurry or recognition is uncertain. This mechanism achieves a dynamic balance between amplification efficiency and risk control within the safe boundary of ensuring cell survival, avoiding catastrophic consequences caused by misjudgment.

[0053] 3. This system is designed with an adaptive correction module, which realizes online calibration of kinetic parameters through a closed-loop feedback mechanism. The system uses high-confidence observation data to calculate the deviation between theoretical activity and actual observation values, and dynamically adjusts the decay coefficient in the kinetic model accordingly. This mechanism enables the system to self-evolve and continuously adapt to the unique physiological decay characteristics of different patient cell lines, thereby eliminating the fixed deviation between model parameters and actual cell characteristics and ensuring the continuous accuracy of state monitoring during long-term culture.

[0054] 4. This system breaks down the barriers between data-driven approaches and mechanistic models by mapping high-dimensional features extracted by spatiotemporal graph neural networks to normalized physical activity indicators. The feature extraction module uses graph structures to capture the spatiotemporal evolution information of cells, while the state assessment and decision-making module uses physical laws to constrain the prediction space. This deeply coupled architecture retains the advantages of deep learning in complex feature extraction and introduces the interpretability and determinism of biophysics, providing a precise and reliable closed-loop intelligent monitoring solution for in vitro immune cell expansion. Attached Figure Description

[0055] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0056] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0058] Example 1:

[0059] Please see Figure 1 An immune cell status analysis system based on image processing technology, comprising:

[0060] The feature extraction module is used to acquire immune cell image sequences and extract classification probability distribution vectors and intermediate layer feature response fields using a spatiotemporal graph neural network.

[0061] The state assessment module is used to calculate the visual semantic entropy based on the classification probability distribution vector and the intermediate layer feature response field, and to map the classification probability distribution vector into a cell normalized activity index, and to establish the dynamic tolerance boundary based on the cell normalized activity index.

[0062] The decision generation module is used to calculate the diffusion index based on the immune cell image sequence, combine the visual semantic entropy and the diffusion index to solve the decision confidence weight, obtain the AI ​​regulation command predicted based on the spatiotemporal graph neural network, generate the conservative command based on the dynamic tolerance boundary, and perform weighted fusion of the AI ​​regulation command and the conservative command based on the decision confidence weight to generate the final regulation command.

[0063] The adaptive correction module is used to calibrate the attenuation coefficient in the dynamic tolerance boundary based on the observation error after executing the final control command.

[0064] This embodiment provides an immune cell state analysis system based on image processing technology. The system is configured to achieve accurate monitoring and robust regulation of immune cell state in a low signal-to-noise ratio microscopic imaging environment by deeply coupling deep learning visual features with biophysical dynamics models. The system mainly includes a feature extraction module, a state assessment module, a decision generation module, and an adaptive correction module.

[0065] The feature extraction module is the system's perception front end, configured to acquire immune cell image sequences and extract classification probability distribution vectors and intermediate layer feature response fields using a pre-trained spatiotemporal graph neural network. Specifically, this module acquires immune cell image sequences during the in vitro expansion phase using a high-sensitivity microscopic imaging device at a set sampling frequency. Considering cell phototoxicity limitations, the acquired images are typically in low-light and low signal-to-noise ratio conditions. The module inputs this image sequence into the spatiotemporal graph neural network for processing, using graph convolution operators or graph attention mechanisms to perform spatiotemporal aggregation and state updates on node features. Prior to this, the preprocessing unit uses the SLIC superpixel segmentation algorithm to segment each frame of the image. Superpixel nodes Construct the adjacency matrix of the graph structure. Among them, spatial edge Based on the Euclidean distance between superpixel centroids being less than a preset threshold Establish connection, time edge Then connect adjacent frames and The superpixel nodes with the highest spatial overlap are used to form a spatiotemporal graph structure. As input to the network, the network outputs two key data streams: one is the classification probability distribution vector. The elements in this vector represent the posterior probabilities of the cell being in different biological states, such as activation, quiescence, or apoptosis; the second is the intermediate layer feature response field. The response field preserves the spatial distribution information of cell edges, textures, and organelle morphology, which is used for subsequent visual uncertainty quantification.

