COPD early screening method based on dynamic dependency graph and self-supervised learning and application

Abstract

The present application belongs to the field of medical image analysis and artificial intelligence, aiming at the problems of insufficient sensitivity and insufficient feature representation, a COPD early screening method based on dynamic dependency graph and self-supervised learning and application are proposed, including: S1: pre-processing the chest CT image data to obtain enhanced image blocks; S2: generating an image block sequence of the enhanced image blocks; S3: inputting the image block sequence into the self-supervised learning Transformer encoder, and constructing the total loss function of the self-supervised learning; S4: determining the COPD risk probability model, constructing the loss function of the supervised classification task to optimize the COPD risk probability model. S5: constructing a comprehensive loss function to optimize the COPD risk probability model as a whole. The method of the present application integrates multiple advanced technologies, aiming to detect and assess the risk of COPD in an automatic, efficient and high-precision manner.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis and artificial intelligence, specifically involving a method and application for early screening of COPD based on dynamic dependency graphs and self-supervised learning. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a common and serious respiratory disease with high morbidity and mortality rates worldwide. According to the World Health Organization, COPD has become one of the leading causes of death globally, and its incidence continues to rise. Early screening is crucial for slowing disease progression and improving patient prognosis, significantly reducing the associated medical burden and socioeconomic costs. Traditional COPD screening methods primarily rely on pulmonary function tests (such as vital capacity measurement) and imaging examinations (such as chest CT scans). While pulmonary function tests are the gold standard for diagnosing COPD, their detection capabilities are limited in the early stages of the disease, especially for patients with mild or no symptoms, leading to frequent missed diagnoses. Imaging results require interpretation by experienced radiologists, which is subjective and costly. Furthermore, the radiation risks and high costs of CT scans limit their large-scale screening application.

[0003] With the development of computer technology, intelligent screening methods based on radiomics and deep learning have gradually emerged. Radiomics extracts a large number of quantitative features (such as shape, texture, grayscale, etc.) from CT images and uses machine learning algorithms to analyze the correlation between these features and COPD, so as to achieve automated disease diagnosis and risk assessment.

[0004] While radiomics has improved the automation and objectivity of screening to some extent, its heavy reliance on manually designed features may overlook the complex spatial relationships and contextual information inherent in images. Furthermore, the variability in image data caused by different scanning devices and parameters limits the generalization ability of radiomics models across different datasets, and the requirement for high-quality labeled data makes the training process time-consuming and labor-intensive, restricting large-scale application. Deep learning techniques, especially convolutional neural networks, have shown great potential in medical image analysis, capable of automatically extracting high-level features from images to achieve disease classification and prediction.

[0005] However, traditional deep learning methods also face challenges in COPD screening, such as large data requirements, insufficient robustness, and the need for extensive annotation. In summary, existing early COPD screening methods all have the following main limitations:

[0006] Insufficient sensitivity: Lung function tests have limited detection capabilities in the early stages of disease and are prone to missed diagnoses.

[0007] Reliance on professional interpretation: Imaging examination results require interpretation by experienced radiologists, which involves strong subjectivity and high interpretation costs.

[0008] Insufficient feature representation: Traditional methods may not be able to fully capture the complex spatial relationships and contextual information in images during feature extraction, which limits the model's recognition ability. Furthermore, they rely entirely on supervised training and do not make full use of the information in the data itself.

[0009] Insufficient robustness: Traditional deep learning and machine learning methods lack robustness when processing CT data from different sources and under different conditions, making it difficult to achieve consistent screening results. Summary of the Invention

[0010] To address the aforementioned problems, the first aspect of this invention proposes an early COPD screening method based on dynamic dependency graphs and self-supervised learning, characterized by comprising:

[0011] S1: Preprocessing chest CT image data to obtain enhanced image blocks;

[0012] S2: Generate an image patch sequence that enhances the image patch;

[0013] S3: Input the elements in the image patch sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image patches, and construct the total loss function of self-supervised learning;

[0014] S4: Input the elements in the image patch sequence into the Transformer encoder of the supervised classification task to extract the supervised classification task features of the image patch. After weighting, the global features are obtained. The global features are mapped to the COPD risk score by the classifier to obtain the COPD risk probability model. The loss function of the supervised classification task is constructed to optimize the COPD risk probability model and realize early screening of COPD.

