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

Abstract

The invention belongs to the field of medical image analysis and artificial intelligence, and provides a COPD early screening method and application based on a dynamic dependency graph and self-supervised learning for the problems of insufficient sensitivity, insufficient feature representation and the like in the prior art, and the method comprises the steps: S1, preprocessing chest CT image data, and obtaining an enhanced image block; s2, generating an image block sequence of the enhanced image blocks; s3, inputting the image block sequence into a self-supervised learning Transformer encoder, and constructing a total loss function of self-supervised learning; and S4, determining a model for predicting the COPD risk probability, and constructing a loss function with a supervised classification task to optimize the model of the COPD risk probability. And S5, constructing a comprehensive loss function, and integrally optimizing the model for predicting the COPD risk probability. According to the method, multiple advanced technologies are fused, and early detection and risk assessment are carried out on COPD in an automatic, efficient and high-precision mode.

Description

Technical Field

[0001] The present invention belongs to the field of medical image analysis and artificial intelligence, and specifically relates to an early screening method for COPD based on dynamic dependency graph and self-supervised learning and its application. Background Art

[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 worldwide, and its incidence continues to rise. Early screening is crucial for delaying disease progression and improving patient prognosis, significantly reducing the medical burden and socioeconomic costs associated with the disease. Traditional COPD screening methods rely primarily on pulmonary function tests (such as spirometry) and imaging studies (such as chest CT scans). Although pulmonary function tests are the gold standard for diagnosing COPD, they have limited detection capabilities in the early stages of the disease, making them prone to missed diagnoses, particularly in patients whose symptoms are not yet apparent. Interpretation of imaging results requires interpretation by experienced radiologists, which can be subjective and costly. Furthermore, the radiation risks and high costs of CT scans limit their application for large-scale screening.

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

[0004] Although radiomics has improved the automation and objectivity of screening to a certain extent, it relies heavily on manually designed features and may overlook the complex spatial relationships and contextual information potentially present in images. Furthermore, differences in imaging data resulting from different scanning devices and parameters limit the generalization ability of radiomics models across different datasets. Furthermore, the requirement for high-quality annotated data makes training time-consuming and labor-intensive, limiting its large-scale application. Deep learning techniques, particularly convolutional neural networks, have shown great potential in medical image analysis, capable of automatically extracting high-level features from images and enabling 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 COPD early screening methods all have the following major limitations: Lack of sensitivity: Pulmonary function tests have limited ability to detect early-stage diseases and are prone to miss diagnoses.

[0006] Reliance on professional interpretation: Imaging examination results need to be interpreted by experienced radiologists, which is subject to high subjectivity and high interpretation costs.

[0007] Insufficient feature representation: Traditional methods may not be able to fully capture the complex spatial relationships and contextual information in the image during the feature extraction process, which limits the recognition ability of the model. They also rely entirely on supervised training and do not fully utilize the information in the data itself.

[0008] Lack of robustness: Traditional deep learning and machine learning methods are not robust enough when processing CT data from different sources and conditions, making it difficult to achieve consistent screening results. Summary of the Invention

[0009] In response to the above-mentioned problems, the first aspect of the present invention proposes a method for early screening of COPD based on a dynamic dependency graph and self-supervised learning, which is characterized by comprising: S1: Preprocess chest CT image data to obtain enhanced image blocks; S2: Generate an image block sequence of enhanced image blocks; S3: Input the elements of the image block sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image block, and construct the total loss function of the self-supervised learning; S4: Input the elements of the image block sequence into the Transformer encoder of the supervised classification task, extract the supervised classification task features of the image blocks, obtain the global features after weighting, use the classifier to map the global features to the COPD risk score, and obtain a model for predicting COPD risk probability. The loss function of the supervised classification task is constructed to optimize the COPD risk probability model and achieve early COPD screening. S5: Combine the total loss function of self-supervised learning and the loss function of supervised classification tasks to construct a comprehensive loss function to achieve overall optimization of the COPD risk probability prediction model.

[0010] Preferably, S2 includes: S2.1: Obtain the centrality score and local feature importance score of the enhanced image block; S2.2: Determine the processing order of the enhanced image blocks to obtain an image block sequence.

[0011] Preferably, S2.1 includes: S2.1.1: Calculate the similarity of the enhanced image blocks using the cosine similarity formula; S2.1.2: Calculate the centrality score of each enhanced image patch; S2.1.5: Calculate the local feature importance score of the enhanced image patch based on the similarity.

