Cornea conus diagnosis model construction method and system

Abstract

The invention relates to a keratoconus diagnosis model construction method and system, and the method comprises the steps: collecting corneal topographic map data of two eyes, and determining corneal measurement features in the corneal topographic map data; preprocessing to obtain a sample set; dividing a training set and a test set; training the training set; respectively inputting the training set with the convergent loss function into a plurality of single models for training, and fusing the plurality of trained single models to obtain a fusion model; performing hyper-parameter optimization on the fusion model; retraining the fusion model; evaluating performance indexes of the fusion model; and determining a final diagnosis model, obtaining a definite diagnosis symptom, and outputting a diagnosis result. The method has great innovation in the aspects of feature extraction and fusion, and can be used for diagnosing the early screening of the keratoconus; the diagnosis method has high robustness and accuracy, is convenient to use, efficient and convenient, saves time and labor in the diagnosis process, can avoid the misdiagnosis rate, effectively improves the diagnosis and treatment efficiency, and has important clinical value and significance.

Description

Technical Field

[0001] This patent application belongs to the field of biomedical technologies. More specifically, it relates to a method and system for constructing a keratoconus diagnosis model. Background Art

[0002] Keratoconus (KC) is a corneal disease characterized by the thinning and forward conical protrusion of the central or paracentral cornea, often causing high myopia and irregular astigmatism, and significant vision loss in the late stage leading to blindness; some patients will experience acute corneal edema, leaving scars after the edema subsides. The disease usually occurs during puberty, with an incidence of approximately 1.38 / 1000 (varying among different ethnic groups), and the lesion progresses progressively, with severe vision loss in the late stage, seriously affecting the quality of life, production capacity, and career choices of patients, and also bringing a heavy burden to families and society.

[0003] The pathogenesis of keratoconus is not yet clear. Studies have shown that the occurrence of the disease is jointly affected by genetic factors and environmental factors. Currently, domestic and foreign scholars have explored the correlations between genetic factors, environmental factors, and KC, and identified multiple genes related to the occurrence of the disease, such as hepatocyte growth factor (HGF), visual system homeobox 1 (VSX1), zinc finger protein 469 (ZNF469), etc. At the same time, environmental factors such as rubbing the eyes, allergic reactions, and ultraviolet exposure have also been found to be closely related to the occurrence of the disease.

[0004] There is currently no effective radical cure for keratoconus. The current clinical prevention and treatment methods mainly include the prevention and treatment of risk factors, surgical treatment, and non-surgical correction. The main measures for preventing and treating risk factors include intervening in eye rubbing behavior, symptomatic treatment of allergic diseases and systemic diseases, and correcting bad sleeping postures; surgical treatments include implanting corneal stromal rings to improve corrected vision, corneal collagen cross-linking to increase corneal biomechanical strength, and corneal transplantation for severe and corneal scar patients; non-surgical correction improves corrected vision by wearing frame glasses or contact lenses. In the early stage of keratoconus, the corneal irregular astigmatism is not obvious, and frame glasses can be used for correction; in the middle and late stages, the irregular astigmatism worsens, and the correction effect of frame glasses is poor. Contact lenses have become a necessary means to improve vision. By using corneal topography and optical coherence tomography imaging, appropriate lenses can be selected according to the type, position, and size of the cone, effectively improving the best corrected vision, comfort, and visual clarity of the wearer. Therefore, developing an effective keratoconus diagnosis model will provide effective guidance and help for ophthalmologists to formulate appropriate treatment plans for patients.

[0005] Currently, the diagnosis of keratoconus mainly relies on the clinical symptoms, signs of the patients and corneal topography. Cases of advanced keratoconus are relatively easy to diagnose. However, for patients with early-stage keratoconus, due to the inconspicuous symptoms and signs, it is more challenging and difficult. Therefore, a more comprehensive analysis of corneal characteristics is required. Research shows that most of the screening and detection of keratoconus utilize some traditional machine learning algorithms, such as decision trees, support vector machines, artificial neural networks, etc., and achieve the diagnosis of the disease by means of corneal parameters obtained from devices. In recent years, with the continuous development of artificial intelligence technology, its ability to analyze and process complex data has also been continuously enhanced. Especially in the field of deep learning, AlexNet, VGGNet, GoogLeNet, ResNet, DenseNet, and other Convolutional Neural Networks (CNNs) have been used in image classification tasks in many scenarios. Currently, deep learning methods have been widely applied in the diagnosis and screening of ophthalmic-related diseases, such as Diabetic Retinopathy (DR), Age-related Macular Degeneration (AMD), Retinopathy of Prematurity (ROP), and other diseases.

