Method and device for detecting anterior chamber angle stenosis or closure disease accompanied with latent lens suspension ligament relaxation

Abstract

The invention provides a method and a device for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens suspension ligament relaxation, and an innovative multi-modal classification model is constructed by combining OCT image features extracted by 3D-CNN and clinical parameter features based on integrated learning. After data preprocessing, key features of an SS-OCT image are automatically extracted by using a 3D-CNN model, and clinical parameters are accurately classified by training a plurality of base machine learning models and integrating results of the models by using a voting classifier. Finally, image features and clinical parameter features are fused, 3D-CNN and MLP parameters are jointly trained, feature combination is optimized, and efficient and accurate diagnosis of the relaxation state of the crystalline lens suspension ligament can be achieved on a computer. According to the method, the comprehensiveness and accuracy of computer-aided diagnosis are improved, subjective errors of manual judgment are reduced, and a powerful auxiliary tool is provided for ophthalmologists.

Description

Technical Field

[0001] The present invention relates to ophthalmic medical treatment and the application of deep learning technology, and in particular to a method and device for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens zonular relaxation. Background Art

[0002] Diagnosis Primary angle closure disease (PACD) is a general term for a series of continuously progressive diseases, which can be divided into suspected angle closure and primary angle closure. When photo-ocular optic nerve changes occur in primary angle closure, it further develops into primary angle closure glaucoma (PACG). Glaucoma is the world's leading irreversible blinding eye disease. PACG and primary open-angle glaucoma are the two most common types of glaucoma, and Asians are more susceptible to PACG. It is estimated that by 2040, the number of PACG patients in the world will reach 32.04 million, of which Asians account for as much as 3 / 4. According to research statistics, the prevalence of PACG in people over 40 years old in my country in the past 20 years was 1.35%, which is higher than the average prevalence of PACG in Asians.

[0003] PACG is caused by increased intraocular pressure due to angle closure, which eventually leads to irreversible damage to the optic nerve. The main function of the lens zonule is to fix the lens between the anterior and posterior chambers of the eyeball, so that the lens remains in the center of the optical axis. Its secondary function is to transmit the accommodative force from the ciliary body to the lens. When the lens zonule is abnormal, such as lens dislocation, it can directly stimulate the ciliary body to increase aqueous humor secretion or cause angle closure. If the lens only moves forward, it will also make the anterior chamber shallower, which is easy to induce pupillary block, angle closure and finally increased intraocular pressure. Therefore, some scholars believe that abnormal lens zonule is one of the pathogenesis of PACG. A study showed that after the use of compound tropicamide to dilate the pupils of PACD patients, the lens zonule showed a drug-induced tension state. It was found that patients with suspected primary angle closure had uneven distribution of zonule relaxation compared with normal people, which also suggested that PACD patients may have hidden lens deviation due to abnormal lens zonule. PACD patients are not detected before surgery, but zonular abnormalities are found during glaucoma cataract surgery (such as wrinkles in the anterior capsule, poor capsule following, unstable anterior capsule floating bag, etc.), which are called occult zonular abnormalities.

[0004] The causes of abnormal lens zonular ligaments can be divided into primary and secondary. The specific mechanism of lens zonular ligament relaxation in PACD patients is still unclear. Some believe that PACG patients have shorter axial lengths, thick and heavy lenses, choroidal expansion, and vitreous pushing the lens-iris septum forward, causing the zonular ligaments to be in a state of tension for a long time and causing abnormalities; others believe that the zonular ligaments of PACG patients are in a state of increased tension for a long time to resist the increased pressure behind the lens, which eventually causes the zonular ligaments to relax or even break, and an imbalance in the pressure between the front and back of the lens.

[0005] At present, the criterion for judging the relaxation of the zonular ligament is to observe radial wrinkles and capsular bag movement in the anterior capsule during continuous circular capsulorhexis during surgery. This criterion is very important in clinical practice because the state of the zonular ligament directly affects the diagnosis and treatment of glaucoma.

[0006] However, clinically, preoperative evaluation of latent zonular relaxation is very challenging. Currently, there is a lack of effective preoperative evaluation methods to diagnose latent zonular relaxation. Existing diagnostic methods such as ultrasound biomicroscopy (UBM) and the use of Pentacam to measure changes in peripheral anterior chamber depth have limitations. UBM requires good cooperation from patients and lacks repeatability; Pentacam lacks accuracy and intuitiveness in measuring lens morphology-related parameters.

[0007] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0008] The main purpose of the present invention is to solve the problems existing in the above-mentioned background technology and to provide a method and device for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens zonular relaxation.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] A method for detecting narrow or closed anterior chamber angle disease with occult zonular relaxation comprises the following steps:

[0011] S1. Preprocessing the collected anterior segment OCT images and clinical parameters of the patient; preferably, the anterior segment OCT images are anterior segment swept frequency optical coherence tomography (SS-OCT) images;

[0012] S2. constructing a three-dimensional convolutional neural network 3D-CNN model, training the 3D-CNN model using the annotated OCT image data, and automatically extracting image features related to lens suspensory ligament relaxation through the trained model;

[0013] S3. Select clinical parameter features related to lens zonular relaxation, train multiple base machine learning models, and integrate the results of multiple base models through a voting classifier;

[0014] S4. Fuse the OCT image features extracted in step S2 with the clinical parameter features obtained in step S3, use the fused features to train a multimodal classification model, and use the multimodal classification model to classify and diagnose the patient's lens suspensory ligament status, wherein the multimodal classification model achieves optimal combination of features and classification decisions by jointly training parameters of 3D-CNN and multi-layer perceptron MLP.

[0015] Furthermore, the data preprocessing of the anterior segment OCT image includes image denoising, normalization and segmentation, and data enhancement of the anterior segment OCT image includes random rotation, translation and flipping; 16 OCT images of the patient acquired from different viewing angles from 0 to 180 degrees enable the model to learn the three-dimensional structural information of the eye tissue.