[0015] S5: Combine the total loss function of self-supervised learning with the loss function of supervised classification tasks to construct a comprehensive loss function, thereby achieving overall optimization of the model for predicting the probability of COPD risk.

[0016] Preferably, S2 includes:

[0017] S2.1: Obtain the centrality score and local feature importance score of the enhanced image patch;

[0018] S2.2: Determine the processing order of the enhanced image blocks to obtain the image block sequence.

[0019] Preferably, S2.1 includes:

[0020] S2.1.1: The similarity of the enhanced image patches is calculated using the cosine similarity formula;

[0021] S2.1.2: Calculate the centrality score for each enhanced image patch;

[0022] S2.1.5: Calculate the importance score of local features of enhanced image patches based on similarity.

[0023] Preferably, S2.2 includes:

[0024] S2.2.1: Combine the centrality score and local feature importance score of the enhanced image patch to calculate the priority score of each enhanced image patch;

[0025] S2.2.2: A priority queue algorithm based on a greedy strategy generates an image block sequence of enhanced image blocks by processing them in descending order of priority score.

[0026] Preferably, S2.2.2 includes:

[0027] S2.2.2.1: Initialize the enhanced image patches to an unprocessed state and construct a set of unprocessed image patches;

[0028] S2.2.2.2: Select the image patch with the highest priority score from the set of unprocessed image patches;

[0029] S2.2.2.3: Add the image patch selected in S2.2.2.2 to the processing sequence and mark it as a processed image patch;

[0030] S2.2.2.4: Update the edge weights in the dynamic dependency graph;

[0031] S2.2.2.5: Repeat S2.2.2.1-S2.2.2.4 until all image blocks in the unprocessed image block set have been processed, generating an image block sequence of enhanced image blocks.

[0032] Preferably, S3 includes:

[0033] S3.1: The first Each image patch is input into a self-supervised learning Transformer encoder to extract the self-supervised learning features of the image patch.

[0034] S3.2: Before aggregation i -1 self-supervised learning features are used to obtain a feature set, and the feature set is weighted to obtain a context vector;

[0035] S3.3: Based on S3.1 and S3.2, we obtain the regularization term for context dependency;

[0036] S3.4: Construct a multi-scale contrastive learning loss function for self-supervised learning features;

[0037] S3.5: Construct the total loss function for self-supervised learning based on regularization terms and multi-scale contrastive learning loss function.

[0038] Preferably, the S3.4 multi-scale contrastive learning loss function for:

[0039] ;

[0040] in, The first character represents the self-supervised learning Transformer encoder. Number of floors Indicates the number to be predicted Image blocks, The number of image patches to be predicted. The total length of the image patch sequence. Indicates the position index of the current image patch. For real image patches at scale The real multi-scale features extracted by the self-supervised learning Transformer encoder. These are the multi-scale features predicted by a Transformer encoder that has undergone self-supervised learning at scale s. j This represents the index in the normalized denominator. Indicates and The features of other image patches are compared, and exp() is an exponential function.

[0041] The preferred model for predicting the probability of COPD risk is:

[0042] ;

[0043] in, For the predicted probability of COPD risk, and These are the trainable weights and biases of the classifier, respectively. It is the Sigmoid activation function. This is a global feature.

[0044] The second aspect of the present invention provides a COPD early screening system based on dynamic dependency graphs and self-supervised learning, which is constructed based on the COPD early screening method based on dynamic dependency graphs and self-supervised learning as described above.

[0045] A third aspect of the present invention provides an electronic device for early COPD screening based on dynamic dependency graphs and self-supervised learning. The electronic device includes a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the COPD early screening method based on dynamic dependency graphs and self-supervised learning as described above.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] Improved the accuracy and robustness of early COPD screening:

[0048] This invention, through dynamic dependency graphs and adaptive sequence generation mechanisms, can more effectively capture complex spatial relationships and contextual information in images, thereby improving the model's sensitivity and accuracy for early small lesions in COPD and overcoming the problem of insufficient sensitivity in traditional methods.

