A strabismus detection method based on causal feature gating

Abstract

The application discloses a strabismus detection method based on a causal feature gate, comprising: modeling and analyzing the causal relationship between key point coordinates and strabismus categories by using a causal discovery algorithm to obtain a causal relationship model for determining the causal relationship, screening the key point coordinates based on the causal relationship model to obtain a causal variable set, and extracting a plurality of causal variables from a plurality of samples in a training sample set according to the causal variable set; introducing a causal feature gate module between adjacent network layers of a classification model, training the classification model by using the plurality of causal variables, and simultaneously adaptively adjusting gate parameters in the training process of the classification model until the classification model converges or a preset training number of rounds is reached to obtain a final classification model. The application embodiment can not only improve the accuracy of strabismus classification, but also significantly enhance the explainability and medical application value of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a strabismus detection method based on causal feature gating. BACKGROUND

[0002] At present, deep learning technology has been widely applied in the field of medical image analysis, among which the model such as convolutional neural network (CNN) has shown high recognition accuracy in the detection and classification of ophthalmic diseases. In the research of strabismus classification, the common method usually relies on the end-to-end feature extraction and classification of eye images, and the model can automatically learn the high-dimensional features of the image and complete the discrimination of strabismus and normal samples. SUMMARY

[0003] The purpose of the present application is to provide a strabismus detection method based on causal feature gating to solve the technical problem that the existing strabismus determination model ignores the causal relationship, resulting in poor reliability in the prior art. The preferred technical solutions in many technical solutions provided by the present application can produce many technical effects, which are described in detail below.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] The first aspect of the present application provides a strabismus detection method based on causal feature gating, comprising: acquiring a plurality of key point coordinates of a face eye region, modeling and analyzing the causal relationship between the key point coordinates and the strabismus category by using a causal discovery algorithm, obtaining a causal relationship model for determining the causal relationship, screening the key point coordinates based on the causal relationship model, obtaining a causal variable set, and extracting a plurality of causal variables from a plurality of samples in a training sample set according to the causal variable set; introducing a causal feature gating module between adjacent network layers of a classification model, the causal feature gating module comprising gating parameters for representing the contribution weight of the corresponding causal variable, the causal feature gating module being used for weighting and modulating each causal variable, training the classification model by using a plurality of causal variables, and simultaneously adapting the gating parameters in the training process of the classification model until the classification model converges or reaches a preset training number of rounds, obtaining a final classification model; predicting a to-be-predicted sample by using the final classification model, obtaining a prediction result, explaining the prediction result by using a SHAP (SHapley Additive exPlanations) explanation method to obtain a visual result, and determining the strabismus category by comprehensively considering the prediction result and the visual result.

[0006] In some embodiments, the modeling and analyzing the causal relationship between the key point coordinates and the strabismus category by using the causal discovery algorithm obtains a causal relationship model for determining the causal relationship, including: generating a plurality of candidate causal relationship models according to the key point coordinates, constructing a corresponding structural equation model for each of the candidate causal relationship models; performing parameter estimation and model fitting on each of the structural equation models, performing conditional independence test based on the fitting result, and screening models passing the conditional independence test; scoring all the models passing the conditional independence test, and obtaining a causal structure diagram of a causal relationship model with the optimal score as the causal structure between the key point coordinates and the strabismus category, for determining the causal relationship.

[0007] In some embodiments, the screening the key point coordinates based on the causal relationship model obtains a causal variable set, including: determining a target node corresponding to the strabismus category from the causal structure diagram, and obtaining a variable set constituting a Markov blanket with the target node, wherein a key point coordinate variable in the variable set is determined as a causal variable related to the strabismus category.

[0008] In some embodiments, the extracting a plurality of causal variables from a plurality of samples in a training sample set according to the causal variable set includes: randomly selecting a sample from the same group of samples in the training sample set as an original sample, and the remaining samples as enhanced samples, and extracting the causal variables from the original sample and the enhanced samples, respectively; and constructing the causal variables into one-dimensional feature arrays, wherein the one-dimensional feature array of the original sample is used as the main input of the classification model, and the one-dimensional feature array of the enhanced sample is used as the auxiliary input of the classification model.

[0009] In some embodiments, the training the classification model by using a plurality of the causal variables while adaptively adjusting the gating parameters in the training process of the classification model includes: sequentially passing a plurality of the causal variables through linear transformation, causal feature gating modulation, and non-linear activation function of each layer of the classification model for forward propagation to obtain a predicted value; calculating a loss function based on the difference between the predicted value and a true label, and constructing a consistency constraint; calculating the gradient of the loss function with respect to the network parameters and the gating parameters in the classification model by using a back propagation algorithm, and jointly updating the network parameters and the gating parameters by using an optimization algorithm.