[0016] Furthermore, in step S2, the 3D-CNN model is a 3DResnet-18 model including multiple convolutional layers, pooling layers and fully connected layers. The convolutional layer extracts feature maps of different scales and depths by applying multiple 3D convolution kernels to the input image. The pooling layer reduces the size of the feature map by downsampling while retaining important spatial information. The fully connected layer integrates and classifies the extracted features. The last layer of the model outputs a vector containing the probability of the suspensory ligament relaxation state.

[0017] Furthermore, in the 3DResnet-18 model, the input three-dimensional image is subjected to preliminary feature extraction through the convolution layer, and the spatial and depth features of the image are captured through multiple 3D convolution kernels; the extracted features are normalized through the batch normalization layer to stabilize the learning process and improve the convergence speed of the model; nonlinear activation is introduced through the rectified linear unit layer ReLU to enhance the model's ability to express complex features; the spatial dimension of the feature map is reduced through the pooling layer to achieve feature selection and extract information; some neurons are randomly discarded through the embedded discard layer to prevent the model from overfitting; in the deep layer of the model, the features extracted by the previous layer are integrated through the fully connected layer, and the final classification decision is made, and the output of the fully connected layer is converted into a probability distribution through the softmax layer to represent the prediction probability of different categories, thereby achieving the classification of the relaxation state of the suspensory ligament; wherein, multiple residual blocks are set, each block is responsible for extracting features at different levels, and the features at different levels are fused through jump connections to effectively transfer gradients in the deep network.

[0018] Furthermore, in step S2, the loss function of the 3DResnet-18 model adopts cross entropy loss, and the optimizer selects Adam optimizer to accelerate the convergence speed of the model; during the training process, the model parameters are continuously adjusted through the back propagation algorithm to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting.

[0019] Furthermore, in step S3, the clinical parameter characteristics related to zonular relaxation of the lens include age (Age), anterior chamber depth (ACD), lens sagittal height (LV), the volume and average value of the angle recess area at 750 mm from the angle of the lens (ARA750), the radius of curvature of the steepest part of the anterior surface of the lens (F-Rs) and its axial direction, the radius of curvature of the flattest part of the anterior surface of the lens (F-Rf) and its axial direction, lens thickness (Thickness) and lens tilt angle (Tilt).

[0020] Furthermore, in step S3, the multiple base models include support vector machine (SVM), random forest (RandomForest), gradient boosting (Gradient Boosting), naive Bayes (Naive Bayes) and K nearest neighbor (KNN), and the multiple base models are combined through the voting classifier.

[0021] Furthermore, in step S4, the multimodal model performs feature splicing on the OCT image features extracted in step S2 and the clinical parameter features obtained in step S3, and inputs the spliced ​​comprehensive feature vector into a multilayer perceptron MLP, wherein the MLP includes a plurality of fully connected layers, and classification is achieved by training and learning the high-order interaction relationship in the feature vector; during the multimodal model training process, a joint training method is used to simultaneously optimize the parameters of the 3D-CNN and MLP, optimize the feature combination, and achieve efficient classification of the relaxation state of the lens suspensory ligament; preferably, the loss function of the multimodal model adopts a cross entropy loss, and the optimizer selects an Adam optimizer to accelerate the convergence speed of the model; during the training process, the model parameters are continuously adjusted through supervised learning to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting

[0022] Furthermore, the method further comprises the following steps:

[0023] S5. Algorithm visualization processing, including:

[0024] The t-SNE method is used to reduce the dimensionality of the high-dimensional features extracted from the 3D-CNN model and map the high-dimensional features to a two-dimensional or three-dimensional space so that the relationship between data points can be intuitively observed and analyzed through visualization, and the distribution boundary between slack and normal data can be separated;

[0025] The Grad-CAM technology is applied to generate a visualized heat map of the model's decision basis. This technology is used to perform gradient weighting on the feature map of the last convolutional layer of CNN to generate a heat map to intuitively display the key areas that the model focuses on on the OCT image.

[0026] A device for detecting narrow or closed angle disease with occult zonular laxity, comprising:

[0027] Memory for storing computer programs;

[0028] A processor, configured to implement the method for detecting narrow or closed anterior chamber angle disease with occult zonular relaxation when executing the computer program;

[0029] Result display module: Feedback the results processed by the processor to the operator in real time through the display screen.

[0030] The present invention has the following beneficial effects:

[0031] The embodiment of the present invention provides a method and device for detecting anterior chamber angle stenosis or closure disease accompanied by occult lens zonular relaxation. Through a deep learning algorithm, especially a convolutional neural network (CNN), the anterior segment swept frequency optical coherence tomography (SS-OCT) image is deeply analyzed, and combined with clinical parameters, an efficient and accurate diagnosis of the state of lens zonular relaxation is achieved. The high-dimensional features of the OCT image are automatically extracted through a 3D-CNN model, and the results of multiple base models are integrated through a voting classifier using multiple trained base machine learning models to accurately classify clinical parameters. The present invention innovatively fuses the high-dimensional features of the OCT image and the clinical parameter classification features integrated by multiple base models to construct a multimodal classification model. This fusion mechanism not only makes full use of the information of image data and clinical data, but also optimizes the feature combination by jointly training the parameters of 3D-CNN and multi-layer perceptron (MLP), significantly improving the comprehensiveness and accuracy of the diagnosis. The method of the present invention ensures the stability of the system under different conditions and the consistency of the diagnostic results, reduces the subjective errors of manual judgment, provides ophthalmologists with a reliable auxiliary diagnostic tool, and effectively improves the accuracy and repeatability of detection and evaluation. The innovative mechanism of multimodal fusion improves the diagnostic efficiency of latent lens zonules relaxation by comprehensively analyzing multi-source data, and provides strong technical support for the early screening and treatment of ophthalmic diseases.

