Coronary artery disease intelligent auxiliary diagnosis system based on feature decoupling and causal intervention

Abstract

The invention discloses a coronary artery disease intelligent auxiliary diagnosis system based on feature decoupling and causal intervention, relates to the technical field of medical image analysis, and aims to solve the problems that an existing coronary artery disease diagnosis system can only carry out coarse-grained narrow detection, cannot meet the fine-grained analysis requirement and is affected by confusion factors. The model comprises a feature decoupling module, a semantic retrieval module and a causal intervention module, the feature decoupling module is formed by stacking a plurality of feature extraction modules and feature separation modules, and is used for carrying out feature extraction and feature separation to obtain narrow features and plaque features and carrying out classification detection on the narrow features and the plaque features; the semantic retrieval module is used for performing semantic retrieval to obtain semantic features and performing stenosis and plaque detection based on the semantic features; and the causal intervention module is used for dynamically updating the confusion factor library according to the narrow features and the plaque features by using a cross-level causal relationship, and establishing a causal relationship between the imaging signal and a prediction result to obtain a detection result.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to an intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention. Background Art

[0002] The growing threat posed by coronary artery disease (CAD) to cardiovascular health worldwide highlights the urgent need to develop reliable automatic diagnostic technology. Coronary CT Angiography (CCTA), as a non-invasive imaging method, can provide high-resolution three-dimensional images that clearly show the internal conditions of the coronary arteries, including lumen stenosis, vascular obstruction, and plaque accumulation, thereby assisting doctors in making accurate clinical judgments. Automatic diagnosis technology for coronary artery disease based on CCTA is of great significance for improving efficiency.

[0003] However, the current automatic diagnosis technology of coronary artery disease based on CCTA images currently faces some challenges. Due to the heterogeneity of stenosis and plaque characteristics, as well as the complex cause-effect relationship in the diagnosis of coronary artery disease, the existing technology is limited in its ability to handle details. This makes the current methods mainly suitable for relatively rough analysis, such as detecting more than 50% luminal stenosis, and is often limited to vascular level assessment, which significantly affects the practicality and effectiveness of the technology in clinical applications. Summary of the invention

[0004] The technical problems to be solved by the present invention are:

[0005] Existing coronary artery disease diagnostic systems are often only able to perform coarse-grained stenosis detection, cannot meet fine-grained analysis needs, and are affected by confounding factors during detection.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The present invention provides an intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention, comprising:

[0008] Data processing module: used to perform surface reconstruction of the coronary centerline of CCTA and perform decoupled driven volume masking to obtain masked vascular surface reconstruction volume data;

[0009] Model construction module: used to construct a feature decoupling and causal intervention model, the feature decoupling and causal intervention model includes a feature decoupling module, a semantic retrieval module and a causal intervention module, the feature decoupling module is formed by stacking multiple feature extraction modules and feature separation modules, and is used to perform feature extraction and feature separation on the masked vascular surface reconstruction volume to obtain stenosis features and plaque features, and classify stenosis and plaques; the semantic retrieval module is used to perform semantic retrieval on the masked vascular surface reconstruction volume data to obtain semantic features, and perform stenosis and plaque detection based on the semantic features; the causal intervention module is used to dynamically update the confusion factor library with the stenosis features and plaque features using cross-level causal relationships, and establish a causal relationship between the imaging signal and the prediction result based on the stenosis and plaque detection results of the semantic retrieval module and the confusion factors in the confusion factor library, so as to obtain stenosis and plaque detection results at the vascular level and patient level;

[0010] Detection module: used to input the masked vascular surface reconstruction into the feature decoupling and causal intervention model to obtain coronary stenosis and plaque detection results.

[0011] Furthermore, the function implementation process of the data processing module includes the following steps:

[0012] Step S11: Perform surface reconstruction on the coronary artery centerline of CCTA to obtain a surface reconstruction volume set

[0013] Step S12: reconstruct the surface into a volume set Perform decoupling-driven volume masking, that is, mask all lesions to create a volume mask x′, and then use x′ as the background to process each lesion separately; for the jth lesion, restore the masked area of ​​the lesion to generate a masked vascular surface reconstruction volume.