[0006] With the emergence of advanced imaging devices, researchers have also begun to study end-to-end automated keratoconus detection algorithms based on corneal topography using deep learning techniques. However, currently, there are relatively few studies on grading the severity of keratoconus using deep learning. Most of the existing methods directly extract features from a single corneal topographic map using some traditional convolutional neural networks such as VGGNet and ResNet, and then obtain the classification results.

[0007] Currently, these keratoconus diagnosis algorithms developed using deep learning still have some limitations. Most studies separately send single corneal topographic maps into the network for processing to obtain their respective classification results, without considering the relationship between multiple corneal topographic maps. In addition, most researchers still use traditional convolutional neural networks to process corneal topographic maps, without significant innovation in feature extraction and fusion, and the performance in the grading task still needs to be improved. Therefore, developing a more effective new method for early screening and diagnosis of keratoconus is of great significance for improving the early diagnosis level of keratoconus and the prognosis of the disease. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a method and system for constructing a keratoconus diagnosis model, which has significant innovation in feature extraction and fusion, can effectively diagnose the early screening of keratoconus, and is more adaptable to the development of modern medicine.

[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0010] A method for constructing a keratoconus diagnosis model, comprising the following steps:

[0011] S1. Collect corneal topographic data of both eyes, and determine corneal measurement features in the corneal topographic data;

[0012] S2. Preprocess the corneal measurement features to obtain a sample set;

[0013] S3. Divide the sample set into a training set and a test set according to a ratio;

[0014] S4. Train the training set until the loss function converges;

[0015] S5. Input the training set with the converged loss function into multiple single models for training respectively, and fuse the trained multiple single models to obtain a Stacking fusion model;

[0016] S6. Use an optimization algorithm to perform hyperparameter optimization on the Stacking fusion model;

[0017] S7. Use the Shapley value for feature filtering, leave the most important multiple features, and label the above features;

[0018] S8. Retrain the Stacking fusion model using the training set;

[0019] S9. Test the retrained Stacking fusion model under the test set, output the results after testing and compare them with the results of the single models to evaluate the performance indicators of the Stacking fusion model;

[0020] S10. The Stacking fusion model after passing the evaluation is used as the final diagnosis model to obtain the confirmed symptoms. A further improvement of the solution of the present invention lies in: "preprocessing the corneal measurement features" in S2 includes

[0021] S21. Missing index processing: Delete the features in the corneal measurement features where the missing index reaches more than half;

[0022] S22. Data filling: Fill the features with less than half of the missing indexes in the initial data set by using the mode or the average to form a data set;

[0023] S23. Outlier removal: Use data inspection to remove the data in the data set that deviates from the average value by more than three standard deviations;

[0024] S24. Data equalization: Use the SMOTE sampling method to equalize the original data.

[0025] A further improvement of the solution of the present invention is that in S3, the sample set is divided into a training set and a test set according to a ratio of 7:3 to 6:4, where the training set is 60% - 70% and the test set is 30% - 40%.

[0026] A further improvement of the solution of the present invention is that in S4, the loss function includes a classification task loss function (entropy and cross - entropy loss, softmax loss, KL divergence, Hinge loss), a regression task loss function (L1 loss, L2 loss, perceptual loss), and a generative adversarial network loss function (LS - GAN, Loss - sensitive GAN).

[0027] A further improvement of the solution of the present invention is that in S5, there are five single models, namely bidirectional LSTM, 3D - CNN, random forest RFC, XGBOOST, and GBDT; the fusion adopts the Voting soft voting method.