[0049] Reduce reliance on professional interpretations:

[0050] This technology, through self-supervised learning and deep learning methods, achieves automated lesion feature extraction and risk assessment, greatly reducing the need for manual interpretation by experienced radiologists, thereby alleviating the workload of medical staff and avoiding the subjectivity and high cost problems of traditional image interpretation.

[0051] Enhancing the robustness of the model under different data sources and conditions:

[0052] By introducing dynamic graph sequence generation and self-supervised learning mechanisms, this invention can adapt to image differences caused by different devices and scanning parameters, improve the model's generalization ability to diverse data sources, enhance the consistency and stability of screening results, and solve the robustness problem of traditional deep learning methods when processing diverse CT data. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0054] In the attached diagram:

[0055] Appendix Figure 1 : Flowchart of the method of the present invention. Detailed Implementation

[0056] The following is in conjunction with the appendix Figure 1 The preferred embodiments of the present invention will be described herein. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0057] Example 1:

[0058] S1: Preprocessing chest CT image data to obtain enhanced image blocks;

[0059] S1.1: Normalize the grayscale values ​​of the acquired chest CT images to... Interval:

[0060] ;

[0061] in, The grayscale values ​​of the normalized chest CT image are... The grayscale values ​​of the acquired chest CT images. and These are the minimum and maximum gray values ​​of the chest CT image, respectively; by normalizing the gray values ​​of the acquired chest CT images, differences caused by different scanning devices and parameters are eliminated.

[0062] S1.2: The normalized chest CT images are then processed according to a fixed size (e.g., ...). The pixel is divided into multiple non-overlapping image patches, denoted as... This is to facilitate subsequent autoregressive modeling, in which... P N For the Nth image patch, the segmentation process ensures that each image patch covers a different lung region in order to capture potential lesion features;

[0063] S1.3: Apply data augmentation methods such as random rotation, translation, scaling, flipping, color dithering, and noise addition to the segmented image blocks in sequence to obtain augmented image blocks. :

[0064] ;

[0065] in, For the i-th enhanced image patch, For the i-th original unenhanced image patch, Add data augmentation functions including random rotation, translation, scaling, flipping, color jitter, and noise;

[0066] All enhanced image patches are combined into an enhanced image patch set. .

[0067] S2: Based on S1, generate an image patch sequence with enhanced image patches;

[0068] S2.1: Obtain the centrality score and local feature importance score of the enhanced image patch;

[0069] S2.1.1: Calculate the similarity of enhanced image patches using the cosine similarity formula:

[0070] ;

[0071] in, It is an enhanced image patch Adjacent nodes in a dynamic dependency graph for The high-dimensional feature vector, for High-dimensional feature vectors;

[0072] Determine the edge set of the dynamic dependency graph based on a similarity threshold. :

[0073] ;

[0074] in, For similarity threshold, To enhance the similarity of image patches;

[0075] Based on enhanced image set and a defined edge set Constructing a dynamic dependency graph for enhanced image patches :

[0076] ;

[0077] S2.1.2: Calculate the centrality score of each enhanced image patch based on the PageRank algorithm. To assess its importance in the dynamic dependency graph:

[0078] ;

[0079] in, The damping factor is usually set to 1. , To and The set of connected neighbor nodes, This represents the total number of nodes. yes The outgoing degree, that is, from The number of edges pointing to other nodes;

[0080] S2.1.3: Determine the importance score of local features for enhanced image patches :

[0081] ;

[0082] yes The neighborhood set, containing [the elements], includes [the elements]. Other image patches with high similarity, It is the size of the neighborhood set, i.e. The number of image patches in the neighborhood.

[0083] S2.2: Based on the centrality score of the dynamic dependency graph and the local features of the image patch, determine the processing order of the enhanced image patch to obtain the image patch sequence.

[0084] S2.2.1: Calculate the priority score for each enhanced image patch by combining the centrality score and the local feature importance score. :

[0085] ;

[0086] in, These are weighting coefficients used to balance global centrality and local importance;

[0087] S2.2.2: A priority queue algorithm based on a greedy strategy is used to score the enhanced image patches according to their priority. Processing sequence from high to low, generating enhanced image patch processing sequence. This indicates the order in which image blocks are processed.

