Method for detecting PD-L1 expression in circulating tumor cells

Abstract

The invention provides a method for detecting PD-L1 expression in circulating tumor cells, and belongs to the technical field of bioinformation.The method comprises the steps that peripheral blood samples are collected from at least 100 groups of patients, the circulating tumor cells are separated through density gradient centrifugation, specific labeling is carried out through anti-epithelial cell adhesion molecule antibodies and anti-PD-L1 fluorescence labeling antibodies, and then the peripheral blood samples are subjected to fluorescence labeling; then, a flow cytometer is used for obtaining multi-dimensional fluorescence signals for quantitative analysis, and a multi-dimensional characteristic data set containing time characteristics, intensity characteristics and spatial distribution characteristics is constructed; based on the historical data, a multi-head attention enhanced adversarial generative network model containing eight independent attention heads is constructed and trained, key hyper-parameters in the training process are dynamically adjusted through a parameter optimization function, and finally the expression level of the circulating tumor cells PD-L1 in a to-be-detected sample is quantitatively analyzed by applying the trained model. The high-precision detection of the expression of the circulating tumor cells PD-L1 is realized.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and specifically relates to a method for detecting PD-L1 expression in circulating tumor cells. Background Technology

[0002] Circulating tumor cells (CTCs) are important biomarkers in liquid biopsies, and the expression level of PD-L1 on their surface is of significant clinical importance for assessing the tumor immune microenvironment, predicting the efficacy of immune checkpoint inhibitors, and monitoring treatment response. Traditional methods for detecting PD-L1 expression in CTCs mainly rely on techniques such as immunohistochemistry, immunofluorescence staining, and flow cytometry. Immunohistochemistry uses specific antibodies to recognize PD-L1 protein and produce a colorimetric reaction for semi-quantitative analysis; immunofluorescence staining uses fluorescently labeled antibodies to bind PD-L1 molecules for observation under a fluorescence microscope; and flow cytometry assesses PD-L1 expression levels based on fluorescence signal intensity.

[0003] However, the above-mentioned detection methods face serious limitations when applied to the analysis of PD-L1 expression in circulating tumor cells: First, circulating tumor cells are extremely rare in peripheral blood (usually only 1-10 per milliliter of blood), and existing detection technologies have limited analytical capabilities for such low-abundance samples; second, traditional flow cytometry and immunofluorescence analysis mainly rely on a single fluorescence intensity threshold to determine PD-L1 positive cells, which cannot comprehensively capture the complex patterns of PD-L1 expression; third, existing analytical methods mostly rely on human experience for interpretation, and the detection results are greatly affected by subjective factors, resulting in poor comparability of results between laboratories.

[0004] With the widespread use of immune checkpoint inhibitors in tumor treatment, accurate assessment of PD-L1 expression status in circulating tumor cells has become particularly important. However, the low detection accuracy of existing technologies has become a key bottleneck restricting the clinical application of liquid biopsy, and there is an urgent need to develop new detection methods to improve the accuracy of PD-L1 expression detection in circulating tumor cells. Summary of the Invention

[0005] In view of this, the present invention provides a method for detecting PD-L1 expression in circulating tumor cells, which can solve the technical problem of low accuracy in detecting PD-L1 expression in circulating tumor cells in the prior art.

[0006] This invention is implemented as follows: A method for detecting PD-L1 expression in circulating tumor cells includes the following steps: collecting peripheral blood samples from patients and separating circulating tumor cells by density gradient centrifugation; labeling specific markers on the surface of circulating tumor cells using an anti-epithelial cell adhesion molecule antibody; adding an anti-PD-L1 fluorescently labeled antibody to bind to PD-L1 molecules on the surface of circulating tumor cells; analyzing the labeled cells using flow cytometry and collecting historical detection data; constructing a multi-head attention-enhanced adversarial generative network (MGGN) model based on the historical detection data, comprising a generator and a discriminator; setting an initial generation function for the MGGN model and optimizing the network parameters using a parameter optimization function and a backpropagation algorithm; using the trained MGGN model to quantitatively analyze the PD-L1 expression level in the tested circulating tumor cell samples; and calculating the PD-L1 expression positivity rate in the tested samples based on the output of the MGGN model.

[0007] In the step of collecting peripheral blood samples from patients and performing density gradient centrifugation to separate circulating tumor cells, the number of peripheral blood samples from patients is at least 100 groups. The at least 100 groups of peripheral blood samples from patients refers to the minimum number of samples collected from patients with different cancer types, ensuring that the dataset has sufficient statistical significance and covers the diversity of PD-L1 expression patterns in circulating tumor cells in different cancer types.

[0008] The steps of analyzing labeled cells and collecting historical detection data using flow cytometry specifically include: setting flow cytometry detection parameters and adjusting the fluorescence compensation matrix to eliminate signal interference between different fluorescence channels; quantifying the duration of cell fluorescence signals using a long fluorescence intensity feature extraction equation; quantifying the peak intensity of cell fluorescence signals using a short fluorescence intensity feature extraction equation; quantifying the continuous distribution of cell fluorescence signals using a continuous fluorescence intensity feature extraction equation; quantifying the discontinuous distribution of cell fluorescence signals using a interval fluorescence intensity feature extraction equation; and integrating all feature data to form a multidimensional feature dataset of PD-L1 expression in circulating tumor cells.

[0009] The fluorescence intensity long feature extraction equation is used to quantify the duration characteristics of PD-L1 fluorescence signal in circulating tumor cells. The inputs include the original fluorescence signal time series data, signal acquisition frequency parameters, excitation wavelength values ​​of fluorescently labeled antibody and emission wavelength values ​​of fluorescently labeled antibody. The output is a feature vector characterizing the duration of PD-L1 fluorescence signal for feature data integration.

[0010] The fluorescence intensity short feature extraction equation is used to quantify the peak intensity characteristics of the PD-L1 fluorescence signal in circulating tumor cells. The inputs include the original fluorescence signal intensity data, signal gain coefficient, background fluorescence threshold and fluorescently labeled antibody concentration value. The output is a feature vector characterizing the peak distribution of the PD-L1 fluorescence signal for feature data integration.

[0011] The fluorescence intensity continuous feature extraction equation is used to quantify the spatial continuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data, spatial resolution parameters, cell surface area values, and continuity threshold parameters. The output is a feature vector characterizing the spatial continuous distribution of PD-L1 fluorescence signals for feature data integration.

[0012] The fluorescence intensity interval feature extraction equation is used to quantify the spatial discontinuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data, signal segmentation threshold, cell morphology parameters and minimum interval distance parameters. The output is a feature vector characterizing the spatial discontinuous distribution of PD-L1 fluorescence signals for feature data integration.

[0013] The specific structure of the multi-head attention-enhanced adversarial generative network model is a deep learning model based on a generator and discriminator architecture. In the generator part, a multi-head attention mechanism is introduced to process the high-dimensional feature information of circulating tumor cells. Eight independent attention heads focus on the PD-L1 expression location distribution features, fluorescence intensity distribution features, cell morphology features, background noise features, surface marker co-expression features, cell population heterogeneity features, fluorescence signal attenuation features, and light scattering features, respectively. Each attention head outputs a 64-dimensional feature vector, which is then connected to form a 512-dimensional feature vector. Subsequently, it is dimensionality reduced by three fully connected layers to generate a PD-L1 expression feature map of circulating tumor cells. The discriminator adopts a five-layer convolutional neural network structure to distinguish between real samples and generated samples. The final layer of the network uses Wasserstein distance as a loss function to evaluate the quality of generated samples.

[0014] The parameter optimization function is used to dynamically adjust the key hyperparameters during the training process of the multi-head attention-enhanced adversarial generative network. The inputs include the current training round number, generator loss function value, discriminator loss function value, sample batch size value, and initial learning rate value. The outputs are the optimized learning rate value used for backpropagation algorithm, the batch weight coefficient value used for model training, the attention head weight coefficient value used for multi-head attention mechanism adjustment, and the regularization coefficient value used to prevent model overfitting.

[0015] The steps for establishing the training dataset during the pre-training process of the multi-head attention-enhanced adversarial generative network model specifically include: collecting peripheral blood samples from 300 patients with different cancer types and isolating circulating tumor cells; using standardized flow cytometry to obtain PD-L1 expression data in the circulating tumor cells of each patient; dividing all patient data into training and validation sets, with the training set containing data from 240 patients for model parameter learning and the validation set containing data from 60 patients for model performance evaluation; preprocessing the training set data including outlier removal, data standardization, data augmentation, and feature selection; and finally constructing a PD-L1 expression pattern dataset for circulating tumor cells containing multidimensional features.

[0016] Compared with existing technologies, this invention provides a method for detecting PD-L1 expression in circulating tumor cells. This invention proposes a method for detecting PD-L1 expression in circulating tumor cells based on a multi-head attention-enhanced adversarial generative network, which integrates the advantages of traditional flow cytometry and advanced deep learning technology. Through a systematically designed feature extraction equation encompassing four categories—long-term, short-term, continuous, and interval features—a comprehensive quantification of the temporal, intensity, and spatial distribution characteristics of PD-L1 expression in circulating tumor cells is achieved, constructing a multi-dimensional feature dataset that provides rich training material for deep learning models.

