Lung elastography and AI-assisted pulmonary nodule benign and malignant identification method and storage medium

Abstract

The invention discloses a lung elastography and AI-assisted pulmonary nodule benign and malignant identification method and a storage medium, and the method comprises the steps: obtaining the individual feature data and pulmonary nodule multi-source detection data of a patient, including lung elastography data and biological digital twin multi-physical field data; preprocessing the multi-source detection data, reducing dimensions by using a principal component analysis method, and screening out a representative feature combination; inputting the representative feature combination into a pulmonary nodule benign and malignant identification model, and outputting a binary classification result and a first confidence coefficient; inputting the individual characteristic data and the representative characteristic combination into a biological digital twin model when the identification is malignant, generating simulation data, and calculating a comprehensive consistency index; inputting the individual feature data, the representative feature combination and the comprehensive consistency index into a correlation model, outputting a malignant risk degree prediction probability, and based on the probability, adjusting the first confidence coefficient to obtain a second confidence coefficient; generating a multi-modal identification report; the method has the advantage of high accuracy of identifying benign and malignant pulmonary

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and artificial intelligence technology, in particular to a lung elastography and AI (Artificial Intelligence) assisted lung nodule benignity and malignancy identification method and storage medium. BACKGROUND

[0002] With the continuous development of medical technology, early detection and benignity and malignancy identification of lung diseases have become increasingly important. Lung nodule is one of the common lung lesions, and its benignity and malignancy identification is of great significance for early detection and intervention of lung-related diseases. However, the existing lung nodule benignity and malignancy identification methods have certain limitations in identification accuracy, which cannot fully meet the clinical demand for high-precision benignity and malignancy identification.

[0003] At present, one of the common lung nodule benignity and malignancy identification methods is to rely solely on medical images such as chest X-ray images, CT (Computed Tomography) images, etc. for benignity and malignancy identification. For example, a visual benignity and malignancy identification method based on CT images, doctors observe the shape, size, edge features, etc. of lung nodules on CT images, and make benignity and malignancy identification combined with clinical experience. This method mainly depends on the subjective judgment of doctors, and has the following technical defects: first, strong subjectivity: different doctors may have significant differences in the benignity and malignancy identification results of the same lung nodule, affecting the identification accuracy; second, limited accuracy: relying only on the features shown on the medical images cannot fully identify the benignity and malignancy of the lung nodules, which may lead to missed diagnosis or misdiagnosis, especially in the early screening of lung-related diseases, this limitation is particularly obvious; third, lack of quantitative indicators: there is a lack of objective quantitative indicators to support the benignity and malignancy identification, making it difficult to accurately classify the risk, further limiting the identification accuracy; fourth, limited clinical experience: for some atypical lung nodules, the identification accuracy is limited due to the limitation of clinical experience.

[0004] In summary, the existing lung nodule benignity and malignancy identification methods have significant shortcomings in identification accuracy, which cannot fully meet the clinical demand for high-precision benignity and malignancy identification. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a lung elastography and AI assisted lung nodule benignity and malignancy identification method and storage medium, which has high identification accuracy.

[0006] The technical solution adopted by the present application to solve the above technical problems is: a lung elastography and AI assisted lung nodule benignity and malignancy identification method, characterized by comprising the following steps: Step 1: Obtain the patient's individual characteristic data and multi-source detection data of lung nodules. The multi-source detection data of lung nodules includes at least lung elastography data and bio-digital twin multi-physics field data with parameters including respiratory airflow, tissue deformation, drug metabolism, and bioelectric propagation. Then, the multi-source detection data of lung nodules is preprocessed to obtain preprocessed multi-source detection data of lung nodules. Step 2: Use principal component analysis to reduce the dimensionality of the parameters of each data in the preprocessed multi-source lung nodule detection data to obtain the principal components corresponding to each data; then combine all principal components into a principal component set; then, according to the preset screening criteria, select a set of principal components from the principal component set that can retain the information related to the differentiation of benign and malignant lung nodules in the preprocessed multi-source lung nodule detection data to the greatest extent, and use it as a representative feature combination. Step 3: Preliminary identification by AI model: Input representative feature combinations into the pre-trained lung nodule benign or malignant identification model, and output a binary classification result of lung nodule as benign or malignant and the first confidence score. If the lung nodule is identified as malignant, continue to execute. Step 4: Input the individual feature data and representative feature combination into the pre-trained bio-digital twin model, perform multiphysics simulation, and output bio-digital twin multiphysics simulation data. The parameters of the bio-digital twin multiphysics simulation data include respiratory airflow dynamics simulation, tissue deformation mechanics simulation, drug metabolism kinetics simulation, and bioelectric propagation simulation. Step 5: Extract all parameters from the multiphysics simulation data of the bio-digital twin and form a simulation feature vector; then, weight each element in the simulation feature vector to obtain a simulation feature weighted vector; next, sample each element in the simulation feature weighted vector at different time points to obtain the time sampling sequence vector corresponding to each element in the simulation feature weighted vector; finally, calculate the cosine similarity between pairwise time sampling sequence vectors to obtain a similarity matrix, where the rows and columns represent the time sampling sequence vectors corresponding to different parameters in the multiphysics simulation data of the bio-digital twin, and the elements are the cosine similarity between the corresponding two time sampling sequence vectors; Step 6: Process the similarity matrix using the comprehensive consistency index calculation method to obtain the comprehensive consistency index of the multiphysics simulation data of the biological digital twin; Step 7: Deep identification by AI model: Input individual feature data, representative feature combination, and comprehensive consistency index into the pre-trained association model, and output the predicted probability of the degree of malignancy risk; Step 8: Predict the probability based on the degree of malignancy risk, and use a confidence adjustment algorithm to adjust the first confidence level to obtain the second confidence level; Step 9: Generate a multimodal identification report containing the predicted probability of malignancy risk level and second confidence level, multiphysics simulation data of the biological digital twin and its comprehensive consistency index, and present it on the display terminal.

[0007] Furthermore, the parameters of the individual characteristic data include the degree of pollution exposure in the living environment, occupational exposure history, past treatment history of lung diseases, and family history of genetic diseases; the multi-source detection data of lung nodules also includes respiratory function data and blood test data.

