Eye index prediction method and device based on optical coherence tomography

Abstract

The invention relates to an eye index prediction method and device based on optical coherence tomography, and the method comprises the steps: obtaining the volume data of optical coherence tomography and a corresponding test report, and constructing a model training data set; carrying out data preprocessing on volume data; constructing an eye index prediction model, wherein the eye index prediction model performs feature extraction on the volume data through a visual encoder; carrying out regression calculation on the photosensitivity threshold value of each test point through a regression head according to the extracted features to obtain a photosensitivity threshold value graph; predicting whether each test point is abnormal or not and the abnormal level through the classification head according to the extracted features, and obtaining a mode deviation probability graph; and training the eye index prediction model based on the model training data set, and performing eye index prediction by adopting the trained eye index prediction model. Compared with the prior art, the mode deviation probability graph can be predicted, and higher prediction accuracy and wider clinical applicability are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an eye index prediction method and device based on optical coherence tomography. Background Art

[0002] The diagnostic approach for glaucoma relies primarily on a combination of structural and functional testing using optical coherence tomography (OCT) and perimetry. Structural-functional mapping is crucial in diagnosing glaucoma because it elucidates the relationship between structural damage and functional loss, allowing early structural changes to be detected before functional deficits become apparent on perimetry. This early detection is critical for timely intervention and potentially better control of glaucoma. OCT is used to detect structural changes in the optic nerve head (ONH) and macula, while perimetry is used to assess visual field function. Structural-functional mapping is key to diagnosing glaucoma because it reveals the relationship between structural damage and functional loss. However, existing technologies have limitations in predicting visual field light sensitivity thresholds and pattern deviation probability maps (PDPs).

[0003] The Humphrey Visual Field (HVF) is a medical test used to assess and record the visual field (VF). Developed by Humphrey Instruments, the HVF test is an integral part of modern ophthalmic clinical practice, providing ophthalmologists with an important tool to assess and understand the patient's visual field status. This test is primarily used to diagnose and monitor glaucoma and other ophthalmic diseases that may cause visual field loss. The HVF test assesses the light sensitivity threshold of the visual field by measuring the retina's response to light stimuli, thereby helping doctors understand potential damage in the patient's visual field. A key feature of the HVF test is that the test results are presented in the form of a visual field map, showing the light sensitivity threshold of different areas in the patient's visual field. These maps typically include total deviation (TD) and pattern deviation (PD) maps, as well as corresponding probability maps. The pattern deviation numerical map corrects for the effects of factors such as age, media opacity, and cataracts that cause a general decrease in retinal sensitivity, and highlights the local decrease in light sensitivity threshold by filtering out the overall decrease in light sensitivity threshold. The pattern deviation probability map describes the pattern deviation results in the form of a probability map, reflecting local visual field defects. It can ignore changes within the normal range and highlight tiny but clinically significant lesions, and can be used to observe early visual field defects.

[0004] Previous studies have demonstrated a strong association between OCT and visual field (VF) defects, suggesting that it is possible to predict VF light sensitivity thresholds from structural OCT scans and to generate pattern deviation probability maps (PDPs).

[0005] OCT and VF findings are not static but evolve over time. Accurately correlating OCT and VF findings in time is crucial to more precisely understand their relationship.

[0006] Furthermore, regional variations in glaucomatous structural and functional changes in the optic nerve head and HVF have not been adequately addressed. Most studies rely on cross-sectional or circular scans rather than volumetric scans, thus neglecting information about other retinal structures outside the scanned area.

[0007] In summary, the shortcomings of the prior art include:

[0008] These studies have limitations such as limited sample diversity, lack of extensive external validation, and difficulty in accurately predicting visual field light sensitivity thresholds and pattern deviation probability maps (PDPs). Specifically:

[0009] (1) Inconsistent OCT and visual field test results at different testing times make it difficult to accurately map the relationship between structural damage and functional loss.

[0010] (2) The data sample size is limited and there is a lack of external medical center verification, which limits the generalizability and reliability of the results.