[0066] The state assessment module is connected to the feature extraction module. Its core function is to quantify the credibility of visual recognition and transform discrete classification results into continuous physical state indicators. This module is configured to calculate visual semantic entropy based on the aforementioned classification probability distribution vector and intermediate layer feature response field to measure the uncertainty of the AI ​​model's judgment of the current image. At the same time, this module maps the multidimensional classification probability distribution vector to a single-dimensional cell normalized activity index and establishes the dynamic tolerance boundary based on the cell normalized activity index. In this way, the system not only identifies the cell category, but also defines the reasonable range of changes in cell life activities through physical constraints.

[0067] The decision generation module is configured to integrate AI visual perception with physical safety constraints to generate robust control strategies. This module calculates the diffusion index based on the input immune cell image sequence to eliminate non-biological mimicry interference. It combines visual semantic entropy and diffusion index to calculate decision confidence weights and quantify the usability of current visual information. It obtains AI control instructions predicted based on spatiotemporal graph neural networks and generates conservative instructions based on dynamic tolerance boundaries. Based on the decision confidence weights, it performs weighted fusion of AI control instructions and conservative instructions to generate the final control instructions.

[0068] The adaptive correction module addresses model mismatch caused by individual differences. This module is configured to calibrate the attenuation coefficient in the kinetic tolerance boundary based on the observation error after executing the final control command. This constitutes a closed-loop feedback mechanism, ensuring that the kinetic model can self-evolve as the cell culture process progresses, always conforming to the true physiological characteristics of the current cell line. Through the synergistic effect of the above modules, this system employs high-precision AI control when the image is clear, and automatically switches to a kinetic safety mode when the image is blurry or there is interference, achieving minimum regret control in high-risk environments.

[0069] Example 2:

[0070] The state assessment module calculates visual semantic entropy, including:

[0071] Call the classification probability distribution vector and the intermediate layer feature response field;

[0072] Calculate the Shannon information entropy based on the classification probability distribution vector;

[0073] Calculate the regularization term based on the gradient of the intermediate layer feature response field;

[0074] Adding the Shannon information entropy to the weighted regularization term yields the visual semantic entropy.

[0075] This embodiment details the specific process by which the state assessment module calculates visual semantic entropy; visual semantic entropy It is a dimensionless physical quantity used to quantify the confidence level of an AI model in judging cell state under low-light interference; the calculation process is as follows:

[0076] The state evaluation module calls the classification probability distribution vector output by the feature extraction module. and intermediate layer characteristic response field Among them, the intermediate layer characteristic response field It involves performing average pooling or computation on the multi-channel feature tensor output by the spatiotemporal graph neural network along the channel dimension. The two-dimensional scalar field obtained after norm is used to characterize spatial location. The overall characteristic response intensity at the location;

[0077] The module calculates the Shannon information entropy based on the classification probability distribution vector; this part is used to measure the degree of disorder in the probability distribution, that is, the degree of hesitation of the model in classifying the class.

[0078] The module computes a regularization term based on the gradient of the intermediate layer feature response field; this step utilizes the gradient's... Norm-constrained feature field spatial continuity aims to identify adversarial noise or artifacts that have high classification confidence but whose feature space is extremely unnatural.