[0010] In some embodiments, the Markov blanket variables of the target node include parent nodes of the target node, child nodes of the target node, and parent nodes of the child nodes of the target node.

[0011] In some embodiments, the classification model is a multilayer perceptron model, and the multilayer perceptron model comprises at least one hidden layer.

[0012] In some embodiments, the structural equation model is represented by the following formula:

[0013]

[0014] wherein, represents the i-th key point coordinate variable, represents a set of parent node variables of the key point coordinate variable, represents a corresponding function relationship, represents a mutually independent noise term. In some embodiments, the structural equation model is parameterized and model fitted by using a least square method or a maximum likelihood estimation method.

[0015] A second aspect of the present application provides a processing device, comprising: one or more processors; a memory for storing one or more computer programs, and one or more processors are configured to execute one or more computer programs stored in the memory to make one or more processors execute the strabismus detection method based on the causal feature gating as described above.

[0016] Implementing one of the technical solutions of the present application has the following advantages or beneficial effects: in the present application, first, a causal relationship model for determining a causal relationship is obtained by a causal discovery algorithm, then the causal relationship model is used to screen a causal variable set from key point coordinates, and these causal vectors are extracted from a training sample set to train a classification model, and in the training process, a causal feature gating module is introduced to make the gating parameters adaptively updated, realizing joint optimization of the classification model network parameters and the gating parameters, obtaining a final classification model for obtaining a prediction result of a strabismus category, and the prediction result is visualized.

[0017] In this case, by performing causal discovery and causal variable construction on eye key points, causal key features with clear medical semantics can be extracted; at the same time, a causal feature gating module is introduced into the classification model, so that the classification model can adaptively assign importance weights to the causal variables, and combined with the SHAP explanation method, the transparency and clinical usability of the prediction result are enhanced. Therefore, the embodiments of the present application can not only improve the accuracy of strabismus classification, but also significantly enhance the explainability and medical application value of the model.

[0018] In this case, by performing causal discovery and causal variable construction on eye key points, causal key features with clear medical semantics can be extracted; at the same time, a causal feature gating module is introduced into the classification model, so that the classification model can adaptively assign importance weights to the causal variables, and combined with the SHAP explanation method, the transparency and clinical usability of the prediction result are enhanced. Therefore, the embodiments of the present application can not only improve the accuracy of strabismus classification, but also significantly enhance the explainability and medical application value of the model. BRIEF DESCRIPTION OF DRAWINGS

[0019] ​In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort, and the drawings are as follows:

[0020] Figure 1 is a flow diagram of the strabismus detection method of the causal feature gating according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the objects, technical solutions and advantages of the present application more clear, the various exemplary embodiments to be described below will be described with reference to the corresponding drawings, which constitute a part of the exemplary embodiments, and various exemplary embodiments that can be used to implement the present application are described. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation described in the following exemplary embodiments does not represent all the implementations consistent with the present disclosure. It should be understood that they are only examples of processes, methods and apparatuses, etc. consistent with some aspects of the present disclosure as described in the appended claims, and other embodiments can be used or structural and functional modifications can be made to the embodiments listed herein without departing from the scope and spirit of the present application.

[0022] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the elements referred to must have a specific orientation, structure and operation. The terms "first", "second" and the like are only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. The term "a plurality of" means two or more. The terms "connected", "connected" should be broadly understood, for example, it can be fixedly connected, detachably connected, integrally connected, mechanically connected, electrically connected, communicatively connected, directly connected, indirectly connected through an intermediate medium, it can be the internal communication of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more related listed items. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0023] In order to illustrate the technical solutions described in the present application, the following will be described by specific embodiments, only showing the parts related to the embodiments of the present application.

[0024] As Figure 1As shown, the present application provides a strabismus detection method based on causal feature gating, including the following steps (steps S1 to S3):

[0025] S1, obtain a plurality of key point coordinates of a face eye region, use a causal discovery algorithm to model and analyze the key point coordinates and strabismus categories, obtain a causal relationship model for determining a causal relationship, filter the key point coordinates based on the causal relationship model, obtain a causal variable set, and extract a plurality of causal variables from a plurality of samples in a training sample set according to the causal variable set.