[0011] (3) Only the numerical view of photosensitivity in VF data is predicted, and the model deviation probability map cannot be predicted, such as Figure 1 As shown in the figure, although the light sensitivity threshold map provides some information about retinal function, it is difficult to highlight local small but clinically significant lesions due to the influence of factors such as age, media turbidity, cataracts, etc. that cause a general decrease in retinal sensitivity, and is not enough to fully assess the risk of glaucoma. In addition, the light sensitivity threshold map has poor readability, and in actual clinical practice, it is difficult for doctors to efficiently understand its content.

[0012] In clinical practice, the pattern deviation probability map (PDP) is often compared with the light field sensitivity threshold Figure 1 In fact, PDP is one of the views that glaucoma doctors pay the most attention to. This view provides clinicians with accurate structure-function relationships in an easy-to-understand way, helps to distinguish visual field damage caused by pathological conditions such as glaucoma from normal variations, and can be used to observe early visual field defects. Summary of the invention

[0013] The purpose of the present invention is to provide an eye index prediction method and device based on optical coherence tomography in order to overcome the defect that the above-mentioned prior art can only generate photosensitivity threshold maps but cannot provide pattern deviation probability maps.

[0014] The purpose of the present invention can be achieved by the following technical solutions:

[0015] A method for predicting eye indicators based on optical coherence tomography comprises the following steps:

[0016] Obtain optical coherence tomography volume data and corresponding test reports to build a model training data set;

[0017] Perform data preprocessing on the volume data in the model training data set;

[0018] An eye index prediction model is constructed, which includes a visual encoder, a regression head and a classification head. The visual encoder is used to extract features from the volume data. The regression head is used to regress and calculate the light sensitivity threshold of each test point based on the extracted features to obtain a light sensitivity threshold map. The classification head is used to predict whether each test point is abnormal and the abnormality level based on the extracted features to obtain a pattern deviation probability map.

[0019] The eye index prediction model is trained based on the model training data set, and the trained eye index prediction model is used to predict the eye index of the optical coherence tomography data of the object to be tested.

[0020] Furthermore, the data preprocessing process includes: normalizing, scaling, adding random noise and flipping the volume data to perform data enhancement.

[0021] Furthermore, the data flipping includes: sequentially flipping each slice image of the left eye or right eye volume data of the same object, and horizontally mirror-flipping each slice image.

[0022] Furthermore, the method further comprises acquiring multiple types of optical coherence tomography volume data, and fusing various types of optical coherence tomography volume data of the same object, thereby constructing the model training data set;

[0023] The fusion method includes horizontally splicing different types of volume data, or merging slice images of each type of volume data after sampling;

[0024] The types of optical coherence tomography volume data include macular volume data, optic disc volume data and wide-field volume data acquired based on optical coherence tomography.

[0025] Furthermore, the method further comprises: counting the abnormal test point data contained in each sample in the statistical model training data set, and expanding the model training data set by copying samples including more abnormal test points.

[0026] Furthermore, the calculation process of the loss function in the training process of the eye index prediction model includes:

[0027] Count the actual number of normal test points and abnormal test points in the current training process, calculate the losses of normal test points and abnormal test points respectively, and divide them by the actual number of corresponding normal test points and abnormal test points respectively to get the final loss result.

[0028] Furthermore, during the regression calculation process of the photosensitivity threshold, the calculated photosensitivity threshold is normalized and restricted to a new range.

[0029] Furthermore, the visual encoder is a 2D encoder or a 3D encoder, the input of the 2D encoder is a plurality of slice images in the volume data, and the features corresponding to the slice images output by the 2D encoder are merged to form features corresponding to the volume data;

[0030] The input of the 3D encoder is volume data, and features corresponding to the volume data are directly acquired.

[0031] Furthermore, the method also includes selecting slice images of each volume data by adopting a regional random sampling strategy, and inputting the selected slice images into an eye index prediction model for prediction.