[0079] Considering that the images processed by computers are discrete pixel matrices, the continuous integral is modified to a discrete summation, and the gradient operator is modified to a discrete difference operator. The modified formula is:

[0080] ;

[0081] in, These are the height and width of the feature map, respectively; Indicates pixel coordinates The square of the gradient magnitude at a given point is specifically determined by applying the Sobel or Laplacian operator to the feature response field. Obtained by performing convolution calculations:

[0082] ;

[0083] in, This indicates the total number of preset cell state categories; This represents the probability distribution vector belonging to the first category. The posterior probability of a class state, satisfying the normalization condition; The regularization weight coefficient is calibrated by benchmark testing of clean samples and noisy samples, and the value that maximizes the distinguishability of the entropy values ​​of the two types of samples is selected. Represents the two-dimensional spatial domain defined by the image; Indicates position Normalized characteristic response intensity at the location; This is a gradient operator; this calculation method can effectively distinguish between normal cell morphological changes and characteristic mutations caused by non-biological noise.

[0084] Example 3:

[0085] The state assessment module maps the classification probability distribution vector to normalized cell activity indices, including:

[0086] Extract the activation state probability and exhaustion state probability from the classification probability distribution vector;

[0087] Use the probability of the active state as the base value;

[0088] Alternatively, calculate the difference between the probability of the active state and the weighted probability of the exhausted state;

[0089] The baseline value or difference was determined as the normalized cell activity index.

[0090] This embodiment details the process by which the state assessment module maps the classification probability distribution vector to a normalized cell activity index; the normalized cell activity index... It is a single-dimensional physical quantity that characterizes the current vitality level of a cell and is used for subsequent dynamic modeling;

[0091] The module extracts the activation state probabilities from the classification probability distribution vector. With exhaustion state probability ;

[0092] The mapping strategy is selected based on the preset system configuration; in one implementation, the activation state probability is directly used as the base value, i.e. In a preferred embodiment, in order to more comprehensively reflect the cell state, the difference between the probability of the activated state and the weighted probability of the exhausted state is calculated.

[0093] This baseline value or difference was determined as an indicator of cell normalization activity. When using difference calculations, to prevent physically meaningless negative values ​​in the activity index due to high depletion probability, a nonlinear correction unit needs to be introduced; the corrected calculation formula is as follows:

[0094] ;

[0095] in, This indicates the operation of obtaining non-negative values; this step ensures... This ensures that the denominator is correct when substituting into the Michaelis equation later. Always greater than zero to avoid calculation divergence; The weighting coefficients, whose values ​​are determined by biological experiments, are used to balance the contribution weights of activating and depleting factors to overall activity; through this mapping, high-dimensional classification probabilities are transformed into scalar indicators suitable for physical modeling.

[0096] Example 4:

[0097] The state assessment module establishes the dynamic tolerance boundary, including:

[0098] Call upon normalized cell activity indices, maximum theoretical activation rate, Michaelis constant, and decay coefficient;

[0099] Based on the Michaelis-Menten kinetic equation, a differential constraint relationship was constructed for the change of normalized cell activity index over time.

[0100] The upper limit of the rate of change determined by the differential constraint relationship is defined as the dynamic tolerance boundary.

[0101] This embodiment details the specific process by which the state assessment module establishes the kinetic tolerance boundary; the kinetic tolerance boundary defines the physical limit of cell activity changes, and any mutation exceeding this boundary is considered abnormal;

[0102] The module calls the cell normalized activity index at the current moment. And pre-stored biophysical parameters: maximum theoretical activation rate Michaelis constant and the current attenuation coefficient ;in, and It is a constant obtained by performing nonlinear least squares fitting on the historical growth curve of the cell line under ideal culture conditions;

[0103] Based on the Michaelis-Menten equation, a differential constraint relationship was constructed to describe the change of normalized cell activity index over time. This relationship describes the competitive mechanism between the cell's ability to increase activity under ideal conditions and its natural decline.

[0104] The upper limit of the rate of change determined by this differential constraint relationship is defined as the dynamic tolerance boundary; specifically, this boundary is described by the following differential inequality:

[0105] ;

[0106] in, The boundary represents time; it acts as the physical fence of the system. When the prediction result of the visual model suggests that the rate of change in cell activity exceeds the calculated value on the right side of the inequality, the system determines that the prediction violates biological common sense, thereby triggering a conservative control strategy.

