An intraocular pressure measurement method and system based on ultrasound and deep learning

Abstract

The present invention provides an intraocular pressure measurement method and system based on ultrasound and deep learning. The method includes: collecting ocular ultrasound data and intraocular pressure data; performing four-dimensional ultrasound reconstruction on the continuously collected ultrasonic intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye; establishing a relationship model between the four-dimensional ocular ultrasound data and the corresponding intraocular pressure data; constructing and training an LSTM neural network intraocular pressure recognition model based on the reconstructed four-dimensional ocular internal structure ultrasound data, intraocular pressure data, and the relationship model; and identifying and displaying the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model. By establishing a relationship model between the four-dimensional ocular ultrasound data and the corresponding intraocular pressure data, the present invention mines ultrasound data features closely related to intraocular pressure, and constructs and trains the LSTM neural network intraocular pressure recognition model for intraocular pressure recognition, thereby achieving dynamic intraocular pressure measurement and improving measurement accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of intraocular pressure measurement, and specifically to an intraocular pressure measurement method and system based on ultrasound and deep learning. Background Art

[0002] Intraocular pressure (IOP) is a key indicator in ophthalmic diagnosis and treatment, crucial for the early detection, disease monitoring, and treatment efficacy assessment of eye diseases such as glaucoma. Glaucoma, a leading cause of irreversible blindness worldwide, is characterized by pathologically elevated IOP, which leads to optic nerve damage and gradual visual field loss. Therefore, accurate and reliable IOP measurement is crucial for the prevention and treatment of glaucoma.

[0003] Traditional methods of measuring intraocular pressure are mainly divided into two categories: contact and non-contact. Among the contact measurement methods, the Goldmann applanation tonometer is regarded as the "gold standard" for intraocular pressure measurement. Its principle is to calculate the intraocular pressure by flattening the cornea to a certain area based on the amount of external force required for applanation. However, this method requires direct contact with the eyeball, which may cause discomfort to the patient and even pose a certain risk of infection, such as corneal abrasion. In addition, the measurement process has high technical requirements for the operator, and the measurement results between different operators may vary. The indentation tonometer is also a contact measurement method that estimates intraocular pressure by measuring the depth of the indented eyeball, but this method also has risks caused by contact, and its accuracy is easily affected by factors such as the hardness of the eyeball.

[0004] Non-contact measurement methods, such as non-contact tonometer (NCT), use airflow pulses to flatten the cornea to a certain degree and calculate intraocular pressure by measuring changes in reflected light. Although direct contact with the eyeball is avoided, it is easily affected by factors such as the surface condition of the eye (such as tear film stability, corneal surface irregularities, etc.), eye movement, and patient cooperation, resulting in relatively low measurement accuracy. Although dynamic contour tonometer has improved measurement accuracy to a certain extent, the equipment is expensive and relatively complex to operate, making it difficult to widely popularize.

[0005] In recent years, the rapid development of ultrasound technology and deep learning algorithms has brought new opportunities for intraocular pressure measurement. Ultrasound technology, with its advantages of being non-invasive and highly penetrating, can obtain detailed information about the internal structures of the eye. Deep learning algorithms have demonstrated remarkable capabilities in processing complex data and pattern recognition. By learning from large amounts of data, they can automatically extract features and build accurate predictive models. The combination of these two methods is expected to overcome the limitations of traditional intraocular pressure measurement methods, enabling continuous, dynamic, and accurate measurement of intraocular pressure, and providing stronger support for clinical diagnosis and research in ophthalmology. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an intraocular pressure measurement method and system based on ultrasound and deep learning, so as to solve the technical problems existing in the existing intraocular pressure measurement described in the background technology.