[0032] In an embodiment of the present invention, a deep learning algorithm is used to train with a large amount of clinical data and images to establish an efficient and accurate multimodal classification model, which can automatically identify and analyze the relaxation state of the lens suspensory ligament. Among them, a convolutional neural network (CNN) is used to process and analyze OCT images (such as SS-OCT images), and a multimodal diagnostic model is constructed by integrating clinical parameter classification features obtained by integrating multiple base machine learning models trained with clinical parameters of patients such as lens thickness, the radius of curvature at the steepest point of the anterior surface of the lens, and the height of the lens arch, thereby improving the accuracy and repeatability of the diagnosis. This method not only improves the diagnostic efficiency, but also reduces the dependence of the operator, providing a powerful auxiliary tool for ophthalmologists.

[0033] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of a detection method according to an embodiment of the present invention.

[0035] Figure 2 It is a structural diagram of the 3dResnet18 model according to an embodiment of the present invention.

[0036] Figure 3It is a structural diagram of a multimodal model according to an embodiment of the present invention.

[0037] Figure 4 1 is an OCT image of the suspensory ligament relaxation according to an embodiment of the present invention and a normal patient.

[0038] Figure 5 It is a t-SNE classification visualization diagram of an embodiment of the present invention.

[0039] Figure 6 It is a Grad-cam classification visualization diagram of an embodiment of the present invention.

[0040] Figure 7 It is a logic diagram of interaction between the present invention and the OCT device. DETAILED DESCRIPTION

[0041] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention.

[0042] When cataract surgery is performed on an eye with a relaxed lens zonule, surgical complications such as posterior capsule rupture, vitreous herniation, postoperative intraocular lens displacement, and large intraocular lens power measurement errors may occur, and hidden lens zonules are more difficult to detect. The present invention aims to provide a method and device for diagnosing hidden lens zonule relaxation based on clinical parameters of deep learning and anterior segment OCT images (such as SS-OCT images). By establishing an efficient deep learning model, combined with the patient's clinical parameters, such as intraocular pressure, age, and anterior chamber depth, and high-resolution SS-OCT images, the relaxed state of the lens zonule is automatically and objectively evaluated. This method can not only improve the accuracy and efficiency of diagnosis and reduce the subjective errors of manual judgment, but also provide analysis reports for preoperative evaluation of lens zonule function in cataract patients, assist doctors in formulating scientific and reasonable surgical strategies, reduce the incidence of intraoperative complications, and ensure high repeatability and consistency of diagnostic results. By achieving these purposes, the present invention helps to significantly improve the diagnosis and treatment effects of patients with PACD combined with cataracts, and improve the overall level of ophthalmic medical services.

[0043] See also Figure 1 The embodiment of the present invention provides a method for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens zonular relaxation, comprising the following steps:

[0044] Step S1. Data preprocessing and enhancement: Data preprocessing is performed on the collected anterior segment OCT images of the patients, including denoising, normalization and segmentation of the anterior segment OCT images, as well as cleaning and standardization of clinical parameters.

[0045] In a preferred embodiment, in step S1, the data preprocessing of the anterior segment OCT image includes image denoising, normalization and segmentation to ensure the quality and consistency of the input data, and the anterior segment OCT image is also data enhanced, including random rotation, translation and flipping, to increase data diversity and prevent model overfitting; 16 OCT images of the patient acquired from different viewing angles of 0 to 180 degrees enable the model to learn the three-dimensional structural information of the eye tissue.

[0046] Step S2. 3D-CNN model training and feature extraction: construct a three-dimensional convolutional neural network 3D-CNN model, use the labeled OCT image data to train the 3D-CNN model, and automatically extract image features related to lens suspensory ligament relaxation through the trained model.

[0047] See also Figure 2 In a preferred embodiment, in step S2, the 3D-CNN model is a 3DResnet-18 model including multiple convolutional layers, pooling layers and fully connected layers. The convolutional layers extract feature maps of different scales and depths by applying multiple 3D convolution kernels to the input image. The pooling layers reduce the size of the feature maps by downsampling operations while retaining important spatial information. The fully connected layers integrate and classify the extracted features. The last layer of the model outputs a vector containing the probability of the suspensory ligament relaxation state.

[0048] See also Figure 2 In a preferred embodiment, in the 3DResnet-18 model, a convolution layer is used to perform preliminary feature extraction on the input three-dimensional image, and multiple 3D convolution kernels are used to capture the spatial and depth features of the image; the extracted features are normalized by a batch normalization layer to stabilize the learning process and improve the convergence speed of the model; nonlinear activation is introduced by a modified linear unit layer ReLU to enhance the model's ability to express complex features; the spatial dimension of the feature map is reduced by a pooling layer to achieve feature selection and extract information; a dropout layer is embedded to randomly drop some neurons to prevent the model from overfitting and enhance the generalization ability of the model; in the deep layer of the model, the features extracted by the previous layer are integrated by a fully connected layer, and the final classification decision is made, and the output of the fully connected layer is converted into a probability distribution by a softmax layer to represent the prediction probability of different categories, so as to achieve accurate classification of the relaxation state of the suspensory ligament; wherein, multiple residual blocks are set, each block is responsible for extracting features at different levels, and features at different levels are fused through jump connections, and gradients are effectively transferred in the deep network, thereby improving the training efficiency and diagnostic accuracy of the model.

[0049] In a preferred embodiment, in step S2, the loss function of the 3DResnet-18 model adopts cross entropy loss, and the optimizer selects Adam optimizer to accelerate the convergence speed of the model; during the training process, the model parameters are continuously adjusted through the back propagation algorithm to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting.

[0050] Step S3. Clinical parameter machine learning classification: select clinical parameter features related to lens zonular relaxation, train multiple base machine learning models, and integrate the results of multiple base models through a voting classifier to improve classification accuracy.