[0107] Example 5:

[0108] The decision generation module calculates the diffusion index, including:

[0109] Calculate the mean square displacement of targets in immune cell image sequences under different time lag windows;

[0110] Establish the functional relationship between the logarithm of the mean square displacement and the logarithm of the time lag window;

[0111] Calculate the derivative of the function to obtain the diffusion index.

[0112] This embodiment details the specific process by which the decision generation module calculates the diffusion index; diffusion index Used to distinguish living cells from non-biological particles by physical movement patterns;

[0113] The module calculates targets in immune cell image sequences at different time lag windows. The mean square displacement below;

[0114] Establish the functional relationship between the logarithm of the mean square displacement and the logarithm of the time lag window;

[0115] Calculate the derivative of this function to obtain the diffusion index. Specifically, a time lag window is selected. logarithm The x-axis represents the logarithm of the mean square displacement. Using the ordinate as the vertical axis, a linear regression is performed on the data points using the least squares method. The slope of the fitted line is the diffusion index; its calculation formula is:

[0116] ;

[0117] in, Indicates the target at time. Position vector; Indicates the ensemble average; and These are the normalization constants per unit length and per unit time, respectively, used to ensure that the domain of the logarithmic function is dimensionless; when the calculated... When the target motion conforms to Brownian motion characteristics, it is determined to be a non-biological impurity; when When the target exhibits active superdiffusion characteristics, it is identified as an active cell.

[0118] Example 6:

[0119] The decision generation module calculates the decision confidence weights, including:

[0120] Invoke visual semantic entropy, entropy threshold, diffusion index, and preset sensitivity coefficient;

[0121] Calculate the first Sigmoid function term;

[0122] The first Sigmoid function term is used to characterize the decreasing credibility trend when the visual semantic entropy exceeds the entropy critical threshold.

[0123] Calculate the second Sigmoid function term;

[0124] The second Sigmoid function term is used to characterize the decreasing credibility trend when the diffusion index is less than the non-biological mimicry threshold.

[0125] Calculate the product of the first Sigmoid function term and the second Sigmoid function term;

[0126] The product is determined as the decision confidence weight.

[0127] This embodiment details the specific process of the decision generation module calculating the decision confidence weights; decision confidence weights Used to determine the extent to which the system adopts the AI's predictions;

[0128] Module calls visual semantic entropy Entropy critical threshold diffusion index and preset sensitivity coefficient and ;in, It is the 95th percentile of the entropy distribution obtained from historical failure case statistics; and It is a positive real number used to adjust the steepness of the Sigmoid function; The value of is set to the reciprocal of the standard deviation of the visual semantic entropy of high signal-to-noise ratio samples in the historical validation set, i.e. , The value of is set to the reciprocal of the standard deviation of the diffusion index of non-biological mimicry interference samples, i.e. This achieves a normalized mapping of the data;

[0129] Calculate the first Sigmoid function term, which is used to characterize the decreasing confidence trend when the visual semantic entropy exceeds the entropy critical threshold;

[0130] Calculate the second Sigmoid function term, which characterizes the decreasing credibility trend when the diffusion index is less than the non-biological mimicry threshold;

[0131] Calculate the product of the first Sigmoid function term and the second Sigmoid function term, and determine this product as the decision confidence weight. The calculation formula is as follows:

[0132] ;

[0133] This weight The range of values ​​is When the image quality is high and the target is confirmed to be a live cell, It approaches 1; conversely, when there is high uncertainty or mimicry interference, Approaching 0.

[0134] Example 7:

[0135] The decision generation module generates the final control instructions, including:

[0136] Invoke decision confidence weights, AI control instructions, and conservative instructions;

[0137] The product of the decision confidence weight and the AI ​​control command is calculated to obtain the aggressive policy component;

[0138] Calculate the difference between 1 and the decision confidence weight;

[0139] Calculate the product of the difference and the conservative instruction to obtain the conservative strategy component;

[0140] The radical strategy component and the conservative strategy component are added together to generate the final control instruction.