[0014] Furthermore, the feature extraction module of the feature decoupling module includes two operation modules consisting of two 3D convolution layers and one maximum pooling layer connected in series, and two operation modules consisting of three 3D convolution layers and one maximum pooling layer connected in series; the feature extraction module is used to extract features from the masked vascular surface reconstruction to obtain the stenosis feature f sten and patch characteristics f plq ;

[0015] The feature separation module is connected to each operation module of the feature extraction module, and first uses a 3D convolution layer and a Softmax activation layer to process the narrow feature f obtained in the feature extraction module. sten and patch characteristics f plq Get the attention distribution α sten and α plq , then fsten and f plq With α sten and α plq Perform the following operations to enhance f sten and f plq The difference between:

[0016]

[0017] Where ° represents the dot product operation;

[0018] The narrow feature f′ sten and patch characteristics f′ plq Linear projection obtains narrow feature e sten and plaque characteristics plq .

[0019] Furthermore, the feature decoupling module further includes a mutually exclusive classification module, which performs classification based on a multi-layer perceptron including a Softmax activation layer to capture narrow features e sten and plaque characteristics plq The heterogeneity of the attributes in the classification is used to obtain the classification prediction results of stenosis and plaque, and the cross entropy loss function of classification is constructed. for:

[0020]

[0021] in is the one-hot encoded vector of the label, and They correspond to e sten and e plq The category prediction results.

[0022] Furthermore, the function implementation process of the semantic retrieval module includes the following steps:

[0023] Step S31: The masked vascular surface reconstruction is passed through a feature extraction module. Get feature e vol ;

[0024] Step S32: vol The input encoder is used for self-attention processing to obtain the semantic features of the surface reconstruction at different positions. cpr ;

[0025] Step S33: Reconstruct the semantic features of the curved surface at different positions cpr With randomly initialized features Input decoder for feature extraction, e qry Using self-attention and cross-attention from e cprRetrieve semantics from the surface and obtain semantic information in the surface reconstruction volume;

[0026] Step S34: The semantic features obtained after semantic retrieval Perform position regression and convert the features of semantic retrieval Respectively with the narrow features sten and plaque characteristics plq Splicing to get splicing features and The splicing feature and Input into the lesion classification module to obtain the stenosis and plaque detection results in the region of interest;

[0027] S35, calculating the objective function based on the obtained stenosis and plaque detection results in the region of interest

[0028]

[0029] Among them, the regression loss of the lesion location and dual-task classification loss The construction method includes the following steps:

[0030] Step S351: Calculate The first sub-stage is to compare the real data set g and the prediction result set Use the Hungarian algorithm for bipartite matching; define each target g i ∈g is (c i ,r i ), where c i represents the category label, r i ∈[0,1] 2 is an image vector defining the center coordinates of the region of interest and its weight (e i ,w i ), and determine the arrangement that minimizes the total cost The formula for bipartite matching is:

[0031]

[0032] in Indicates that it belongs to category c i The probability of Indicates that there is no lesion category, r i ∈[0,1] 2 , RoI loss It is a linear combination of absolute error loss and intersection-over-union loss:

[0033]

[0034] where λ iou and Is the impact The hyperparameters of is the predicted value, It is the intersection and ratio loss;

[0035] Step S352: The second sub-stage is based on pairing calculation, combining negative log-likelihood and RoI loss to achieve:

[0036]

[0037] in Indicates that the prediction result belongs to category c i The probability of

[0038] Step S353: The cross entropy loss including stenosis degree and plaque composition:

[0039]

[0040] where y sten and plq are the one-hot encoded vectors of plaque and stenosis labels, respectively, and They are the category detection results of plaque and stenosis respectively.