[0008] Preferably, the preprocessing process is as follows: first, each type of data in the multi-source detection data of lung nodules is automatically cleaned, and then the automatically cleaned data is standardized. The automatic cleaning includes deduplication, error correction, and completion.

[0009] Preferably, the pre-trained lung nodule benign or malignant identification model is a pre-trained machine learning model or a deep learning model. The training process of the lung nodule benign / malignant differentiation model is as follows: a) Obtain multi-source detection data of lung nodules from several patients. First, obtain the preprocessed multi-source detection data of lung nodules according to the process in step 1. Then, obtain the representative feature combination corresponding to each patient according to the process in step 2. Finally, combine the representative features corresponding to all patients to form a training set. b) Initialize the network parameters of the lung nodule benign / malignant differentiation model, including weights and biases; c) In each training round, a batch of data samples is drawn from the training set, wherein the data samples are representative feature combinations; d) Input this batch of data samples into the lung nodule benign and malignant differentiation model, and use the preset feature extraction method to extract modal features from the data samples; then calculate the attention weights of different modal features through the attention mechanism; then combine the attention weights of different modal features to perform weighted fusion of the extracted modal features to obtain comprehensive features; and then obtain the binary classification result and the first confidence score based on the comprehensive features. e) After all data samples in the training set have undergone one training round, calculate the cross-entropy loss function value between the binary classification result and the true label; f) Calculate the gradient of the cross-entropy loss function value with respect to the parameters of each layer of the lung nodule benign and malignant identification model using the backpropagation algorithm; then use the Adam optimizer to update the network parameters of the lung nodule benign and malignant identification model based on the calculated gradient and the dynamically adjusted learning rate. g) Proceed to the next training round and continuously monitor the cross-entropy loss function value. When the decrease in the cross-entropy loss function value within a preset number of consecutive training rounds is less than a preset threshold, the training ends.

[0010] Further specifying, the process of acquiring the bio-digital twin model is as follows: using a bio-digital twin modeling algorithm, a bio-digital twin model is constructed on a local server or cloud server based on the collected patient biomedical data, wherein the patient biomedical data includes physiological data and medical imaging data.

[0011] Preferably, the correlation model is a machine learning model or a deep learning model, used to analyze the correlation between individual characteristic data, representative feature combinations, comprehensive consistency indicators and the predicted probability of malignancy risk.

[0012] Further specifying, the correlation between the individual characteristic data and the bio-digital twin multiphysics data is as follows: the degree of pollution exposure in the living environment is correlated with respiratory airflow; occupational exposure history is correlated with tissue deformation; past lung disease treatment history is correlated with drug metabolism; and family history of genetic diseases is correlated with bioelectric propagation. The process of obtaining the simulated feature weighting vector is as follows: the i-th element in the simulated feature weighting vector is represented as A. wi A wi = A i ×F i , where, i=1,2,3,4, A i F represents the i-th element in the simulated feature vector. i This represents the i-th element in the individual feature vector, which is composed of individual feature data.

[0013] Further specifying, the process of obtaining the second confidence level is as follows: Step 8.1: Quantify the values ​​of each parameter in the individual characteristic data to obtain the quantified values ​​of each parameter; then apply a weighted summation method to the quantified values ​​of all parameters in the individual characteristic data, and use the obtained value as the patient's risk assessment value; then determine whether the patient's risk assessment value exceeds the preset risk score threshold. If it exceeds, proceed to step 8.2; if it does not exceed, proceed to step 8.3. Step 8.2: Confidence Adjustment Algorithm. An adaptive confidence adjustment algorithm based on a deep reinforcement learning model is selected to adjust the first confidence level to obtain the second confidence level. The training process of the deep reinforcement learning model is as follows: a) Initialize the network parameters of the deep reinforcement learning model, including weights and biases; initialize the ε value in the ε-greedy policy; set up an experience replay pool to store experience samples, which include the current state, action, reward, and next state; b) The deep reinforcement learning model selects an action based on the current state using an ε-greedy policy. After executing the action, it obtains the corresponding reward and the next state. When selecting an action using the ε-greedy policy, the action with the largest Q value is selected with a preset probability value, while the remaining probability is used to randomly select an action. Then, the current state, action, reward, and next state are combined into an experience sample and stored in the experience replay pool. c) During the training phase, a predetermined number of experience samples are randomly selected from the experience replay pool to train the deep reinforcement learning model. During the training process, the gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the network parameters are updated according to the gradient descent method. The ε value is gradually decreased during the training process. d) When the decrease in the loss function over multiple rounds is less than a preset threshold and the reward reaches or exceeds a preset reward level, the deep reinforcement learning model is considered to have converged, and training ends. Step 8.3: Combine the individual characteristic data, representative characteristic combinations, and comprehensive consistency index from Step 7 to form the current sample; then, for the same patient, repeat steps 1 to 7. N -1 times, to get the th Predicted probability of the degree of malignancy risk for each sample; Step 8.4: Determine the second to the third N If the standard deviation of the predicted probability of malignancy risk level for each sample is greater than the set fluctuation threshold, then the confidence adjustment algorithm will use an uncertainty-weighted adjustment algorithm, calculated as follows: ,in, This indicates the second confidence level. Indicates the first confidence level. This indicates the predicted probability of the severity of the current sample. Represents the trade-off coefficients used for balancing. and , If the absolute value is used, then the confidence adjustment algorithm uses a trend-weighted adjustment algorithm, calculated as follows: ,in, Indicates the trend influence coefficient. Represents a first-order difference sequence The average value, , , Indicates the first Predicted probability of malignancy risk level for each sample hour Predict the probability of the severity of the current sample. hour Indicates the first The predicted probability of the degree of malignancy risk corresponding to each sample.

[0014] Further specifying, the formula for calculating the comprehensive consistency index is: ,in, This indicates a comprehensive consistency index. This indicates the number of parameters in the multiphysics simulation data of a biological digital twin. The first in the multiphysics simulation data of a biological digital twin m The parameter and the first j Weights among the parameters The first in the multiphysics simulation data of a biological digital twin m The time sampling sequence vector corresponding to the parameter and the first parameter j Cosine similarity between the time-sampled sequence vectors corresponding to each parameter.

[0015] A storage medium, characterized in that the storage medium stores a computer program, which, when executed by a processor, implements the lung elastography and AI-assisted method for differentiating benign and malignant lung nodules as described above.