[0141] This embodiment details the specific process by which the decision generation module generates the final control instruction; this process is based on the Bayesian model averaging concept to achieve risk minimization control;

[0142] Module call decision confidence weight AI control instructions and conservative instructions Among them, the control command is defined as 3D control vector These correspond to the culture medium pump speed, gas flow rate, and stirring speed, respectively.

[0143] It is the output of the neural network. dimensional vector; Based on current activity The minimum metabolic requirement can be obtained by looking up a table or by calculation. dimensional reference vector; where, Based on current image features Radical strategies predicted by neural networks;

[0144] It is a safety strategy that maintains the minimum metabolic requirements of cells, calculated based on the kinetic tolerance boundary.

[0145] The product of the decision confidence weight and the AI ​​control command is calculated to obtain the aggressive policy component;

[0146] Simultaneously, the difference between 1 and the decision confidence weight is calculated, and the product of this difference and the conservative instruction is calculated to obtain the conservative policy component;

[0147] The aggressive strategy component and the conservative strategy component are added together to generate the final regulatory instruction. The calculation formula is:

[0148] ;

[0149] This formula allows the system to prioritize AI instructions to improve amplification efficiency when confidence is high, and smoothly transition to conservative instructions when confidence is low to ensure cell survival, thereby avoiding catastrophic consequences caused by misjudgment.

[0150] Example 8:

[0151] The adaptive correction module calibrates the attenuation coefficient in the dynamic tolerance boundary, including:

[0152] Obtain the decay coefficient, adaptive learning rate, theoretical activity value for the next time step, and actual observed activity value for the next time step at the current time step;

[0153] Calculate the deviation between the theoretical activity value and the actual observed activity value;

[0154] The correction amount is obtained by multiplying the bias by the adaptive learning rate;

[0155] Add the current attenuation coefficient to the correction amount to generate the updated attenuation coefficient.

[0156] This embodiment details the specific process by which the adaptive correction module calibrates the attenuation coefficient in the kinetic tolerance boundary; this process is used to eliminate the deviation between model parameters and actual cell characteristics.

[0157] The module obtains the attenuation coefficient at the current moment. and preset adaptive learning rate The system utilizes the differential dynamics relationship established in Example 4 to construct a discretized state predictor; it assumes that under ideal conditions without abnormal disturbances, the evolution of cell activity follows the equality boundary of the differential equation; discretization is performed using the Euler method, and the theoretical activity value at the next time step is obtained. The calculation is as follows:

[0158] ;

[0159] in, The current observation value. The sampling time interval; the theoretical activity value for the next time step based on the current model prediction. and the actual observed activity value at the next moment. Among them, the actual observed activity value The acquisition employs a self-supervised pseudo-label strategy: the system filters out... Visual semantic entropy at time step Below the preset minimum threshold and diffusion index High-confidence activity frames; for these specific frames, the system determines that the AI ​​model's predictions accurately reflect the real biological state, and directly uses the normalized cell activity index output by the model at that moment as the benchmark true value for physical model calibration. ;

[0160] Calculate the deviation between the theoretical activity value and the actual observed activity value;

[0161] Calculate the product of this bias and the adaptive learning rate to obtain the correction amount;

[0162] Add the current decay coefficient to the correction amount to generate the updated decay coefficient. The parameter update law is as follows:

[0163] ;

[0164] like This indicates that the model prediction was too optimistic and the actual decay was faster, so the system automatically increases the decay coefficient; conversely, it decreases it. Through this iterative update, the system can learn online and adapt to the specific natural decay rate of different patient cells.

