AI model construction method for diagnosing keratoconus
By processing Corvis ST examination images using incremental digital image correlation methods and machine learning models, internal corneal information is obtained, and an AI model is established. This solves the problem that existing technologies cannot comprehensively assess corneal biomechanical properties, thereby improving the diagnostic accuracy and treatment efficacy of keratoconus.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE EYE HOSPITAL OF WENZHOU MEDICAL UNIVERSITY
- Filing Date
- 2023-09-20
- Publication Date
- 2026-04-17
AI Technical Summary
Current technology cannot obtain information about the inside of the cornea in vivo, which makes it impossible to fully assess the biomechanical properties of the cornea and affects the diagnosis and treatment of keratoconus.
Incremental digital image correlation was used to process Corvis ST examination images to obtain displacement and strain data inside the cornea. This data was then combined with a machine learning model for classification, and an AI model was established to diagnose keratoconus.
It enables comprehensive analysis of corneal biomechanical information, improves the diagnostic accuracy of keratoconus, and enhances the treatment process.
Smart Images
Figure CN121885147A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image-assisted diagnosis technology, and more specifically, to a method for constructing an AI model for diagnosing keratoconus. Background Technology
[0002] The cornea is a unique, aspherical, biological soft tissue. To maintain its normal biological functions and capabilities, its morphology needs to provide sufficient rigidity to resist occasional external injuries and withstand intraocular pressure (IOP). The histological structure of the cornea reveals its complex viscoelastic biomechanical properties, playing a crucial role in mechanical responses (e.g., creep, stress relaxation, and hysteresis).
[0003] Early studies demonstrated that collagen fibers are the primary load-bearing component and concluded that there is a close relationship between corneal biomechanical behavior and the content and distribution of stromal fibers. However, over time, the overall biomechanical behavior of the cornea changes at the cellular, tissue, and suborgan levels, which may lead to serious corneal diseases.
[0004] Among these corneal diseases, keratoconus (KC) is a primary degenerative disease with a prevalence of 20.6‰ in men and 18.33‰ in women. It is the most common ectopic corneal disease, with a prevalence of approximately 1 / 2000, and is one of the most common causes of corneal transplantation. Common clinical manifestations of KC include high irregular myopic astigmatism and varying degrees of visual impairment, which severely impacts patients' vision and quality of life. Current research suggests that in some areas of the KC stroma, there is a reduction in collagen fiber cross-linking, disordered fiber arrangement, and extracellular matrix degradation, leading to weakened local mechanical properties. Under intraocular pressure, the cornea gradually bulges outward, resulting in localized thinning and steepening, which is a significant cause of KC progression. The occurrence and development of KC are closely related to changes in corneal biomechanical properties.
[0005] During diagnosis and subsequent treatment, it is necessary to understand the internal information of the cornea in real time. However, current methods can only obtain information about the surface of the cornea in vivo, but cannot obtain information about the internal cornea in vivo. Therefore, it is not possible to comprehensively assess the biomechanical properties of the cornea in vivo. Thus, it is necessary to improve auxiliary diagnostic techniques to improve the subsequent treatment process. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the present invention aims to provide an AI model construction method for diagnosing keratoconus. This AI model construction method is based on the displacement and strain inside the cornea obtained through the incremental DIC method, and uses an artificial intelligence model to assist in the diagnosis of keratoconus.
[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI model construction method for diagnosing keratoconus, comprising the following steps:
[0008] S1, Obtain the patient's Corvis ST examination images;
[0009] S2, inspects images by processing them using incremental digital image correlation methods;
[0010] S3, obtains corneal displacement data, strain data, velocity data and strain rate data in the horizontal direction;
[0011] S4, obtains corneal displacement data, strain data, velocity data and strain rate data in the vertical direction;
[0012] S5. Establish an AI model, train the model based on the displacement, strain, velocity and strain rate data of the cornea in the horizontal direction and the displacement, strain, velocity and strain rate data of the cornea in the vertical direction, and use the model to classify normal cornea and keratoconus.