[0028] A further improvement of the solution of the present invention is that in S6, the optimization algorithms used for hyperparameter optimization are Gradient Descent, Stochastic Gradient Descent (SGD), Adagrad, RMSprop, or Adam. These algorithms have different update rules, and appropriate algorithms can be selected according to different problems and requirements.

[0029] A further improvement of the solution of the present invention is that in S7, the seven most important features are: the thickness of the thinnest point of the cornea, the anterior surface height of the thinnest point of the cornea, the posterior surface height of the thinnest point of the cornea, the deviation of the anterior surface height difference map, the deviation of the posterior surface height difference map, the deviation of the average thickness progression, and the minimum thickness deviation.

[0030] A keratoconus diagnosis model construction system that utilizes the method according to any one of claims 1 - 7, comprising the following modules:

[0031] A pre - processing module, used to obtain corneal topographic data of both eyes, determine corneal measurement features, perform pre - processing, and obtain a sample set;

[0032] A sample set division and training module, used to divide the training set and the test set, and train the training set until the loss function converges;

[0033] The fusion training module is used to input the training sets with converged loss functions into multiple single models respectively for training and fuse them to obtain a Stacking fusion model;

[0034] The hyperparameter optimization module is used to optimize the hyperparameters of the Stacking fusion model using an optimization algorithm;

[0035] The Shap feature filtering module is used to perform feature filtering using the Shapley value and retain the most important multiple features;

[0036] The retraining module is used to retrain the Stacking fusion model;

[0037] The performance evaluation module further tests the retrained Stacking fusion model and evaluates the performance metrics of the Stacking fusion model to form a final diagnostic model;

[0038] The symptom diagnosis module is used to obtain the diagnosed symptoms;

[0039] The above-mentioned preprocessing module, sample set division and training module, fusion training module, hyperparameter optimization module, Shap feature filtering module, retraining module, performance evaluation module, and symptom diagnosis module are connected with information in sequence.

[0040] A further improvement of the solution of the present invention is that the symptom diagnosis module can perform visual output, and the visual output includes text, pictures, videos, charts, and animation demonstrations.

[0041] A further improvement of the solution of the present invention is that the data visualization tool used for visual output is Echarts or Matplotlib.

[0042] Due to the adoption of the above technical solution, the beneficial effects obtained by the present invention are:

[0043] 1) Starting from the fusion of multiple single models, the present invention combines the advantages and disadvantages of various single models, makes multi-dimensional comprehensive judgments on the incidence of keratoconus in patients as a unit, combines binocular data, includes both manually selected features and features learned by the deep network from big data by itself, and the diagnostic method has stronger robustness and accuracy.

[0044] 2) The present invention can associate corneal images at different times, so that the neural network model during training can fully learn the dynamic change process of the state reflected by keratoconus images in the time dimension, thereby improving the learning effect of the model and the diagnostic accuracy of the model for early keratoconus.

[0045] 3) There are significant innovations in feature extraction and fusion, which can effectively diagnose the early screening of keratoconus and better adapt to the development of modern medicine.

[0046] 4) The present invention is convenient to use, efficient and convenient. The diagnostic process not only saves time and effort, but also can avoid the occurrence of misdiagnosis rate, effectively improving the diagnostic and treatment efficiency, and has important clinical value and significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of the method of the present invention.

[0048] Figure 2 It is a connection block diagram of the system of the present invention.

[0049] Figure 3 It is a process diagram for preprocessing corneal measurement features in the method of the present invention.

[0050] Figure 4 It is a usage process diagram of the system of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] It should be understood that when used in this specification and the appended claims, the terms "including" and "comprising" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0053] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in this specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0054] It should be further understood that the term " / and / " used in this specification of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0055] Next, the present invention will be further described in detail in conjunction with the embodiments.