[0165] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An immune cell status analysis system based on image processing technology, characterized in that, include: The feature extraction module is used to acquire immune cell image sequences and extract classification probability distribution vectors and intermediate layer feature response fields using a spatiotemporal graph neural network. The state assessment module is used to calculate the visual semantic entropy based on the classification probability distribution vector and the intermediate layer feature response field, and to map the classification probability distribution vector into a cell normalized activity index, and to establish the dynamic tolerance boundary based on the cell normalized activity index. The decision generation module is used to calculate the diffusion index based on the immune cell image sequence, combine the visual semantic entropy and the diffusion index to solve the decision confidence weight, obtain the AI ​​regulation command predicted based on the spatiotemporal graph neural network, generate the conservative command based on the dynamic tolerance boundary, and perform weighted fusion of the AI ​​regulation command and the conservative command based on the decision confidence weight to generate the final regulation command. The adaptive correction module is used to calibrate the attenuation coefficient in the dynamic tolerance boundary based on the observation error after executing the final control command. The state assessment module calculates visual semantic entropy, including: Call the classification probability distribution vector and the intermediate layer feature response field; Calculate the Shannon information entropy based on the classification probability distribution vector; Calculate the regularization term based on the gradient of the intermediate layer feature response field; Adding the Shannon information entropy to the weighted regularization term yields the visual semantic entropy; The decision generation module calculates the diffusion index, including: Calculate the mean square displacement of targets in immune cell image sequences under different time lag windows; Establish the functional relationship between the logarithm of the mean square displacement and the logarithm of the time lag window; Calculate the derivative of the function to obtain the diffusion index.

2. The immune cell status analysis system based on image processing technology according to claim 1, characterized in that, The state assessment module maps the classification probability distribution vector to a normalized cell activity index, including: Extract the activation state probability and exhaustion state probability from the classification probability distribution vector; Use the probability of the active state as the base value; Alternatively, calculate the difference between the probability of the active state and the weighted probability of the exhausted state; The baseline value or difference was determined as the normalized cell activity index.

3. The immune cell status analysis system based on image processing technology according to claim 1, characterized in that, The state assessment module establishes the dynamic tolerance boundary, including: Call upon normalized cell activity indices, maximum theoretical activation rate, Michaelis constant, and decay coefficient; Based on the Michaelis-Menten kinetic equation, a differential constraint relationship was constructed for the change of normalized cell activity index over time. The upper limit of the rate of change determined by the differential constraint relationship is defined as the dynamic tolerance boundary.

4. The immune cell status analysis system based on image processing technology according to claim 1, characterized in that, The decision generation module calculates the decision confidence weights, including: Invoke visual semantic entropy, entropy threshold, diffusion index, and preset sensitivity coefficient; Calculate the first Sigmoid function term; The first Sigmoid function term is used to characterize the decreasing credibility trend when the visual semantic entropy exceeds the entropy critical threshold. Calculate the second Sigmoid function term; The second Sigmoid function term is used to characterize the decreasing credibility trend when the diffusion index is less than the non-biological mimicry threshold. Calculate the product of the first Sigmoid function term and the second Sigmoid function term; The product is determined as the decision confidence weight.

5. The immune cell status analysis system based on image processing technology according to claim 1, characterized in that, The decision generation module generates the final control instruction, including: Invoke decision confidence weights, AI control instructions, and conservative instructions; The product of the decision confidence weight and the AI ​​control command is calculated to obtain the aggressive policy component; Calculate the difference between 1 and the decision confidence weight; Calculate the product of the difference and the conservative instruction to obtain the conservative strategy component; The radical strategy component and the conservative strategy component are added together to generate the final control instruction.

6. The immune cell status analysis system based on image processing technology according to claim 1, characterized in that, The adaptive correction module calibrates the attenuation coefficient in the dynamic tolerance boundary, including: Obtain the decay coefficient, adaptive learning rate, theoretical activity value for the next time step, and actual observed activity value for the next time step at the current time step; Calculate the deviation between the theoretical activity value and the actual observed activity value; The correction amount is obtained by multiplying the bias by the adaptive learning rate; Add the current attenuation coefficient to the correction amount to generate the updated attenuation coefficient.

Images