[0051] In a preferred embodiment, in step S3, the clinical parameter characteristics related to zonular relaxation of the lens include age (Age), anterior chamber depth (ACD), lens sagittal height (LV), the volume and average value of the angle recess area (ARA750) at 750 mm from the angle of the lens, the radius of curvature of the steepest part of the anterior surface of the lens (F-Rs) and its axial direction, the radius of curvature of the flattest part of the anterior surface of the lens (F-Rf) and its axial direction, lens thickness (Thickness) and lens tilt angle (Tilt).

[0052] In a preferred embodiment, in step S3, the multiple base models include support vector machine (SVM), random forest (RandomForest), gradient boosting (Gradient Boosting), naive Bayes (Naive Bayes) and K nearest neighbor (KNN), and the multiple base models are combined through the voting classifier.

[0053] Step S4. Multimodal model classification: The OCT image features extracted in step S2 are fused with the clinical parameter features obtained in step S3, and a multimodal classification model is trained using the fused features. The multimodal classification model is used to classify and diagnose the patient's lens suspensory ligament status, wherein the multimodal classification model achieves optimal combination of features and classification decisions by jointly training parameters of 3D-CNN and multilayer perceptron (MLP).

[0054] See also Figure 3In a preferred embodiment, in step S4, the multimodal model performs feature splicing on the OCT image features extracted in step S2 and the clinical parameter features obtained in step S3, and inputs the spliced ​​comprehensive feature vector into a multilayer perceptron MLP, which includes a plurality of fully connected layers, and achieves classification by training and learning the high-order interaction relationship in the feature vector; during the multimodal model training process, a joint training method is used to simultaneously optimize the parameters of 3D-CNN and MLP, optimize the feature combination, and achieve efficient classification of the relaxation state of the lens suspensory ligament; preferably, the loss function of the multimodal model adopts a cross entropy loss, and the optimizer selects an Adam optimizer to accelerate the convergence speed of the model; during the training process, the model parameters are continuously adjusted through supervised learning to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting

[0055] In the above steps of the present invention, high-dimensional features are automatically extracted from OCT images by using a 3D-CNN model. At the same time, multiple trained base machine learning models are used to accurately classify clinical parameters, and the results of these base models are integrated through a voting classifier to improve the accuracy and robustness of classification. The OCT image features extracted by 3D-CNN are combined with the clinical parameter features obtained by ensemble learning to construct a multimodal classification model. This fusion strategy fully utilizes the advantages of image analysis and clinical data analysis, and optimizes the combination of features by jointly training the parameters of 3D-CNN and multi-layer perceptron (MLP), thereby significantly improving the comprehensiveness and accuracy of the diagnosis of the relaxation state of the lens zonules.

[0056] In a preferred embodiment, the method further comprises the following steps:

[0057] Step S5: Algorithm visualization processing, specifically including:

[0058] The t-SNE method is used to reduce the dimensionality of the high-dimensional features extracted from the 3D-CNN model and map the high-dimensional features to a two-dimensional or three-dimensional space so that the relationship between data points can be intuitively observed and analyzed through visualization, and the distribution boundary between slack and normal data can be separated;

[0059] The Grad-CAM technology is applied to generate a visualized heat map of the model's decision basis. This technology is used to perform gradient weighting on the feature map of the last convolutional layer of CNN to generate a heat map to intuitively display the key areas that the model focuses on on the OCT image.

[0060] An embodiment of the present invention also provides a device for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens zonular relaxation, comprising: a memory for storing a computer program; a processor for implementing the method for detecting anterior chamber angle stenosis or closure disease accompanied by latent lens zonular relaxation when executing the computer program; and a result display module for feeding back the results processed by the processor to an operator in real time through a display screen.

[0061] The embodiment of the present invention provides an automated auxiliary diagnosis method and device based on deep learning technology, which can efficiently and accurately assist in the evaluation of lens zonular function and early screening of latent lens zonular relaxation by combining clinical parameters and SS-OCT images. This method can not only improve the accuracy and repeatability of detection and evaluation, but also assist doctors in making rapid and reliable preoperative evaluations and treatment decisions.

[0062] The specific embodiments and experimental verification of the present invention are further described below.

[0063] A multimodal lens zonular relaxation diagnostic method based on deep learning combines OCT images and clinical parameters for comprehensive analysis. First, a three-dimensional convolutional neural network (3D-CNN) was used to classify OCT images and extract the characteristic information of the lens zonular ligament. At the same time, a machine learning method was used to classify and analyze clinical parameters. The specific method included importing clinical data, selecting key features, standardizing data, and dividing the data into training sets and test sets. Multiple base models (SVM, random forest, gradient boosting, naive Bayes, and KNN) were trained and combined into a voting classifier for integration. Finally, the 3D-CNN image classification results were combined with the machine learning clinical parameter classification results to construct a multimodal classification model and integrate it into the software to achieve full process automation from data acquisition, processing to diagnosis, significantly improving the accuracy and efficiency of diagnosis.

[0064] 3D-CNN

[0065] In the present invention, 3D convolutional neural network (3D-CNN) is the core technology for processing and analyzing OCT images. 3D-CNN can capture the spatial relationship and depth information of image data by applying convolution operations in three-dimensional space, thereby more accurately extracting features related to lens zonules relaxation. The following are the specific implementation steps and contents of 3D-CNN technology in the present invention:

[0066] Data preprocessing and enhancement

[0067] To improve the robustness and accuracy of the model, the OCT images were first preprocessed and data augmented. The preprocessing steps included image denoising, normalization, and segmentation to ensure the quality and consistency of the input data. Data augmentation techniques such as random rotation, translation, and flipping were applied to the training dataset to increase the diversity of the data and prevent the model from overfitting. Because OCT images are two-dimensional, but because they are taken by rotation, they are processed into a three-dimensional tensor to fully utilize spatial information for analysis. By utilizing 16 OCT images of each patient acquired from different viewing angles from 0 to 180 degrees, the model is able to learn the three-dimensional structural information of the ocular tissue from these sequences.