[0012] Preferably, S2.2 includes: S2.2.1: Calculate the priority score of each enhanced image block by combining the centrality score and local feature importance score of the enhanced image block; S2.2.2: Based on the greedy strategy, the priority queue algorithm processes the enhanced image blocks in descending order of priority scores to generate an image block sequence of the enhanced image blocks.

[0013] Preferably, S2.2.2 includes: S2.2.2.1: Initialize the enhanced image blocks to an unprocessed state and construct a set of unprocessed image blocks; S2.2.2.2: Select the image block with the highest priority score from the set of unprocessed image blocks; S2.2.2.3: Add the image blocks selected in S2.2.2.2 to the processing sequence and mark them as processed image blocks; S2.2.2.4: Update 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 are processed, generating an image block sequence of enhanced image blocks.

[0014] Preferably, S3 includes: S3.1: The image blocks are input into the self-supervised learning Transformer encoder to extract the self-supervised learning features of the image blocks; S3.2: Before aggregation i -1 self-supervised learning feature to obtain a feature set, and the feature set is weighted to obtain a context vector; S3.3: Based on S3.1 and S3.2, a context-dependent regularization term is obtained; S3.4: Constructing 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 the regularization term and the multi-scale contrastive learning loss function.

[0015] Preferably, S3.4 multi-scale contrastive learning loss function for: ; in, represents the first Number of layers, Indicates the first image blocks, is the number of image blocks to be predicted, is the total length of the image block sequence, Indicates the position index of the current image block, For real image patches at scale The real multi-scale features extracted by the Transformer encoder under self-supervised learning, is the multi-scale feature predicted by the Transformer encoder of self-supervised learning at scale s, j represents the index in the normalized denominator, Represents The features of other image blocks for comparison, exp( ) is an exponential function.

[0016] Preferably, the model for predicting COPD risk probability is: ; in, is the predicted COPD risk probability, and are the trainable weights and biases of the classifier, is the Sigmoid activation function, It is a global feature.

[0017] A second aspect of the present invention provides a COPD early screening system based on a dynamic dependency graph and self-supervised learning, which is constructed based on the COPD early screening method based on a dynamic dependency graph and self-supervised learning as described above.

[0018] A third aspect of the present invention provides an electronic device for early screening of COPD based on a dynamic dependency graph and self-supervised learning. The electronic device includes a processor and a memory, wherein a computer program is stored in the memory. The computer program is loaded and executed by the processor to implement the above-mentioned method for early screening of COPD based on a dynamic dependency graph and self-supervised learning.

[0019] Compared with the prior art, the present invention has the following beneficial effects: Improved accuracy and robustness of early COPD screening: Through dynamic dependency graphs and adaptive sequence generation mechanisms, the present invention can more effectively capture complex spatial relationships and contextual information in images, thereby improving the model's sensitivity and accuracy for early COPD microlesions and overcoming the problem of insufficient sensitivity of traditional methods.

[0020] Reduce reliance on professional interpretation: This technology uses self-supervised learning and deep learning methods to achieve automated lesion feature extraction and risk assessment, greatly reducing the need for manual interpretation by experienced radiologists, thereby alleviating the medical workload and avoiding the subjectivity and high cost issues in traditional imaging interpretation.

[0021] Enhance the robustness of the model under different data sources and conditions: By introducing dynamic image sequence generation and self-supervised learning mechanisms, the present invention can adapt to image differences caused by different devices and scanning parameters, improve the model's generalization ability for diverse data sources, enhance the consistency and stability of screening effects, and solve the robustness problem of traditional deep learning methods when processing diverse CT data. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0023] In the attached figure: Attachment Figure 1 : Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0024] The following is combined with Figure 1 The preferred embodiments of the present invention are described. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0025] Example 1: S1: Preprocess chest CT image data to obtain enhanced image blocks; S1.1: Normalize the grayscale value of the acquired chest CT image to Range: ; in, is the gray value of the normalized chest CT image, is the grayscale value of the acquired chest CT image, and are the minimum and maximum grayscale values ​​of the chest CT image, respectively. The grayscale values ​​of the acquired chest CT images are normalized to eliminate the differences caused by different scanning devices and parameters. S1.2: The normalized chest CT image is converted to a fixed size (e.g. pixels) is divided into multiple non-overlapping image blocks (patch), denoted as , for subsequent autoregressive modeling, where P NFor the Nth image patch, the segmentation process ensures that each image patch covers different lung areas to capture potential lesion characteristics; S1.3: The segmented image blocks are subjected to random rotation, translation, scaling, flipping, color jittering, and noise addition data enhancement methods in sequence to obtain enhanced image blocks. : ; in, is the i-th enhanced image block, is the i-th original unenhanced image block, Data augmentation functions added for random rotation, translation, scaling, flipping, color jittering, and noise; All enhanced image blocks constitute an enhanced image block set .