[0016] Compared with the prior art, the advantages of the present invention are as follows: 1) The acquired multi-source detection data of lung nodules includes at least lung elastography data and bio-digital twin multi-physics field data. This allows the lung nodule benign and malignant differentiation model to assess the characteristics of lung nodules from multiple perspectives, thereby improving the accuracy of benign and malignant differentiation.

[0017] 2) Principal component analysis (PCA) was used to reduce the dimensionality of the parameters of each data in the preprocessed multi-source lung nodule detection data, and principal components related to the differentiation of benign and malignant lung nodules were selected. This not only reduced the complexity of the data, but also improved the model's ability to identify key features, and further improved the accuracy of benign and malignant differentiation.

[0018] 3) The final multimodal identification report not only includes the predicted probability of malignancy risk and the second confidence level, but also covers the multiphysics simulation data of the bio-digital twin and its comprehensive consistency index. This comprehensive report format provides doctors with rich identification information, helping to more comprehensively assess the patient's condition.

[0019] 4) An AI model is used for initial identification, quickly providing a binary classification of lung nodules as benign or malignant with a first confidence level. If the initial identification indicates malignancy, further in-depth identification is performed to assess the degree of malignancy risk. This phased identification process can improve the accuracy of identification.

[0020] 5) Incorporating patients’ individual characteristics data (such as the degree of pollution exposure in the living environment, occupational exposure history, past treatment history of lung diseases, family history of genetic diseases, etc.) allows the method of the present invention to better adapt to the specific circumstances of different patients. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall implementation process of the method of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0023] The present invention proposes a method for differentiating benign and malignant pulmonary nodules using lung elastography and AI assistance, the overall implementation flowchart of which is shown below. Figure 1 As shown, it includes the following steps: Step 1: Obtain the patient's individual characteristic data and multi-source detection data of lung nodules. The multi-source detection data of lung nodules includes at least lung elastography data and bio-digital twin multi-physics field data with parameters including respiratory airflow, tissue deformation, drug metabolism, and bioelectric propagation. Then, the multi-source detection data of lung nodules is preprocessed to obtain preprocessed multi-source detection data of lung nodules.

[0024] To further define the parameters, individual characteristic data include the level of pollution exposure in the living environment, occupational exposure history, past treatment history of lung diseases, and family history of genetic diseases; multi-source detection data for lung nodules also includes respiratory function data and blood test data.

[0025] The above-mentioned pollution exposure level in the living environment refers to the pollutant concentration at the patient's residence obtained through connection with environmental monitoring departments; occupational exposure history refers to the exposure information matched by the database of the medical information management system or electronic health record (EHR) system based on the patient's occupational information; family history of genetic diseases refers to family disease information extracted after authorization through connection with the electronic medical record system; past lung disease treatment history refers to lung treatment records obtained after authorization through connection with the integrated medical information system; lung elastography data is received in real time through interaction with medical imaging equipment specifically used to acquire lung elastography data; respiratory function data is obtained by automatically uploading test indicators through connection with the detection instrument system; blood test data is obtained by synchronous detection through connection with the laboratory information system; and bio-digital twin multiphysics field data is generated by simulation based on physiological and imaging data modeling.

[0026] Here, respiratory airflow, tissue deformation, drug metabolism, and bioelectrical propagation in the bio-digital twin multiphysics data are all generated through simulation. The simulation program of the bio-digital twin model is initiated, simulation parameters (such as simulation duration and time step) are set, and the simulation is run to generate respiratory airflow, tissue deformation, drug metabolism, and bioelectrical propagation. The structure of lung elastography data, respiratory function data, and blood test data is analyzed, such as data dimensions and data types. Based on these structural characteristics, the parameters in the bio-digital twin multiphysics data are processed, such as adjusting data dimensions and converting data types, to ensure structural compatibility. Lung elastography data, respiratory function data, and blood test data, together with the structurally matched bio-digital twin multiphysics data, constitute the multi-source detection data for lung nodules, ensuring the integrity and consistency of all types of data in the multi-source detection data for lung nodules.

[0027] As a preferred approach, the preprocessing procedure is as follows: first, each data point in the multi-source lung nodule detection data is automatically cleaned; then, the automatically cleaned data is standardized. Automatic cleaning includes deduplication, error correction, and data completion. Standardization uses the Z-score method. Deduplication removes duplicate data records, ensuring data uniqueness; error correction corrects erroneous values ​​through data validation and verification algorithms, ensuring data accuracy; data completion fills in missing data values ​​through interpolation or model-based prediction methods, ensuring data integrity; and standardization transforms the data into a form with zero mean and unit variance for subsequent analysis and processing.

[0028] Step 2: Use Principal Component Analysis (PCA) to reduce the dimensionality of the parameters of each data point in the preprocessed multi-source lung nodule detection data (dimensionality reduction reduces the number of features and complexity of the data without losing too much information, making it easier for subsequent analysis and processing), and obtain the principal components corresponding to each data point; then combine all principal components into a principal component set; then, according to the preset screening criteria, select a set of principal components from the principal component set that can retain the most relevant information for differentiating benign and malignant lung nodules in the preprocessed multi-source lung nodule detection data as a representative feature combination.

[0029] Principal Component Analysis (PCA) is a commonly used data dimensionality reduction technique. It transforms multiple correlated variables into a set of uncorrelated new variables called principal components through linear transformation. These principal components are arranged in descending order of variance, and the first few principal components typically contain most of the information from the original data. The PCA process is as follows: 1) Calculate the covariance matrix: Calculate the covariance matrix between each variable in the data (e.g., the parameters of each data point in multi-source lung nodule detection data). The covariance matrix reflects the correlation between variables; 2) Solve for eigenvalues ​​and eigenvectors: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​represent the magnitude of the variance of the principal components, and the eigenvectors represent the direction of the principal components; 3) Select principal components: Sort the eigenvectors in descending order of eigenvalues ​​and select the principal components corresponding to the first k eigenvectors, such that the sum of the variances of these principal components reaches a preset proportion (e.g., 80% or 90%), thereby achieving data dimensionality reduction.

[0030] Here, the parameters of lung elastography data include texture features, edge features, and intensity features; the parameters of respiratory function data include respiratory rate, tidal volume, and vital capacity; and the parameters of blood test data include white blood cell count, red blood cell count, and hemoglobin concentration.