[0041] Furthermore, the functional implementation process of the causal intervention module includes the following steps:

[0042] Step S41: the narrow feature e sten and plaque characteristics plq Execute the confusion factor writing operation respectively to obtain the confusion factor library;

[0043] Step S42: performing a confounding factor reading operation on all the vascular branch stenosis features and plaque features of the patient through the confounding factor library, finding the most similar confounding factor features at the patient level in the library, and participating in the causal intervention at the patient level;

[0044] Step S43: input the most similar confounding factor feature into a causal intervention module, establish a causal relationship between the imaging signal and the prediction result, and obtain the lesion results at the vessel level and the patient level.

[0045] Furthermore, step S41 of the function implementation process of the causal intervention module includes the following steps:

[0046] Step S411, construct two confounding factor libraries for two lesions, stenosis and plaque: and Where K 1 and K2 The number of categories representing stenosis and plaque attributes, each library contains K 1 / K 2 A mixed queue, represented by Where M is each The maximum number of promiscuous vectors in ;

[0047] Step S412: Input the feature e extracted in step S2 sten and e plq Through multi-layer perceptron dimensionality reduction and And according to the classification results c obtained in the mutually exclusive classification sten and c plq , the reduced vector v sten With v plq Write the corresponding confusion factor queue;

[0048] Step S413: Check the current confusion factor library. If the number of confusion factors of the current category is less than the maximum limit M, directly set the feature e atr Add to Queue (atr∈{sten,plq}); if If the size has reached M, update the elements in the queue and calculate the new e atr The similarity between the stored features is calculated as follows:

[0049]

[0050] Update the most similar existing confounding factor with the weighted average of the new confounding factor.

[0051] Furthermore, step S42 of the function implementation process of the causal intervention module includes the following steps:

[0052] Step S421: Set the narrow feature e sten and plaque characteristics plq Dimensionality reduction and Select from the confounding factor library and The most similar confusion factor generates attribute focus mediator m sten and m plq ;

[0053] Step S422: generate m sten and m plq Intervention mediators that tandemly generate the complete coronary tree participate in causal interventions at the patient level.

[0054] Furthermore, step S43 of the function implementation process of the causal intervention module includes the following steps:

[0055] Step S431: Add the intervention medium e of the complete coronary tree med =[m sten ,m plq ] and the semantic information of each branch Transformed into query, key, and value representations through linear projection and layer normalization;

[0056] Step S432: Introduce cross-level causal intervention to obtain the final features of all vascular branches to form a feature set Right now:

[0057]

[0058] in is the causal intervention affine matrix of the i-th branch, d is the dimension of the feature, For query, K med is the key, is the value;

[0059] Step S433: Automatic diagnosis of cross-level coronary artery disease. The features corresponding to each branch in the image are used to regress the lesion location and classify the degree of stenosis and plaque composition for vessel-level evaluation, and the results of the degree of stenosis and plaque composition of each vessel are obtained, which are then aggregated for patient-level evaluation.

[0060] Furthermore, the loss function used by the feature decoupling and causal intervention model is

[0061]

[0062] and denote the regression loss of the lesion location and the dual-task classification loss of the ith branch, respectively, η is a hyperparameter, and the condition and Used to evaluate the corresponding branch p l Whether it affects the current category of reporting and data systems and overall plaque characteristics.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The system of the present invention performs surface reconstruction and decoupling-driven volume masking on CCTA images and coronary artery centerlines obtained by coronary CT angiography technology to perform feature decoupling and causal intervention, which can effectively capture the heterogeneous features of stenosis and plaques with different constraints, analyze the lesions more carefully, and eliminate the interference of confounding factors to ensure that the system remains effective and robust. The automatic diagnosis technology of coronary artery disease based on feature decoupling and causal intervention proposed by the present invention is a non-interference framework specially customized for cross-level, fine-grained automatic diagnosis of coronary artery disease, which solves the problem that the current diagnostic system only performs well on coarse-grained features and can be widely used in clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a structural schematic diagram of an intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention in an embodiment of the present invention;

[0066] Figure 2 It is a schematic diagram of the structure of a characteristic decoupling module in an embodiment of the present invention;