[0026] Specifically, first, a face image of a to-be-detected object can be obtained, and a plurality of key point coordinates of a face eye region can be extracted from the face image based on a preset face key point algorithm. In addition, face images can also be obtained from public general face recognition data sets, such as the CelebA data set, the LFW data set, etc., and face images generated by AI (artificial intelligence) technology can also be used. The face eye region can be located from the face image, and eye key point information can be extracted from the face eye region, wherein the eye key point information can include key point coordinates, and specifically, the xy coordinates of all key points of the eye region can be extracted using the mediapiple framework, which is a multi-modal machine learning model application framework and can be used to perform visual perception tasks, such as face detection.

[0027] Subsequently, the causal relationship between the key point coordinates and the strabismus categories can be modeled and analyzed based on a training sample set using a causal discovery algorithm, a causal relationship model is obtained, causal variables having a direct causal relationship with the strabismus categories are determined, and the key point coordinates are filtered based on the causal relationship model using Markov blanket theory to obtain a causal variable set related to the strabismus categories. Finally, the corresponding causal variables can be extracted according to the causal variable set to construct the input of the classification model.

[0028] In some embodiments, the modeling and analysis of the key point coordinates and the strabismus categories using the causal discovery algorithm to obtain the causal relationship model for determining the causal relationship can include: generating a plurality of candidate causal relationship models according to the key point coordinates, constructing a corresponding structural equation model for each candidate causal relationship model; performing parameter estimation and model fitting on each structural equation model, performing conditional independence test based on the fitting result, and filtering the models that pass the conditional independence test; scoring all models that pass the conditional independence test, and obtaining the causal structure diagram of the causal relationship model with the optimal score as the causal structure between the key point coordinates and the strabismus categories, which is used to determine the causal relationship.

[0029] In some embodiments, the structural equation model can be represented by the following formula:

[0030]

[0031] in, It can represent the first Key point coordinate variables, It can represent the coordinate variables of key points. The set of parent node variables, It can represent the corresponding functional relationship. It can represent mutually independent noise terms. Structural equation modeling can be used to describe the dependencies between causal variables.

[0032] Specifically, the least squares method or the maximum likelihood estimation method can be used to estimate parameters and fit the model of the structural equation model. The least squares method is used to minimize the sum of squared prediction errors, while the maximum likelihood estimation method is used to maximize the probability of observed data.

[0033] In some embodiments, parameter estimation and model fitting can be performed on each structural equation model, and a conditional independence test can be performed based on the fitting results to screen candidate causal relationship models that meet preset independence constraints. The candidate causal relationship models that pass the conditional independence test are scored, and the causal relationship model with the best score is selected according to a preset scoring criterion. The causal structure diagram corresponding to the best-scoring causal relationship model is used as the causal structure between keypoint coordinates and the squint category to determine the causal relationship. Further, it can be tested whether the error terms of the fitted model meet the independence assumption. If there is a statistical dependency between the error terms, it indicates that the causal structure diagram corresponding to that model is invalid. This is used to determine which models pass the independence test, thereby obtaining the models that meet the preset independence constraints, i.e., those that pass the conditional independence test.

[0034] In some embodiments, filtering keypoint coordinates based on a causal relationship model to obtain a set of causal variables may include: determining the target node corresponding to the strabismus category from the causal structure graph, and obtaining a set of variables that form a Markov blanket with the target node, wherein the keypoint coordinate variables in the variable set are determined as causal variables related to the strabismus category. Specifically, the Markov blanket variables of the target node corresponding to the strabismus category may include the parent node of the target node, the child nodes of the target node, and the parent nodes of the child nodes of the target node. Unlike existing high-dimensional features that rely on CNNs (Convolutional Neural Networks), the causal variables obtained through causal relationship filtering are low-dimensional, semantically clear, and medically interpretable. Furthermore, by eliminating redundant features through causal relationships, the consistency between features and disease representation and the model's generalization ability can be improved.

[0035] In some embodiments, extracting the plurality of causal variables from the plurality of samples of the training sample set according to the set of causal variables can include: randomly selecting a sample from the same group of samples of the training sample set as an original sample, and the remaining samples as augmented samples, and extracting the causal variables from the original sample and the augmented samples respectively; and constructing the causal variables into one-dimensional feature arrays, wherein the one-dimensional feature array of the original sample is used as the main input of the classification model, and the one-dimensional feature array of the augmented sample is used as the auxiliary input of the classification model. Wherein, a plurality of face image samples belonging to the same to-be-detected object and acquired within the same time period can be divided into the same group of samples.