[0017] Compared to traditional detection methods, this invention significantly improves the accuracy of PD-L1 expression detection in circulating tumor cells. The multi-head attention mechanism enables the model to simultaneously focus on different dimensions of PD-L1 expression features, including eight aspects such as location distribution, intensity distribution, and cell morphology, greatly enhancing its ability to recognize complex expression patterns. The adversarial generative network architecture continuously optimizes feature extraction through an adversarial learning process between the generator and discriminator, enabling the model to accurately distinguish PD-L1 expression signals with different intensities and distribution patterns, effectively reducing false positives and false negatives.

[0018] By introducing an automated deep learning analysis process, this invention eliminates the subjective factors of manual interpretation, achieving high consistency and reproducibility of results. The application of parameter optimization functions ensures that the model maintains stable performance even when faced with complex clinical samples, adapting to biological differences among different patient samples. This method successfully solves the technical problem of low accuracy in detecting PD-L1 expression in circulating tumor cells, providing reliable technical support for precision medicine in immunotherapy. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2This is a graph showing the trend of loss changes during the training process of the multi-head attention-enhanced adversarial generative network model in Example 2.

[0021] Figure 3 This is a comparison chart of the weight coefficients of each attention head in the multi-head attention mechanism of Example 2.

[0022] Figure 4 The graph shows the correlation between fluorescence intensity and PD-L1 expression intensity in Example 2. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shows a flowchart of a method for detecting PD-L1 expression in circulating tumor cells provided by the present invention. This method includes the following steps:

[0025] S01. Collect at least 100 sets of peripheral blood samples from patients and perform density gradient centrifugation to separate circulating tumor cells;

[0026] S02. Using anti-epithelial cell adhesion molecule antibodies to label specific markers on the surface of circulating tumor cells;

[0027] S03, Add anti-PD-L1 fluorescently labeled antibody to bind to PD-L1 molecules on the surface of circulating tumor cells;

[0028] S04. Use flow cytometry to analyze the labeled cells and collect historical test data;

[0029] S05. Construct a multi-head attention-enhanced adversarial generative network model based on historical detection data, which includes a network structure consisting of a generator and a discriminator.

[0030] S06. Set an initial generation function for the multi-head attention-enhanced adversarial generative network model and apply a parameter optimization function to optimize the network parameters through the backpropagation algorithm.

[0031] S07. Quantitative analysis of PD-L1 expression level in the tested circulating tumor cell samples was performed using a trained multi-head attention-enhanced adversarial generative network model.

[0032] S08. Calculate the positive rate of PD-L1 expression in circulating tumor cells in the test sample based on the output results of the multi-head attention-enhanced adversarial generative network model.

[0033] Among them, at least 100 peripheral blood samples from patients refers to the minimum number of samples collected from patients with different cancer types to ensure that the dataset has sufficient statistical significance and covers the diversity of PD-L1 expression patterns in circulating tumor cells in different cancer types.

[0034] Circulating tumor cells (CTCs) are tumor cells that detach from the primary tumor or metastatic lesions and enter the peripheral blood circulation. They are important biomarkers in liquid biopsy.

[0035] PD-L1 refers to programmed death-ligand 1, an immune checkpoint molecule expressed on the surface of tumor cells. It inhibits T cell immune function after binding to PD-1 on the surface of T cells.

[0036] The steps of analyzing labeled cells and collecting historical detection data using flow cytometry specifically include: setting flow cytometry detection parameters and adjusting the fluorescence compensation matrix to eliminate signal interference between different fluorescence channels; quantifying the duration of cell fluorescence signals using a long fluorescence intensity feature extraction equation; quantifying the peak intensity of cell fluorescence signals using a short fluorescence intensity feature extraction equation; quantifying the continuous distribution of cell fluorescence signals using a continuous fluorescence intensity feature extraction equation; quantifying the discontinuous distribution of cell fluorescence signals using a interval fluorescence intensity feature extraction equation; and integrating all feature data to form a multidimensional feature dataset of PD-L1 expression in circulating tumor cells.

[0037] The fluorescence intensity long feature extraction equation is used to quantify the duration characteristics of the PD-L1 fluorescence signal in circulating tumor cells. The inputs include the original fluorescence signal time series data obtained from step S04.1, the signal acquisition frequency parameters set from step S04.1, the excitation wavelength values ​​of the fluorescently labeled antibody used in step S03, and the emission wavelength values ​​of the fluorescently labeled antibody used in step S03. The output is a feature vector characterizing the duration of the PD-L1 fluorescence signal, which is used for feature data integration in step S04.6.

[0038] The fluorescence intensity short feature extraction equation is used to quantify the peak intensity characteristics of the PD-L1 fluorescence signal in circulating tumor cells. The inputs include the raw fluorescence signal intensity data obtained from step S04.1, the signal gain coefficient set from step S04.1, the background fluorescence threshold set from step S04.1, and the concentration value of the fluorescently labeled antibody used in step S03. The output is a feature vector characterizing the peak distribution of the PD-L1 fluorescence signal, which is used for feature data integration in step S04.6.

[0039] The fluorescence intensity continuous feature extraction equation is used to quantify the spatial continuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data obtained from step S04.1, spatial resolution parameters set from step S04.1, cell surface area values ​​marked from step S02, and continuity threshold parameters set from step S04.1. The output is a feature vector characterizing the spatial continuous distribution of PD-L1 fluorescence signals, which is used for feature data integration in step S04.6.

[0040] The fluorescence intensity interval feature extraction equation is used to quantify the spatial discontinuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data obtained from step S04.1, signal segmentation threshold set from step S04.1, cell morphology parameters marked from step S02, and minimum interval distance parameters set from step S04.1. The output is a feature vector characterizing the spatial discontinuous distribution of PD-L1 fluorescence signals, which is used for feature data integration in step S04.6.

[0041] The specific structure of the multi-head attention-enhanced adversarial generative network model is a deep learning model based on a generator and discriminator architecture. In the generator part, a multi-head attention mechanism is introduced to process the high-dimensional feature information of circulating tumor cells. Eight independent attention heads focus on the PD-L1 expression location distribution features, fluorescence intensity distribution features, cell morphology features, background noise features, surface marker co-expression features, cell population heterogeneity features, fluorescence signal attenuation features, and light scattering features, respectively. Each attention head outputs a 64-dimensional feature vector, which is then connected to form a 512-dimensional feature vector. Subsequently, it is dimensionality reduced by three fully connected layers to generate a PD-L1 expression feature map of circulating tumor cells. The discriminator adopts a five-layer convolutional neural network structure to distinguish between real samples and generated samples. The final layer of the network uses Wasserstein distance as a loss function to evaluate the quality of generated samples.

[0042] The parameter optimization function is used to dynamically adjust the key hyperparameters during the training process of the multi-head attention-enhanced adversarial generative network. The inputs include the current training round number obtained from step S05, the generator loss function value obtained from step S05, the discriminator loss function value obtained from step S05, the sample batch size value determined from step S01, and the initial learning rate value set from step S06. The outputs are the optimized learning rate value used for the backpropagation algorithm in step S06, the batch weight coefficient value used for model training in step S05, the attention head weight coefficient value used for multi-head attention mechanism adjustment in step S05, and the regularization coefficient value used to prevent model overfitting in step S05.

[0043] The steps for establishing the training dataset in the pre-training process of the multi-head attention-enhanced adversarial generative network model specifically include: collecting peripheral blood samples from 300 patients with different cancer types and isolating circulating tumor cells; using standardized flow cytometry to obtain PD-L1 expression data in the circulating tumor cells of each patient; dividing all patient data into training and validation sets, with the training set containing data from 240 patients for model parameter learning and the validation set containing data from 60 patients for model performance evaluation; and preprocessing the training set data, including outlier removal, data standardization, data augmentation, and feature selection, to finally construct a PD-L1 expression pattern dataset for circulating tumor cells containing multidimensional features.

[0044] The pre-training steps of the multi-head attention-enhanced adversarial generative network model specifically include setting initial training parameters such as batch size of 32, learning rate of 0.0001, training epochs of 500, number of attention heads of 8, and attention layer depth of 2. The Adam optimizer is used to update the parameters. During training, the model performance is evaluated every 50 epochs and the loss function values ​​of the generator and discriminator are recorded. When the model performance does not improve significantly after 5 consecutive evaluations, a learning rate decay strategy is adopted. After training, the model accuracy, precision, recall and F1 score are calculated using validation set data to evaluate the overall performance of the model. The optimal model parameters are saved as the final pre-trained model.

[0045] The initial generation function refers to the initial probability distribution function set for the generator before training the multi-head attention-enhanced adversarial generative network, which is used to guide the generator to produce initial sample data.

[0046] Among them, the backpropagation algorithm refers to an optimization algorithm that adjusts the network parameters layer by layer from the output layer to the input layer by calculating the gradient of the loss function with respect to the network parameters.

[0047] The PD-L1 expression positivity rate refers to the percentage of cells with PD-L1 molecules detected on their surface out of the total number of circulating tumor cells detected.

[0048] Among them, historical detection data refers to the PD-L1 expression-related data of circulating tumor cells collected and processed from steps S01 to S04, including multidimensional information such as fluorescence intensity value, fluorescence distribution pattern, cell morphology characteristics and background signal.

[0049] Among them, the sample to be tested refers to the peripheral blood sample of a new patient who needs to be tested for PD-L1 expression, which is a circulating tumor cell sample obtained after undergoing the same processing procedure as S01 to S03.

[0050] The specific implementation methods of the above steps are described in detail below.

