Artificial intelligence algorithm for myopia risk prediction and clinical application thereof

Abstract

The invention belongs to the technical field of myopia risk prediction, and particularly relates to an artificial intelligence algorithm for myopia risk prediction and clinical application thereof, and the algorithm comprises the following steps: S1, strain field calculation; s2, carrying out strain prediction; and S3, carrying out binary classification prediction based on risk levels. According to the method, artificial intelligence and a mechanical algorithm are combined, the AI model capable of non-invasively and conveniently evaluating the sclera biomechanical characteristics is innovatively developed, the sclera biomechanical characteristics can be evaluated only by performing conventional imaging on the posterior portion of the eye, and the accuracy and convenience of myopia risk prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of myopia risk prediction, and particularly relates to an artificial intelligence algorithm for myopia risk prediction and its clinical application. Background Art

[0002] In terms of eye structure, the most relevant change associated with the occurrence of myopia is the outward expansion of the posterior pole of the eyeball, which will directly lead to a mismatch between the lengthened eye axis and refraction. The light entering the eye is focused in front of the retina, resulting in blurred vision, that is, myopia. A large number of basic and clinical studies have shown that the changes in the macroscopic biomechanical properties of the sclera and the remodeling of the extracellular matrix and the transformation of cell types at the microscopic level are the main reasons for the elongation of the eye axis. A review article named "Scleral structureandbiomechanics" pointed out that the in vivo measurement of scleral biomechanical properties can be achieved by inverse calculation methods. First, one of the known loads acting on the sclera needs to be changed, while continuously monitoring the scleral tissue and measuring the resulting local deformation. Only in this way can the stiffness (or biomechanical properties) of the sclera be estimated, that is, the ratio of load change to deformation. However, these "in vivo biomechanical property tests" need to be carried out within a safe physiological range and strictly control variables, that is, only one load can be changed at a time, while other loads should remain unchanged. GIRARD MJ et al. developed a three-dimensional tracking algorithm named "In vivo opticnerve headbiomechanics:performance testingofathree-dimensional tracking algorithm", which can track the displacement and strain of the posterior sclera after the change of intraocular pressure. This algorithm scans the volume of the target position using optical coherence tomography (OCT) technology before and after the change of intraocular pressure, then performs mechanical conversion on it, and optimally matches the deformed OCT volume after the change of intraocular pressure to output a three-dimensional displacement field, from which the tensile, compressive and effective strain components are deduced. ZHONG F et al. proposed a high-accuracy and high-efficiency digital volume correlation (DVC) method named "Ahigh-accuracy and highefficiencydigital volumecorrelation methodto characterize invivo optic nerve headbiomechanicsfromoptical coherence tomog-raphy" to characterize the in vivo ONH deformation of the volume collected by optical coherence tomography (OCT). By combining comprehensive tests and the analysis of OCT of the ONHs of monkeys subjected to acute and chronic intraocular pressure elevation, it was demonstrated that the proposed method overcame several challenges of the traditional DVC method.First, it takes into account the large ONH rigid body motion in the OCT volume, which may otherwise lead to analysis failure; second, although some OCT volumes have high noise and low image contrast, displacement with sub-voxel accuracy can be guaranteed; third, the computational efficiency is greatly improved, so the memory consumption of this method is much lower than that of traditional methods; fourth, a parameter for measuring displacement confidence is introduced. The image noise effect test shows that the proposed DVC method has a displacement error of less than 0.028 voxels for speckle noise and less than 0.037 voxels for Gaussian noise; the absolute (relative) strain errors in three directions are all lower than 0.0018 (4%) under speckle noise and less than 0.0045 (8%) under Gaussian noise. Compared with the traditional DVC method, the proposed DVC method greatly improves the overall displacement and strain errors under large object motion (up to 70% reduction), reduces the calculation time by 75%, and saves about 30% of the memory at the same time. Therefore, this study demonstrates the potential of the proposed technology in studying ONH biomechanics.