[0036] The original sample and the augmented sample can be used jointly in the training process of the classification model. Compared with the high-dimensional image features extracted by the CNN, the causal variables of the present application can represent more compact and high-level semantic information.

[0037] For example, after the causal discovery algorithm and the causal relationship, X causal variables having a causal relationship with the strabismus category are screened out to form a set of causal variables. When the plurality of causal variables are extracted from the plurality of samples of the training sample set, at least the X causal variables in the set of causal variables can be extracted.

[0038] S2, introducing a causal feature gating module between adjacent network layers of the classification model, the causal feature gating module including gating parameters for representing contribution weights of corresponding causal variables, the causal feature gating module being used for weighting and modulating each causal variable, the plurality of causal variables being used for training the classification model, and the gating parameters being adaptively adjusted in the training process of the classification model until the classification model converges or reaches a preset training number of rounds, to obtain a final classification model.

[0039] Specifically, the causal feature gating module can include learnable gating parameters corresponding to the causal variables, for weighting and modulating the causal feature representation input to the current network layer.

[0040] In the forward propagation process of the classification model, the causal feature gating module can perform element-by-element or channel-by-channel gating modulation operation on the feature vector from the previous network layer to obtain a gated feature representation, and the gated feature representation is used as the input of the current network layer.

[0041] In the training process of the classification model, the gating parameters can be used as trainable parameters to participate in back propagation, and the gating parameters and network parameters are jointly optimized by minimizing a preset loss function, so that the gating parameters are adaptively adjusted in the training process, thereby dynamically enhancing the features having causal contribution to the classification result and suppressing the features having weak or redundant relevance to the classification task, until the classification model converges or reaches a preset training number of rounds, to obtain a final classification model.

[0042] To solve the problem of data distribution imbalance caused by relatively small number of strabismus samples and large number of non-strabismus samples in actual application scenarios, the embodiment of the present application adopts a training strategy based on group data in the model training stage:

[0043] Specifically, multiple face images of the same to-be-detected object collected in the same time period form a group of samples, wherein the samples in the group differ in shooting device type, shooting angle or shooting environment.

[0044] In each training iteration process, one sample is selected from each group of samples as an original sample, and the remaining samples are selected as enhanced samples, and group training data is constructed based on the original sample and the enhanced sample.

[0045] In the model training process, the causal variables corresponding to the same group of samples are respectively input into the classification model to obtain multiple prediction results, and a group consistency loss function is constructed based on the prediction results, and the network parameters in the classification model and the gating parameters in the causal feature gating module are jointly optimized by minimizing the group consistency loss function, so that the gating parameters are adaptively adjusted in the training process to enhance the response ability to the features with stable causal contribution and suppress the non-causal interference features introduced by the change of shooting conditions.

[0046] The range of the gating parameter can be 0 to 1, which can be regarded as multiplying a soft mask by each causal variable, i.e., a causal mask, which is used to add a weight or probability to the original sample, and the original information and a general "masking representation" can be weighted and mixed. Through adaptive updating of the gating parameter, the classification model can allocate different weights according to the importance of the features.

[0047] In some embodiments, the classification model is trained using multiple causal variables, and the gating parameter is adaptively adjusted in the training process of the classification model, which can include:

[0048] The multiple causal variables are sequentially subjected to linear transformation, causal feature gating modulation and non-linear activation function of each layer of the classification model to perform forward propagation to obtain prediction values; a loss function is calculated based on the difference between the prediction values and the true labels, and a group consistency constraint is constructed; the gradients of the loss function with respect to the network parameters in the classification model and the gating parameters are calculated by a back propagation algorithm, and the network parameters and the gating parameters are jointly updated by an optimization algorithm. The causal variables are sequentially subjected to linear transformation and non-linear activation function of each layer of the classification model, and the features are gradually extracted, transformed and combined in each layer to obtain higher-level and more abstract representations.

[0049] Specifically, causal variables from the main input and auxiliary input can be input into the classification model, allowing them to pass through linear transformations, causal feature gating modulation, and nonlinear activation functions at each layer of the classification model in sequence, performing forward propagation to obtain the corresponding predicted values. Then, the loss function can be calculated based on the difference between the predicted values ​​and the true labels, and group consistency constraints can be constructed based on the prediction results of samples in the same group. Finally, the gradient of the loss function with respect to the network parameters and causal feature gating parameters in the classification model can be calculated using the backpropagation algorithm, and the network parameters and causal feature gating parameters can be jointly updated using an optimization algorithm.