[0056] A method for constructing a keratoconus diagnosis model, asFigure 1 , including the following steps:

[0057] S1. Collect corneal topographic data of both eyes, and determine corneal measurement features in the corneal topographic data, which generally include eye data such as the thickness of the thinnest point of the cornea, the anterior surface height of the thinnest point of the cornea, the posterior surface height of the thinnest point of the cornea, the average thickness progression index, Ambrosio-related thickness, the deviation of the anterior surface height difference map, the deviation of the posterior surface height difference map, the average thickness progression deviation, the minimum thickness deviation, the ARTmax deviation, etc., and also include more than a dozen or even dozens of corneal measurement features such as time, region, age, gender, family medical history, etc. The overall classification of the severity of the diagnosis results is divided into five categories: extremely severe, severe, moderate, mild, and observation degree. Mild means there are already mild symptoms, and the observation degree means there are no specific symptoms yet, but it is at the critical value edge.

[0058] S2. Preprocess the corneal measurement features to obtain a sample set;

[0059] Combined with Figure 3 , "preprocessing the corneal measurement features" includes the following steps:

[0060] S21. Missing index processing: Delete the features in the corneal measurement features where the missing index reaches more than half.

[0061] S22. Data filling: Fill the features with less than half of the missing indexes in the initial data set by using the mode or average method to form a data set;

[0062] S23. Outlier removal: Use data testing to remove the data in the data set that differs from the average value by more than three standard deviations;

[0063] S24. Data balancing: Use the SMOTE sampling method to balance the original data.

[0064] After the above steps, the data is cleaned to obtain the required data, which is convenient for the subsequent steps.

[0065] S3. Divide the sample set into a training set and a test set according to a ratio;

[0066] In this step, divide the sample set into a training set and a test set according to a ratio of 7:3 to 6:4. For example, the training set is 60% - 70% and the test set is 30% - 40%. This ratio is more appropriate and will not result in too few training sets or too high test sets.

[0067] S4. Train the training set until the loss function converges;

[0068] In this step, the loss function includes the loss functions for classification tasks (entropy and cross-entropy loss, softmax loss, KL divergence, Hinge loss), regression tasks (L1 loss, L2 loss, perceptual loss), and generative adversarial network loss functions (LS-GAN, Loss-sensitive GAN).

[0069] For entropy and cross-entropy loss, entropy is used to measure the uncertainty of a random variable, while cross-entropy is used to measure the similarity between two probability distributions.

[0070] Entropy is a fundamental concept in information theory, which represents the uncertainty of a random variable. In machine learning, entropy is often used as part of the loss function, especially in classification problems. For example, in a binary classification problem, if we have a random variable X that takes values 0 or 1 with corresponding probabilities p and 1 - p respectively, then the entropy H(X) of X can be expressed as: H(X) = -p * log2(p) - (1 - p) * log2(1 - p).

[0071] Cross-Entropy is also an important concept in information theory, which measures the difference between two probability distributions. In machine learning, cross-entropy is often used as a loss function, especially when training classification models. Suppose we have a true target probability distribution P and a predicted probability distribution Q, then the cross-entropy between these two distributions can be expressed as: H(P, Q) = -∑P(i) * log(Q(i)).

[0072] It should be noted that cross-entropy is not a symmetric metric, that is, H(P, Q) is not equal to H(Q, P). In machine learning, we usually use cross-entropy as the loss function to optimize the model parameters so that the predicted distribution Q is as close as possible to the true distribution P.

[0073] Taking KL divergence as another example, also known as relative entropy, it is a method to describe the difference between two probability distributions P and Q. KL divergence is asymmetric and satisfies the triangle inequality.

[0074] Other types of loss functions will not be elaborated further.

[0075] S5. Input the training sets for which the loss function converges into multiple single models respectively for training, and fuse the trained multiple single models to obtain a Stacking fusion model;

[0076] In this step, there are five single models, namely bidirectional LSTM, 3D-CNN, random forest RFC, XGBOOST, and GBDT. The performances of these five single models are all excellent, but they still have their own deficiencies and disadvantages. Therefore, they are fused to make the overall effect after fusion better and improve the diagnosis rate.

[0077] The fusion adopts the Voting soft voting method. The voting method is a combination strategy for classification problems in ensemble learning. It is an ensemble learning model that follows the principle of the minority obeying the majority. By integrating multiple models, the variance is reduced, thereby improving the robustness of the model (how tolerant the algorithm is to data changes). In an ideal situation, the prediction effect of the voting method will be better than that of any single base model.