[0032] The present invention also provides an eye index prediction device based on optical coherence tomography, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] (1) The present invention directly uses the original 3D OCT scan volume data for processing to achieve the prediction of the pattern deviation probability map. The original 3D OCT scan volume data more accurately predicts the 52 visual field light sensitivity thresholds and pattern deviation probability maps of the VF, providing a more useful tool for the diagnosis of glaucoma;

[0035] Compared with 2D images, OCT scan volume data contains more key information about the retina, but the processing and analysis of OCT images involve multiple technical difficulties such as high-dimensional data processing, feature extraction, and computational efficiency. In contrast, the present invention proposes multiple data enhancement, fusion and sampling methods, and proposes a neural network model as the basic architecture, which combines regression and classification prediction heads and can predict the light sensitivity threshold and pattern deviation probability points in the visual field test point by point; the pattern deviation probability map can perform more detailed spatial analysis than the light sensitivity threshold, provides a closer connection between structure and function, and can highlight local tiny but clinically significant lesions.

[0036] (2) The present invention utilizes multiple OCT scans (macula, optic disc, wide field of view and their fusion) for more detailed structural analysis, and the deep learning-based system can more accurately predict the light sensitivity threshold.

[0037] (3) Compared with generating a photosensitivity threshold map, generating a pattern deviation probability map is a more difficult task. One of the main challenges is the imbalance of the data set. To this end, the present invention first uses an oversampling technique to count the abnormal test point data contained in each sample in the training data set, and by replicating samples including more abnormal test points, its representativeness in the training set is increased;

[0038] In addition, a cost-sensitive learning method is introduced, which dynamically adjusts the weights in each iteration to allow the model to adjust the importance of learning according to the actual number of samples in each category. Through this method, the learning process of the model can be adjusted without changing the data set to adapt to the inherent imbalance of the data and pay more attention to those samples that are more critical to clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an example diagram of a visual field report provided in the background technology of the present invention.

[0040] Figure 2 A schematic flow chart of an eye index prediction method based on optical coherence tomography provided in an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of a data processing flow of an eye index prediction method based on optical coherence tomography provided in an embodiment of the present invention;

[0042] Figure 4 It is a structural block diagram of an eye index prediction device based on optical coherence tomography provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0044] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0046] Example 1

[0047] like Figure 2 and Figure 3 As shown, this embodiment provides an eye index prediction method based on optical coherence tomography, comprising the following steps:

[0048] S1: Obtain the volume data of optical coherence tomography and the corresponding test report to build a model training data set;

[0049] S2: perform data preprocessing on the volume data in the model training data set;

[0050] S3: Construct an eye index prediction model, which includes a visual encoder, a regression head and a classification head. The visual encoder is used to extract features from the volume data. The regression head is used to regress and calculate the light sensitivity threshold of each test point based on the extracted features to obtain a light sensitivity threshold map. The classification head is used to predict whether each test point is abnormal and the abnormality level based on the extracted features to obtain a pattern deviation probability map.

[0051] S4: Train the eye index prediction model based on the model training data set, and use the trained eye index prediction model to predict the eye index of the optical coherence tomography data of the object to be tested.

[0052] Various aspects of this method are described in detail below.

[0053] 1. Model training data set

[0054] For the above model training data set, optionally, the volume data of optical coherence tomography (OCT) includes optic disc OCT, macula OCT, wide field OCT, etc. The test report may be a VF test report, obtained using a Humphrey perimeter manufactured by Zeiss.

[0055] 2. Data Preprocessing Process

[0056] The process of data preprocessing includes the following aspects:

[0057] 2.1 Data Flip

[0058] Each slice image of the left eye or right eye volume data of the same subject is sequentially flipped, and each slice image is horizontally mirrored and flipped.

[0059] Specifically, three-dimensional OCT volume data is used as data, and one volume data contains 256 slice images. In order to align the left eye and right eye data, the right eye OCT data is flipped in this embodiment, and the specific implementation includes the order of the 256 flipped slice images (from 0, 1...255 to 255, 254...0) and horizontal mirror flipping of each slice image.

[0060] 2.2 Data Augmentation

[0061] During the training phase, a variety of image preprocessing methods were used to enhance the OCT data, including normalization, scaling, and adding random noise.