[0067] Figure 3 1 and 2 are diagrams of lesion results at the vessel level and patient level in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0070] Specific implementation plan 1: Combine Figure 1 to Figure 2 As shown, the present invention provides an intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention, comprising:

[0071] Data processing module: used to perform surface reconstruction of the coronary centerline of CCTA and perform decoupled driven volume masking to obtain masked vascular surface reconstruction volume data;

[0072] Model construction module: used to construct a feature decoupling and causal intervention model, the feature decoupling and causal intervention model includes a feature decoupling module, a semantic retrieval module and a causal intervention module, the feature decoupling module is formed by stacking multiple feature extraction modules and feature separation modules, and is used to perform feature extraction and feature separation on the masked vascular surface reconstruction volume to obtain stenosis features and plaque features, and classify stenosis and plaques; the semantic retrieval module is used to perform semantic retrieval on the masked vascular surface reconstruction volume data to obtain semantic features, and perform stenosis and plaque detection based on the semantic features; the causal intervention module is used to dynamically update the confusion factor library with the stenosis features and plaque features using cross-level causal relationships, and establish a causal relationship between the imaging signal and the prediction result based on the stenosis and plaque detection results of the semantic retrieval module and the confusion factors in the confusion factor library, so as to obtain stenosis and plaque detection results at the vascular level and patient level;

[0073] Detection module: used to input the masked vascular surface reconstruction into the feature decoupling and causal intervention model to obtain coronary stenosis and plaque detection results.

[0074] Specific implementation scheme 2: The function implementation process of the data processing module includes the following steps:

[0075] Step S11: Perform curved planar reformation (CPR) on the coronary artery centerline of CCTA to obtain a curved reconstructed volume set. It consists of a total of 16 major coronary artery branches at the patient level;

[0076] Step S12: reconstruct the surface into a volume set Decoupled Volume Masking (DVM) is performed, that is, all l lesions are masked with a Hu value of -1024 to create a volume mask x′, and then x′ is used as the background to process each of the l lesions separately; for the jth lesion, the masked area of ​​the lesion is restored to its original Hu value to generate a masked vascular surface reconstruction volume. The rest of this embodiment is the same as the specific embodiment 1.

[0077] Specific implementation plan three: Figure 2 As shown, the feature extraction module of the feature decoupling module includes four operation modules, the first two operation modules are composed of two 3D convolution layers and a maximum pooling layer in series, and the last two neural network blocks are composed of three 3D convolution layers and a maximum pooling layer in series; the feature extraction module is used to extract features from the masked vascular surface reconstruction to obtain the stenosis feature f sten and patch characteristics f plq ;

[0078] The feature separation module is connected to each operation module of the feature extraction module, and first uses a 3D convolution layer and a Softmax activation layer to process the narrow feature f obtained in the feature extraction module. sten and patch characteristics f plq Get the attention distribution α sten and α plq , then f sten and f plq With α sten and α plq Perform the following operations to enhance f sten and f plq The difference between:

[0079]

[0080] Where ° represents the dot product operation;

[0081] The narrow feature f′ sten and patch characteristics f′ plq Linear projection obtains narrow feature e sten and plaque characteristics plq The rest of this implementation is the same as the second specific implementation.

[0082] Specific implementation scheme 4: The feature decoupling module also includes a mutually exclusive classification module, which performs classification based on a multi-layer perceptron including a Softmax activation layer, and captures the narrow feature e sten and plaque characteristics plq The heterogeneity of the attributes in the classification is used to obtain the classification prediction results of stenosis and plaque, and the cross entropy loss function of classification is constructed. for:

[0083]

[0084] in is the one-hot encoded vector of the 10 concatenated labels, and They correspond to e sten and e plq The category prediction result, that is, the probability distribution of category prediction. The rest of this implementation plan is the same as the specific implementation plan three.