[0031] Here, preset screening criteria can be set according to specific business needs and experience. The purpose is to screen out the principal components that can retain the information related to the differentiation of benign and malignant lung nodules in the preprocessed multi-source lung nodule detection data to the greatest extent. For example, the variance contribution rate of the principal component is higher than the preset contribution rate or the correlation with the differentiation of benign and malignant lung nodules is higher than the preset correlation threshold as preset screening criteria. The preset contribution rate and preset correlation threshold can be preset by the user.

[0032] Step 3: Preliminary identification by AI model: The representative feature combination is input into the pre-trained lung nodule benign / malignant identification model, which outputs a binary classification result of lung nodule as benign or malignant and a first confidence score. If the lung nodule is identified as malignant, the process continues. The first confidence score represents the degree of credibility of the pre-trained lung nodule benign / malignant identification model for the binary classification result, which is usually represented by a probability value. The higher the first confidence score, the greater the confidence of the pre-trained lung nodule benign / malignant identification model in the binary classification result.

[0033] Preferably, the pre-trained lung nodule benign or malignant differentiation model is a pre-trained machine learning model or deep learning model. It can output a binary classification result based on the combination of representative features of the input. The machine learning model or deep learning model is trained with a large number of samples and calculates and analyzes the combination of representative features of the input based on the relationship between the learned weights and the binary classification result.

[0034] Further defining the training process for the lung nodule benign / malignant differentiation model is as follows: a) Obtain multi-source detection data of lung nodules from several patients. First, obtain the preprocessed multi-source detection data of lung nodules according to the process in step 1. Then, obtain the representative feature combination corresponding to each patient according to the process in step 2. Finally, combine the representative features corresponding to all patients to form a training set.

[0035] b) Initialize the network parameters of the lung nodule benign / malignant differentiation model, including weights and biases.

[0036] c) In each training round, a batch of data samples is drawn from the training set, wherein the data samples are representative feature combinations.

[0037] Here, data sample extraction: A batch of data samples is extracted according to preset sampling rules (such as random sampling, stratified sampling, etc.). The sampling rules ensure that the extracted data samples are representative and can cover the data distribution of different features in the training set.

[0038] d) This batch of data samples is input into a lung nodule benign / malignant differentiation model. A pre-defined feature extraction method is used to extract modal features from the data samples. Then, an attention mechanism is used to calculate the attention weights of different modal features. The extracted modal features are then weighted and fused together using these attention weights to obtain a comprehensive feature. Finally, based on the comprehensive feature, a binary classification result and a first confidence score are obtained. Here, the fusion of multimodal features and the attention mechanism improves the accuracy of feature extraction.

[0039] Here, modal feature extraction: For each data sample in this batch of extracted data samples, according to the data modality type it contains, the corresponding preset feature extraction method is called to extract modal features, such as extracting texture features from lung elastography data, and extracting key indicator numerical features from respiratory function data, etc.

[0040] Here, the preset feature extraction methods are: based on the characteristics of different modalities of data, feature extraction methods are preset. For example, for image modalities (such as lung elastography data), feature extraction methods related to convolutional neural networks (CNN) can be used; for numerical data (such as respiratory function data and blood test data), statistical analysis, principal component analysis and other methods can be used to extract features.

[0041] Here, the attention weights of different modal features are calculated using an attention mechanism network, which is a special type of neural network that can automatically learn the importance of different modal features. Since different modal features contribute differently to the final identification, the attention mechanism can highlight important modal features and suppress unimportant modal features, thereby improving the performance of the model.

[0042] Attention weights: These represent the relative importance of different modal features in the fusion process. The higher the attention weight, the greater the influence of that modal feature on the final result; the lower the attention weight, the smaller the influence. These attention weights are learned through an attention mechanism network.

[0043] Here, the weighted fusion to obtain the comprehensive feature is achieved through a multimodal fusion network set up in the lung nodule benign / malignant differentiation model. A multimodal fusion network is a neural network capable of processing multiple different types of data (modalities). It can fuse lung elastography data, respiratory function data, blood test data, and bio-digital twin multiphysics data. In the multimodal fusion network, network branches are specifically designed to handle different modal features; for example, a convolutional neural network branch for processing lung elastography data and a fully connected neural network branch for processing respiratory function data. The weighted fusion process is as follows: based on the calculated attention weights, the features output by each network branch are weighted, that is, the feature vector or matrix output by each network branch is multiplied by its corresponding attention weight. Then, the weighted features of each modality are added or concatenated to obtain the comprehensive feature. For example, the network branch output features for lung elastography data are multiplied by their attention weights, and the network branch output features for respiratory function data are multiplied by their attention weights. Then, these two weighted feature vectors are added dimensionally to obtain a comprehensive feature.

[0044] Multimodal fusion networks consist of multiple layers, each containing a series of parameters such as weights and biases. Weights determine the degree of influence of input data across different dimensions as it passes through that layer, while biases adjust the activation thresholds of neurons. These parameters are continuously updated during network training to optimize network performance. Before training begins, initial values ​​are assigned to the parameters of each layer in the multimodal fusion network. Proper initialization helps the network converge and learn faster, avoiding problems such as vanishing or exploding gradients during training. The initialization process is as follows: Determine the initialization method: Based on the structure and data characteristics of the multimodal fusion network, select a suitable initialization method from the preset options. For example, if the network structure is complex and contains multiple ReLU activation functions, the Kaiming initialization method can be chosen. Parameter initialization operation: For each layer in the multimodal fusion network, generate corresponding initial parameter values ​​according to the selected initialization method. For example, for weight parameters, when using normal distribution initialization, random numbers are generated as initial weight values ​​based on the set mean and standard deviation; for bias parameters, they can generally be initialized to 0 or a small constant. The generated initial parameter values ​​are assigned to the corresponding parameter variables of each layer of the network through program code, thus completing parameter initialization and preparing for the subsequent training process.

[0045] e) After all data samples in the training set have undergone one training round, calculate the cross-entropy loss function value between the binary classification result and the true label.