[0068] 3DResnet-18 model architecture

[0069] The 3DResnet-18 model consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer extracts feature maps of different scales and depths by applying multiple 3D convolution kernels to the input image. The pooling layer reduces the size of the feature map by downsampling while retaining important spatial information. The fully connected layer integrates and classifies the extracted features. The last layer of the model outputs a vector containing the probability of the relaxation state of the suspensory ligament.

[0070] Model training and optimization

[0071] During the model training process, a large amount of annotated OCT image data was used for supervised learning. The loss function used cross entropy loss, and the optimizer used Adam optimizer to speed up the convergence of the model. During the training process, the model parameters were continuously adjusted through the back propagation algorithm to minimize the loss function value. In order to prevent overfitting, early stopping techniques and regularization methods such as Dropout and L2 regularization were used. In this study, the model was trained for a total of 60 epochs to ensure that the data features were fully learned.

[0072] Feature extraction and classification

[0073] The trained 3D-CNN model can automatically extract the key features of the zonular ligament from OCT images. By analyzing and classifying these features, the relaxation state of the zonular ligament can be effectively determined.

[0074] Result verification and evaluation

[0075] The performance of the model is verified by evaluating it on an independent test set. Common evaluation indicators include accuracy, loss, recall, and F1 score. In this study, the performance of the 3D-CNN model on different datasets is as follows:

[0076] Training set: accuracy 0.835, loss 0.3536, recall 0.9773, F1 score 0.8350

[0077] Validation set: accuracy 0.75, loss 0.5569, recall 0.8571, F1 score 0.8000

[0078] Test set: accuracy 0.77, loss 0.4908, recall 1.0000, F1 score 0.8333

[0079] From these results, we can see that the 3D-CNN model performs relatively well on the training set, with an accuracy of 0.8350 and a recall of 0.9773, indicating that the model has a high sensitivity in identifying the relaxation state of the suspensory ligament. The performance of the validation set and the test set is slightly worse than that of the training set, but still maintains a high accuracy and consistency, with an accuracy of 0.7500 for the validation set and 0.8000 for the test set, showing the generalization ability and robustness of the model. Although the recall rates of the validation set and the test set are different, especially the recall rate of the test set reaches 1.0000, showing that the model can fully identify all instances of relaxation state on the test data, which may be due to the distribution characteristics of the test set samples.

[0080] Machine learning classification based on clinical parameters

[0081] The present invention also uses clinical parameters for machine learning classification to assist in the diagnosis of lens zonular relaxation. In this study, multiple machine learning models were used and integrated through a voting classifier to improve the accuracy and robustness of the classification. The specific steps are as follows:

[0082] Data import and preprocessing

[0083] First, clinical data were imported from Excel files. The selected features included age (Age), anterior chamber depth (ACD), lens sagittal height (LV), volume and average of the angle recess area at 750 mm from the angle (ARA750), the radius of curvature of the steepest part of the anterior surface of the lens (F-Rs) and its axial direction, the radius of curvature of the flattest part of the anterior surface of the lens (F-Rf) and its axial direction, lens thickness (Thickness) and lens tilt angle (Tilt), etc. These features were standardized to eliminate the impact of dimensional differences on model training.

[0084] Dataset Partitioning

[0085] The standardized dataset is divided into a training set and a test set, with the training set accounting for 80% and the test set accounting for 20%. This division ensures that the performance of the model on unseen data is reasonably evaluated.

[0086] Model selection and integration

[0087] A variety of base models were selected, including support vector machine (SVM), random forest, gradient boosting, naive Bayes and K-nearest neighbor (KNN). These models have their own advantages and disadvantages. Combining them through voting classifiers can comprehensively utilize the advantages of each model and improve classification performance.

[0088] Model training and prediction

[0089] The voting classifier is trained using the training set and predictions are made on the test set. The performance of the model is judged by evaluating its accuracy, precision, recall, and F1 score.

[0090] In this study, the voting classifier performed as follows on the test set:

[0091] Accuracy: 0.7824

[0092] Precision: 1.00 (normal - 1 category), 0.75 (relaxed - 2 categories)

[0093] Recall: 0.82 (normal - 1 class), 1.00 (relaxed - 2 classes)

[0094] F1 Score: 0.90 (Normal - Class 1), 0.86 (Relaxed - Class 2)

[0095] The overall accuracy reached 88.24%, showing the high efficiency of the model in clinical parameter classification. The model performed well in identifying normal and relaxed lenses, especially in the relaxed state, with a recall rate of 100%, indicating that the model can fully identify all instances of the relaxed state. These results prove the effectiveness and reliability of machine learning classification using clinical parameters in the diagnosis of lens zonular relaxation.

[0096] Multimodal model classification

[0097] In the present invention, in order to further improve the diagnostic accuracy of the relaxation state of the lens zonular ligament, it is planned to fuse the OCT image analysis model based on 3D-CNN with the machine learning model based on clinical parameters to construct a multimodal classification model. The multimodal model can comprehensively utilize image data and clinical data to achieve a more comprehensive and accurate diagnosis. This section describes the ideas and methods of multimodal model fusion.

[0098] Fusion of multimodal data

[0099] Multimodal data fusion refers to combining data from different sources and types to improve the performance of the model. In the present invention, the goal of fusion is to combine the OCT image features extracted by 3D-CNN with the clinical parameter features, and achieve more efficient classification of the lens zonules relaxation state by jointly training a deep learning model.

[0100] Feature extraction and fusion methods

[0101] (1) Feature extraction:

[0102] OCT image features: 3D-CNN is used to extract deep features from OCT images. These features are high-dimensional abstract feature vectors that reflect the structure and texture information in the image. Although these features do not directly correspond to specific physical quantities such as curvature, length, and thickness, they can effectively capture relevant information about the relaxation of the lens zonules.