[0003] Although some progress has been made in the research on myopic scleral biomechanics in terms of structure, composition, and measurement methods, there are still many deficiencies that need to be improved. First, there is no good animal model (including transgenic models) that can be used to study the changes in myopic scleral biomechanics. The myopia in existing animal models is induced rapidly in a short time, and there are still significant differences in the scleral biomechanical properties between them and human sclera. The development of new animal models will help to deeply understand the important role of biomechanics in myopia. Second, the current measurement methods of scleral biomechanics are mainly limited to in vitro. In vivo biomechanical measurement is bound to become the mainstream method for future biomechanical measurement. To meet this requirement, advanced imaging technology, post-processing technology for image data, and more complex finite element modeling are necessary to make the measurement results of biomechanics more accurate. Finally, although scleral remodeling is the key factor determining the changes in myopic scleral biomechanics, recent research data have highlighted the role of scleral cells, especially myofibroblasts. The role played by myofibroblasts in the occurrence and development of myopia still needs further study. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an artificial intelligence algorithm for myopia risk prediction and its clinical application. By combining artificial intelligence and mechanical algorithms, an AI model that can non-invasively and conveniently evaluate scleral biomechanical properties is innovatively developed. The scleral biomechanical properties can be evaluated only by performing routine imaging on the posterior part of the eye.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] The present invention provides an artificial intelligence algorithm for myopia risk prediction, including the following steps:

[0007] S1: Strain field calculation;

[0008] S2: Strain prediction;

[0009] S3: Binary classification prediction based on risk level.

[0010] Furthermore, the strain field calculation includes the following steps:

[0011] S11: First, the spatial alignment technique is used to ensure the consistency of the spatial positions of the baseline image and the mechanical loading image;

[0012] S12: By using the data splicing and singular value decomposition (SVD) algorithm, combined with the Bhattacharyya distance feature pattern selection method, the key feature images are successfully extracted, and the strain field is calculated.

[0013] Furthermore, the strain prediction includes the following steps:

[0014] S21: Use the VAE model to extract the deep features of the image and use them as the input of the downstream regression model;

[0015] S22: Calculate the feature difference between the baseline image and the mechanical loading image through the paired image processing model, and then predict the strain.

[0016] Furthermore, the binary classification prediction includes the following steps:

[0017] S31: First, use statistical methods to classify the data, and then use LDA for dimensionality reduction after grouping the data;

[0018] S32: Then use the classification model to train the dimensionality-reduced data to achieve the binary classification prediction effect;

[0019] S33: Verify the trained model and compare multiple machine learning models, and the one with the best performance is selected as the final classification model.

[0020] The beneficial effects of the present invention are:

[0021] By applying a variety of mathematical models and artificial intelligence models, such as data filtering, feature extraction, paired image processing model, variational autoencoder (VAE), support vector machine regression (SVR), and LightGBM, the present invention deeply analyzes the posterior eye strain field and improves the accuracy of myopia risk prediction.

[0022] Other advantages, objects, and features of the present invention will be set forth in part in the description which follows, and in part will be obvious to those skilled in the art upon examination of the following or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the means of the instrumentalities and combinations particularly pointed out hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to make the objectives, technical solutions, and advantages of the present invention more clear, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0024] Figure 1 is the schematic diagram of the strain field calculation of the present invention;

[0025] Figure 2 is the schematic diagram of the strain field prediction of the present invention;

[0026] Figure 3 is the schematic diagram of the binary classification prediction of the compliance of the posterior part of the eyeball of the present invention;

[0027] Figure 4 is the schematic diagram of the program package of the present invention;

[0028] Figure 5 is the schematic diagram of the jupyter notebook page of the present invention;

[0029] Figure 6 is the schematic diagram of the modification path of the present invention;

[0030] Figure 7 is the schematic diagram of the image folder of the present invention;

[0031] Figure 8 is the schematic diagram of the operation of the code block of Task 1 of the present invention;

[0032] Figure 9 is the schematic diagram of the path modification of Task 2 of the present invention;

[0033] Figure 10 is the schematic diagram of the label path of the present invention;

[0034] Figure 11 is the schematic diagram of the path modification of Task 3 of the present invention;