[0078] Taking the Voting soft voting method as an example, for a certain sample:

[0079] The prediction result of Model 1 is that the probability of class A is 99%;

[0080] The prediction result of Model 2 is that the probability of class A is 49%;

[0081] The prediction result of Model 3 is that the probability of class A is 59%;

[0082] Finally, the average of the prediction probabilities for class A is (99 + 49 + 59) / 3 = 69, so the prediction result is class A.

[0083] S6. Use an optimization algorithm to perform hyperparameter tuning on the Stacking fusion model;

[0084] Since the Stacking fusion model can be obtained from the previous step, in order to verify the performance and performance of the Stacking fusion model, such as accuracy, ROC, etc., it is necessary to perform tuning on it. Combining the hyperparameter tuning of the optimization algorithm can improve the overall performance of the fusion model. In this step, the optimization algorithms used for hyperparameter tuning are Gradient Descent, Stochastic Gradient Descent (SGD), Adagrad, RMSprop, or Adam. These algorithms all have different update rules, and the appropriate algorithm can be selected according to different problems and requirements, without specific requirements. Of course, it can also be other optimization algorithms.

[0085] S7. Use the Shapley value to perform feature filtering, leave the most important multiple features, and label the above features;

[0086] In this step, after feature filtering, the seven most important features left are: the thickness of the thinnest point of the cornea, the anterior surface height of the thinnest point of the cornea, the posterior surface height of the thinnest point of the cornea, the deviation of the anterior surface height difference map, the deviation of the posterior surface height difference map, the deviation of the average thickness progression, and the minimum thickness deviation. Originally there were ten to dozens of features, and this step can select the best from the best, so as to select the most important features as the key basis for diagnosis and judgment.

[0087] S8. Retrain the Stacking fusion model using the training set;

[0088] S9. Test the retrained Stacking fusion model under the test set, output the results after testing and compare them with the results of the single model to evaluate the performance indicators of the Stacking fusion model;

[0089] S10. The qualified Stacking fusion model is used as the final diagnosis model to obtain the confirmed symptoms.

[0090] Compare the prediction effect of the Stacking fusion model with that of bidirectional LSTM, 3D-CNN, random forest RFC, XGBOOST, and GBDT. The scores of each model under the evaluation indicators are shown in Table 1:

[0091] Table 1 Algorithm comparison

[0092]

[0093] It can be seen from the simulation results that the performance of the model of the present invention has been improved in each performance index compared with other basic models, and the accuracy rate, precision rate, recall rate, and f1 score have reached 0.89, 0.91, 0.94, and 0.95 respectively. It can be seen that each parameter is relatively ideal.

[0094] Example: Select a 25-year-old female patient with no previous medical history and keratoconus found during ophthalmic screening. Input her eye data into the model before and after fusion. The diagnosis result of the model before fusion is extremely severe, and the diagnosis result of the model after fusion is severe. After expert consultation, it is determined that it is a severe case, proving that the model after fusion is more accurate.

[0095] Subsequently, another 100 patients were selected, regardless of age, gender, region, and previous medical history, and verified one by one. After statistics, the accuracy rate of the model after fusion is as high as over 90%, reaching a satisfactory accuracy rate, and it can be promoted in hospitals, medical institutions, medical approval institutions, etc.

[0096] In addition to the above-mentioned method for constructing a keratoconus diagnosis model, the present invention also discloses a system for constructing a keratoconus diagnosis model, see Figure 2: including

[0097] A preprocessing module, which is used to obtain corneal topographic data of both eyes, determine corneal measurement features, perform preprocessing, and obtain a sample set;

[0098] A sample set division and training module, which is used to divide the training set and the test set, and train the training set until the loss function converges;

[0099] A fusion training module, which is used to input the training set with the converged loss function into multiple single models for training and fuse them to obtain a Stacking fusion model;

[0100] A hyperparameter optimization module, which is used to optimize the hyperparameters of the Stacking fusion model using an optimization algorithm;

[0101] A Shap feature filtering module, which is used to perform feature filtering using the Shapley value and retain multiple most important features;

[0102] A retraining module, which is used to retrain the Stacking fusion model;

[0103] A performance evaluation module, which further tests the retrained Stacking fusion model and evaluates the performance metrics of the Stacking fusion model to form a final diagnosis model;

[0104] A symptom diagnosis module, which is used to obtain the diagnosed symptoms;

[0105] The above-mentioned preprocessing module, sample set division and training module, fusion training module, hyperparameter optimization module, Shap feature filtering module, retraining module, performance evaluation module, and symptom diagnosis module are sequentially connected by information.