[0103] Clinical parameter features: Clinical parameter features are extracted through standardization and machine learning methods. These features include age, anterior chamber depth, lens sagittal height, etc.

[0104] (2) Feature Fusion:

[0105] Feature splicing: Splice the image features extracted by 3D-CNN with the clinical parameter features to form a high-dimensional feature vector. In this way, the information of image and clinical data can be integrated.

[0106] Multilayer Perceptron (MLP): The concatenated feature vector is input into a multilayer perceptron (MLP) for further processing and classification. MLP consists of several fully connected layers. Through training, it can learn the high-order interactions in the feature vector and improve the classification performance.

[0107] Model training and optimization

[0108] In the process of multimodal model training, the joint training method is adopted to optimize the parameters of 3D-CNN and MLP at the same time. The loss function still uses the cross entropy loss, and the optimizer uses the Adam optimizer. During the training process, the model parameters are continuously adjusted through supervised learning to minimize the loss function value. In order to prevent overfitting, early stopping techniques and regularization methods such as Dropout and L2 regularization are still used.

[0109] Result verification and evaluation

[0110] In this study, the performance of the multimodal model on the test set is as follows:

[0111] Accuracy: 0.8023

[0112] Precision: 1.00 (normal - 1st class), 1.00 (relaxed - 2nd class)

[0113] Recall: 0.74 (normal - 1 class), 0.90 (relaxed - 2 classes)

[0114] F1 Score: 0.90 (Normal - Class 1), 0.86 (Relaxed - Class 2)

[0115] The overall accuracy rate reached 80.23%, indicating that the multimodal fusion model can effectively combine OCT image features and clinical parameter features to achieve high diagnostic accuracy. The F1 scores of both categories are high, indicating that the model performs well and reliably on different categories.

[0116] Visual analysis

[0117] t-SNE (t-Distributed Stochastic Neighbor Embedding) is a method for exploring local structures and similarities of data in high-dimensional space. In image classification tasks, t-SNE can map high-dimensional features extracted by convolutional neural networks to low-dimensional space, and visually observe the relationship between data points through visualization. In this study, 3dresnet18 was used to extract features from the fully connected layers or specific layers of these networks, and t-SNE dimensionality reduction visualization analysis was performed on the semantic features of all images in the test set. The analysis results show that the proposed method can clearly separate the distribution boundaries of relaxed and normal data, further verifying the effectiveness of the method.

[0118] Grad-CAM (Gradient-weighted Class Activation Mapping) is a method for generating a visual heat map of the basis for model decisions, which is particularly suitable for convolutional neural networks (CNN). In the analysis of OCT (optical coherence tomography) images in ophthalmology, Grad-CAM can help explain the model's focus on specific areas, thereby revealing which image areas have the greatest impact on classification decisions. By gradient weighting the last convolutional layer of the CNN and generating the corresponding heat map, the key areas that the model focuses on in the anterior segment OCT images can be intuitively displayed. In this study, the Grad-CAM method successfully generated a heat map for the anterior segment images, clearly showing the areas that the model focused on in the image, providing a valuable reference for further clinical decision-making.

[0119] Algorithms integrated into OCT devices

[0120] The present invention not only includes a multimodal classification model for analyzing OCT images and clinical parameters, but also further integrates the algorithm into the OCT device to achieve full-process automated diagnosis. This section describes the specific method and steps for integrating the algorithm into the OCT device.

[0121] Hardware Interface

[0122] In order to integrate the algorithm into the OCT device, it is first necessary to establish a hardware interface with the OCT device. Through this interface, it is possible to collect raw image data from the OCT device and feed back the processing results to the device. The specific implementation methods include:

[0123] (1) Data acquisition module: The image data of the OCT device is transmitted to the processing unit in real time through a high-speed data interface (such as USB 3.0 or Ethernet interface).

[0124] (2) Processing unit: A computer or embedded system equipped with a high-performance GPU for running deep learning models and multimodal fusion algorithms.

[0125] (3) Result display module: The processed results are fed back to the operator in real time through a display screen or touch screen.

[0126] Software Architecture

[0127] In order to integrate the algorithm into the OCT device, an efficient software architecture needs to be designed. The software architecture mainly includes the following modules:

[0128] (1) Data preprocessing module: responsible for performing preprocessing operations such as denoising, normalization, and segmentation on the image data collected from the OCT device to ensure the quality and consistency of the input data.

[0129] (2) Feature extraction module: Run the 3D-CNN model to extract deep features from the preprocessed images.

[0130] (3) Clinical parameter analysis module: Combine clinical parameters obtained from the electronic medical record system or other input devices for standardization and feature extraction.

[0131] (4) Multimodal fusion module: The image features are fused with the clinical parameter features, classified through a multi-layer perceptron (MLP), and the prediction results of the relaxation state of the suspensory ligament are output.

[0132] (5) Result feedback module: The classification results and diagnostic recommendations are displayed in real time on the screen of the OCT device, and a detailed report is generated for the doctor's reference.

[0133] Integration and Optimization

[0134] After completing the hardware interface and software architecture design, integration and optimization are required to ensure the efficient operation of the algorithm on the OCT device.

[0135] (1) Algorithm optimization: Optimize 3D-CNN and multimodal fusion models, including model compression, pruning, and quantization, to reduce computing resource usage and improve running speed.

[0136] (2) System integration: Integrate the optimized algorithm into the embedded system of the OCT device to ensure seamless connection between hardware and software.

[0137] (3) Testing and verification: The integrated system is tested and verified in an actual clinical environment, including the evaluation of algorithm accuracy, real-time performance, and robustness, to ensure that it meets the requirements for clinical use.