[0085] Specific implementation scheme 5: The function implementation process of the semantic retrieval module includes the following steps:

[0086] Step S31: The masked vascular surface reconstruction is passed through a feature extraction module. Get feature e vol , where the feature extraction module The structure is the same as the feature extraction module structure described in step S211;

[0087] Step S32: vol The input is processed into the Transformer encoder for self-attention to obtain the semantic features of the surface reconstruction at different positions. cpr ;

[0088] Step S33: Reconstruct the semantic features of the curved surface at different positions cpr With randomly initialized features Input into Transformer decoder for feature extraction, e qry Using self-attention and cross-attention from e cpr Retrieve semantics from the surface and obtain semantic information in the surface reconstruction volume;

[0089] Step S34: The semantic features obtained after semantic retrieval Input position regression module (MLP with Sigmoid activation function) to locate the lesion; the semantic retrieval features Respectively with the narrow features sten and plaque characteristics plq Splicing to get splicing features and The splicing feature and The input is fed into the lesion classification module (MLP with Softmax activation function) to obtain the stenosis and plaque detection results in the region of interest (RoI);

[0090] S35, calculating the objective function based on the obtained stenosis and plaque detection results in the region of interest

[0091]

[0092] Among them, the regression loss of the lesion location and dual-task classification loss The construction method includes the following steps:

[0093] Step S351: Calculate The first sub-stage is to compare the real data set g and the prediction result set Use the Hungarian algorithm for bipartite matching; define each target g i ∈g is (c i ,r i ), where c i represents the category label, r i ∈[0,1] 2is an image vector defining the center coordinates of the region of interest and its weight (e i ,w i ), and determine the arrangement that minimizes the total cost The formula for bipartite matching is:

[0094]

[0095] in Indicates that it belongs to category c i The probability of Indicates that there is no lesion category, r i ∈[0,1] 2 , RoI loss It is a linear combination of absolute error loss and intersection-over-union loss:

[0096]

[0097] where λ iou and Is the impact The hyperparameters of is the predicted value, It is the intersection and ratio loss;

[0098] Step S352: The second sub-stage is based on pairing calculation, combining negative log-likelihood and RoI loss to achieve:

[0099]

[0100] in Indicates that the prediction result belongs to category c i The probability of

[0101] Step S353: The cross entropy loss including stenosis degree and plaque composition:

[0102]

[0103] where y sten and plq are the one-hot encoded vectors of plaque and stenosis labels, respectively, and The other aspects of this embodiment are the same as those of the fourth embodiment.

[0104] Specific implementation scheme 6: The functional implementation process of the causal intervention module includes the following steps:

[0105] Step S41: the narrow feature e sten and plaque characteristics plqExecute the Confounder Writing operation respectively to obtain the Confounder Writing library;

[0106] Step S42, performing a confounder reading operation on all vascular branch stenosis features and plaque features of the patient through the confounder library, finding the most similar confounder features at the patient level in the library, and participating in the causal intervention at the patient level;

[0107] Step S43: Input the most similar confounding factor feature into the causal intervention module, establish the causal relationship between the imaging signal and the prediction result, and obtain the lesion results at the vessel level and the patient level, such as Figure 3 As shown, the vessel-level results include the location of the lesion, stenosis degree and plaque composition of each coronary branch of the current patient, and the patient-level results include the overall CAD-RADS (Coronary Artery Disease Reporting and Data System) and the overall amount of plaque rating of the current patient. The rest of this embodiment is the same as the specific embodiment five.

[0108] Specific implementation scheme seven: Step S41 of the function implementation process of the causal intervention module includes the following steps:

[0109] Step S411, construct two confounding factor libraries for two lesions, stenosis and plaque: and Where K 1 and K 2 The number of categories representing stenosis and plaque attributes, each library contains K 1 / K 2 A mixed queue, represented by Where M is each The maximum number of promiscuous vectors in ;

[0110] Step S412: Input the feature e extracted in step S2 sten and e plq Through multi-layer perceptron dimensionality reduction and And according to the classification results c obtained in the mutually exclusive classification sten and c plq , the reduced vector v sten With v plq Write the corresponding confusion factor queue;

[0111] Step S413: Check the current confusion factor library. If the number of confusion factors of the current category is less than the maximum limit M, directly set feature e atr Add to Queue In (atr∈{sten,plq}). If If the size has reached M, update the elements in the queue and calculate the new e atr The similarity between the stored features is calculated as follows:

[0112]

[0113] The weighted average of the new confusion factors is used to update the most similar existing confusion factors. The rest of this embodiment is the same as the specific embodiment 6.