[0088] The specific steps of S2.2.2 are as follows:

[0089] S2.2.2.1: Initialize the enhanced image patches to an unprocessed state and construct a set of unprocessed image patches. ;

[0090] S2.2.2.2: Set of unprocessed image patches Select the one with the highest priority rating , Indicates the first In the next iteration, the image patch with the highest priority score among the unprocessed enhanced image patches;

[0091] S2.2.2.3: Select Add it to the processing sequence and mark it as a processed image patch;

[0092] S2.2.2.4: Update the edge weights in the dynamic dependency graph;

[0093] S2.2.2.4.1: For each enhanced image patch... Treating them as nodes in a dynamic dependency graph, the initial feature vector of the enhanced image patch is calculated when initializing the edge weights. and the initial feature vectors of adjacent nodes cosine similarity ;

[0094] ;

[0095] in, Represents the norm function; and They represent and The norm of is used for normalization.

[0096] S2.2.2.4.2: Will As a node in the dynamic dependency graph and Initial edge weights between them;

[0097] In the subsequent Next, select and process the image block with the highest priority score in this round. Afterwards, regarding Extracting new feature vectors from connected neighboring nodes Then, the following dynamic update formula is used to obtain the first... Edge weights after the next iteration :

[0098] ;

[0099] in, This is a decay factor used for edge weights in the previous iteration. The new cosine similarity is weighted to reflect the latest associations between image patches after each update; The edge weight is the weight at the th The value after the nth iteration , Representing nodes respectively and In the The feature vector at the next iteration;

[0100] S2.2.2.5: Repeat S2.2.2.1-S2.2.2.4, when At that time, until all image patches in the set of unprocessed image patches have been processed, a sequence of image patches is generated to produce enhanced image patches: This indicates the order in which image blocks are processed.

[0101] S3: Input the elements in the image patch sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image patches, and construct the total loss function of self-supervised learning based on the extracted features.

[0102] S3.1: Take the first image patch from the image patch sequence... Image blocks Input a self-supervised learning Transformer encoder to extract self-supervised learning features from image patches. ;

[0103] S3.2: Aggregate the previous part of S3.1 The feature set is obtained by self-supervised learning. The context vector is obtained by weighting the feature set. :

[0104] ;

[0105] in, These are the weighting coefficients; It is the index of the image patch currently being computed. It is the index of the currently processed image patch;

[0106] The calculation formula is as follows:

[0107] ;

[0108] in, and These are the trainable weight vector and matrix, respectively. It is the bias vector; yes transpose, It is the hyperbolic tangent activation function. It is an exponential function, used to convert scalars into non-negative values. Indicates the number to be predicted One characteristic, Indicates all with Indexes of other related image patches, Indicates the index number.

[0109] S3.3: Based on S3.1 and S3.2, obtain the context dependency regularization term. :

[0110] ;

[0111] in, For dynamic dependency graphs and image patches The set of neighboring image patches with the highest dependency weight;

[0112] S3.4: Constructing a multi-scale contrastive learning loss function for self-supervised learning features :

[0113] ;

[0114] in: ;

[0115] It is a scale The self-supervised learning Transformer encoder is used, where s represents the s-th layer of the self-supervised learning Transformer encoder. Indicates the number to be predicted Image blocks, The number of image patches to be predicted. It is a real image patch Real multi-scale features extracted by a Transformer encoder that has undergone self-supervised learning at scale s; represent A continuous patch of real images, This represents the total length of the image patch sequence. This represents the position index of the current image patch, which is usually the position of the image patch in the sequence. One location, This represents the index in the normalized denominator. Indicates and Features of other image patches for comparison. It is an exponential function.

[0116] S3.5: Construct the total loss function for self-supervised learning based on regularization terms and multi-scale contrastive learning loss function:

[0117] ;

[0118] in, Representing different scale levels. Indicates the total number of scale levels. For multi-scale contrastive learning loss function, yes The weighting coefficients, Regularization terms representing context dependencies constrain the context vector. Consistency between features of high-centrality score image patches in the dynamic dependency graph yes The weighting coefficients.