[0138] Figure 1 The complete process of the classification algorithm technology of the present invention is demonstrated. First, the data is collected and labeled by Shenzhen Eye Hospital, and the samples are divided into two states: relaxed and normal. Then, a variety of classification networks are designed, including 3D convolutional neural network (3DCNN) for analyzing images from multiple angles, a binary classification model based on clinical parameters, and a multimodal classification model combining 3DCNN image features and clinical parameter features. Subsequently, appropriate loss functions and evaluation indicators are selected to train the model and optimize parameters. Finally, the optimized model is tested to ensure its effectiveness and reliability in practical applications.

[0139] Figure 2 The model diagram of 3dresnet18 is shown.

[0140] Figure 3The overall framework and implementation process of the multimodal classification model for detecting the relaxation state of the lens suspensory ligament in the present invention are demonstrated. First, the system collects two types of data, OCT images and clinical parameters, of which OCT images are used to capture detailed information of the eye structure, while clinical parameters provide more comprehensive diagnostic information. In the feature extraction and fusion stage, the OCT images are used to extract features related to the relaxation state and the normal state through a deep learning network, and the clinical parameters are also subjected to feature engineering, optimization and algorithm selection through an automated machine learning system to finally obtain the corresponding features. These features are then fused through classification algorithms such as logistic regression, support vector machine, and random forest to improve the accuracy and reliability of the model. In the model evaluation and interpretation stage, methods such as t-SNE and Grad-CAM are used to visualize the decision-making process of the model, and the performance of the model is comprehensively evaluated through indicators such as AUC, F1 score, precision, recall, and accuracy, ensuring the robustness and interpretability of the model, thereby verifying its practicality in clinical diagnosis. Figure 5 and Figure 6 That is, the visualization model of the two methods t-SNE and Grad-CAM.

[0141] Figure 7 The interactive logic flow between the present invention and the OCT device is demonstrated. First, the OCT device captures images of the anterior segment of the eye and transmits these images and diagnosed clinical parameters to the Redis database and stores them in Byte format. These stored OCT image files are then passed to the Django framework, which integrates the diagnostic algorithm for the zonules of the lens. After the algorithm processes the data, it returns the analysis results to the OCT device for visual analysis of the measurement results, ultimately achieving the diagnosis and display of the eye condition.

[0142] Table 1 Parameter comparison of three different types of models

[0143]

[0144] It can be seen that by integrating the multimodal classification algorithm of the present invention into the OCT device, the whole process from image acquisition, processing to result feedback can be automated, greatly improving the efficiency and accuracy of diagnosis. The integrated system can analyze the patient's OCT images and clinical parameters in real time, provide accurate judgment of the relaxation state of the lens suspensory ligament, assist doctors in diagnosis and treatment decisions, and improve the overall level of clinical work and the treatment effect of patients.

[0145] In general, the present invention has the following outstanding advantages and effects:

[0146] (1) Comprehensive analysis of multimodal data: The present invention achieves a comprehensive analysis of the relaxation state of the lens zonules by integrating the OCT image features extracted by 3D-CNN and the machine learning model based on clinical parameters. This multimodal classification method can make full use of the information of image data and clinical data to improve the comprehensiveness and accuracy of diagnosis.

[0147] (2) Improve the accuracy and repeatability of diagnosis: The present invention uses deep learning and machine learning algorithms to automatically extract key features from large amounts of data and perform accurate classification. Through joint training and multimodal fusion, the accuracy and repeatability of diagnosis are significantly improved, the subjective errors of manual judgment are reduced, and a reliable auxiliary diagnostic tool is provided for clinicians.

[0148] (3) Enhance the robustness of the system: The present invention designs efficient data preprocessing and enhancement methods to ensure the stability of the system under different environmental conditions. Through multimodal fusion, the system can effectively cope with data from different sources and qualities, maintaining the consistency and reliability of the diagnosis results.

[0149] (4) Automated full-process diagnosis: The present invention integrates the algorithm into the OCT device to achieve full-process automation from data acquisition, preprocessing, feature extraction, classification to result feedback. This greatly improves the efficiency of diagnosis, reduces the workload of operators, and enhances the practicality and ease of use of the system.

[0150] (5) Strong adaptability: The present invention has strong adaptability and can be applied to different models and brands of OCT equipment, and can process OCT images of different resolutions and qualities. The modular design of the system facilitates integration and expansion, and meets the needs of different clinical environments.

[0151] (6) Significant clinical application effect: Through system verification and clinical trials, the present invention has shown significant clinical application effects in the diagnosis of ophthalmic diseases such as cataracts and glaucoma. The multimodal classification method provides more accurate and reliable diagnostic results, which helps to improve the diagnostic efficiency and accuracy of doctors and improve the treatment effect and quality of life of patients.

[0152] (7) By comparing with traditional diagnostic methods, the significant advantages of the present invention in improving diagnostic accuracy and efficiency were verified.

[0153] In summary, the present invention provides a method and device for detecting occult zonular relaxation associated with anterior chamber angle stenosis or closure disease, and constructs an innovative multimodal classification model by combining OCT image features extracted by 3D-CNN and clinical parameter features based on ensemble learning. After data preprocessing, the 3D-CNN model is used to automatically extract the key features of the SS-OCT image. At the same time, the clinical parameters are accurately classified by training multiple base machine learning models and integrating the results of these models using a voting classifier. Finally, the image features are fused with the clinical parameter features, and by jointly training 3D-CNN and MLP parameters and optimizing the feature combination, an efficient and accurate diagnosis of the state of zonular relaxation of the lens can be achieved on a computer. This method not only improves the comprehensiveness and accuracy of computer-aided diagnosis, but also reduces the subjective errors of manual judgment, providing a powerful auxiliary tool for ophthalmologists.