[0050] In some embodiments, the classification model can be a multilayer perceptron (MLP) model, which includes at least one hidden layer.

[0051] During each training iteration, multiple sets of samples from the training sample set can be input simultaneously, from which original samples and augmented samples can be randomly selected, thereby enriching the diversity of the training data. In other embodiments, the classification model can be logistic regression, random forest, support vector machine, or a graph neural network-based model.

[0052] During training, the classification model can simultaneously utilize both original and augmented samples to optimize the gating parameters. The causal feature gating module can adjust the gating parameters based on the consistency constraints of the group data to ensure that the prediction results corresponding to the same group of samples remain consistent within a preset range. The consistency constraints make the high-level feature representations of the original and augmented samples as close as possible. The causal feature gating module can adjust the gating weight of each causal variable according to the synergistic effect of the consistency constraints and the task objective, thereby highlighting key causal features while maintaining feature consistency and improving the robustness and generalization ability of the final classification model.

[0053] The causal feature gating module can adjust the gating parameters based on the consistency constraints of group data during model training, so that the prediction results corresponding to the same group of samples remain consistent within a preset range. This guides the gating parameters to enhance the response to stable causal features and suppress non-causal features introduced by changes in collection conditions.

[0054] S3. Use the final classification model to predict the sample to be predicted and obtain the prediction results. At the same time, use the SHAP interpretation method to interpret the prediction results and obtain the visualization results. Combine the prediction results and visualization results to determine the strabismus category.

[0055] Specifically, after the model training is completed, the final classification model trained can be used to predict the to-be-predicted sample, obtain the corresponding strabismus classification prediction result, and perform feature contribution analysis on the prediction result based on the SHAP explanation method to generate a visual result of the influence degree of the causal variables on the classification result. The prediction result and the visual result are comprehensively used to determine the strabismus category of the to-be-detected object.

[0056] Specifically, the SHAP method (SHapley Additive exPlanations) is a unified method for explaining the prediction result of any machine learning model, which can be used to calculate the contribution value of each causal variable to the prediction result, and the visual result can be used to show the proportion weight of each causal variable in the final classification decision. In this case, the model judgment basis can be intuitively understood through the visual result, and the credibility of the model result is improved.

[0057] In the present application, firstly, a causal relationship model for determining a causal relationship is obtained through a causal discovery algorithm, and then the causal relationship model is used to screen a causal variable set from the key point coordinates, and the causal vectors are extracted from the training sample set to train a classification model. In the training process, the causal feature gating module is introduced to adaptively update the gating parameters, realize the joint optimization of the classification model network parameters and the gating parameters, obtain the final classification model for obtaining the prediction result of the strabismus category, and perform visual processing on the prediction result.

[0058] In this case, by performing causal discovery and causal variable construction on the eye key points, causal key features with clear medical semantics can be extracted. Meanwhile, the causal feature gating module is introduced into the classification model, so that the classification model can adaptively assign importance weights to the causal variables, and the SHAP explanation method is combined to enhance the transparency and clinical usability of the prediction result. Therefore, the embodiments of the present application can not only improve the accuracy of strabismus classification, but also significantly enhance the explainability and medical application value of the model.

[0059] Those skilled in the art can understand that all or part of the features / steps of the above-mentioned method embodiments can be implemented by a method, a data processing system or a computer program, which can be implemented in a hardware manner, a software manner or a combination of hardware and software. The aforementioned computer program can be stored in one or more computer-readable storage media, and the storage medium stores the computer program. When the computer program is executed (such as a processor), the steps of the strabismus detection method embodiment based on the causal feature gating are executed.

[0060] The aforementioned storage medium that can store program codes includes a static hard disk, a solid state hard disk, a random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), an optical storage device, a magnetic storage device, a flash memory, a magnetic disk or an optical disk, and / or a combination of the above devices, i.e., can be implemented by any type of volatile or non-volatile storage device or a combination thereof.

[0061] The application also provides a processing device embodiment, comprising one or more processors and a memory; wherein the memory is used to store one or more computer programs, and the one or more processors are used to execute the one or more computer programs stored in the memory, so that the processor executes the features / steps of the aforementioned strabismus detection method embodiment based on the causal feature gating.

[0062] The above only describes the preferred embodiments of the present application, and those skilled in the art know that various changes or equivalent replacements can be made to the features and embodiments without departing from the spirit and scope of the present application. In addition, the features and embodiments can be modified to adapt to specific conditions and materials under the guidance of the present application without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the protection scope of the present application.