[0026] S2: Based on S1, generate an image block sequence of enhanced image blocks; S2.1: Obtain the centrality score and local feature importance score of the enhanced image block; S2.1.1: Use the cosine similarity formula to calculate the similarity of the enhanced image blocks: ; in, is the enhanced image block Adjacent nodes in the dynamic dependency graph, for The high-dimensional feature vector of for High-dimensional feature vector of ; Determine the edge set of the dynamic dependency graph based on the similarity threshold : ; in, is the similarity threshold, To enhance the similarity of image blocks; Based on enhanced image collection and a certain set of edges , building a dynamic dependency graph of enhanced image patches : ; S2.1.2: Calculate the centrality score of each enhanced image block based on the PageRank algorithm , to assess its importance in the dynamic dependency graph: ; in, is the damping factor, usually set to , For The set of connected neighbor nodes, is the total number of nodes; yes The out-degree, that is, from The number of edges pointing to other nodes; S2.1.3: Determine the local feature importance score of the enhanced image patch : ; yes The neighborhood set of Other image blocks with higher similarity, is the size of the neighborhood set, i.e. The number of image patches in the neighborhood.

[0027] S2.2: Based on the centrality scores of the dynamic dependency graph and the local features of the image blocks, the processing order of the enhanced image blocks is determined to obtain an image block sequence.

[0028] S2.2.1: Combine the centrality score and local feature importance score of the enhanced image block to calculate the priority score of each enhanced image block : ; in, is the weight coefficient, which is used to balance the global centrality and local importance; S2.2.2: Use the priority queue algorithm based on the greedy strategy to score the enhanced image blocks according to their priority From high to low processing order, generate enhanced image block processing sequence Indicates the order in which image blocks are processed.

[0029] The specific steps of S2.2.2 are: S2.2.2.1: Initialize the enhanced image block to an unprocessed state and construct a set of unprocessed image blocks ; S2.2.2.2: Unprocessed Image Block Set Select the one with the highest priority score , Indicates in The image block with the highest priority score among the unprocessed enhanced image blocks in the iteration; S2.2.2.3: Select Add to the processing sequence and mark as processed image block; S2.2.2.4: Update edge weights in the dynamic dependency graph; S2.2.2.4.1: Each enhanced image block As a node in the dynamic dependency graph, when initializing the edge weight, the initial feature vector of the enhanced image block is calculated and the initial feature vectors of adjacent nodes Cosine similarity of ; ; in, represents the norm function; and Respectively and The norm of is used for normalization; S2.2.2.4.2: As a node in a dynamic dependency graph and The initial edge weights between In the subsequent Select and process the image block with the highest priority score this time Afterwards, Connected neighbor nodes extract new feature vectors , then use the following dynamic update formula to get the Edge weights after iterations : ; in, is the decay factor used for the edge weight in the previous iteration The weighted cosine similarity is used to reflect the latest correlation between image blocks after each update. The edge weight is The value after iterations, 、 Represents nodes respectively and In the The eigenvector at the iteration; S2.2.2.5: Repeat S2.2.2.1-S2.2.2.4, when Until all the image blocks in the set of unprocessed image blocks are processed, an image block sequence of enhanced image blocks is generated: Indicates the order in which image blocks are processed.

[0030] S3: Input the elements in the image block sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image blocks, and construct the total loss function of the self-supervised learning based on the extracted features.

[0031] S3.1: The first Image blocks Input the self-supervised learning Transformer encoder to extract the self-supervised learning features of the image block ; S3.2: Aggregate the front of S3.1 self-supervised learning features to obtain feature sets , weight the feature set to obtain the context vector : ; in, is the weight coefficient; is the index of the image block currently being calculated, is the index of the currently processed image block; The calculation formula is as follows: ; in, and are trainable weight vectors and matrices, respectively. is the bias vector; yes The transpose of is the hyperbolic tangent activation function, is the exponential function, used to convert scalars into non-negative values. Indicates the first Features, Indicates all The indices of other related image blocks, Indicates the index number.