[0046] f) Calculate the gradient of the cross-entropy loss function value with respect to the parameters of each layer of the lung nodule benign / malignant identification model using the backpropagation algorithm; then use the Adam optimizer to update the network parameters of the lung nodule benign / malignant identification model based on the calculated gradient and the dynamically adjusted learning rate.

[0047] g) Proceed to the next training round and continuously monitor the cross-entropy loss function value. When the decrease in the cross-entropy loss function value within a preset number of consecutive training rounds is less than a preset threshold, the training ends.

[0048] Step 4: Input the individual feature data and representative feature combination into the pre-trained biological digital twin model, and use the preset physical equations and algorithms to perform multiphysics simulation (simulate multiple interacting physical fields in the organism), and output the biological digital twin multiphysics simulation data. The parameters of the biological digital twin multiphysics simulation data include respiratory airflow dynamics simulation, tissue deformation mechanics simulation, drug metabolism kinetics simulation, and bioelectric propagation simulation.

[0049] As a preferred approach, the process of acquiring a bio-digital twin model is as follows: using bio-digital twin modeling algorithms such as finite element analysis and computational fluid dynamics, a bio-digital twin model is constructed on a local server or cloud server (such as a medical-specific server with a bio-digital twin modeling engine deployed) based on the collected patient biomedical data. The patient biomedical data includes physiological data (such as respiratory rate, heart rate, etc.) and medical imaging data (such as CT images, MRI images, etc.).

[0050] Step 5: Extract all parameters from the multiphysics simulation data of the bio-digital twin and form a simulated feature vector; then, use an existing weighting method to weight each element in the simulated feature vector to obtain a simulated feature weighted vector; then, sample at different time points for each element in the simulated feature weighted vector to obtain the time sampling sequence vector corresponding to each element in the simulated feature weighted vector; finally, use the cosine similarity algorithm to calculate the cosine similarity between pairwise time sampling sequence vectors to obtain a similarity matrix, which is a square matrix. The rows and columns of the matrix represent the time sampling sequence vectors corresponding to different parameters in the multiphysics simulation data of the bio-digital twin, and the elements in the matrix are the cosine similarity between the corresponding two time sampling sequence vectors. The similarity matrix provides an intuitive basis for analyzing the similarity and synergistic relationship between parameters in the multiphysics simulation data of the bio-digital twin, and helps in the identification and judgment of lung nodules; Here, the sampling frequency for obtaining the time-sampled sequence vector can be set by the user; the cosine similarity algorithm is a method for calculating the similarity between two vectors. It measures the degree of similarity by calculating the cosine value of the angle between the two vectors. The closer the cosine value is to 1, the more similar the two vectors are; the closer it is to -1, the less similar the two vectors are.

[0051] Furthermore, the correlation between individual characteristic data and bio-digital twin multiphysics data is as follows: the degree of pollution exposure in the living environment is associated with respiratory airflow; occupational exposure history is associated with tissue deformation; past lung disease treatment history is associated with drug metabolism; and family history of genetic diseases is associated with bioelectric propagation. The process of obtaining the simulated feature weighting vector is as follows: the i-th element in the simulated feature weighting vector is represented as A. wi A wi = A i ×F i , where, i=1,2,3,4, A i F represents the i-th element in the simulated feature vector. i This represents the i-th element in the individual feature vector, which is composed of individual feature data.

[0052] Step 6: The similarity matrix is ​​processed using the comprehensive consistency index calculation method to obtain the comprehensive consistency index of the bio-digital twin multiphysics simulation data. The comprehensive consistency index is used to measure the overall consistency between parameters in the bio-digital twin multiphysics simulation data. The higher the index, the better the consistency between parameters, and vice versa. This index comprehensively considers the relationship between the parameters in the bio-digital twin multiphysics simulation data at different disease stages.

[0053] Here, the formula for calculating the overall consistency index is: ,in, This indicates a comprehensive consistency index. This indicates the number of parameters in the multiphysics simulation data of a biological digital twin. The first in the multiphysics simulation data of a biological digital twin m The parameter and the first j Weights among the parameters The first in the multiphysics simulation data of a biological digital twin m The time sampling sequence vector corresponding to the parameter and the first parameter j The cosine similarity between the time-sampled sequence vectors corresponding to each parameter, when calculating the comprehensive consistency index, integrates the weights with the cosine similarity, which can comprehensively and accurately reflect the degree of coordination between parameters in the multi-physics simulation data of biological digital twins, providing a more scientific quantitative basis for the differentiation of benign and malignant lung nodules.

[0054] Step 7: Deep identification by AI model: Input individual feature data, representative feature combination, and comprehensive consistency index into the pre-trained association model, and output the predicted probability of malignancy risk. The predicted probability of malignancy risk ranges from 0 to 1. The closer it is to 1, the greater the possibility of malignancy of the lung nodule.

[0055] Preferably, the correlation model is a machine learning model or a deep learning model, used to analyze the correlation between individual characteristic data, representative feature combinations, comprehensive consistency indicators and the predicted probability of malignancy risk.

[0056] The training process of the association model is as follows: During the training phase, a large number of samples containing individual feature data, representative feature combinations and comprehensive consistency indicators, as well as real labels, are collected. The model is trained using a machine learning model or a deep learning model. The model parameters are adjusted to minimize the error between the prediction results and the real labels until the model converges, and the trained association model is obtained.

[0057] Step 8: Predict the probability based on the degree of malignancy risk, and use a confidence adjustment algorithm to adjust the first confidence level to obtain a second confidence level. Here, the confidence adjustment algorithm is used to calculate an adjusted confidence level that better reflects the actual situation, thus making the results more reliable.

[0058] To further specify, the process for obtaining the second confidence level is as follows: Step 8.1: Quantify the parameters of the individual characteristic data numerically. For example, the pollution exposure level in the living environment is quantified using the PM2.5 cumulative exposure index to obtain the numerical quantified values ​​of each parameter of the individual characteristic data. Then, apply a weighted summation method to the numerical quantified values ​​of all parameters of the individual characteristic data and use the obtained value as the patient's risk assessment value. Then, determine whether the patient's risk assessment value exceeds the preset risk score threshold. If it exceeds, proceed to step 8.2; if it does not exceed, proceed to step 8.3.