[0119] S4: Input the elements in the image patch sequence into the Transformer encoder for a supervised classification task, extract the supervised classification task features of the image patches, and weight them to obtain the global features. Using a classifier to By mapping the model to the COPD risk score, a model for predicting the probability of COPD risk is obtained. A loss function for a supervised classification task is constructed, and the model for predicting the probability of COPD risk is optimized to achieve early screening of COPD.

[0120] S4.1: Input the elements in the image patch sequence into the Transformer encoder for the supervised classification task, extract the supervised classification task features of the image patches, and use a hierarchical multi-head self-attention mechanism to dynamically fuse the self-supervised learning features of the image patches with the supervised classification task features of the image patches.

[0121] S4.1.1: The first image patch in the sequence... Image blocks Inputting the data into a Transformer encoder for a supervised classification task extracts features from the i-th image patch. ;

[0122] In each layer of the Transformer encoder for supervised classification tasks Supervised learning features are incorporated through the multi-head self-attention mechanism MHSA. As the query (Q), the features of the supervised classification task in the upper layer. As keys (K) and values ​​(V), queries (Q), keys (K), and values ​​(V) are mapped to the same vector space and divided into... There are attention heads, each attention head at the _ ... Each layer has independent projection parameters ,in, All are learnable parameter matrices; for a single attention head, first calculate... Then, the attention weight A is obtained using the following attention formula:

[0123] ,

[0124] in, Given K as the vector dimension, softmax normalizes the similarity to an attention distribution;

[0125] The sum of values ​​is obtained by summing the values ​​based on the attention weights A. :

[0126] ;

[0127] S4.1.2: Put all After concatenation, a linear transformation is performed and the image is projected back to the original dimension to obtain the single-layer multi-head attention output. By using residual connections and LayerNorm, a feature is formed that combines "self-supervised learning features" and "supervised classification task features". :

[0128] ;

[0129] S4.2: Based on S4.1.2, priority scoring based on image patches. Generate a comprehensive global feature representation. :

[0130] ;

[0131] in, This represents the total number of layers in the Transformer encoder. Indicates the fused first Layer Features of an image patch The total number of image patches is denoted by , and Concat is the concatenation operation. The MLP consists of multiple fully connected layers and non-linear activation functions, which are responsible for non-linearly integrating multi-layer features to generate a global feature representation with high discriminative power.

[0132] S4.3: Based on a classifier consisting of fully connected layers and sigmoid activation function layers, Mapping this to a COPD risk score, the model for predicting the probability of COPD risk is as follows:

[0133] ;

[0134] in, For the predicted probability of COPD risk, and These are the trainable weights and biases of the classifier, respectively. Using the Sigmoid activation function ensures that the output probability is within... between;

[0135] S4.4: Construct a loss function for a supervised classification task based on the predicted COPD risk probability. :

[0136] ;

[0137] in, Indicates an index. The total number of image blocks;

[0138] S5: Combine the total loss function of self-supervised learning and the loss function of supervised classification tasks to construct a comprehensive loss function. This achieves overall optimization of the probability model for predicting COPD risk:

[0139] ;

[0140] in, Let be the total loss function for self-supervised learning. Let the loss function be for the supervised classification task. and These are weighting coefficients used to balance the influence of self-supervised learning tasks and supervised classification tasks, and are determined through cross-validation or dynamic adjustment strategies. and The value is adjusted to optimize overall performance.

[0141] By combining supervised classification tasks and self-supervised learning tasks, more robust general features are extracted using self-supervised learning, thereby enhancing the model's generalization ability and representation learning performance.

[0142] During the optimization process, the Adam optimizer is used to minimize the overall loss function:

[0143] ;

[0144] in, For model parameters, This is the learning rate.

[0145] Example 2:

[0146] An early COPD screening system based on dynamic dependency graphs and self-supervised learning is constructed based on the early COPD screening method based on dynamic dependency graphs and self-supervised learning as described in Example 1.

[0147] Example 3:

[0148] An electronic device for early COPD screening based on dynamic dependency graphs and self-supervised learning is disclosed. The device includes a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the early COPD screening method based on dynamic dependency graphs and self-supervised learning as described in Example 1.