[0051] The specific implementation of step S01 involves collecting a peripheral blood sample from the patient and separating circulating tumor cells using density gradient centrifugation. This step begins by drawing 10 ml of peripheral blood from the patient's vein using a vacuum blood collection tube pre-added with EDTA anticoagulant to prevent blood clotting. The collected blood sample is then mixed thoroughly with an equal volume of phosphate-buffered saline (PBS), and the mixture is carefully and slowly spread onto a density gradient separation solution with a density of 1.077 g / ml. The mixture is centrifuged at 800 g for 30 minutes at room temperature, using slow acceleration and slow deceleration modes to maintain a clear interface. After centrifugation, a multilayer structure is formed, with a mononuclear cell layer located between the plasma layer and the separation solution. This cell layer is carefully aspirated using a sterile pipette and transferred to a new centrifuge tube. The cells are washed twice with PBS (centrifuged at 300 g for 10 minutes each time) to remove platelets and residual separation solution. The resulting cell pellet is resuspended in 2 ml of PBS containing 2% bovine serum albumin, and the cell concentration is adjusted to 1 × 10⁻⁶ cells / ml. 6 The goal of this step is to isolate a population of mononuclear cells containing circulating tumor cells from the blood, laying the foundation for subsequent labeling and analysis.

[0052] The specific implementation of step S02 involves labeling circulating tumor cell surface-specific markers with an anti-epithelial cell adhesion molecule antibody. The cell suspension obtained in S01 is aliquoted into multiple flow cytometry tubes, approximately 100 μl per tube. A pre-optimized concentration (typically 5 μg / ml) of anti-human EpCAM (epithelial cell adhesion molecule) monoclonal antibody, conjugated with PE (phycoerythrin) fluorescent dye, is added to the cell suspension. The cell-antibody mixture is incubated at 4°C in the dark for 30 minutes, gently mixing every 10 minutes to ensure adequate antibody contact with the cell surface. After incubation, 2 ml of PBS containing 0.1% bovine serum albumin is added to each tube, and the cells are centrifuged at 300g for 5 minutes to remove unbound antibodies. The supernatant is discarded, and the cells are resuspended in 100 μl of fresh buffer. To distinguish between live and dead cells, 7-AAD (7-aminoactinomycin D) live / dead cell dye is added to a final concentration of 1 μg / ml, and the cells are incubated at room temperature in the dark for 5 minutes. This step specifically labels circulating tumor cells with epithelial cell characteristics while excluding normal leukocytes in the blood, providing a clear target cell population for subsequent PD-L1 expression analysis.

[0053] The specific implementation of step S03 involves adding an anti-PD-L1 fluorescently labeled antibody to bind to PD-L1 molecules on the surface of circulating tumor cells. An appropriate amount (usually 2.5 μg / ml) of anti-human PD-L1 monoclonal antibody, pre-conjugated with FITC (fluorescein isothiocyanate) fluorescent dye, is added to the EpCAM-labeled cell suspension. After antibody addition, the cells are incubated at 4°C in the dark for 30 minutes to ensure sufficient binding of the antibody to PD-L1 molecules on the cell surface. During incubation, the cells are gently mixed every 10 minutes to improve labeling efficiency. After incubation, the cells are washed twice with 2 ml of PBS containing 0.1% BSA (centrifuged at 300g for 5 minutes each time) to remove unbound antibody and reduce background fluorescence. Finally, the cells are resuspended in 300 μl of PBS containing 0.1% paraformaldehyde to fix the cells and prevent fluorescence decay. A negative control (using an isotype control antibody instead of the anti-PD-L1 antibody) and a positive control (using a known PD-L1 high-expression cell line) are also provided to ensure experimental accuracy. The purpose of this step is to specifically detect the expression of PD-L1 molecules on the surface of circulating tumor cells, providing a fluorescence signal for subsequent quantitative analysis.

[0054] The specific implementation of step S04 involves using a flow cytometer to analyze labeled cells and collect historical detection data. First, the flow cytometer parameters are set, including laser power adjustment (FITC channel excitation wavelength 488nm, emission wavelength 530±15nm; PE channel excitation wavelength 561nm, emission wavelength 585±15nm; 7-AAD channel excitation wavelength 561nm, emission wavelength 650±20nm), voltage parameter settings (FSC voltage set to 350V, SSC voltage set to 300V, FITC voltage set to 500V, PE voltage set to 450V, 7-AAD voltage set to 600V), and flow rate control (set to a low flow rate, approximately 60μl / min). Fluorescent microspheres are used for fluorescence channel compensation; the compensation value of FITC to PE is approximately 3–5%, and the compensation value of PE to FITC is approximately 1–2%, ensuring no crosstalk between multicolor fluorescence. The duration of cellular fluorescence signals is quantitatively analyzed using a long-feature extraction equation based on fluorescence intensity. This equation employs a time-domain signal integration algorithm, Gaussian filtering the original fluorescence signal time-series data to calculate the fluorescence peak width, and converting it to actual time values ​​based on the signal acquisition frequency, outputting a feature vector of cellular fluorescence duration. A short-feature extraction equation based on fluorescence intensity is used to quantitatively analyze the peak intensity of cellular fluorescence signals. This equation employs a peak detection algorithm, performing background correction on the original fluorescence intensity data and standardizing it according to the signal gain coefficient, outputting a feature vector of cellular fluorescence peak distribution. A continuous-feature extraction equation based on fluorescence intensity is used to quantitatively analyze the continuous distribution of cellular fluorescence signals. This equation employs a spatial autocorrelation algorithm to analyze the continuous pattern of fluorescence distribution on the cell surface, calculate the gradient of fluorescence intensity change on the cell surface, and output a feature vector of continuous cellular fluorescence distribution. A discontinuous-feature extraction equation based on fluorescence intensity intervals is used to quantitatively analyze the discontinuous distribution of cellular fluorescence signals. This equation employs a spatial breakpoint detection algorithm, identifying discontinuous regions of fluorescence intensity by setting a signal segmentation threshold, and calculating the relative interval distance by combining cell morphology parameters, outputting a feature vector of cellular fluorescence interval distribution. Finally, all the above feature data are integrated to form a dataset containing multi-dimensional features for subsequent machine learning model training. The purpose of this step is to comprehensively collect multidimensional characteristic data related to PD-L1 expression in circulating tumor cells using flow cytometry, laying a data foundation for building an accurate analytical model.

[0055] The specific implementation of step S05 involves constructing a multi-head attention-enhanced adversarial generative network model based on historical detection data. This model architecture consists of a generator and a discriminator. The generator employs an encoder-decoder structure. The encoder contains three convolutional layers with kernel sizes of 5×5, 3×3, and 3×3, each with a stride of 2, and filter numbers of 64, 128, and 256, respectively. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function (α = 0.2). The encoder output is then fed into a multi-head attention module, which contains eight independent attention heads, each containing a query transformation matrix. Key transformation matrix W k Sum transformation matrix W v All features are 64×64 in dimension. Each attention head focuses on different features: the first head focuses on the distribution of PD-L1 expression location using a spatial attention mechanism; the second head focuses on the distribution of fluorescence intensity using an intensity-weighted attention mechanism; the third head focuses on cell morphology features using a morphology-related attention mechanism; the fourth head focuses on background noise features using a noise suppression attention mechanism; the fifth head focuses on surface marker co-expression features using a channel cross-attention mechanism; the sixth head focuses on cell population heterogeneity features using a heterogeneity-sensing attention mechanism; the seventh head focuses on fluorescence signal attenuation features using a time-dependent attention mechanism; and the eighth head focuses on light scattering features using a scattering-related attention mechanism. The multi-head attention outputs are concatenated to form a 512-dimensional feature vector, which is then dimensionality-reduced through three fully connected layers (512×256, 256×128, and 128×64) to finally generate a 64×64×3 feature map. The discriminator employs a five-layer convolutional neural network structure, with all convolutional kernels measuring 4×4 and a stride of 2. The number of filters is 64, 128, 256, 512, and 1, respectively. Each convolutional layer is followed by an instance normalization layer and a LeakyReLU activation function (α = 0.2). The final layer uses a linear activation function to output the discrimination result. Network training utilizes the WGAN-GP (Wasserstein Generative Adversarial Network-Gradient Penalty) framework. The loss function combines Wasserstein distance and a gradient penalty term, with the gradient penalty coefficient λ set to 10. The aim of this step is to construct a deep learning model capable of accurately capturing the complex expression patterns of PD-L1 in circulating tumor cells, enhancing the perception of different feature dimensions through a multi-head attention mechanism.

[0056] The specific implementation of step S06 involves setting an initial generator function for the multi-head attention-enhanced adversarial generative network model and applying a parameter optimization function to optimize the network parameters through backpropagation. The initial generator function uses a truncated normal distribution (mean μ = 0, standard deviation σ = 0.02, cutoff range [-0.05, 0.05]) to initialize the generator network parameters, ensuring network stability in the early stages of training. The parameter optimization function employs an adaptive moment estimation (Adam) optimizer combined with a learning rate scheduling strategy, with an initial learning rate set to 0.0001, β1 = 0.5, and β2 = 0.999. This function comprehensively analyzes the current training epoch, generator loss, discriminator loss, batch size, and initial learning rate, dynamically adjusting the following parameters: the learning rate is updated exponentially according to the training progress, with a decay rate of 0.95 / 50 epochs; the batch weight coefficient is adjusted according to the sample class distribution, with the weight ratio of imbalanced classes ranging from 1:5; the attention head weight coefficient is determined by evaluating the attention contribution based on validation set performance, initially all being 0.125; the regularization coefficient increases with the number of training epochs, gradually increasing from 0.0001 to 0.001. The backpropagation algorithm uses gradient descent, first calculating the gradient of each layer's output with respect to the loss function, then propagating the gradient layer by layer from the output layer to the input layer using a chain rule to update the parameters. To prevent gradient explosion, a gradient clipping threshold of 10 is set. Batch gradient descent updates parameters using 32 samples per batch, with a 5:1 update ratio (5 updates for the discriminator, 1 update for the generator). The early stopping strategy is set to stop training and save the optimal model parameters when the validation set loss shows no improvement for 20 consecutive epochs. The purpose of this step is to efficiently adjust the network parameters through algorithm optimization, so that the model can accurately learn the PD-L1 expression pattern of circulating tumor cells and has good generalization ability.