[0059] Step 8.2: Confidence Adjustment Algorithm. An adaptive confidence adjustment algorithm based on a deep reinforcement learning model is selected to adjust the first confidence level to obtain the second confidence level. The training process of the deep reinforcement learning model is as follows: a) Initialize the network parameters of the deep reinforcement learning model, including weights and biases; initialize the ε value in the ε-greedy policy; set up an experience replay pool to store experience samples, which include the current state, action, reward and next state, and use the experience replay pool to improve training efficiency.

[0060] b) The deep reinforcement learning model selects an action based on the current state using an ε-greedy policy. After executing the action, it obtains the corresponding reward and the next state. When selecting an action using the ε-greedy policy, the action with the largest Q value is selected with a preset probability value, while the remaining probability is used to randomly select an action. The ε value determines the balance between exploration (random action selection) and exploitation (selecting the action with the largest Q value) in the deep reinforcement learning model. Then, the current state, action, reward, and next state are combined into an experience sample and stored in the experience replay pool to ensure the accumulation of experience samples and provide data for subsequent training.

[0061] Current State: This refers to the current state information obtained from the environment, such as the patient's risk assessment value and first confidence level. Action: This could be an adjustment action, such as increasing, decreasing, or maintaining the first confidence level. Reward: This is the feedback signal from the environment after the model performs the action, used to measure the quality of the action. If the second confidence level is closer to reality, the model receives a positive reward; conversely, if the second confidence level deviates further from reality, it receives a negative reward. Setting the reward value helps guide the model to learn the correct strategy. Next State: The new state of the environment after the model performs the action. For example, after adjusting the confidence level, the new state formed by combining new relevant patient information (such as newly added test results), including the updated risk assessment value and the adjusted confidence level.

[0062] Q-value: In reinforcement learning, the Q-value represents the expected cumulative reward obtained by taking a certain action in a given state. The model learns and attempts to find the action that maximizes the Q-value in each state, thereby achieving the optimal policy.

[0063] c) During the training phase, a predetermined number of experience samples are randomly selected from the experience replay pool to train the deep reinforcement learning model. The selected experience samples contain information on the state, action, reward, and next state at different times, providing diverse data for model training. During training, the gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the network parameters are updated according to the gradient descent method. The ε value is gradually decreased during training. As training progresses, the deep reinforcement learning model gradually shifts from exploration to utilization, improving the certainty of decision-making.

[0064] Dynamically adjusting the ε value and updating the network parameters helps the deep reinforcement learning model converge, adaptively adjusts the first confidence level, and improves the accuracy of distinguishing between benign and malignant lung nodules.

[0065] Here, ε decays: After each training step or a certain number of training steps (e.g., 100), the system reduces the ε value according to a pre-defined ε decay rule (e.g., exponential decay, linear decay, etc.). For example, using linear decay, the initial ε value is 0.9, and the decay rate is 0.001. After each training step, the ε value is updated to ε = ε - 0.001. As training progresses, the ε value becomes smaller and smaller, and the probability of the model randomly exploring new actions gradually decreases, relying more on the learned optimal actions.

[0066] d) When the decrease in the loss function over multiple rounds is less than a preset threshold and the reward reaches or exceeds a preset reward level, the deep reinforcement learning model is considered to have converged, and training ends.

[0067] Convergence Judgment: Check whether the decrease in the loss function over multiple consecutive training epochs is less than a preset threshold. For example, assuming a judgment period of 5 consecutive training epochs, trace back 5 epochs from the current epoch and calculate the decrease in the loss function between each adjacent epoch. If the decrease in each of these 5 epochs is less than the preset threshold of 0.001, simultaneously check whether the current reward value reaches or exceeds the preset reward level of 0.8.

[0068] Deep reinforcement learning models: These are models that allow agents to interact with their environment and continuously learn optimal strategies based on reward signals received from the environment. In the scenario of adjusting confidence levels for lung nodules, this model utilizes a deep neural network structure to process complex state information, enabling adaptive adjustment of confidence levels. Network parameters (weights and biases): Deep neural networks consist of multiple neurons. Weights determine the strength of signal transmission between neurons, while biases are used to adjust the activation thresholds of neurons. These parameters are continuously updated during model training to optimize model performance. ε-greedy policy: A decision-making strategy used to balance exploration and exploitation in the model. In this policy, the agent randomly selects actions with probability ε to explore new possibilities; it then selects the action currently considered optimal (maximum Q-value) with probability 1-ε to exploit, i.e., utilizing existing experience. Initial ε value: An initial value of ε is set at the start of the ε-greedy policy. This value determines the model's tendency to explore new actions in the early stages of training. It is usually set to a large value to encourage more exploration in the early stages of training, gradually decreasing as training progresses. Experience Replay Pool: This pool stores experiential data generated by the agent during its interaction with the environment. This includes the current state (e.g., a patient's risk assessment value, current confidence level), actions (e.g., adjustments to the confidence level), rewards (feedback on actions, such as a positive reward for a more accurate adjusted confidence level), and the next state (the new state entered after performing an action). Randomly sampling samples from the experience replay pool for training breaks down correlations between data points, improving the stability and efficiency of model training. Preset Number of Experience Samples: This is the maximum number of experience samples the experience replay pool can store. Once this preset number of experience samples is reached, newly generated experience samples will overwrite the oldest samples, ensuring the data in the experience replay pool remains up-to-date and representative.

[0069] Step 8.3: Combine the individual characteristic data, representative characteristic combinations, and comprehensive consistency index from Step 7 to form the current sample; then, for the same patient, repeat steps 1 to 7. N -1 times, to get the th The predicted probability of the degree of malignancy risk corresponding to each sample.

[0070] Step 8.4: Determine the second to the third NIf the standard deviation of the predicted probability of malignancy risk level for each sample is greater than the set fluctuation threshold, then the confidence adjustment algorithm will use an uncertainty-weighted adjustment algorithm, calculated as follows: ,in, This indicates the second confidence level. Indicates the first confidence level. This indicates the predicted probability of the severity of the current sample. Represents the trade-off coefficients used for balancing. and , If the absolute value is used, then the confidence adjustment algorithm uses a trend-weighted adjustment algorithm, calculated as follows: ,in, Indicates the trend influence coefficient. Represents a first-order difference sequence The average value, , , Indicates the first Predicted probability of malignancy risk level for each sample hour Predict the probability of the severity of the current sample. hour Indicates the first The predicted probability of malignancy risk level corresponds to each sample. By analyzing the fluctuation of the predicted probability, the confidence adjustment algorithm can be intelligently selected. When the standard deviation is large, the adjustment algorithm based on uncertainty weighting can balance the first confidence level and the predicted probability of malignancy risk level. When the standard deviation is small, the adjustment algorithm based on trend weighting can be optimized according to the changing trend of the predicted probability of malignancy risk level, making the second confidence level more realistic and improving the reliability of distinguishing between benign and malignant lung nodules.