[0035] Figure 12 is the schematic diagram of the label path of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only schematically illustrate the basic concept of the present invention. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0037] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams rather than physical diagrams, and should not be construed as a limitation to this patent; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted. In addition, the present invention may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0038] As Figures 1-3 shown, the feasibility of the present invention in clinical resources will be elaborated from four aspects: the overall package structure, the feasibility of task one program design, the feasibility of task two program design, and the feasibility of task three program design. In order to enable the normal operation of the code, the IDE for code operation is Jupyter Notebook / Jupyter Lab, and the libraries that need to be downloaded are as follows in the table:

[0039]

[0040]

[0041] The program package of the present invention is as Figure 4 shown:

[0042] As Figure 4 shown in the file structure, our program package structure consists of a main match directory folder, two folders named array and image slices under the match folder, a csv file for all data, three code files in ipynb format: challenge_cup_task1, challenge_cup_task2, challenge_cup_task3, four trained model files, and an array file for the best threshold.

[0043] (1) The array file is used to save the first principal strain array generated during the operation of challenge_cup_task1 (task one), and we perform relevant calculations by reading this array.

[0044] (2) The image slice folder stores the baseline images of the left and right eyes of each photographer and the best slices extracted from the mechanical loading images.

[0045] (3) The files "Challenge Cup_task1", "Challenge Cup_task2", and "Challenge Cup_task3" correspond to the main codes of our three tasks respectively. They can be directly run in Jupyter Notebook, and then the corresponding results can be obtained.

[0046] (4) The remaining model files are the model files trained by running our existing data in our program. Without training, these model files can be directly called to test the data in our test set.

[0047] In the program package, the file "Challenge Cup_task1.ipynb" is the main code for Task 1. This code realizes the spatial alignment of the baseline and mechanical loading images, removes interference values, performs SVD dimensionality reduction, feature pattern selection, correlation calculation, calculation of the best slice pair, and displacement field, strain field, and the upper limit value of the first principal stress. The following introduces the feasibility of the program for Task 1.

[0048] First, take the "match" file as the main directory file and open it through Jupyter Notebook to enter the following Figure 5 page. Find the file "Challenge Cup_task1.ipynb" on this page and double-click to open it to enter the code page. Before running, we need to adjust the path where the image data is located in the code. Therefore, we need to first find the following Figure 6 location of the code block.

[0049] After finding the above code block in the file "Challenge Cup_task1.ipynb", we need to change the path of the variable "path" to the path where our own images are located. The following Figure 7 content should be in this path file:

[0050] As shown in the figure file, all numbered files to be processed should exist in the path file. The numbered files store the images (Dicom format) of the left and right eyes corresponding to our numbers. We can change the "path" variable to the absolute path of the image folder.

[0051] After changing the path, directly run the code blocks one by one from top to bottom or run all the code blocks at once, such as finding and clicking "Restart&Run All" in the following Figure 8 to run all the code blocks.

[0052] After the code finishes running, the image information.csv and task1 image data.csv files will be automatically saved. Among them, the task1 image data.csv is the file we need to use later, which contains number information, the gender and age of the photographer, the baselines of the left and right eyes and the paths of the mechanical loading images, and the positions of the best slices of the baseline and mechanical loading images. The challenge_cup_task2.ipynb file is the main code for our task 2. In it, we used multiple models to achieve the prediction of the upper limit value of the first principal stress and saved the trained VAE and SVR models (no further training is required). Through these two models, we can directly predict other data. The following introduces the feasibility of the task 2 program.

[0053] First, adjust the image file path. In the challenge_cup_task2 program, find the following Figure 9 code block:

[0054] Replace the path of the global_flie variable in Figure 9 with the path where our image data is located. This path is the same as the path modified in task 1. Then run all the code blocks. For example, in Figure 8 , click Restart&Run All to run all the code blocks. After the program finishes running, it will automatically save the test image basic information.csv, test image slice data.csv, the best slice pairs of all images, and the prediction results.csv file. Among them, the prediction results.csv saves the strain prediction results of the images corresponding to the numbers.

[0055] MAE score verification. To facilitate the MAE score between the prediction results of the test code and the true labels, the task2_MAE verification.ipynb code file was designed. Considering the case where the true label is a 5D or 3D array, first, we need to save our labels in the form of numbered_OD.npy or numbered_OS.npy arrays. The number should be the same as the number of the image folder. For example, in the folder numbered 1171, the baseline and mechanical loading images of the right eye (OD) are stored. Then our label for 1171 should be 1173_OD.npy. Under the condition that our file format and the label array meet the above conditions, open the task2_MAE verification.ipynb file and find the Figure 10 code block shown.