[0149] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for early screening of COPD based on dynamic dependency graphs and self-supervised learning, characterized in that, include: S1: Preprocessing chest CT image data to obtain enhanced image blocks; S2: A sequence of image patches used to generate enhanced image patches; S2 includes: S2.1: Obtain the centrality score and local feature importance score of the enhanced image patch; S2.2: Determine the processing order of the enhanced image patches to obtain the image patch sequence; S2.2 includes: S2.2.1: Combine the centrality score and local feature importance score of the enhanced image patch to calculate the priority score of each enhanced image patch; S2.2.2: A priority queue algorithm based on a greedy strategy generates an image block sequence of enhanced image blocks by processing them in descending order of priority score. S2.2.2 includes: S2.2.2.1: Initialize the enhanced image patches to an unprocessed state and construct a set of unprocessed image patches; S2.2.2.2: Select the image patch with the highest priority score from the set of unprocessed image patches; S2.2.2.3: Add the image patch selected in S2.2.2.2 to the processing sequence and mark it as a processed image patch; S2.2.2.4: Update the edge weights in the dynamic dependency graph; S2.2.2.5: Repeat S2.2.2.1-S2.2.2.4 until all image blocks in the unprocessed image block set have been processed, generating an image block sequence of enhanced image blocks; S3: Input the elements in the image patch sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image patches, and construct the total loss function for self-supervised learning; S3 includes: S3.1: The first Image patches are input into a self-supervised learning Transformer encoder to extract self-supervised learning features from the image patches; S3.2: Aggregate the first i-1 self-supervised learning features to obtain a feature set, and weight the feature set to obtain a context vector; S3.3: Based on S3.1 and S3.2, we obtain the regularization term for context dependency; S3.4: Construct a multi-scale contrastive learning loss function for self-supervised learning features; S3.5: Construct the total loss function for self-supervised learning based on regularization terms and multi-scale contrastive learning loss function; S4: Input the elements in the image patch sequence into the Transformer encoder of the supervised classification task to extract the supervised classification task features of the image patch. After weighting, the global features are obtained. The global features are mapped to the COPD risk score by the classifier to obtain the COPD risk probability model. The loss function of the supervised classification task is constructed to optimize the COPD risk probability model and realize early screening of COPD. The model for predicting the probability of COPD risk is: ； in, For the predicted probability of COPD risk, and These are the trainable weights and biases of the classifier, respectively. It is the Sigmoid activation function. For global features; S5: Combine the total loss function of self-supervised learning with the loss function of supervised classification tasks to construct a comprehensive loss function, thereby achieving overall optimization of the model for predicting the probability of COPD risk.

2. The COPD early screening method based on dynamic dependency graphs and self-supervised learning according to claim 1, characterized in that, S2.1 includes: S2.1.1: The similarity of the enhanced image patches is calculated using the cosine similarity formula; S2.1.2: Calculate the centrality score for each enhanced image patch; S2.1.5: Calculate the importance score of local features of enhanced image patches based on similarity.

3. The COPD early screening method based on dynamic dependency graphs and self-supervised learning according to claim 2, characterized in that, S3.4 Multi-scale contrastive learning loss function for: ； Where s represents the s-th layer number of the self-supervised Transformer encoder, Indicates the number to be predicted Image blocks, The number of image patches to be predicted. The total length of the image patch sequence. Indicates the position index of the current image patch. The real multi-scale features extracted by a Transformer encoder that is self-supervised learning at scale s for real image patches. These are the multi-scale features predicted by a Transformer encoder that has undergone self-supervised learning at scale s. j This represents the index in the normalized denominator. Indicates and Features of other image patches for comparison. It is an exponential function.

4. A COPD early screening system based on dynamic dependency graphs and self-supervised learning, characterized in that, This method is constructed based on the COPD early screening method based on dynamic dependency graph and self-supervised learning as described in any one of claims 1-3.

5. An electronic device for early COPD screening based on dynamic dependency graphs and self-supervised learning, characterized in that: The electronic device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the COPD early screening method based on dynamic dependency graphs and self-supervised learning as described in any one of claims 1-3.

Images