[0071] Step 9: Generate a multimodal identification report containing the predicted probability of malignancy risk level and second confidence level, multiphysics simulation data of the biological digital twin and its comprehensive consistency index, and present it on the display terminal.

[0072] Multimodal processing includes, but is not limited to, text, charts, and images, which are presented on the display terminal in a multimodal manner to facilitate doctors' intuitive and comprehensive access and analysis.

[0073] The general process of presenting data in a multimodal manner on a display terminal is as follows: 1) Data preparation: The predicted probability of malignancy risk and the second confidence level, the multiphysics simulation data of the bio-digital twin and its comprehensive consistency index are prepared according to the needs of multimodal presentation. For example, the parameters in the multiphysics simulation data of the bio-digital twin are converted into a format suitable for chart display; 2) Presentation format design: The presentation format is determined according to the characteristics of various types of data. For example, the predicted probability of malignancy risk is highlighted with large font text, the second confidence level is displayed with a progress bar or percentage number, the multiphysics simulation data of the bio-digital twin is presented with charts such as line graphs and bar charts to show the changes of different physical fields over time or other variables, and the comprehensive consistency index is presented in the form of numerical values ​​combined with simple text descriptions; 3) Display terminal display: The designed multimodal report is transmitted to the display terminal (such as the doctor's workstation monitor in the hospital) for display. Doctors can directly view this information to assist in the analysis and diagnosis of the patient's condition.

[0074] The present invention also discloses a storage medium storing a computer program thereon, which, when executed by a processor, implements the lung elastography and AI-assisted method for differentiating benign and malignant lung nodules as described above.

[0075] The embodiments described herein are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with this application should be included within the scope of protection of this application.

Claims

1. A method for differentiating benign and malignant pulmonary nodules using pulmonary elastography and AI-assisted methods, characterized in that... Includes the following steps: Step 1: Obtain the patient's individual characteristic data and multi-source detection data of lung nodules. The multi-source detection data of lung nodules includes at least lung elastography data and bio-digital twin multi-physics field data with parameters including respiratory airflow, tissue deformation, drug metabolism, and bioelectric propagation. Then, the multi-source detection data of lung nodules is preprocessed to obtain preprocessed multi-source detection data of lung nodules. Step 2: Principal component analysis is used to reduce the dimensionality of the parameters of each data point in the preprocessed multi-source lung nodule detection data to obtain the principal components corresponding to each data point; Then, all principal components are combined into a principal component set; then, according to the preset screening criteria, a set of principal components that can retain the information related to the differentiation of benign and malignant lung nodules in the preprocessed multi-source lung nodule detection data to the greatest extent are selected from the principal component set as representative feature combination. Step 3: Preliminary identification by AI model: Input representative feature combinations into the pre-trained lung nodule benign or malignant identification model, and output a binary classification result of lung nodule as benign or malignant and the first confidence score. If the lung nodule is identified as malignant, continue to execute. Step 4: Input the individual feature data and representative feature combination into the pre-trained bio-digital twin model, perform multiphysics simulation, and output bio-digital twin multiphysics simulation data. The parameters of the bio-digital twin multiphysics simulation data include respiratory airflow dynamics simulation, tissue deformation mechanics simulation, drug metabolism kinetics simulation, and bioelectric propagation simulation. Step 5: Extract all parameters from the multiphysics simulation data of the bio-digital twin and form a simulation feature vector; then, weight each element in the simulation feature vector to obtain a simulation feature weighted vector; next, sample each element in the simulation feature weighted vector at different time points to obtain the time sampling sequence vector corresponding to each element in the simulation feature weighted vector; finally, calculate the cosine similarity between pairwise time sampling sequence vectors to obtain a similarity matrix, where the rows and columns represent the time sampling sequence vectors corresponding to different parameters in the multiphysics simulation data of the bio-digital twin, and the elements are the cosine similarity between the corresponding two time sampling sequence vectors; Step 6: Process the similarity matrix using the comprehensive consistency index calculation method to obtain the comprehensive consistency index of the multiphysics simulation data of the biological digital twin; Step 7: Deep identification by AI model: Input individual feature data, representative feature combination, and comprehensive consistency index into the pre-trained association model, and output the predicted probability of the degree of malignancy risk; Step 8: Predict the probability based on the degree of malignancy risk, and use a confidence adjustment algorithm to adjust the first confidence level to obtain the second confidence level; Step 9: Generate a multimodal identification report containing the predicted probability of malignancy risk level and second confidence level, multiphysics simulation data of the biological digital twin and its comprehensive consistency index, and present it on the display terminal.

2. The method for differentiating benign and malignant pulmonary nodules using pulmonary elastography and AI-assisted methods according to claim 1, characterized in that... The parameters of the individual characteristic data include the degree of pollution exposure in the living environment, occupational exposure history, past treatment history of lung diseases, and family history of genetic diseases; the multi-source detection data of lung nodules also includes respiratory function data and blood test data.

3. The method for lung elastography and AI-assisted differentiation of benign and malignant lung nodules according to claim 1 or 2, characterized in that... The preprocessing process is as follows: first, each type of data in the multi-source detection data of lung nodules is automatically cleaned, and then the automatically cleaned data is standardized. The automatic cleaning includes deduplication, error correction, and completion.