[0056] Replace the path variable in Figure 10 with the location of the file where the label is located (it can be in the form of an absolute path). Then run all the program blocks according to the method in Figure 8 . After running, our MAE value will be printed. At the same time, the task 2 upper_bound label.csv file is saved, which is used to make the binary classification labels of the test data in task 3.

[0057] The main code for Task 3 in the match folder is the Challenge Cup_task3, which implements functions such as label production, dimensionality reduction, model construction, and prediction. In the match folder, all models for Task 3 have been saved. Therefore, during use, training does not need to be performed, and predictions can be directly made on the test data as needed. The following introduces the feasibility of Task 3.

[0058] First, open the Challenge Cup_task3 code file and find the Figure 11 location of the code block shown. As shown in the code block in the Figure 11 path of Program 3, the path of the variable FILE_PATH needs to be changed to the absolute path where your image data is located, and then click Restart&Run All to run all code blocks. After running is completed, the classification result

[0059] result.csv file will be obtained. The content shown in the file includes the image number and the predicted result (high or low).

[0060] After obtaining the binary classification prediction results of the test image dataset, in order to obtain the ACU score of the prediction results and the true results, the Challenge Cup_task3_AUC verification.ipynb code file was designed. Since it is uncertain whether there are corresponding labels, two considerations were made. The first is that the image data has no classification labels and classification needs to be calculated by oneself. The Figure 6 Challenge Cup_task2 upper_bound label.csv file generated in the feasibility of Program 2 of Task 2 is used to make classification labels. The second is that there are classification labels, and the labels exist in the form of csv or excel file formats, with the file numbers in the form of number_OD and number_OS (e.g., 1173_OD and 1173_OS) corresponding to the corresponding number high or low. After meeting one of the above conditions, the AUC score results can be obtained by running the Challenge Cup_task3_AUC verification.ipynb code.

[0061] For example Figure 12 , if the image data has corresponding labels and the labels are saved in the form of csv or excel files, meeting the label data form in Table 2.

[0062] Table 2 Label Data

[0063] Number Label 1171_OD 0.562 1171_OS 0.4632 1173_OD 0.256 1173_OS 03686

[0064] We need to Figure 12Change the value of the file_name variable in it to the path where the corresponding csv or excel file is located, and change the number variable and the target variable to the corresponding column names. As shown in Table 4.1, my number should be the serial number and the target should be the label. After making these changes, you can run all the code blocks, and finally the AUC result will be printed out.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and all of them should be covered by the scope of the claims of the present invention.

Claims

1. An artificial intelligence algorithm for myopia risk prediction, characterized in that: The following steps are involved: S1: strain field calculation; S2: strain prediction; S3: Binary classification prediction based on risk level.

2. The artificial intelligence algorithm for myopia risk prediction according to claim 1, characterized in that: The strain field calculation comprises the following steps: S11: First, the consistency of the baseline image and the mechanical loading image in spatial position was ensured by using the spatial alignment technology; S12: Using data splicing and singular value decomposition (SVD) algorithm, combined with Bhattacharyya distance feature pattern selection method, the key feature images were successfully extracted and the strain field was calculated.

3. The artificial intelligence algorithm for myopia risk prediction according to claim 1, characterized in that: The strain prediction comprises the following steps: S21: Use the VAE model to extract deep features of the image and use it as input to the downstream regression model; S22: The characteristic difference between the baseline image and the mechanical loading image is calculated through a paired image processing model to predict the strain.

4. The artificial intelligence algorithm for myopia risk prediction according to claim 1, characterized in that: The binary classification prediction includes the following steps: S31: First, use statistical methods to classify the data, and then use LDA to reduce the dimension after grouping the data; S32: Then use the classification model to train the reduced-dimensional data to achieve binary classification prediction effect; S33: The trained model was verified and compared with various machine learning models, and the one with the best performance was selected as the final classification model.

5. Clinical application of an artificial intelligence algorithm for myopia risk prediction as described in any one of claims 1 to 4.

Images