[0062] 2.3. Multiple OCT Fusion

[0063] Preferably, the method further comprises acquiring multiple types of optical coherence tomography volume data, and fusing the various types of optical coherence tomography volume data of the same object, thereby constructing a model training data set;

[0064] The fusion methods include horizontal splicing of different types of volume data, or merging slice images of each type of volume data after sampling;

[0065] The types of optical coherence tomography volume data include macular volume data, optic disc volume data and wide-field volume data acquired based on optical coherence tomography.

[0066] For example, to fuse optic disc OCT and macula OCT for prediction, one implementation scheme is to horizontally stitch the optic disc slice image and the macula slice image, and another implementation scheme is to sample S / 2 slice images for each OCT and merge them to obtain S slice images.

[0067] 2.4 Regional random sampling strategy

[0068] Preferably, the method of the present invention further comprises selecting slice images of each volume data by adopting a regional random sampling strategy, and inputting the selected slice images into an eye index prediction model for prediction.

[0069] Specifically, using OCT data including 256 slices for training requires very large computing resources and training time. In view of the small interval between OCT scans, the present invention compares the data with different numbers of slices for training and testing, including 8, 16, 32, ... 256, etc. For this purpose, a regional random sampling strategy is designed to select slice images. Taking 16 slice images for training as an example, the 256 slice images are first divided into 16 groups, for example, the first group is [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15], and then a slice image is randomly selected from the first group as the sample of the first group in each training iteration. Similarly, the other 15 groups use the same sampling strategy, and finally 16 groups of slice images are obtained as input data.

[0070] 3. Data processing of the eye index prediction model

[0071] The data processing process for the above eye index prediction model includes:

[0072] 3.1. The visual encoder is a 2D encoder or a 3D encoder. The input of the 2D encoder is a plurality of slice images in the volume data. The features corresponding to the slice images output by the 2D encoder are merged to form the features corresponding to the volume data.

[0073] The input of the 3D encoder is volume data, and the features corresponding to the volume data are directly obtained.

[0074] Here we take the Transformer model as an example: when using a 2D encoder, the basic skeleton of the Swin-Transformer, ViT and VIT-base series models can be directly used as the encoder. OCT data is sent to the network in the form of 2D slice images. S slice images of an OCT data are encoded and calculated at the same time. The size of each slice image is H*W. Finally, S slice image embeddings will be obtained. By merging the S slice image embeddings, the OCT embedding is finally obtained. When a 3D encoder is used, the feature embedding, position encoding and other components of the Transformer encoder are adjusted according to the 3D data processing requirements. The OCT data is sent to the network in the form of 3D volume data with a size of H*W*S. In fact, H and W are the height and width of the slice image, respectively, and S is the number of slice images used. The encoding operation obtains the OCT embedding.

[0075] 3.2. The goal of the present invention is to predict the point-by-point photosensitivity threshold and pattern deviation probability of the VF test.

[0076] For the task of predicting the light sensitivity threshold, it is regarded as a continuous regression task. According to the extracted features, the light sensitivity threshold of each test point is regressed and calculated to obtain the light sensitivity threshold map;

[0077] For the task of predicting the probability of pattern deviation, it is regarded as a classification task. If it is predicted whether a test point is abnormal, it is a 2-classification task; if it is necessary to predict the abnormality level at the same time, it is a 5-classification task.

[0078] To this end, the present invention designs two prediction heads related to the task. The first prediction head is a regression head, which is used to regress the light sensitivity threshold of each test point. The second prediction head is a classification head, which is used to determine whether each test point of the pattern deviation probability map is abnormal and the abnormality level.

[0079] Preferably, in the calculation of the photosensitivity threshold map, the photosensitivity threshold interval is usually [0, 40]; for training stability and efficiency, the network prediction target is transformed to be restricted to a specific range and normalized to a new range, for example, adjusted to [-0.5, 0.5].