[0154] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0155] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0156] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0157] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0158] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0159] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0160] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0161] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0162] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0163] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0164] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0165] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0166] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for detecting anterior chamber angle stenosis or closure disease with latent zonular relaxation, characterized in that: The following steps are involved: S1. Preprocessing the collected anterior segment OCT images and clinical parameters of the patient; preferably, the anterior segment OCT images are anterior segment swept frequency optical coherence tomography (SS-OCT) images; S2. constructing a three-dimensional convolutional neural network 3D-CNN model, training the 3D-CNN model using the annotated OCT image data, and automatically extracting image features related to lens suspensory ligament relaxation through the trained model; S3. Select clinical parameter features related to lens zonular relaxation, train multiple base machine learning models, and integrate the results of multiple base models through a voting classifier; S4. Fuse the OCT image features extracted in step S2 with the clinical parameter features obtained in step S3, use the fused features to train a multimodal classification model, and use the multimodal classification model to classify and diagnose the patient's lens suspensory ligament status, wherein the multimodal classification model achieves optimal combination of features and classification decisions by jointly training parameters of 3D-CNN and multi-layer perceptron MLP.

2. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to claim 1, characterized in that: In step S1, the data preprocessing of the anterior segment OCT image includes image denoising, normalization and segmentation, and data enhancement of the anterior segment OCT image is performed, including random rotation, translation and flipping; 16 OCT images of the patient acquired from different viewing angles from 0 to 180 degrees enable the model to learn the three-dimensional structural information of the eye tissue.

3. The method for detecting anterior chamber angle stenosis or closure disease accompanied by latent zonular relaxation according to claim 1 or 2, characterized in that: In step S2, the 3D-CNN model is a 3DResnet-18 model including multiple convolutional layers, pooling layers and fully connected layers. The convolutional layer extracts feature maps of different scales and depths by applying multiple 3D convolution kernels to the input image. The pooling layer reduces the size of the feature map by downsampling while retaining key spatial information. The fully connected layer integrates and classifies the extracted features. The last layer of the model outputs a vector containing the probability of the relaxation state of the suspensory ligament.

4. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation as claimed in claim 3, characterized in that: In the 3DResnet-18 model, the convolution layer is used to perform preliminary feature extraction on the input three-dimensional image, and the spatial and depth features of the image are captured by multiple 3D convolution kernels; the extracted features are normalized by the batch normalization layer to stabilize the learning process and improve the convergence speed of the model; nonlinear activation is introduced by the rectified linear unit layer ReLU to enhance the model's ability to express complex features; the spatial dimension of the feature map is reduced by the pooling layer to achieve feature selection and extract information; By embedding a dropout layer to randomly discard some neurons, the model can be prevented from overfitting. In the deep layer of the model, the features extracted by the previous layer are integrated through the fully connected layer, and the final classification decision is made. The output of the fully connected layer is converted into a probability distribution through the softmax layer to represent the prediction probability of different categories, so as to realize the classification of the relaxation state of the suspensory ligament. Among them, multiple residual blocks are set, each block is responsible for extracting features at different levels, and the features at different levels are fused through jump connections to effectively transfer gradients in the deep network.

5. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to any one of claims 3 to 4, characterized in that: In step S2, the loss function of the 3DResnet-18 model adopts cross entropy loss, and the optimizer selects Adam optimizer to speed up the convergence of the model; during the training process, the model parameters are continuously adjusted through the back propagation algorithm to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting.

6. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to any one of claims 1 to 5, characterized in that: In step S3, the clinical parameter characteristics related to lens zonular relaxation include age (Age), anterior chamber depth (ACD), lens sagittal height (LV), the volume and average value of the angle recess area (ARA750) at 750 mm from the angle of the lens, the radius of curvature of the steepest part of the anterior surface of the lens (F-Rs) and its axial direction, the radius of curvature of the flattest part of the anterior surface of the lens (F-Rf) and its axial direction, lens thickness (Thickness) and lens tilt angle (Tilt).

7. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to any one of claims 1 to 6, characterized in that: In step S3, the multiple base models include support vector machine (SVM), random forest (RandomForest), gradient boosting (Gradient Boosting), naive Bayes (Naive Bayes) and K nearest neighbor (KNN), and the multiple base models are combined through the voting classifier.

8. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to any one of claims 1 to 7, characterized in that: In step S4, the multimodal model performs feature splicing on the OCT image features extracted in step S2 and the clinical parameter features obtained in step S3, and inputs the spliced ​​comprehensive feature vector into the multilayer perceptron MLP, which includes multiple fully connected layers, and achieves classification by training and learning the high-order interaction relationship in the feature vector; in the multimodal model training process, a joint training method is used to simultaneously optimize the parameters of 3D-CNN and MLP, optimize the feature combination, and achieve efficient classification of the relaxation state of the lens suspensory ligament; preferably, the loss function of the multimodal model adopts cross entropy loss, and the optimizer selects Adam optimizer to speed up the convergence of the model; during the training process, the model parameters are continuously adjusted through supervised learning to minimize the loss function value; early stopping technology and regularization methods such as Dropout and L2 regularization are used to prevent overfitting.

9. The method for detecting anterior chamber angle stenosis or closure disease accompanied by occult zonular relaxation according to any one of claims 1 to 8, characterized in that: The following steps are also included: S5. Algorithm visualization processing, including: The t-SNE method is used to reduce the dimensionality of the high-dimensional features extracted from the 3D-CNN model and map the high-dimensional features to a two-dimensional or three-dimensional space so that the relationship between data points can be intuitively observed and analyzed through visualization, and the distribution boundary between slack and normal data can be separated; The Grad-CAM technology is applied to generate a visualized heat map of the model's decision basis. This technology is used to perform gradient weighting on the feature map of the last convolutional layer of CNN to generate a heat map to intuitively display the key areas that the model focuses on on the OCT image.

10. A device for detecting narrow or closed anterior chamber angle disease accompanied by latent zonular relaxation, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the method for detecting anterior chamber angle narrowing or closure disease with occult lens zonular relaxation as described in any one of claims 1 to 9 when executing the computer program; Result display module: Feedback the results processed by the processor to the operator in real time through the display screen.

Images