[0114] Specific implementation scheme eight: Step S42 of the function implementation process of the causal intervention module includes the following steps:

[0115] Step S421: Set the narrow feature e sten and plaque characteristics plq Dimensionality reduction and Select from the confounding factor library and The most similar confusion factor generates attribute focus mediator m sten and m plq ; The definitions are as follows:

[0116]

[0117] in Indicates selection The operation of the element with the highest similarity to v;

[0118] Step S422: generate m sten and m plq The intervention mediators of the complete coronary tree are generated in series and participate in the causal intervention at the patient level. The rest of this implementation plan is the same as the specific implementation plan seven.

[0119] Specific implementation scheme nine: Step S43 of the function implementation process of the causal intervention module includes the following steps:

[0120] Step S431: Add the intervention medium e of the complete coronary tree med =[m sten ,m plq ] and the semantic information of each branch Transformed into query, key, and value representations through linear projection and layer normalization;

[0121] Step S432: Calculation query and key K med We use the attention between , to obtain an affine matrix that captures the distribution difference between the imaging signal and the confusion factor. Multiply with the affine matrix and add the query Integrate the broker into In the process, cross-level causal intervention is introduced to obtain the final features of all vascular branches to form a feature set Right now:

[0122]

[0123] in is the causal intervention affine matrix of the i-th branch, d is the dimension of the feature, For query, K med is the key, is the value;

[0124] Step S433: Automatic diagnosis of cross-level coronary artery disease. The features corresponding to each branch in the image were used to regress the lesion location and classify the degree of stenosis and plaque composition for vessel-level assessment, and the results of stenosis degree and plaque composition of each vessel were obtained; they were then aggregated for patient-level assessment, in which CAD-RADS (Coronary Artery Disease Reporting and Data System) was calculated based on the maximum degree of stenosis of all stenoses, and the overall amount of plaque was calculated based on the number of plaques in the patient.

[0125] For a given image x and corresponding lesion category c, the diagnostic process of predicting y through causal intervention is expressed as:

[0126] P(Y|do(X))=∑(P(y c |x,d))p(d)

[0127] Where d is the interference factor, and do(.) refers to the intervention of the variable;

[0128] The Softmax activation function is used to predict the probability of different lesion categories, that is, Based on the normalized weighted geometric mean, the intervention expectation is approximated as The rest of this implementation plan is the same as the specific implementation plan eight.

[0129] Specific implementation plan 10: The loss function used in the feature decoupling and causal intervention model is

[0130]

[0131] and denote the regression loss of the lesion location and the dual-task classification loss of the ith branch, respectively, η is a hyperparameter, and the condition and Used to evaluate the corresponding branch p l Whether to affect the current category and overall plaque characteristics of the reporting and data systems. Returns 1 if the condition is met; otherwise returns 0.

[0132] Based on the loss function The model is trained until the maximum number of iterations is reached and the training is stopped to obtain a trained automatic diagnosis model for coronary artery disease. The rest of this embodiment is the same as the ninth embodiment.