[0080] In traditional weighted methods, the identification of abnormal test points often relies on a preset threshold range, which may result in insufficient sensitivity in the identification of abnormal conditions, especially when the threshold range is wide. However, in the diagnosis of ophthalmic diseases such as glaucoma, the identification of abnormal test points is crucial because these test points may indicate the early development of the disease. The method of the present invention can more accurately identify these abnormal conditions by conducting an in-depth analysis of the light sensitivity threshold of the abnormal test points. This method not only improves the ability to identify abnormal test points, but also can more accurately reflect the functional status of the visual field, thereby providing a more valuable reference for the diagnosis and treatment of the disease.

[0081] Optionally, the above-mentioned eye index prediction model uses a neural network model as the basic architecture, combining regression and classification prediction heads, and is able to predict the light sensitivity threshold and pattern deviation probability points in the visual field test point by point.

[0082] The neural network models mentioned above include Transformer, CNN and other network structures. An implementation case is to use the EVA02 model as the basic framework to extract the representation of OCT data, and use the multi-layer perceptron MLP as the regression and classification prediction head network. The two tasks use the prediction head network dedicated to the task respectively.

[0083] IV. Training process of eye index prediction model

[0084] For the training process of the above eye index prediction model:

[0085] The present invention improves the accuracy of photosensitivity threshold maps and pattern deviation probability maps through advanced image processing and machine learning techniques, which helps to diagnose and monitor ophthalmic diseases more accurately. Compared with generating photosensitivity threshold maps, generating pattern deviation probability maps is a more difficult task. One of the main challenges is the imbalance of the data set: in the data set, the number of samples of normal test points far exceeds the number of samples of abnormal test points. Specifically, normal test points account for approximately more than 80% of the total samples, while abnormal test points are scattered in the remaining minority categories. This imbalance causes the model to more easily learn the characteristics of normal test points during training, but perform poorly in identifying abnormal test points.

[0086] The long-tail problem is a common problem in the field of machine learning, especially in medical image analysis. It stems from the uneven distribution of real-world data, which makes it difficult for the model to capture the key features of minority categories. Through case analysis, the long-tail problem reduces the accuracy of the model in identifying abnormal test points in early and middle stages of glaucoma, thus affecting the diagnosis and treatment decisions of patients.

[0087] The strategy for solving the long tail problem of the present invention is as follows:

[0088] 4.1. First, this embodiment uses oversampling technology to count the abnormal test point data contained in each sample in the training data set, and increases its representativeness in the training set by replicating samples including more abnormal test points.

[0089] However, when training deep learning models, it was observed that even with oversampling, the problem of imbalanced categories between normal and abnormal test points in a training batch could not be solved. This is because in patients with early and middle stage glaucoma, there are relatively few abnormal test points, while normal test points account for the vast majority. Even with oversampling, the problem of imbalanced categories between normal and abnormal test points in a training batch could not be solved. In addition, it was observed that in this extreme case, oversampling easily leads to overfitting of the model, resulting in poor performance on the external test set.

[0090] 4.2. In order to solve this problem, this embodiment introduces a cost-sensitive learning method. In the diagnosis of glaucoma, the cost of misclassifying an abnormal test point as normal (false negative) may be much higher than the cost of misclassifying a normal test point as abnormal (false positive). Therefore, a strategy is needed to adjust the loss function so that the model pays more attention to those more costly errors during training. This embodiment designs a specific loss function that dynamically adjusts the weight of each category during training to reflect its relative cost, so as to punish the model's errors in predicting abnormal categories and improve its performance in identifying abnormal test points. Specifically, in each iteration, the actual number of normal test points and abnormal test points in the current minibatch is counted. Then, the losses of normal test points and abnormal test points are calculated separately, and these losses are divided by the corresponding number of test points.

[0091] Dynamically adjusting the weights in each iteration allows the model to adjust the importance of its learning according to the actual number of samples in each category. In this way, the model's learning process can be adjusted without changing the data set, adapting to the imbalance of the data and focusing more on samples that are more critical to clinical decision-making. This method not only improves the accuracy of the model in identifying abnormal test points, but also reduces the risk of overfitting and improves the model's generalization ability on external test sets.