[0057] The specific implementation of step S07 involves using a trained multi-head attention-enhanced adversarial generative network model to quantitatively analyze the PD-L1 expression level in the circulating tumor cell samples to be tested. First, the samples are processed according to the same procedures as in S01 to S04 to obtain multidimensional feature data. The feature data is then input into the trained model generator, which generates a PD-L1 expression feature map of the samples. The map has dimensions of 64×64×3, where the first channel represents fluorescence intensity distribution, the second channel represents spatial location information, and the third channel represents cellular heterogeneity information. Post-processing of the feature map includes Gaussian filtering (kernel size 5×5, σ=1.5) to remove noise and median filtering (kernel size 3×3) to enhance edge features. A threshold segmentation method (using the Otsu adaptive thresholding algorithm) is used to binarize the feature map. The threshold is set to three times the background fluorescence intensity or a threshold automatically determined by the Otsu algorithm, taking the larger of the two. Morphological operations were performed on the binarized images, including opening operations (structuring element: 3×3 disk) to remove small noise and closing operations (structuring element: 5×5 disk) to fill small holes. Connected component analysis was used to label each positive cell region, and the area, perimeter, average fluorescence intensity, and maximum fluorescence intensity of each region were calculated. Based on the feature importance weights learned by the model, and considering fluorescence intensity, distribution pattern, and spatial features, a PD-L1 expression intensity score was calculated for each circulating tumor cell, ranging from 0 to 100. The purpose of this step is to use a trained deep learning model to accurately quantify the PD-L1 expression level in new samples, overcoming the limitations of traditional flow cytometry in rare cell analysis.

[0058] Step S08 involves calculating the PD-L1 expression positivity rate of circulating tumor cells in the test sample based on the output of the multi-head attention-enhanced adversarial generative network model. First, a PD-L1 expression threshold is set, and the optimal cutoff value is determined based on ROC curve analysis of the model on the validation set. Typically, a PD-L1 expression intensity score ≥25 is considered positive (this threshold corresponds to 95% specificity and 85% sensitivity). All circulating tumor cells analyzed in step S07 are judged as positive / negative, and cells with an expression intensity score ≥ the threshold are classified as PD-L1 positive. The number of PD-L1 positive circulating tumor cells (N-positive) and the total number of circulating tumor cells (N-total) in the sample are counted. The PD-L1 expression positivity rate is calculated as (N-positive / N-total) × 100%. To assess the reliability of the results, a 95% confidence interval is calculated, and the upper and lower limits of the confidence interval are calculated using the Wilson scoring method. For samples with fewer than 10 detected circulating tumor cells, the report is marked as "insufficient cell count, limited reliability of results". Furthermore, the PD-L1 expression positivity rate was classified into four categories: negative (<1%), low expression (1%–5%), moderate expression (5%–50%), and high expression (>50%). The quantitative results were compared with clinical reference ranges. The reference range for PD-L1 expression positivity rate in circulating tumor cells in the normal population is 0–0.5%, while the reference ranges for different tumor types were determined based on historical databases. The purpose of this step is to provide an accurate quantitative indicator of PD-L1 expression in circulating tumor cells, providing a basis for clinical immunotherapy decisions.

[0059] The mathematical model or calculation process involved in this invention will be described in detail below.

[0060] In step S04, the process of analyzing labeled cells and collecting historical detection data using flow cytometry involves the calculation of multiple feature extraction equations.

[0061] The specific expression of the fluorescence intensity long feature extraction equation is as follows:

[0062]

[0063] In the formula, F L (t) represents the fluorescence signal duration feature vector; I(t) represents the original fluorescence signal time series data; G(t, σ) is the Gaussian filter function, where σ is typically 1.5; t0 is the fluorescence signal start time point; Δt is the fluorescence signal duration window, typically 10 ms; f s The signal acquisition frequency parameter is typically set to 20kHz; λ e λ represents the excitation wavelength of the fluorescently labeled antibody; for FITC-labeled antibody, it is 488 nm. m The emission wavelength of the fluorescently labeled antibody is 530 nm for FITC-labeled antibody; εL This is the error term, and its range is 0.01 to 0.05.

[0064] This equation is constructed based on the principle of time-domain signal integration. It obtains a feature vector reflecting the duration characteristics of the fluorescence signal by integrating the original fluorescence signal after Gaussian filtering within its duration window. The Gaussian filter function is introduced to reduce the influence of random noise on the integration result. The reciprocal term of the signal acquisition frequency parameter is used to convert the integration result from the number of sampling points to the actual time value. The 0.5 power term of the excitation wavelength and the -0.2 power term of the emission wavelength are used to compensate for the differences in photophysical properties of different fluorescent dyes under different wavelength conditions. This power relationship is determined based on experimental observations of the relationship between dye excitation efficiency and wavelength in fluorescence physics.

[0065] The specific equation for extracting short fluorescence intensity features is as follows:

[0066]

[0067] In the formula, F S (I) represents the peak intensity feature vector of the fluorescence signal; I(i) represents the original fluorescence signal intensity data points; I bg This is the background fluorescence threshold, typically set to 1.5 times the average fluorescence value of the negative control samples; G a This is the signal gain coefficient, typically ranging from 1 to 10, adjusted according to the instrument's sensitivity; C ab This refers to the concentration of the fluorescently labeled antibody, expressed in μg / ml, typically ranging from 1 to 10 μg / ml; I peak I represents the peak signal strength. mean ε is the average signal strength; S This is the error term, and its range is 0.05 to 0.1.

[0068] This equation is constructed based on the principles of peak detection and signal normalization. First, baseline noise is removed by subtracting a background fluorescence threshold. Then, the signal is normalized under different detection conditions using a signal gain coefficient. The -0.3 power term of antibody concentration compensates for nonspecific binding and fluorescence quenching effects that may result from high antibody concentrations; this power relationship is obtained by fitting experimental data on the nonlinear relationship between antibody concentration and signal intensity. The logarithmic term enhances the peak-to-valley ratio characteristic of the signal, helping to distinguish between true fluorescence signals and random fluctuations.

[0069] The equation for continuous feature extraction of fluorescence intensity is expressed as follows:

[0070]

[0071] In the formula, F C (x, y) is the spatially continuous distribution feature vector of the fluorescence signal; I(x i yj ) is the image data of cell surface fluorescence intensity distribution at location (x i y j The value at (); N p This represents the number of image sampling points, typically 64; d ij R is the distance between adjacent sampling points; α is the smoothing factor, usually taken as 0.1; s This is a spatial resolution parameter, measured in μm / pixel, typically ranging from 0.2 to 1.0; A c This represents the cell surface area, in μm. 2 ;T c This is a continuity threshold parameter, typically ranging from 0.5 to 2.0; G xy ε is the spatial gradient value; C This is the error term, and its range is 0.02 to 0.08.

[0072] This equation is constructed based on the principle of spatial autocorrelation analysis, quantifying the continuity of fluorescence distribution by calculating the correlation of fluorescence intensity at different locations on the cell surface. The first part of the fraction calculates the weighted product of fluorescence intensity at adjacent points, divided by the square of the distance to reflect the spatial attenuation effect. The ratio of spatial resolution to cell surface area is used to normalize the results for cells of different sizes. The exponential term is used to transform the continuity threshold parameter into a nonlinear weight, making feature extraction more sensitive to intensity changes.

[0073] The fluorescence intensity interval feature extraction equation is specifically expressed as follows:

[0074]

[0075] In the formula, F D (x, y) is the spatial discontinuous distribution feature vector of the fluorescence signal; N c The number of fluorescent signal breakpoint clusters; A k S is the area of ​​the k-th fluorescent breakpoint cluster; k S represents the intensity of the k-th fluorescent breakpoint cluster; total d represents the total fluorescence intensity; min This is the minimum interval distance parameter, in μm, typically ranging from 0.5 to 2.0; T s The signal segmentation threshold is typically set to twice the background intensity; I k P represents the average intensity of the k-th fluorescent breakpoint cluster; m A is a cell morphology parameter representing cell perimeter in μm. c Cell surface area, in μm 2 ;f var ε is the coefficient of morphological variation, ranging from 0 to 1; D This is the error term, and its range is 0.03 to 0.1.

[0076] This equation is constructed based on the principles of spatial breakpoint detection and morphological analysis, identifying and quantifying discontinuous regions in fluorescence distribution. The first part of the fraction calculates the characteristics of fluorescence breakpoint clusters, considering the effects of area, relative intensity, and spacing. The Logistic term weights the fluorescence intensity of the breakpoints based on a signal segmentation threshold, making the algorithm more accurate in determining signal strength. The ratio of the square of the cell perimeter to the surface area is used to quantify the influence of cell morphology on the breakpoint distribution. The sine term introduces periodic modulation of the morphological variation coefficient, making feature extraction more sensitive to irregular morphologies.