[0032] S3.3: Based on S3.1 and S3.2, we get the regularization term for context dependence : ; in, For the image block in the dynamic dependency graph The set of neighboring image patches with the highest dependency weight; S3.4: Constructing a Multi-Scale Contrastive Learning Loss Function for Self-Supervised Learning Features : ; in: ; It's a scale The Transformer encoder for self-supervised learning is s, and s represents the s-th layer of the Transformer encoder for self-supervised learning. Indicates the first image blocks, is the number of image blocks to be predicted, is the real image patch True multi-scale features extracted by the Transformer encoder through self-supervised learning at scale s; represent continuous real image blocks, is the total length of the image block sequence; Indicates the position index of the current image block, usually the first positions, represents the index in the normalized denominator, Represents The features of other image blocks for comparison, is an exponential function.

[0033] S3.5: Based on the regularization term and the multi-scale contrastive learning loss function, construct the total loss function for self-supervised learning: ; in, Indicates different scale levels, represents the total number of scale levels, is the multi-scale contrastive learning loss function, yes The weight coefficient of The regularization term representing context dependency constrains the context vector The consistency between the features of the high centrality scoring image patches in the dynamic dependency graph, yes The weight coefficient of .

[0034] S4: Input the elements in the image block sequence into the Transformer encoder of the supervised classification task, extract the supervised classification task features of the image block and weight them, and obtain the global features after weighting. , using the classifier Mapped to the COPD risk score, a model for predicting COPD risk probability is obtained, a loss function for the supervised classification task is constructed, the model for predicting COPD risk probability is optimized, and early screening for COPD is achieved.

[0035] S4.1: Input the elements of the image patch sequence into the Transformer encoder of the supervised classification task, extract the supervised classification task features of the image patch, and use a hierarchical multi-head self-attention mechanism to dynamically fuse the self-supervised learning features of the image patch with the supervised classification task features of the image patch; S4.1.1: The first Image blocks Input to the Transformer encoder of the supervised classification task to extract the features of the i-th image block ; In each layer of the Transformer encoder for supervised classification tasks , through the multi-head self-attention mechanism MHSA, the supervised learning features As query (Q), the supervised classification task features of the previous layer As keys (K) and values ​​(V), queries (Q), keys (K) and values ​​(V) are mapped to the same vector space and divided into attention heads, each attention head in the Each layer has its own projection parameters ,in, Both are learnable parameter matrices; for a single attention head, first calculate , and then obtain the attention weight A through the following attention formula: , in, is the vector dimension of K, and softmax normalizes the similarity into attention distribution; Sum the values ​​based on the attention weight A and get the sum of the values : ; S4.1.2: All After splicing, the single-layer multi-head attention output can be obtained by linearly transforming and projecting back to the original dimension: , through residual connection and LayerNorm to form the features after integrating "self-supervised learning features + supervised classification task features" : ; S4.2: Based on S4.1.2, priority scoring based on image blocks , generating a comprehensive global feature representation : ; in, is the total number of layers of the Transformer encoder, After fusion, Tier The features of the image blocks, is the total number of image blocks, Concat is the concatenation operation, and MLP consists of multiple fully connected layers and nonlinear activation functions, which is responsible for nonlinearly integrating multi-layer features to generate a global feature representation with high discriminative ability; S4.3: Based on the classifier composed of a fully connected layer and a sigmoid activation function layer, Mapped to the COPD risk score, the model for predicting COPD risk probability is: ; in, is the predicted COPD risk probability, and are the trainable weights and biases of the classifier, is the Sigmoid activation function, ensuring that the output probability is between; S4.4: Construct a loss function for the supervised classification task based on the predicted COPD risk probability : ; in, Represents an index, is the total number of image blocks; S5: Combine the total loss function of self-supervised learning and the loss function of supervised classification tasks to construct a comprehensive loss function , to achieve overall optimization of the COPD risk probability model: ; in, is the total loss function for self-supervised learning, is the loss function for supervised classification tasks, and is a weight coefficient used to balance the impact of self-supervised learning tasks and supervised classification tasks, determined by cross-validation or dynamic adjustment strategies and to optimize overall performance.