[0133] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. An intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention, characterized in that: include: Data processing module: used to perform surface reconstruction of the coronary centerline of CCTA and perform decoupled driven volume masking to obtain masked vascular surface reconstruction volume data; Model construction module: used to construct a feature decoupling and causal intervention model, the feature decoupling and causal intervention model includes a feature decoupling module, a semantic retrieval module and a causal intervention module, the feature decoupling module is formed by stacking multiple feature extraction modules and feature separation modules, and is used to perform feature extraction and feature separation on the masked vascular surface reconstruction volume to obtain stenosis features and plaque features, and classify stenosis and plaques; the semantic retrieval module is used to perform semantic retrieval on the masked vascular surface reconstruction volume data to obtain semantic features, and perform stenosis and plaque detection based on the semantic features; the causal intervention module is used to dynamically update the confusion factor library with the stenosis features and plaque features using cross-level causal relationships, and establish a causal relationship between the imaging signal and the prediction result based on the stenosis and plaque detection results of the semantic retrieval module and the confusion factors in the confusion factor library, so as to obtain stenosis and plaque detection results at the vascular level and patient level; Detection module: used to input the masked vascular surface reconstruction into the feature decoupling and causal intervention model to obtain coronary stenosis and plaque detection results.

2. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 1 is characterized in that: The function implementation process of the data processing module includes the following steps: Step S11: Perform surface reconstruction on the coronary artery centerline of CCTA to obtain a surface reconstruction volume set Step S12: reconstruct the surface into a volume set Perform decoupling-driven volume masking, that is, mask all lesions to create a volume mask x′, and then use x′ as the background to process each lesion separately; for the jth lesion, restore the masked area of ​​the lesion to generate a masked vascular surface reconstruction volume.

3. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 2 is characterized in that: The feature extraction module of the feature decoupling module includes two operation modules consisting of two 3D convolution layers and one maximum pooling layer connected in series, and two operation modules consisting of three 3D convolution layers and one maximum pooling layer connected in series; the feature extraction module is used to extract features from the masked vascular surface reconstruction to obtain the stenosis feature f sten and patch characteristics f plq ; The feature separation module is connected to each operation module of the feature extraction module, and first uses a 3D convolution layer and a Softmax activation layer to process the narrow feature f obtained in the feature extraction module. sten and patch characteristics f plq Get the attention distribution α sten and α plq , then f sten and f plq With α sten and α plq Perform the following operations to enhance f sten and f plq The difference between: f′ sten =f sten oh sten the(1-a plq ) f′ plq =f plq oh plq the(1-a sten ) Where ο represents the dot product operation; The narrow feature f′ sten and patch characteristics f′ plq Linear projection obtains narrow feature e sten and plaque characteristics plq .

4. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 3 is characterized in that: The feature decoupling module also includes a mutually exclusive classification module, which performs classification based on a multi-layer perceptron including a Softmax activation layer to capture narrow features e sten and plaque characteristics plq The heterogeneity of the attributes in the classification is used to obtain the classification prediction results of stenosis and plaque, and the cross entropy loss function of classification is constructed. for: in is the one-hot encoded vector of the label, and They correspond to e sten and e plq The category prediction results.

5. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 4 is characterized in that: The function implementation process of the semantic retrieval module includes the following steps: Step S31: The masked vascular surface reconstruction is passed through a feature extraction module. Get feature e vol ; Step S32: vol The input encoder is used for self-attention processing to obtain the semantic features of the surface reconstruction at different positions. cpr ; Step S33: Reconstruct the semantic features of the curved surface at different positions cpr With randomly initialized features Input decoder for feature extraction, e qry Using self-attention and cross-attention from e cpr Retrieve semantics from the surface and obtain semantic information in the surface reconstruction volume; Step S34: The semantic features obtained after semantic retrieval Perform position regression and convert the features of semantic retrieval Respectively with the narrow features sten and plaque characteristics plq Splicing to get splicing features and The splicing feature and Input into the lesion classification module to obtain the stenosis and plaque detection results in the region of interest; S35, calculating the objective function based on the obtained stenosis and plaque detection results in the region of interest Among them, the regression loss of the lesion location and dual-task classification loss The construction method includes the following steps: Step S351: Calculate The first sub-stage is to compare the real data set g and the prediction result set Use the Hungarian algorithm for bipartite matching; define each target g i ∈g is (c i ,r i ), where c i represents the category label, r i ∈[0,1] 2 is an image vector defining the center coordinates of the region of interest and its weight (e i ,w i ), and determine the arrangement that minimizes the total cost The formula for bipartite matching is: in Indicates that it belongs to the category i The probability of Indicates that there is no lesion category, r i ∈[0,1] 2 , RoI loss It is a linear combination of absolute error loss and intersection-over-union loss: where λ ou and Is the impact The hyperparameters of is the predicted value, It is the intersection and ratio loss; Step S352: The second sub-stage is based on pairing calculation, combining negative log-likelihood and RoI loss to achieve: in Indicates that the prediction result belongs to category c i The probability of Step S353: The cross entropy loss including stenosis degree and plaque composition: where y sten and plq are the one-hot encoded vectors of plaque and stenosis labels, respectively, and They are the category detection results of plaque and stenosis respectively.

6. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 5, characterized in that: The functional implementation process of the causal intervention module includes the following steps: Step S41: the narrow feature e sten and plaque characteristics plq Execute the confusion factor writing operation respectively to obtain the confusion factor library; Step S42: performing a confounding factor reading operation on all the vascular branch stenosis features and plaque features of the patient through the confounding factor library, finding the most similar confounding factor features at the patient level in the library, and participating in the causal intervention at the patient level; Step S43: input the most similar confounding factor feature into a causal intervention module, establish a causal relationship between the imaging signal and the prediction result, and obtain the lesion results at the vessel level and the patient level.

7. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 6, characterized in that: Step S41 of the function implementation process of the causal intervention module includes the following steps: Step S411, construct two confounding factor libraries for two types of lesions, stenosis and plaque: and Where 1 and 2 represent the number of categories of stenosis and plaque attributes, and each library contains 1 / 2 mixed cohorts, expressed as Where M is each The maximum number of promiscuous vectors in ; Step S412: Input the feature e extracted in step S2 sten and e plq Through multi-layer perceptron dimensionality reduction and And according to the classification results c obtained in the mutually exclusive classification sten and c plq , the reduced vector v sten With v plq Write the corresponding confusion factor queue; Step S413: Check the current confusion factor library. If the number of confusion factors of the current category is less than the maximum limit M, directly set the feature e atr Add to queue Q[ atr ](atr∈{sten,plq}); If Q[c atr ] has reached M, then update the elements in the queue and calculate the new e atr The similarity between the stored features is calculated as follows: Update the most similar existing confounding factor with the weighted average of the new confounding factor.

8. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 7, characterized in that: Step S42 of the function implementation process of the causal intervention module includes the following steps: Step S421: Set the narrow feature e sten and plaque characteristics plq Dimensionality reduction and Select from the confounding factor library and The most similar confusion factor generates attribute focus mediator m sten and m plq ; Step S422: generate m sten and m plq Intervention mediators that tandemly generate the complete coronary tree participate in causal interventions at the patient level.

9. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 8, characterized in that: Step S43 of the function implementation process of the causal intervention module includes the following steps: Step S431: Add the intervention medium e of the complete coronary tree med =[m sten ,m plq ] and the semantic information of each branch Transformed into query, key, and value representations through linear projection and layer normalization; Step S432: Introduce cross-level causal intervention to obtain the final features of all vascular branches to form a feature set Right now: in is the causal intervention affine matrix of the th branch, d is the dimension of the feature, For query, K med is the key, is the value; Step S433: Automatic diagnosis of cross-level coronary artery disease. The features corresponding to each branch in the image are used to regress the lesion location and classify the degree of stenosis and plaque composition for vessel-level evaluation, and the results of the degree of stenosis and plaque composition of each vessel are obtained, which are then aggregated for patient-level evaluation.

10. The intelligent auxiliary diagnosis system for coronary artery disease based on feature decoupling and causal intervention according to claim 9, characterized in that: The loss function used by the feature decoupling and causal intervention model is and denote the regression loss of the lesion location and the dual-task classification loss of the ith branch, respectively, η is a hyperparameter, and the condition and Used to evaluate the corresponding branch p l Whether it affects the current category of reporting and data systems and overall plaque characteristics.

Images