[0092] V. Model Evaluation and Validation

[0093] The experimental results of the present invention show that by adopting the above strategy, the performance of the model of the present invention in dealing with long-tail problems has been significantly improved. This method not only avoids the problem of overfitting, but also improves the generalization ability of the model on the external test set.

[0094] The present invention significantly improves the accuracy of generating light sensitivity thresholds by optimizing both the tasks of generating light sensitivity thresholds (regression) and generating pattern deviation probability maps (classification). For the task of generating pattern deviation maps, the model can not only accurately identify normal test points, but also effectively identify abnormal test points, reducing the occurrence of false negatives, which is particularly important in the field of medical diagnosis, because incorrect diagnosis may lead to delayed treatment or inappropriate treatment plans, which will have a serious impact on the health and well-being of patients.

[0095] In summary, the present invention not only significantly improves the accuracy of generating photosensitivity thresholds and the accuracy of pattern deviation probability maps, but also shows great potential in clinical application value. These achievements strongly support the innovation and practicality of the present invention, and provide important technical support for the diagnosis and treatment of ophthalmic diseases such as glaucoma.

[0096] VI. Clinical significance

[0097] The most critical core of the present invention is to generate a pattern deviation probability map, however, the verification of its ultimate value must be completed through rigorous clinical research.

[0098] In actual clinical applications, compared with the light sensitivity threshold map, the pattern deviation probability map (PDP) is one of the views that glaucoma doctors pay the most attention to. This view provides clinicians with accurate structure-function relationships in an easy-to-understand way, has good readability, and is in line with clinicians' practice processes and habits. By interpreting the pattern deviation probability map, doctors can diagnose and monitor diseases such as glaucoma more accurately and efficiently, and develop more reasonable treatment plans for patients.

[0099] The present invention conducted a clinical comparative study: the first group of clinicians used the pattern deviation map predicted by the present invention and the OCT report to diagnose glaucoma, and the other group used the pattern deviation probability map of the real clinical VF report and the OCT report to diagnose glaucoma. The two groups of diagnostic conclusions reached a high consistency, which shows that when there is a lack of VF testing or it is difficult to obtain a reliable VF test report, using the pattern deviation probability map predicted by the present invention for auxiliary diagnosis is a method with clinical application prospects. In addition, the present invention has demonstrated its reliability and applicability in different clinical environments through external verification.

[0100] Through these clinical studies, not only can the clinical effectiveness of the pattern deviation probability map be demonstrated, but it can also provide a powerful and user-friendly tool for glaucoma diagnosis, helping doctors make more accurate and timely treatment decisions. This will hopefully improve the treatment effect and quality of life of glaucoma patients, and also provide new perspectives and methods for the management of ophthalmic diseases.

[0101] In summary, the solution of the present invention has the following characteristics:

[0102] 1. Complexity of technical methods: 1) Compared with 2D images, OCT volume data contains more key information about the retina, but the processing and analysis of OCT images involve multiple technical difficulties such as high-dimensional data processing, feature extraction, and computational efficiency. 2) The sample ratio of normal and abnormal test points in the data set is very different, and there is a serious long-tail problem. This imbalance makes the model perform poorly for minority categories during learning and prediction, and methods such as cost-sensitive learning need to be used to solve it.

[0103] 2. Recognition of clinical needs: Recognition of the importance of pattern deviation probability maps for glaucoma diagnosis is a clinical insight rather than a direct technical derivation. This in-depth understanding of clinical needs prompted the development of methods for generating pattern deviation probability maps using OCT.

[0104] 3. Strategies for solving the long-tail problem: This paper adopts strategies such as sampling and cost-sensitive learning to solve the long-tail problem of the data set. These strategies require a deep understanding of the long-tail problem and the development of effective solutions.

[0105] 4. Task collaborative optimization: For the two subtasks of predicting photosensitivity threshold and predicting pattern deviation probability map, the present invention optimizes the two tasks simultaneously and adopts the same loss weighting strategy. This method ensures that the loss of each category (classification) or value range (regression) is given proper attention, thereby improving the classification performance and regression performance of the model in identifying abnormal test points, and providing a more valuable reference for the diagnosis and treatment of diseases. This demonstrates the innovation and practicality of the present invention in processing photosensitivity threshold maps and pattern deviation probability maps.