[0077] In step S05, the multi-head attention mechanism involves matrix calculation, specifically as follows:

[0078]

[0079] In the formula, Q is the query matrix with dimensions N×d. k N is the batch size; K is the key matrix with dimension N×d. k V is a value matrix with dimensions N×d. v ;d k The attention head feature dimension is set to 64; d v The output dimension is set to 64; softmax is the softmax normalization function.

[0080] This attention calculation formula is the core of the multi-head attention mechanism. It calculates the attention weights based on the similarity between the query and the key, then sums them with the value matrix using a weighted method to obtain the output. (Divided by...) The term is used to scale the dot product result and prevent gradient vanishing. In the actual implementation, the eight attention heads are calculated independently and then concatenated. The calculation formula is as follows:

[0081] MultiHead(Q,K,V)=Concat(head1,...,head8)W O ;

[0082] In the formula, W i Q , W i V These are the query, key, and value transformation matrices for the i-th attention head, respectively, each with a dimension of 64×64; W O The output transformation matrix has a dimension of 512×512.

[0083] In step S06, the parameter optimization function involves a calculation process, specifically as follows:

[0084]

[0085] In the formula, η t Let η be the learning rate in round t; η0 be the initial learning rate, set to 0.0001; t be the current training round; L g L represents the generator loss function value. d α is the discriminator loss function value; α is the generator influence factor, usually set to 0.1; β is the early learning rate boosting factor, usually set to 0.5; γ is the decay rate, usually set to 0.01. This represents the floor function.

[0086] This learning rate optimization formula combines an exponential decay strategy and a dynamic adjustment strategy based on the loss function. The exponential decay term reduces the learning rate by 5% every 50 epochs, which helps with convergence in the later stages of training. The generator-to-discriminator loss ratio term dynamically adjusts the learning rate according to the balance between the two, appropriately reducing the learning rate when the generator loss is too large. The exponential term provides a slightly higher learning rate in the early stages of training to accelerate convergence, and gradually decays to the base value as training progresses.

[0087] The formula for calculating the batch weighting coefficient is as follows:

[0088]

[0089] In the formula, w c N represents the batch weight coefficient for category c; total C is the total number of samples; N is the number of categories; c N is the number of samples in category c; max λ is the maximum number of samples in each class; λ is the imbalance adjustment coefficient, which is usually set to 4.

[0090] The weighting formula is based on the inverse frequency principle. First, it calculates the base weight by dividing the total number of samples by the product of the number of classes and the number of samples in the current class. Then, it further adjusts the weighting using a sample imbalance weighting term. This design gives minority class samples higher weights, helping to address the class imbalance problem.

[0091] The formula for calculating the attention head weighting coefficient is as follows:

[0092]

[0093] In the formula, α i s represents the weight coefficient of the i-th attention head; i h represents the performance score of the i-th attention head on the validation set. i is the output of the i-th attention head; L is the loss function; δ is the gradient influence factor, which is usually taken as 0.3.

[0094] This weighting formula combines performance-based softmax normalization with gradient-based importance assessment. Performance scores are calculated using the accuracy of each attention head acting individually on the validation set. The gradient term assesses the impact on model performance by calculating the partial derivatives of the loss function with respect to the outputs of each attention head, ensuring that weight allocation better reflects the importance of different features.

[0095] The formula for calculating the regularization coefficient is as follows:

[0096]

[0097] In the formula, λ t Let be the regularization coefficient for round t; λ0 is the initial regularization coefficient, set to 0.0001; λ max is the maximum regularization coefficient, set to 0.001; k is the growth rate, usually taken as 3; T max The maximum number of training rounds is set to 500.

[0098] The regularization coefficient adjustment formula adopts an exponential saturation growth model, causing the regularization strength to gradually increase as training progresses. This design allows the model to focus primarily on fitting the data in the early stages of training, and gradually strengthens the regularization as training deepens to prevent overfitting, which aligns with commonly used regularization strategies in deep learning.

[0099] In step S07, the formula for calculating the PD-L1 expression intensity score is as follows:

[0100]

[0101] In the formula, S PD-L1 The expression intensity score for PD-L1 ranges from 0 to 100; I mean I represents the average fluorescence intensity of the cells. bg I represents the background fluorescence intensity. max The maximum reference fluorescence intensity is usually the average intensity of the positive control; A pos Area of ​​the fluorescent positive region; A total The total cell area; α i The weight coefficients for the i-th feature are determined during model training; f i (x) is the i-th feature function, which includes the quantification results of features such as intensity distribution and spatial pattern.

[0102] The intensity score calculation formula comprehensively considers fluorescence intensity, the proportion of positive areas, and multiple feature weights. The intensity normalization term linearly maps fluorescence intensity to a standard range. The area proportion term quantifies the area ratio. The feature-weighted summation term integrates multi-dimensional feature information, and the weight coefficients of different features are automatically optimized through model training, making the calculation results more consistent with expert evaluation standards.

[0103] In step S08, the formula for calculating the 95% confidence interval for PD-L1 expression positivity rate is as follows:

[0104]

[0105] In the formula, CI lower CI is the lower limit of the confidence interval. upper 1 is the upper limit of the confidence interval; P is the observed PD-L1 expression positivity rate; n is the total number of circulating tumor cells; z is the critical value of the standard normal distribution, and the 95% confidence interval is 1.96.

[0106] The confidence interval calculation formula is based on the Wilson scoring method, which is an interval estimation method for the proportion of a binomial distribution. It is particularly suitable for situations with small sample sizes or proportions close to 0 or 1. Compared with the traditional normal approximation method, the Wilson scoring method provides more accurate confidence intervals in small sample cases, which is especially important for the study of rare cells such as circulating tumor cells. The square term in the numerator and the correction term in the denominator make the estimation results more robust, avoiding the problem of intervals exceeding the [0, 1] range that may occur with traditional methods under extreme proportions.

[0107] Specifically, the principle of this invention is as follows: The core technology of this invention is based on the multi-dimensional and precise capture and deep learning analysis of the PD-L1 expression characteristics of circulating tumor cells. First, circulating tumor cells are enriched from peripheral blood using density gradient centrifugation, and these cells are specifically labeled using anti-epithelial cell adhesion molecule antibodies to ensure the purity of the target cell population; simultaneously, anti-PD-L1 fluorescently labeled antibodies are used to bind to PD-L1 molecules on the cell surface, providing a basis for subsequent quantitative analysis.

[0108] Building upon traditional flow cytometry which focuses solely on fluorescence intensity peaks, this invention innovatively designs four types of feature extraction equations to systematically quantify the multidimensional characteristics of PD-L1 fluorescence signals. The long fluorescence intensity feature extraction equation analyzes signal duration characteristics, capturing the temporal continuity of PD-L1 expression; the short fluorescence intensity feature extraction equation focuses on the precise quantification of signal peak intensity; the continuous fluorescence intensity feature extraction equation analyzes the spatially continuous distribution pattern of PD-L1 molecules on the cell surface; and the interval fluorescence intensity feature extraction equation focuses on the spatially discontinuous distribution characteristics of PD-L1 molecules. This multidimensional feature extraction strategy enables the system to comprehensively capture the complex patterns of PD-L1 expression, rather than relying solely on a single intensity value.

[0109] In terms of model construction, this invention employs an adversarial generative network architecture, continuously improving feature extraction and classification performance through a game-like learning process between the generator and the discriminator. The introduction of a multi-head attention mechanism is a key innovation for improving detection accuracy. Eight independent attention heads focus on feature information of different dimensions, forming a 512-dimensional feature vector, enabling the model to simultaneously process and integrate multiple expressive features. This divide-and-conquer attention mechanism significantly enhances the model's ability to distinguish subtle differences in expression, greatly improving detection accuracy.

[0110] The design of the parameter optimization function ensures the stability of the model training process. By dynamically adjusting hyperparameters such as the learning rate and batch weights, common problems during training, such as gradient vanishing and overfitting, are avoided. Using Wasserstein distance as the loss function more accurately measures the distribution difference between generated and real samples, improving the model's generalization ability. These technological innovations collectively constitute an accurate and stable PD-L1 expression detection system for circulating tumor cells, effectively solving the technical problem of low detection accuracy.

[0111] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0112] The specific implementation methods of steps S01-S03 are the same as those described above, and will not be repeated here.