Claims

1. A strabismus detection method based on causal feature gating, characterized in that, include: The coordinates of several key points in the eye region of a face are obtained. The causal relationship between the key point coordinates and the strabismus category is modeled and analyzed using a causal discovery algorithm to obtain a causal relationship model for determining the causal relationship. The key point coordinates are filtered based on the causal relationship model to obtain a set of causal variables. Multiple causal variables are extracted from several samples in the training sample set according to the set of causal variables. A causal feature gating module is introduced between adjacent network layers of the classification model. The causal feature gating module includes gating parameters for characterizing the contribution weights of corresponding causal variables. The causal feature gating module is used to perform weighted modulation on each causal variable and train the classification model using multiple causal variables. At the same time, the gating parameters are adaptively adjusted during the training process of the classification model until the classification model converges or reaches a preset number of training rounds to obtain the final classification model. The final classification model is used to predict the sample to be predicted to obtain the prediction result. At the same time, the SHAP interpretation method is used to interpret the prediction result to obtain the visualization result. The strabismus category is determined by combining the prediction result and the visualization result. The step of training the classification model using multiple causal variables while adaptively adjusting the gating parameters during the training process includes: forward propagating the multiple causal variables sequentially through linear transformations, causal feature gating modulation, and nonlinear activation functions at each layer of the classification model to obtain predicted values; calculating a loss function based on the difference between the predicted values ​​and the true labels, and constructing consistency constraints; calculating the gradient of the loss function with respect to the network parameters and the gating parameters in the classification model using a backpropagation algorithm, and jointly updating the network parameters and the gating parameters using an optimization algorithm. The causal feature gating module adjusts the gating parameters based on the consistency constraints of group data to ensure that the prediction results of the same group of samples remain consistent within a preset range.

2. The strabismus detection method based on causal feature gating according to claim 1, characterized in that, The step of using a causal discovery algorithm to model and analyze the causal relationship between the keypoint coordinates and the strabismus category to obtain a causal relationship model for determining the causal relationship includes: generating several candidate causal relationship models based on the keypoint coordinates; constructing a corresponding structural equation model for each candidate causal relationship model; performing parameter estimation and model fitting on each structural equation model; performing a conditional independence test based on the fitting results; and selecting models that pass the conditional independence test; scoring all models that pass the conditional independence test; and obtaining the causal structure diagram of the causal relationship model with the best score as the causal structure between the keypoint coordinates and the strabismus category, which is used to determine the causal relationship.

3. The strabismus detection method based on causal feature gating according to claim 2, characterized in that, The step of filtering the key point coordinates based on the causal relationship model to obtain a set of causal variables includes: determining the target node corresponding to the strabismus category from the causal structure diagram, and obtaining a set of variables that form a Markov blanket with the target node, wherein the key point coordinate variables in the set of variables are determined as causal variables related to the strabismus category.

4. The strabismus detection method based on causal feature gating according to claim 1, characterized in that, The step of extracting multiple causal variables from several samples in the training sample set based on the set of causal variables includes: randomly selecting one sample from the same group of samples in the training sample set as the original sample, and the remaining samples as augmented samples; extracting the causal variables from the original sample and the augmented samples respectively; constructing the causal variables into a one-dimensional feature array, wherein the one-dimensional feature array of the original sample is used as the main input of the classification model, and the one-dimensional feature array of the augmented sample is used as the auxiliary input of the classification model.

5. The strabismus detection method based on causal feature gating according to claim 3, characterized in that, The Markov blanket variables of the target node include the parent node of the target node, the child nodes of the target node, and the parent node of the child nodes of the target node.

6. The strabismus detection method based on causal feature gating according to claim 1, characterized in that, The classification model is a multilayer perceptron model, which includes at least one hidden layer.

7. The strabismus detection method based on causal feature gating according to claim 2, characterized in that, The structural equation model is expressed by the following formula: in, Indicates the first Key point coordinate variables, Represents the coordinate variables of the key points The set of parent node variables, This represents the corresponding functional relationship. This represents mutually independent noise terms.

8. The strabismus detection method based on causal feature gating according to claim 2, characterized in that, include: The least squares method or the maximum likelihood estimation method are used to estimate the parameters and fit the model of the structural equation model.

9. A processing device, characterized in that, include: One or more processors; A memory for storing one or more computer programs, and one or more processors for executing the one or more computer programs stored in the memory to cause the one or more processors to perform the strabismus detection method based on causal feature gating as described in any one of claims 1-8.

Images