[0013] In summary, this invention offers the following advantages: It uses incremental digital image correlation (DIR) methods to process examination images of normal and patient corneas (Corvis ST), obtaining corresponding displacement and strain information to establish an AI classification model. Then, the AI model is trained and tested using the corneal biomechanical property information obtained in the previous step to classify healthy corneas, atrophic keratoconus, and keratoconus. This AI model employs a multi-modal approach, possessing excellent diagnostic efficacy and capable of analyzing comprehensive corneal biomechanical information, encompassing both the internal and surface layers of the cornea. Attached Figure Description
[0014] Figure 1 A schematic diagram illustrating the steps involved in building an AI model. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] Reference Figure 1 As shown, to achieve the above objectives, the present invention provides the following technical solution: an AI model construction method for diagnosing keratoconus, comprising the following steps: S1, acquiring the patient's Corvis ST examination image;
[0017] S2, inspects images by processing them using incremental digital image correlation methods;
[0018] S3, obtains corneal displacement data, strain data, velocity data and strain rate data in the horizontal direction;
[0019] S4, obtains corneal displacement data, strain data, velocity data and strain rate data in the vertical direction;
[0020] S5. Establish an AI model, train the model based on the displacement, strain, velocity and strain rate data of the cornea in the horizontal direction and the displacement, strain, velocity and strain rate data of the cornea in the vertical direction, and use the model to classify normal cornea and keratoconus.
[0021] This invention uses incremental digital image correlation (DIR) to process examination images of normal and patient corneas (Corvis ST), obtaining corresponding displacement and strain information to establish an AI classification model. The AI model is then trained and tested using the corneal biomechanical property information obtained in the previous step to classify healthy corneas, atrophic keratoconus, and keratoconus. This AI model employs a multi-modal approach, possessing excellent diagnostic efficacy and capable of analyzing comprehensive corneal biomechanical information, including both internal and surface corneal data.
[0022] Macroscopic mechanical response of the cornea was determined through image analysis. The rapid evolution of corneal displacement, strain, velocity, and strain rate across the entire field of view was reconstructed using incremental digital image correlation (DIC) method.
[0023] The incremental DIC method was combined with the Corvis ST tonometer for in vivo high-speed corneal deformation measurement. By using image registration through incremental DIC analysis, spatiotemporal dynamic strain / strain rate maps of the cornea can be estimated at the tissue level.
[0024] Step S5 specifically includes:
[0025] S51, use a machine learning classification model to classify normal cornea and keratoconus, and calculate its classification accuracy on the test set;
[0026] Specifically, common machine learning classification models can be used for classification operations.
[0027] S52 integrates the machine learning models using the Voting Classifier to obtain a comprehensive classification result, and calculates its classification accuracy on the test set.
[0028] The machine learning classification models include Decision Tree, Random Forest, Support Vector Machine, Naive Bayes, KNN, Neural Network, XGBoost, and LightGBM.
[0029] The data acquisition method in step S5 includes:
[0030] B1, parameters obtained from the results of processing Corvis ST images using the incremental DIC method: time, max-v, max-exy, max-vr, max-exyr, ave-v, ave-exy, ave-vr, ave-exyr;
[0031] B2 uses the parameters of the normal corneal group, the abrupt keratoconus group, and the keratoconus group as input data.
[0032] Then (1). Define a function: a function to read data from a file;
[0033] (2). Define functions: Functions used to preprocess data and labels;
[0034] 1) Merge all data into one array
[0035] 2) Convert labels to numbers
[0036] (3). Define the data path and labels
[0037] directory_paths=['sample_data / ffkc','sample_data / kc','sample_data / normal']
[0038] indexed_labels=['ffkc','kc','normal']
[0039] (4) Reading data from a file
[0040] (5) Preprocessing data and labels
[0041] The data processing method in step S5 includes:
[0042] B3 splits the input data into a training set and a test set;
[0043] B4, initialize Decision Tree, Random Forest, SVM, Naive Bayes, KNN, Neural Network, XGBoost, and LightGBM models;
[0044] B5 creates a voting classifier based on several individual models with high classification accuracy, and trains both the individual models and the voting classifier while performing 10-fold cross-validation.