[0106] The symptom diagnosis module can perform visual output, and the visual output includes text, pictures, videos, charts, and animation demonstrations. The data visualization tool used for visual output is Echarts or Matplotlib.

[0107] The visual outputs supported by the symptom diagnosis module include text, pictures, videos, charts, and animation demonstrations. For example, pictures or videos of the lesion progression can be output to help patients understand and analyze the data. Data visualization tools such as Echarts or Matplotlib can also be used to generate intuitive and shocking charts, such as line charts and bar charts, to display information such as lesions, colors, textures, and later developments. In addition, Web application display can be implemented: A simple Web application can be built using the Flask framework to present the data analysis results to users. Users can obtain more personalized services through the login and registration functions, including functions such as saving the hospitals and doctors they are concerned about, setting reminders for expert consultations, etc., and can log in using WeChat mini-programs for system push, etc.

[0108] Interactive visualization can also be supported. By using visualization libraries (such as Matplotlib, Plotly, etc.), future rehabilitation data can be displayed, and interactive charts can be provided for patients to explore the data. Generate rehabilitation prediction charts for the next few days / weeks / months / years to intuitively display the prediction results and reduce the psychological burden of patients.

[0109] The visual user interface of the present invention is user-friendly. A simple and clear user interface is designed to facilitate users to query disease information and view prediction results, and supports custom queries, allowing users to query medical treatment data and predictions according to hospitals and dates.

[0110] In addition, the present invention also provides a computer device, which includes a memory and a processor. A computer program is stored on the memory, and when the processor executes the computer program, the method described in any one of claims 1 to 7 is implemented.

[0111] The processor obtains the binocular data of each ophthalmology hospital department in a sub-category statistical manner; performs data cleaning on more than a dozen or dozens of corneal measurement characteristics of the diseased eye, and conducts extreme value quality control and time consistency quality control.

[0112] The processor can be a Central Processing Unit (CPU), and this processor can also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or this processor can also be any conventional processor, etc.

[0113] Those of ordinary skill in the art can understand that all or part of the processes in the methods of implementing the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0114] Meanwhile, the present invention also provides a storage medium, which stores a computer program. When the computer program is executed by a processor, the above methods can be implemented. When the computer program is executed by the processor, the processor executes the following steps:

[0115] Combined with Figure 4 , it shows the usage process diagram of the system of the present invention. Based on the corneal topographic data of both eyes, corneal measurement features are obtained according to the corneal topographic data, and a sample set is obtained after preprocessing; the training set in the sample set is subjected to loss function convergence. After the convergence meets the requirements, multiple single models are fused to obtain a Stacking fusion model. After the hyperparameters of the Stacking fusion model are optimized by an optimization algorithm, the most important multiple features are left. After the performance indicators of the Stacking fusion model are evaluated as qualified, the final diagnosis model is determined. When medical staff actually use it or during routine examinations, they only need to input the corneal pictures, videos or parameters of both eyes into the final diagnosis model for disease diagnosis and analysis. After the program runs, the keratoconus score and diagnosis result can be obtained, and the diagnosis result is output to the terminal for medical staff to refer to. According to the keratoconus score, it can be divided into five grades: extremely severe, severe, moderate, mild, and observation degree. For example, 90-100 points is extremely severe, 80-90 is severe, less than 5 points is the observation degree, and the moderate and mild degrees can be assigned scores as appropriate.

[0116] The above diagnosis process not only saves time and effort, but also avoids the occurrence of misdiagnosis rate, improves the diagnosis and treatment efficiency, has important clinical value and significance, can effectively diagnose the early screening of keratoconus, and is more adaptable to the development of modern medicine.