4. The method for differentiating benign and malignant pulmonary nodules using lung elastography and AI-assisted methods according to claim 1, characterized in that... The pre-trained lung nodule benign / malignant differentiation model is a pre-trained machine learning model or deep learning model. The training process of the lung nodule benign / malignant differentiation model is as follows: a) Obtain multi-source detection data of lung nodules from several patients. First, obtain the preprocessed multi-source detection data of lung nodules according to the process in step 1. Then, obtain the representative feature combination corresponding to each patient according to the process in step 2. Finally, combine the representative features corresponding to all patients to form a training set. b) Initialize the network parameters of the lung nodule benign / malignant differentiation model, including weights and biases; c) In each training round, a batch of data samples is drawn from the training set, wherein the data samples are representative feature combinations; d) Input this batch of data samples into the lung nodule benign and malignant differentiation model, and use the preset feature extraction method to extract modal features from the data samples; then calculate the attention weights of different modal features through the attention mechanism; then combine the attention weights of different modal features to perform weighted fusion of the extracted modal features to obtain comprehensive features; and then obtain the binary classification result and the first confidence score based on the comprehensive features. e) After all data samples in the training set have undergone one training round, calculate the cross-entropy loss function value between the binary classification result and the true label; f) Calculate the gradient of the cross-entropy loss function value with respect to the parameters of each layer of the lung nodule benign and malignant identification model using the backpropagation algorithm; then use the Adam optimizer to update the network parameters of the lung nodule benign and malignant identification model based on the calculated gradient and the dynamically adjusted learning rate. g) Proceed to the next training round and continuously monitor the cross-entropy loss function value. When the decrease in the cross-entropy loss function value within a preset number of consecutive training rounds is less than a preset threshold, the training ends.

5. The method for differentiating benign and malignant pulmonary nodules using pulmonary elastography and AI-assisted methods according to claim 1, characterized in that... The process of acquiring the bio-digital twin model is as follows: using a bio-digital twin modeling algorithm, a bio-digital twin model is constructed on a local server or cloud server based on the collected patient biomedical data, wherein the patient biomedical data includes physiological data and medical imaging data.

6. The method for lung elastography and AI-assisted differentiation of benign and malignant lung nodules according to claim 1, characterized in that... The correlation model is a machine learning model or a deep learning model, used to analyze the correlation between individual characteristic data, representative feature combinations, comprehensive consistency indicators and the predicted probability of malignancy risk.

7. The method for differentiating benign and malignant pulmonary nodules using pulmonary elastography and AI-assisted methods according to claim 1, characterized in that... The correlation between the individual characteristic data and the bio-digital twin multiphysics data is as follows: the degree of pollution exposure in the living environment is correlated with respiratory airflow; occupational exposure history is correlated with tissue deformation; past lung disease treatment history is correlated with drug metabolism; and family history of genetic diseases is correlated with bioelectric propagation. The process of obtaining the simulated feature weighting vector is as follows: the i-th element in the simulated feature weighting vector is represented as A. wi A wi = A i ×F i , where, i=1,2,3,4, A i F represents the i-th element in the simulated feature vector. i This represents the i-th element in the individual feature vector, which is composed of individual feature data.

8. The method for lung elastography and AI-assisted differentiation of benign and malignant lung nodules according to claim 1 or 7, characterized in that... The process of obtaining the second confidence level is as follows: Step 8.1: Quantify the values ​​of each parameter in the individual characteristic data to obtain the quantified values ​​of each parameter; then apply a weighted summation method to the quantified values ​​of all parameters in the individual characteristic data, and use the obtained value as the patient's risk assessment value; then determine whether the patient's risk assessment value exceeds the preset risk score threshold. If it exceeds, proceed to step 8.2; if it does not exceed, proceed to step 8.

3. Step 8.2: Confidence Adjustment Algorithm. An adaptive confidence adjustment algorithm based on a deep reinforcement learning model is selected to adjust the first confidence level to obtain the second confidence level. The training process of the deep reinforcement learning model is as follows: a) Initialize the network parameters of the deep reinforcement learning model, including weights and biases; initialize the ε value in the ε-greedy policy; set up an experience replay pool to store experience samples, which include the current state, action, reward, and next state; b) The deep reinforcement learning model selects an action based on the current state using an ε-greedy policy. After executing the action, it obtains the corresponding reward and the next state. When selecting an action using the ε-greedy policy, the action with the largest Q value is selected with a preset probability value, while the remaining probability is used to randomly select an action. Then, the current state, action, reward, and next state are combined into an experience sample and stored in the experience replay pool. c) During the training phase, a predetermined number of experience samples are randomly selected from the experience replay pool to train the deep reinforcement learning model. During the training process, the gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the network parameters are updated according to the gradient descent method. The ε value is gradually decreased during the training process. d) When the decrease in the loss function over multiple rounds is less than a preset threshold and the reward reaches or exceeds a preset reward level, the deep reinforcement learning model is considered to have converged, and training ends. Step 8.3: Combine the individual characteristic data, representative characteristic combinations, and comprehensive consistency index from Step 7 to form the current sample; then, for the same patient, repeat steps 1 to 7. N -1 times, to get the first Predicted probability of the degree of malignancy risk for each sample; Step 8.4: Determine the second to the third N If the standard deviation of the predicted probability of malignancy risk level for each sample is greater than the set fluctuation threshold, then the confidence adjustment algorithm will use an uncertainty-weighted adjustment algorithm, calculated as follows: ,in, This indicates the second confidence level. Indicates the first confidence level. This indicates the predicted probability of the severity of the current sample. Represents the trade-off coefficients used for balancing. and , If the absolute value is used, then the confidence adjustment algorithm uses a trend-weighted adjustment algorithm, calculated as follows: ,in, Indicates the trend influence coefficient. Represents a first-order difference sequence The average value, , , Indicates the first Predicted probability of malignancy risk level for each sample hour Predict the probability of the severity of the current sample. hour Indicates the first The predicted probability of the degree of malignancy risk corresponding to each sample.

9. The method for lung elastography and AI-assisted differentiation of benign and malignant lung nodules according to claim 1, characterized in that... The formula for calculating the overall consistency index is as follows: ,in, This indicates a comprehensive consistency index. This indicates the number of parameters in the multiphysics simulation data of a biological digital twin. The first in the multiphysics simulation data of a biological digital twin m The parameter and the first j Weights among the parameters The first in the multiphysics simulation data of a biological digital twin m The time sampling sequence vector corresponding to the parameter and the first parameter j Cosine similarity between the time-sampled sequence vectors corresponding to each parameter.

10. A storage medium, characterized in that... The storage medium stores a computer program, which, when executed by a processor, implements the lung elastography and AI-assisted method for differentiating benign and malignant lung nodules as described in claim 9.

Images