[0045] Finally, the accuracy of the individual model and the accuracy of the voting classifier are output.
[0046] The machine learning models mentioned above can achieve an accuracy of over 80% in distinguishing between normal cornea and atrophic keratoconus, with the best results reaching 95%. The voting classifier can achieve an accuracy of 90%.
[0047] In step S1, dynamic images of the cornea are recorded on a horizontal cross-section generated by an air thruster using a high-speed Scheimpflug camera with a Corvis ST tonometer.
[0048] Corvis ST can acquire surface biomechanical information of the cornea in vivo, and evaluate the biomechanical properties of the cornea through the obtained parameters SSI or SSI 2, thereby assisting in the diagnosis of keratoconus.
[0049] In summary, by combining the Corvis ST and incremental DIC methods to obtain intracorneal displacement and strain, an artificial intelligence model is used to assist in the diagnosis of keratoconus. Continuous training and testing of the AI model results in a final model with excellent diagnostic efficacy. This model can analyze more comprehensive corneal biomechanical information, including both the internal and surface layers, thus effectively improving the auxiliary diagnostic technique and consequently improving the subsequent treatment process.
[0050] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for constructing an AI model for diagnosing keratoconus, characterized by: The steps include: S1, obtaining Corvis ST examination images of the patient; S2, inspects images by processing them using incremental digital image correlation methods; S3, obtains corneal displacement data, strain data, velocity data and strain rate data in the horizontal direction; S4, obtains corneal displacement data, strain data, velocity data and strain rate data in the vertical direction; S5. Establish an AI model, train the model based on the displacement, strain, velocity and strain rate data of the cornea in the horizontal direction and the displacement, strain, velocity and strain rate data of the cornea in the vertical direction, and use the model to classify normal cornea and keratoconus.
2. The method for constructing an AI model for diagnosing keratoconus according to claim 1, characterized in that: Step S5 specifically includes: S51, use a machine learning classification model to classify normal cornea and keratoconus, and calculate its classification accuracy on the test set; S52 integrates the machine learning models using the Voting Classifier to obtain a comprehensive classification result, and calculates its classification accuracy on the test set.
3. The method for constructing an AI model for diagnosing keratoconus according to claim 2, characterized in that: The machine learning classification models include Decision Tree, Random Forest, Support Vector Machine, Naive Bayes, KNN, Neural Network, XGBoost, and LightGBM.
4. The method for constructing an AI model for diagnosing keratoconus according to claim 2, characterized in that: The data acquisition method in step S5 includes: B1, parameters obtained from the results of processing Corvis ST images using the incremental DIC method: time, max-v, max-exy, max-vr, max-exyr, ave-v, ave-exy, ave-vr, ave-exyr; B2 uses the parameters of the normal corneal group, the abrupt keratoconus group, and the keratoconus group as input data.
5. The method for constructing an AI model for diagnosing keratoconus according to claim 4, characterized in that: The data processing method in step S5 includes: B3 splits the input data into a training set and a test set; B4, initialize Decision Tree, Random Forest, SVM, Naive Bayes, KNN, Neural Network, XGBoost, and LightGBM models; B5 creates a voting classifier based on several individual models with high classification accuracy, and trains both the individual models and the voting classifier while performing 10-fold cross-validation.
6. The method for constructing an AI model for diagnosing keratoconus according to claim 1, characterized in that: In step S1, dynamic images of the cornea are recorded on a horizontal cross-section generated by an air thruster using a high-speed Scheimpflug camera with a Corvis ST tonometer.