[0117] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of each unit is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0118] The steps in the method of the embodiments of the present invention can be adjusted in order, combined, and deleted according to actual needs. The units in the device of the embodiments of the present invention can be combined, divided, and deleted according to actual needs. In addition, in each embodiment of the present invention, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0119] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present invention.

[0120] The method and system for constructing a keratoconus diagnosis model provided by the present invention have been introduced in detail above. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for constructing a keratoconus diagnostic model, characterized in that The steps include: S1. collecting corneal topography data of both eyes, and determining corneal measurement features in the corneal topography data; S2, preprocessing corneal measurement features to obtain a sample set; S3, divide the sample set into training set and test set according to the proportion; S4, train the training set until the loss function converges; S5, inputting the training sets with converged loss functions into multiple single models for training respectively, and fusing the trained multiple single models to obtain a Stacking fusion model; S6. Use optimization algorithms to optimize the hyperparameters of the Stacking fusion model; S7. Use Shapley value to filter features, keep the most important features, and mark the above features; S8. Retrain the Stacking fusion model using the training set; S9. Test the retrained Stacking fusion model on the test set, output the test results and compare them with the results of the single model to evaluate the performance indicators of the Stacking fusion model; S10. The Stacking fusion model that has passed the evaluation is used as the final diagnostic model to obtain the confirmed symptoms.

2. A method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: "Preprocessing of corneal measurement features" in S2 includes S21, missing index processing: delete the features with more than half of the missing indexes in the corneal measurement features; S22, data filling: For the features with less than half of the missing indicators in the initial data set, fill them by mode or average to form a data set; S23, outlier elimination: using data testing, the data in the data set that differs from the mean by more than three times the standard deviation are eliminated; S24, data equalization: Use the SMOTE sampling method to perform data equalization on the original data.

3. The method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: In S3, the training set is 60% to 70% and the test set is 30% to 40%.

4. The method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: In S4, the loss function is a classification task loss function, a regression task loss function, or a generative adversarial network loss function.

5. The method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: In S5, there are five single models, namely bidirectional LSTM, 3D-CNN, random forest RFC, XGBOOST, and GBDT; the fusion adopts the Voting soft voting method.

6. The method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: In S6, the optimization algorithm used for hyperparameter optimization is gradient descent, stochastic gradient descent, dynamic gradient descent, RMSprop or Adam.

7. The method for constructing a keratoconus diagnostic model according to claim 1, characterized in that: In S7, there are seven most important features, namely: thickness at the thinnest point of the cornea, height of the anterior surface at the thinnest point of the cornea, height of the posterior surface at the thinnest point of the cornea, deviation of the anterior surface height difference map, deviation of the posterior surface height difference map, deviation of the average thickness progression, and deviation of the minimum thickness.

8. A keratoconus diagnostic model construction system, using the method according to any one of claims 1 to 7, characterized in that include: A preprocessing module, used to obtain corneal topography data of both eyes, determine corneal measurement features, perform preprocessing, and obtain a sample set; The sample set partitioning training module is used to partition the training set and the test set, and train the training set until the loss function converges; The fusion training module is used to input the training sets with converged loss functions into multiple single models for training and fuse them to obtain a Stacking fusion model; Hyperparameter optimization module, used to optimize the hyperparameters of the Stacking fusion model using optimization algorithms; Shap feature filtering module, used to filter features using Shapley values, leaving the most important features; Retraining module, used to retrain the Stacking fusion model; The performance evaluation module further tests the retrained Stacking fusion model and evaluates the performance indicators of the Stacking fusion model to form the final diagnosis model; Symptom diagnosis module, used to obtain confirmed symptoms; The above-mentioned preprocessing module, sample set partitioning training module, fusion training module, hyperparameter optimization module, Shap feature filtering module, retraining module, performance evaluation module, and symptom diagnosis module are sequentially connected.

9. A keratoconus diagnosis model construction system according to claim 8, characterized in that: The symptom diagnosis module can perform visual output, including text, pictures, videos, charts, and animation demonstrations.

10. A keratoconus diagnosis model construction system according to claim 9, characterized in that: The data visualization tools used for visualization output are Echarts or Matplotlib.

Images