[0036] By combining supervised classification tasks and self-supervised learning tasks, self-supervised learning is used to extract more robust general features, thereby enhancing the generalization ability of the model and the representation learning effect.

[0037] During the optimization process, the Adam optimizer is used to minimize the comprehensive loss function: ; in, are model parameters, is the learning rate.

[0038] Example 2: A COPD early screening system based on dynamic dependency graph and self-supervised learning is constructed based on the COPD early screening method based on dynamic dependency graph and self-supervised learning as described in Example 1.

[0039] Example 3: An electronic device for early screening of COPD based on a dynamic dependency graph and self-supervised learning, the planning device comprising a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the method for early screening of COPD based on a dynamic dependency graph and self-supervised learning as described in Example 1.

[0040] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for early COPD screening based on dynamic dependency graph and self-supervised learning, characterized by: include: S1: Preprocess chest CT image data to obtain enhanced image blocks; S2: Generate an image block sequence of enhanced image blocks; S3: Input the elements of the image block sequence into the self-supervised learning Transformer encoder, extract the self-supervised learning features of the image block, and construct the total loss function of the self-supervised learning; S4: Input the elements of the image block sequence into the Transformer encoder of the supervised classification task, extract the supervised classification task features of the image blocks, obtain the global features after weighting, use the classifier to map the global features to the COPD risk score, and obtain a model for predicting COPD risk probability. The loss function of the supervised classification task is constructed to optimize the COPD risk probability model and achieve early COPD screening. S5: Combine the total loss function of self-supervised learning and the loss function of supervised classification tasks to construct a comprehensive loss function to achieve overall optimization of the COPD risk probability prediction model.

2. A method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 1, characterized in that S2 include: S2.1: Obtain the centrality score and local feature importance score of the enhanced image block; S2.2: Determine the processing order of the enhanced image blocks to obtain an image block sequence.

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

4. The method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 3, characterized in that: S2.2 includes: S2.2.1: Calculate the priority score of each enhanced image block by combining the centrality score and local feature importance score of the enhanced image block; S2.2.2: Based on the greedy strategy, the priority queue algorithm processes the enhanced image blocks in descending order of priority scores to generate an image block sequence of the enhanced image blocks.

5. The method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 4, characterized in that: S2.2.2 includes: S2.2.2.1: Initialize the enhanced image blocks to an unprocessed state and construct a set of unprocessed image blocks; S2.2.2.2: Select the image block with the highest priority score from the set of unprocessed image blocks; S2.2.2.3: Add the image blocks selected in S2.2.2.2 to the processing sequence and mark them as processed image blocks; S2.2.2.4: Update 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 are processed, generating an image block sequence of enhanced image blocks.

6. The method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 5, characterized in that S3 include: S3.1: The image blocks are input into the self-supervised learning Transformer encoder to extract the self-supervised learning features of the image blocks; S3.2: Before aggregation i -1 self-supervised learning feature to obtain a feature set, and the feature set is weighted to obtain a context vector; S3.3: Based on S3.1 and S3.2, a context-dependent regularization term is obtained; S3.4: Constructing 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 the regularization term and the multi-scale contrastive learning loss function.

7. The method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 6, characterized in that: S3.4 Multi-scale contrastive learning loss function for: ； in, represents the first Number of layers, Indicates the first image blocks, is the number of image blocks to be predicted, is the total length of the image block sequence, Indicates the position index of the current image block, For real image patches at scale The real multi-scale features extracted by the Transformer encoder under self-supervised learning, is the multi-scale feature predicted by the Transformer encoder of self-supervised learning at scale s, j represents the index in the normalized denominator, Represents The features of other image blocks for comparison, exp( ) is an exponential function.

8. The method for early COPD screening based on dynamic dependency graph and self-supervised learning according to claim 7, characterized in that: The model for predicting the probability of COPD risk is: ； in, is the predicted COPD risk probability, and are the trainable weights and biases of the classifier, is the Sigmoid activation function, It is a global feature.

9. A COPD early screening system based on dynamic dependency graph and self-supervised learning, characterized by: It 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 1-8.

10. An electronic device for early COPD screening based on dynamic dependency graph and self-supervised learning, characterized by: The electronic device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the COPD early screening method based on dynamic dependency graph and self-supervised learning as described in any one of claims 1-8.

Images