[0113] The specific implementation of step S04 involves using a flow cytometer to analyze labeled cells and collect historical detection data. First, the flow cytometer parameters are set, including laser power adjustment (FITC channel excitation wavelength 488nm, emission wavelength 530±15nm; PE channel excitation wavelength 561nm, emission wavelength 585±15nm; 7-AAD channel excitation wavelength 561nm, emission wavelength 650±20nm), voltage parameter settings (FSC voltage set to 350V, SSC voltage set to 300V, FITC voltage set to 500V, PE voltage set to 450V, 7-AAD voltage set to 600V), and flow rate control (set to a low flow rate, approximately 60μl / min). Fluorescent microspheres are used for fluorescence channel compensation; the compensation value of FITC to PE is approximately 3–5%, and the compensation value of PE to FITC is approximately 1–2%, ensuring no crosstalk between multicolor fluorescence. The duration of cellular fluorescence signals was quantitatively analyzed using a long-term fluorescence intensity feature extraction equation. This equation employs a time-domain signal integration algorithm and is specifically expressed as follows: In the formula, F L(t) represents the fluorescence signal duration feature vector; i(t) represents the original fluorescence signal time series data; G(t, σ) is the Gaussian filter function, where σ is typically 1.5; t0 is the fluorescence signal start time point; Δt is the fluorescence signal duration window, typically 10 ms; f s The signal acquisition frequency parameter is typically set to 20kHz; λ e λ represents the excitation wavelength of the fluorescently labeled antibody; for FITC-labeled antibody, it is 488 nm. m为荧 The emission wavelength of the photolabeled antibody is 530 nm for FITC-labeled antibody; ε L The error term ranges from 0.01 to 0.05. A short feature extraction equation based on fluorescence intensity is used to quantify the peak intensity of cell fluorescence signals. This equation employs a peak detection algorithm and is specifically expressed as follows: In the formula, F S (I) represents the peak intensity feature vector of the fluorescence signal; I(i) represents the original fluorescence signal intensity data points; I bg This is the background fluorescence threshold, typically set to 1.5 times the average fluorescence value of the negative control samples; G a This is the signal gain coefficient, typically ranging from 1 to 10, adjusted according to the instrument's sensitivity; C ab This refers to the concentration of the fluorescently labeled antibody, expressed in μg / ml, typically ranging from 1 to 10 μg / ml; I peak I represents the peak signal strength. mean ε is the average signal strength; S The error term ranges from 0.05 to 0.1. A fluorescence intensity continuous feature extraction equation is used to quantify the continuous distribution of cell fluorescence signals. This equation employs a spatial autocorrelation algorithm and is specifically expressed as follows: In the formula, F C (x, y) is the spatially continuous distribution feature vector of the fluorescence signal; I(x i y j ) is the image data of cell surface fluorescence intensity distribution at location (x i y j The value at (); N p This represents the number of image sampling points, typically 64; d ij R is the distance between adjacent sampling points; α is the smoothing factor, usually taken as 0.1; s This is a spatial resolution parameter, measured in μm / pixel, typically ranging from 0.2 to 1.0; A c This represents the cell surface area, in μm. 2 ;T c This is a continuity threshold parameter, typically ranging from 0.5 to 2.0; G xy ε is the spatial gradient value;C The error term ranges from 0.02 to 0.08. A fluorescence intensity interval feature extraction equation is used to quantify the discontinuous distribution of cell fluorescence signals. This equation employs a spatial breakpoint detection algorithm and is specifically expressed as follows: In the formula, F D (x, y) is the spatial discontinuous distribution feature vector of the fluorescence signal; N c The number of fluorescent signal breakpoint clusters; A k S is the area of ​​the k-th fluorescent breakpoint cluster; k S represents the intensity of the k-th fluorescent breakpoint cluster; total d represents the total fluorescence intensity; min This is the minimum interval distance parameter, in μm, typically ranging from 0.5 to 2.0; T s The signal segmentation threshold is typically set to twice the background intensity; I k P represents the average intensity of the k-th fluorescent breakpoint cluster; m A is a cell morphology parameter representing cell perimeter in μm. c Cell surface area, in μm 2 ;f var ε is the coefficient of morphological variation, ranging from 0 to 1; D The error term ranges from 0.03 to 0.1. Finally, all the aforementioned feature data are integrated to form a dataset containing multidimensional features for subsequent machine learning model training. The purpose of this step is to comprehensively collect multidimensional feature data related to PD-L1 expression in circulating tumor cells using flow cytometry, laying a data foundation for building an accurate analytical model.

[0114] The specific implementation of step S05 is to construct a multi-head attention-enhanced adversarial generative network model based on historical detection data. This model architecture consists of a generator and a discriminator. The generator adopts an encoder-decoder structure. The encoder contains three convolutional layers with kernel sizes of 5×5, 3×3, and 3×3, each with a stride of 2, and filter numbers of 64, 128, and 256, respectively. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function (α = 0.2). The encoder output is then fed into a multi-head attention module, which employs the attention calculation principle, specifically as follows: In the formula, Q is the query matrix with dimensions N×d. k N is the batch size; K is the key matrix with dimension N×d. k V is a value matrix with dimensions N×d. v ;d k The attention head feature dimension is set to 64; d vThe output dimension is set to 64; softmax is the softmax normalization function. The calculation of the multi-head attention mechanism is represented as: MultiHead(Q, K, V) = Concat(head1, ..., head8)W O In the formula, W i Q , W i V These are the query, key, and value transformation matrices for the i-th attention head, respectively, each with a dimension of 64×64; W O The output transformation matrix has a dimension of 512×512. This module contains 8 independent attention heads, each containing a query transformation matrix W. Q Key transformation matrix W K Sum transformation matrix W V All features are 64×64 in dimension. Each attention head focuses on different features: the first head focuses on the distribution of PD-L1 expression location using a spatial attention mechanism; the second head focuses on the distribution of fluorescence intensity using an intensity-weighted attention mechanism; the third head focuses on cell morphology features using a morphology-related attention mechanism; the fourth head focuses on background noise features using a noise suppression attention mechanism; the fifth head focuses on surface marker co-expression features using a channel cross-attention mechanism; the sixth head focuses on cell population heterogeneity features using a heterogeneity-sensing attention mechanism; the seventh head focuses on fluorescence signal attenuation features using a time-dependent attention mechanism; and the eighth head focuses on light scattering features using a scattering-related attention mechanism. The multi-head attention outputs are concatenated to form a 512-dimensional feature vector, which is then dimensionality-reduced through three fully connected layers (512×256, 256×128, and 128×64) to finally generate a 64×64×3 feature map. The discriminator employs a five-layer convolutional neural network structure, with all convolutional kernels measuring 4×4 and a stride of 2. The number of filters is 64, 128, 256, 512, and 1, respectively. Each convolutional layer is followed by an instance normalization layer and a LeakyReLU activation function (α = 0.2). The final layer uses a linear activation function to output the discrimination result. Network training utilizes the WGAN-GP (Wasserstein Generative Adversarial Network-Gradient Penalty) framework. The loss function combines Wasserstein distance and a gradient penalty term, with the gradient penalty coefficient λ set to 10. The aim of this step is to construct a deep learning model capable of accurately capturing the complex expression patterns of PD-L1 in circulating tumor cells, enhancing the perception of different feature dimensions through a multi-head attention mechanism.

[0115] The specific implementation of step S06 involves setting an initial generator function for the multi-head attention-enhanced adversarial generative network model and applying a parameter optimization function to optimize the network parameters using the backpropagation algorithm. The initial generator function uses a truncated normal distribution (mean μ = 0, standard deviation σ = 0.02, cutoff range [-0.05, 0.05]) to initialize the generator network parameters, ensuring network stability in the early stages of training. The parameter optimization function uses an adaptive moment estimation (Adam) optimizer combined with a learning rate scheduling strategy; the learning rate optimization is expressed as: In the formula, η t Let η be the learning rate in round t; η0 be the initial learning rate, set to 0.0001; t be the current training round; L g L represents the generator loss function value. d α is the discriminator loss function value; α is the generator influence factor, usually set to 0.1; β is the early learning rate boosting factor, usually set to 0.5; γ is the decay rate, usually set to 0.01. This represents the floor function. The batch weighting coefficient is calculated as follows: In the formula, w c N represents the batch weight coefficient for category c; total C is the total number of samples; N is the number of categories; c N is the number of samples in category c; max λ represents the maximum number of samples in the largest class; λ is the imbalance adjustment coefficient, typically set to 4. The attention head weight coefficient is calculated as follows: In the formula, α i s represents the weight coefficient of the i-th attention head; i h represents the performance score of the i-th attention head on the validation set. i Let L be the output of the i-th attention head; L be the loss function; and δ be the gradient influence factor, typically set to 0.3. The regularization coefficient is calculated as follows: In the formula, λ t Let be the regularization coefficient for round t; λ0 is the initial regularization coefficient, set to 0.0001; λ max κ is the maximum regularization coefficient, set to 0.001; κ is the growth rate, usually taken as 3; T maxThe maximum number of training epochs is set to 500. The backpropagation algorithm uses gradient descent. First, it calculates the gradient of each layer's output with respect to the loss function. Then, it propagates the gradient layer by layer from the output layer to the input layer using the chain rule, updating the parameters accordingly. To prevent gradient explosion, a gradient clipping threshold of 10 is set. Batch gradient descent updates parameters using 32 samples per batch, with a 5:1 update ratio (5 updates for the discriminator, 1 update for the generator). The early stopping strategy stops training and saves the optimal model parameters when the validation set loss shows no improvement for 20 consecutive epochs. The purpose of this step is to efficiently adjust the network parameters through optimization algorithms, enabling the model to accurately learn the PD-L1 expression pattern of circulating tumor cells and possess good generalization ability.