[0106] 5. Requirements for experimental validation: Generating pattern deviation probability maps and validating their clinical value requires a large amount of experimental data and clinical studies. The collection and analysis of these data is a time-consuming process that requires careful design of experiments and accurate interpretation of results.

[0107] Expertise and experience: To ensure that this technology can realize its potential in actual medical settings, it is necessary to work closely with clinicians to design experiments, collect data, and perform statistical analysis on the results to prove the clinical validity of the pattern deviation probability map. This interdisciplinary collaboration not only promotes the optimization of the technology, but also strengthens the trust and acceptance of the technology by clinicians.

[0108] In addition to the basic technical solution, the present invention can also realize different variants by changing the model architecture, data preprocessing method or training strategy. In addition, the present invention can be applied to the diagnosis of other types of ophthalmic diseases, as long as these diseases are related to the structural and visual field function assessment.

[0109] The present invention also provides an eye index prediction device based on optical coherence tomography, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the eye index prediction method based on optical coherence tomography as described above.

[0110] Specifically, this embodiment also provides Figure 4 A schematic structural diagram of an eye index prediction device based on optical coherence tomography is shown in FIG. Figure 4 As mentioned above, at the hardware level, the optical coherence tomography-based eye index prediction device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 2Of course, in addition to the software implementation, the present invention does not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0111] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting eye indicators based on optical coherence tomography, characterized in that: The following steps are involved: Obtain optical coherence tomography volume data and corresponding test reports to build a model training data set; Perform data preprocessing on the volume data in the model training data set; An eye index prediction model is constructed, which includes a visual encoder, a regression head and a classification head. The visual encoder is used to extract features from the volume data. The regression head is used to regress and calculate the light sensitivity threshold of each test point based on the extracted features to obtain a light sensitivity threshold map. The classification head is used to predict whether each test point is abnormal and the abnormality level based on the extracted features to obtain a pattern deviation probability map. The eye index prediction model is trained based on the model training data set, and the trained eye index prediction model is used to predict the eye index of the optical coherence tomography data of the object to be tested.

2. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The data preprocessing process includes: normalizing, scaling, adding random noise and flipping the volume data to perform data enhancement.

3. The method for predicting eye indicators based on optical coherence tomography according to claim 2, characterized in that: The data flipping includes: sequentially flipping each slice image of the left eye or right eye volume data of the same object, and horizontally mirror-flipping each slice image.

4. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The method further includes acquiring multiple types of optical coherence tomography volume data, and fusing various types of optical coherence tomography volume data of the same object to construct the model training data set; The fusion method includes horizontally splicing different types of volume data, or merging slice images of each type of volume data after sampling; The types of optical coherence tomography volume data include macular volume data, optic disc volume data and wide-field volume data acquired based on optical coherence tomography.

5. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The method further includes: counting the abnormal test point data contained in each sample in the model training data set, and expanding the model training data set by copying samples including more abnormal test points.

6. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The calculation process of the loss function in the training process of the eye index prediction model includes: Count the actual number of normal test points and abnormal test points in the current training process, calculate the losses of normal test points and abnormal test points respectively, and divide them by the actual number of corresponding normal test points and abnormal test points respectively to get the final loss result.

7. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: During the regression calculation process of the photosensitivity threshold, the calculated photosensitivity threshold is normalized and limited to a new range.

8. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The visual encoder is a 2D encoder or a 3D encoder, the input of the 2D encoder is a plurality of slice images in the volume data, and the features corresponding to the slice images output by the 2D encoder are merged to form features corresponding to the volume data; The input of the 3D encoder is volume data, and features corresponding to the volume data are directly acquired.

9. The method for predicting eye indicators based on optical coherence tomography according to claim 1, characterized in that: The method also includes selecting slice images of each volume data by adopting a regional random sampling strategy, and inputting the selected slice images into an eye index prediction model for prediction.

10. An eye index prediction device based on optical coherence tomography, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any one of the methods according to claims 1 to 9.

Images