[0116] The specific implementation of step S07 involves using a trained multi-head attention-enhanced adversarial generative network model to quantitatively analyze the PD-L1 expression level in the circulating tumor cell samples to be tested. First, the samples are processed according to the same procedures as in S01 to S04 to obtain multidimensional feature data. The feature data is then input into the trained model generator, which generates a PD-L1 expression feature map of the samples. The map has dimensions of 64×64×3, where the first channel represents fluorescence intensity distribution, the second channel represents spatial location information, and the third channel represents cellular heterogeneity information. Post-processing of the feature map includes Gaussian filtering (kernel size 5×5, σ=1.5) to remove noise and median filtering (kernel size 3×3) to enhance edge features. A threshold segmentation method (using the Otsu adaptive thresholding algorithm) is used to binarize the feature map. The threshold is set to three times the background fluorescence intensity or a threshold automatically determined by the Otsu algorithm, taking the larger of the two. Morphological operations were performed on the binarized image, including opening operations (structuring element: 3×3 disk) to remove small noise and closing operations (structuring element: 5×5 disk) to fill small holes. Connected component analysis was used to label each positive cell region, and the area, perimeter, average fluorescence intensity, and maximum fluorescence intensity of each region were calculated. Based on the feature importance weights learned by the model, and considering fluorescence intensity, distribution pattern, and spatial features, a PD-L1 expression intensity score was calculated for each circulating tumor cell, specifically expressed as follows: In the formula, S PD-L1 The expression intensity score for PD-L1 ranges from 0 to 100; I mean I represents the average fluorescence intensity of the cells. bg I represents the background fluorescence intensity. max The maximum reference fluorescence intensity is usually the average intensity of the positive control; A pos Area of ​​the fluorescent positive region; A total The total cell area; α i The weight coefficients for the i-th feature are determined during model training; f i(x) represents the i-th feature function, including the quantitative results of features such as intensity distribution and spatial pattern. The purpose of this step is to use a trained deep learning model to accurately quantify the PD-L1 expression level of new samples, overcoming the limitations of traditional flow cytometry in rare cell analysis.

[0117] The specific implementation of step S08 involves calculating the PD-L1 expression positivity rate of circulating tumor cells in the test sample based on the output of the multi-head attention-enhanced adversarial generative network model. First, a PD-L1 expression threshold is set, and the optimal cutoff value is determined based on ROC curve analysis of the model on the validation set. Typically, a PD-L1 expression intensity score ≥25 is considered positive (this threshold corresponds to 95% specificity and 85% sensitivity). All circulating tumor cells analyzed in step S07 are then classified as positive / negative, and cells with an expression intensity score ≥ the threshold are classified as PD-L1 positive. The number of PD-L1 positive circulating tumor cells (N) in the sample is counted. 阳性 ) and total circulating tumor cell count (N 总 ). Calculate the PD-L1 expression positivity rate = (N 阳性 / N 总 ()×100%. To assess the reliability of the results, a 95% confidence interval was also calculated. The Wilson scoring method was used to calculate the confidence interval, which is specifically expressed as follows: and In the formula, CI lower CI is the lower limit of the confidence interval. upper The upper limit of the confidence interval is given by p; the observed PD-L1 expression positivity rate is given by n; the total number of circulating tumor cells is given by n; and z is the cutoff value of the standard normal distribution, with a 95% confidence interval of 1.96. Samples with fewer than 10 detected circulating tumor cells are marked as "insufficient cell count, limited reliability of results" in the report. Furthermore, the PD-L1 expression positivity rate is classified into four categories: negative (<1%), low expression (1%–5%), moderate expression (5%–50%), and high expression (>50%). The quantitative results are compared with clinical reference ranges. The reference range for PD-L1 expression positivity rate in circulating tumor cells in the normal population is 0–0.5%, and the reference range for patients with different tumor types is determined based on historical databases. The purpose of this step is to provide an accurate quantitative indicator of PD-L1 expression in circulating tumor cells, providing a basis for clinical immunotherapy decisions.

[0118] To better understand and implement this invention, the following is an example 2 of a specific application scenario: Researchers collected peripheral blood samples from 320 patients with non-small cell lung cancer, including 180 patients with adenocarcinoma, 105 patients with squamous cell carcinoma, and 35 patients with other types of cancer. The samples were processed according to step S01 of the protocol: First, 10 ml of peripheral blood was drawn from each patient using a vacuum blood collection tube, and EDTA anticoagulant was added to prevent coagulation. After mixing the blood with an equal volume of PBS, it was slowly spread onto a density gradient separation solution with a density of 1.077 g / ml and centrifuged at 800 g for 30 minutes at room temperature. After separating the mononuclear cell layer, the cells were washed twice with PBS, and finally, the cell pellet was resuspended in PBS containing 2% bovine serum albumin, with the concentration adjusted to 1×10⁻⁶. 6 The sample size was [number] cells / ml. In step S02, anti-human EpCAM monoclonal antibody (conjugated with PE fluorescent dye) at a concentration of 5 μg / ml was used for labeling, and the sample was incubated at 4°C in the dark for 30 minutes. In step S03, anti-human PD-L1 monoclonal antibody (conjugated with FITC fluorescent dye) at a concentration of 2.5 μg / ml was added, and the sample was also incubated at 4°C in the dark for 30 minutes, followed by necessary washing and fixation steps. Flow cytometry parameters were set according to protocol S04 for detection, and historical detection data were obtained.

[0119] From 320 patients, researchers randomly selected 240 patients' data as the training set, and the remaining 80 patients' data as the validation and test sets. The training set data underwent preprocessing, including outlier removal, data normalization, data augmentation, and feature selection, to construct a dataset containing multidimensional features of circulating tumor cell PD-L1 expression patterns. Table 1 shows the detection of circulating tumor cells and PD-L1 expression in patients with different cancer types in the training set.

[0120] Table 1. Statistics on the detection of circulating tumor cells and PD-L1 expression in the training set of patients.

[0121]

[0122] In step S04, the researchers used the aforementioned four feature extraction equations to perform quantitative analysis of the fluorescence signal. Figure 1 The average and standard deviation of the four feature extraction equations in the training set are shown.

[0123] Table 2 shows the mean and standard deviation of the four feature extraction equations in the training set.

[0124] Feature extraction equation PD-L1 positive CTC PD-L1 negative CTC p-value Long fluorescence intensity characteristics (F_L) 0.753±0.124 0.326±0.085 <0.001 Short fluorescence intensity characteristic (F_S) 0.842±0.156 0.284±0.073 <0.001 Continuous fluorescence intensity characteristics (F_C) 0.687±0.138 0.412±0.096 <0.001 Fluorescence intensity interval characteristics (F_D) 0.594±0.142 0.365±0.088 <0.001

[0125] In steps S05 and S06, the researchers constructed a multi-head attention-enhanced adversarial generative network model based on historical detection data. The model employs the aforementioned structure, comprising eight independent attention heads, each focusing on different feature dimensions. Initial training parameters were set as follows: batch size 32, initial learning rate 0.0001, training epochs 500, and gradient clipping threshold 10. Model performance was evaluated every 50 epochs during training, and the loss function value was recorded. Table 3 shows the changes in generator and discriminator losses during model training.

[0126] Table 3. Loss variation during training of the multi-head attention-enhanced adversarial generative network model.

[0127]

[0128]

[0129] Table 4 shows the final values ​​of the eight attention head weight coefficients after training, reflecting the importance of different features in the PD-L1 expression determination. Figure 2 This demonstrates the generator loss (L) over 500 training rounds. g ), discriminator loss (L) d ) and learning rate (η) t The graph shows the changing trend of the generator loss. As the number of training epochs increases, the generator loss rapidly decreases from the initial 2.754 to 0.558, exhibiting a typical exponential decay characteristic. The decrease is most pronounced in the first 200 epochs, after which it tends to flatten out. The discriminator loss also decreases from 1.138 to 0.586, but the change is less significant than that of the generator loss. Simultaneously, the learning rate gradually decreases from 0.000097 to 0.000061 according to the preset decay strategy. This intuitively reflects the convergence characteristics of the model during training, indicating that after 500 training epochs, the model has reached a relatively stable state.

[0130] Table 4. Weight coefficients of the eight attention heads in the multi-head attention mechanism.

[0131] Attention Focus on features Initial weights Final weight (α_i) Head 1 PD-L1 expression location distribution 0.125 0.186 First 2 Fluorescence intensity distribution 0.125 0.214 First 3 Cell morphology 0.125 0.153 First 4 Background noise 0.125 0.068 First 5 Surface marker co-expression 0.125 0.132 First 6 Cell population heterogeneity 0.125 0.127 First 7 Fluorescence signal attenuation 0.125 0.075 First 8 Light scattering characteristics 0.125 0.045

[0132] After training, the researchers evaluated the model's performance using a test set (40 patient samples). Figure 3The comparison shows the changes in the weight coefficients of the eight attention heads in the initial state and after training. As can be seen from the figure, initially all attention heads had a weight of 0.125. After training, the importance of different features diverged significantly. The attention head for fluorescence intensity distribution had the highest weight, reaching 0.214; followed by "PD-L1 expression location distribution" with a weight of 0.186; while light scattering features and background noise had the lowest weights, only 0.045 and 0.068, respectively. This change in weight distribution reflects that the model automatically learned the importance of different features in determining PD-L1 expression. Table 5 shows the model's performance metrics on the test set.

[0133] Table 5 Performance metrics of the multi-head attention-enhanced adversarial generative network model on the test set.

[0134] Performance indicators numerical values accuracy 0.924 Accuracy 0.913 Recall rate 0.896 F1 score 0.904 AUC 0.936 Specificity 0.942 Sensitivity 0.896

[0135] To verify the practicality of the method, researchers selected 20 newly diagnosed non-small cell lung cancer patients, collected their peripheral blood, and processed it according to steps S01 to S03. The samples were analyzed using both conventional flow cytometry and this method. Table 6 shows a comparison of the detection results of the two methods on the same batch of samples.

[0136] Table 6 Comparison of detection results of the same batch of samples by traditional flow cytometry method and this method.

[0137]

[0138] Figure 4The relationship between different fluorescence intensity ranges and PD-L1 expression intensity scores and positive rates is illustrated. The bar chart represents the average PD-L1 expression intensity score and its standard deviation, while the line graph represents the percentage of positive rates for each intensity range. The figure clearly shows that PD-L1 expression intensity scores increase significantly with increasing fluorescence intensity. The average expression score is only 12.6±5.3 in the low intensity range (≤500), rising to 35.7±12.8 in the medium intensity range (501-1500), and reaching as high as 72.4±18.6 in the high intensity range (>1500). Simultaneously, the PD-L1 positive rate also increases significantly with increasing fluorescence intensity, rising from 15.8% in the low intensity range to 89.6% in the high intensity range, indicating a strong positive correlation between fluorescence intensity and PD-L1 expression levels. The figure also shows the confidence intervals for each expression intensity score, reflecting the distribution characteristics and reliability of the data. Traditional detection of PD-L1 expression in circulating tumor cells (CTCs) primarily relies on conventional flow cytometry and immunofluorescence staining techniques. These methods have several limitations when detecting rare cell populations such as CTCs. First, conventional flow cytometry has a low detection rate for CTCs, especially when the cell count is extremely low (<5 cells / 7.5 ml blood), where the detection sensitivity drops significantly. Second, the analysis of fluorescence signals using conventional methods is relatively simple, mainly relying on thresholding to distinguish between positive and negative cells, neglecting the multidimensional characteristics of fluorescence signals, such as duration and spatial distribution patterns. Third, conventional methods are easily affected by background fluorescence, non-specific binding, and operator subjectivity, leading to insufficient accuracy and reproducibility. Finally, when the number of CTCs detected in a sample is extremely low, conventional methods cannot provide reliable estimates of the PD-L1 expression positivity rate and its confidence interval.

[0139] In contrast, this invention significantly improves the performance of PD-L1 expression detection in circulating tumor cells in several aspects by introducing a multi-head attention-enhanced adversarial generative network model. First, multi-dimensional feature extraction of the fluorescence signal is performed using four feature extraction equations, capturing signal details that are difficult to obtain using traditional methods and improving the richness of feature expression. As shown in Table 2, these four features show significant differences between PD-L1-positive and negative circulating tumor cells (P<0.001). Second, the multi-head attention mechanism can simultaneously focus on different feature dimensions and automatically adjust the weight coefficients of each feature through training. Table 4 shows that fluorescence intensity distribution (head 2, weight 0.214) and position distribution (head 1, weight 0.186) are the most important features, while background noise (head 4, weight 0.068) and light scattering features (head 8, weight 0.045) contribute less. Third, the adversarial generative network architecture, through adversarial training between the generator and discriminator, enables the model to better learn the data distribution, improving generalization ability. Table 5 shows that the model achieved an accuracy of 0.924 and an AUC of 0.936 on the test set, demonstrating its excellent performance. Finally, this method also outperforms traditional methods in terms of circulating tumor cell detection rate. Table 6 shows that this method detected an average of 8.45 circulating tumor cells, while the traditional method detected only 5.4, representing an improvement of 56.5%. Furthermore, this method can provide a 95% confidence interval for the PD-L1 expression positivity rate, enhancing the reliability of the results.

[0140] It should be noted that the variables involved in this invention are explained in detail in Tables 8 and 9 below.

[0141] Table 8. Variable Explanation Table (Part 1)

[0142]

[0143]

[0144] Table 9. Variable Explanation Table (Part Two)

[0145]

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting PD-L1 expression in circulating tumor cells, characterized in that, Includes the following steps: Peripheral blood samples were collected from the patient and circulating tumor cells were separated by density gradient centrifugation. Using anti-epithelial cell adhesion molecule antibodies to label specific biomarkers on the surface of circulating tumor cells; Anti-PD-L1 fluorescently labeled antibody was added to bind to PD-L1 molecules on the surface of circulating tumor cells; the labeled cells were analyzed using flow cytometry, and historical detection data were collected; a multi-head attention-enhanced adversarial generative network (MGGN) model was constructed based on the historical detection data, comprising a generator and a discriminator; an initial generation function was set for the MGGN model, and the network parameters were optimized using a parameter optimization function and a backpropagation algorithm; the trained MGGN model was used to quantitatively analyze the PD-L1 expression level in the tested circulating tumor cell samples; the PD-L1 expression positivity rate in the tested samples was calculated based on the output results of the MGGN model.

2. The method for detecting PD-L1 expression in circulating tumor cells according to claim 1, characterized in that, In the step of collecting peripheral blood samples from patients and performing density gradient centrifugation to separate circulating tumor cells, the number of peripheral blood samples from patients is at least 100 groups. The at least 100 groups of peripheral blood samples from patients refers to the minimum number of samples collected from patients with different cancer types, ensuring that the dataset has sufficient statistical significance and covers the diversity of PD-L1 expression patterns in circulating tumor cells in different cancer types.

3. The method for detecting PD-L1 expression in circulating tumor cells according to claim 2, characterized in that, The steps of analyzing labeled cells and collecting historical detection data using flow cytometry specifically include: setting flow cytometry detection parameters and adjusting the fluorescence compensation matrix to eliminate signal interference between different fluorescence channels; quantifying the duration of cell fluorescence signals using a long fluorescence intensity feature extraction equation; quantifying the peak intensity of cell fluorescence signals using a short fluorescence intensity feature extraction equation; quantifying the continuous distribution of cell fluorescence signals using a continuous fluorescence intensity feature extraction equation; quantifying the discontinuous distribution of cell fluorescence signals using a interval fluorescence intensity feature extraction equation; and integrating all feature data to form a multidimensional feature dataset of PD-L1 expression in circulating tumor cells.

4. The method for detecting PD-L1 expression in circulating tumor cells according to claim 3, characterized in that, The fluorescence intensity long feature extraction equation is used to quantify the duration characteristics of PD-L1 fluorescence signal in circulating tumor cells. The input includes raw fluorescence signal time series data, signal acquisition frequency parameters, excitation wavelength values ​​of fluorescently labeled antibody and emission wavelength values ​​of fluorescently labeled antibody. The output is a feature vector characterizing the duration of PD-L1 fluorescence signal for feature data integration.

5. The method for detecting PD-L1 expression in circulating tumor cells according to claim 4, characterized in that, The fluorescence intensity short feature extraction equation is used to quantify the peak intensity characteristics of the PD-L1 fluorescence signal in circulating tumor cells. The inputs include the original fluorescence signal intensity data, signal gain coefficient, background fluorescence threshold and fluorescently labeled antibody concentration value. The output is a feature vector characterizing the peak distribution of the PD-L1 fluorescence signal for feature data integration.

6. The method for detecting PD-L1 expression in circulating tumor cells according to claim 5, characterized in that, The fluorescence intensity continuous feature extraction equation is used to quantify the spatial continuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data, spatial resolution parameters, cell surface area values, and continuity threshold parameters. The output is a feature vector characterizing the spatial continuous distribution of PD-L1 fluorescence signals for feature data integration.

7. The method for detecting PD-L1 expression in circulating tumor cells according to claim 6, characterized in that, The fluorescence intensity interval feature extraction equation is used to quantify the spatial discontinuous distribution characteristics of PD-L1 fluorescence signals in circulating tumor cells. The inputs include cell surface fluorescence intensity distribution image data, signal segmentation threshold, cell morphology parameters and minimum interval distance parameters. The output is a feature vector characterizing the spatial discontinuous distribution of PD-L1 fluorescence signals for feature data integration.

8. The method for detecting PD-L1 expression in circulating tumor cells according to claim 7, characterized in that, The specific structure of the multi-head attention-enhanced adversarial generative network model is a deep learning model based on a generator and discriminator architecture. In the generator part, a multi-head attention mechanism is introduced to process the high-dimensional feature information of circulating tumor cells. Eight independent attention heads focus on the PD-L1 expression location distribution features, fluorescence intensity distribution features, cell morphology features, background noise features, surface marker co-expression features, cell population heterogeneity features, fluorescence signal attenuation features, and light scattering features, respectively. Each attention head outputs a 64-dimensional feature vector, which is then connected to form a 512-dimensional feature vector. Subsequently, it is dimensionality reduced by three fully connected layers to generate a PD-L1 expression feature map of circulating tumor cells. The discriminator adopts a five-layer convolutional neural network structure to distinguish between real samples and generated samples. The final layer of the network uses Wasserstein distance as a loss function to evaluate the quality of generated samples.

9. The method for detecting PD-L1 expression in circulating tumor cells according to claim 8, characterized in that, The parameter optimization function is used to dynamically adjust the key hyperparameters during the training process of the multi-head attention-enhanced adversarial generative network. The inputs include the current training round number, the generator loss function, the discriminator loss function, the batch size, and the initial learning rate. The outputs are the optimized learning rate for backpropagation, the batch weight coefficient for model training, the attention head weight coefficient for adjusting the multi-head attention mechanism, and the regularization coefficient to prevent overfitting.

10. The method for detecting PD-L1 expression in circulating tumor cells according to claim 9, characterized in that, The steps for establishing the training dataset during the pre-training process of the multi-head attention-enhanced adversarial generative network model specifically include: collecting peripheral blood samples from 300 patients with different cancer types and isolating circulating tumor cells; using standardized flow cytometry to obtain PD-L1 expression data in the circulating tumor cells of each patient; dividing all patient data into training and validation sets; the training set containing data from 240 patients for model parameter learning; and the validation set containing data from 60 patients for model performance evaluation. Preprocessing of the training set data includes outlier removal, data standardization, data augmentation, and feature selection. Finally, a dataset containing multi-dimensional features of circulating tumor cell PD-L1 expression patterns is constructed.

Images