[0007] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0008] A method for measuring intraocular pressure based on ultrasound and deep learning, comprising the following steps:

[0009] Collect ocular ultrasound data and intraocular pressure data through head-mounted ultrasound equipment and intraocular pressure measurement equipment;

[0010] Performing four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye;

[0011] Establish a relationship model between four-dimensional ocular ultrasound data and corresponding intraocular pressure data;

[0012] By reconstructing four-dimensional ultrasound data of the internal structure of the eye, intraocular pressure data and relationship models, an LSTM neural network intraocular pressure recognition model is constructed and trained;

[0013] Based on the trained LSTM neural network intraocular pressure recognition model, the patient's intraocular pressure is classified, identified and displayed.

[0014] Optionally, the continuously collected ultrasonic intraocular pressure data sequence is subjected to four-dimensional ultrasonic reconstruction, specifically: an image processing algorithm based on spatiotemporal features is used to perform four-dimensional ultrasonic reconstruction on the data sequence, assuming that the two-dimensional ultrasonic image sequence has T frames, spatial coordinates (x, y), and the ultrasonic signal intensity of the t-th frame image at point (x, y) is I t (x, y), and the eye tissue displacement vector field (Δx t (x,y),Δy t (x,y)), the reconstruction formula is:

[0015]

[0016] Where ω(tt′) is the time interpolation weight function.

[0017] Optionally, the establishing of the relationship model between the four-dimensional ultrasound data and the corresponding intraocular pressure data is specifically as follows:

[0018] Data annotation: annotate the four-dimensional ultrasound data and the corresponding intraocular pressure data, and divide the samples into normal intraocular pressure group and patient intraocular pressure group according to the intraocular pressure value range;

[0019] Feature extraction and analysis: extract features from four-dimensional ultrasound data, analyze the characteristic differences of ultrasound data of internal eye structures under different intraocular pressure conditions, and perform correlation analysis with intraocular pressure values ​​to determine the characteristic vector related to intraocular pressure;

[0020] A relationship model was constructed, and the multivariate linear regression method was used to construct a mathematical model between the ultrasound data characteristics and the intraocular pressure value. The characteristic vector was used as the independent variable and the intraocular pressure value was used as the dependent variable. The model parameters were determined by training the model, and a quantitative correspondence between ultrasound data and intraocular pressure was established.

[0021] Optionally, the LSTM neural network intraocular pressure recognition model is constructed and trained by reconstructing the four-dimensional intraocular structure ultrasound data, intraocular pressure data and relationship model, specifically:

[0022] The model was trained using ultrasound data and intraocular pressure data samples with established correspondence, and the sample data was divided into training set, validation set, and test set in a ratio of 7:2:1 or 8:1:1;

[0023] Construct an LSTM neural network model with 1-3 LSTM layers, 30-100 hidden units per layer, followed by a fully connected layer for classification output. The model is trained using a 4D ultrasound data sequence of the internal structure of the eye as input and the corresponding intraocular pressure data as output.

[0024] The stochastic gradient descent optimization algorithm was used to adjust the model weights and bias parameters through the back propagation algorithm according to the difference between the model prediction results and the actual intraocular pressure data to minimize the loss function.

[0025] Optionally, the patient's intraocular pressure is identified based on the trained LSTM neural network intraocular pressure recognition model, specifically: the ultrasonic intraocular pressure data is collected and reconstructed in real time, and the trained LSTM neural network intraocular pressure recognition model is input to determine whether the intraocular pressure is normal and whether there is any abnormal change trend, and at the same time, the dynamic change indicators of intraocular pressure are calculated, including the intraocular pressure fluctuation amplitude and the intraocular pressure change frequency.

[0026] Optionally, the feature extraction of the four-dimensional ultrasound data and analysis of the feature differences of the ultrasound data of the internal structure of the eye under different intraocular pressure states are specifically as follows:

[0027] The division and location of ocular tissue regions uses an image segmentation algorithm to analyze features of ultrasound images, not limited to grayscale and texture, to identify the boundaries and ranges of ocular structures, including but not limited to the cornea, anterior chamber, lens, vitreous body, and retina. The cornea is located based on its arc structure and echo intensity; the anterior chamber is determined based on echo characteristics and its positional relationship with the cornea and lens; the lens is identified by its shape and echo performance; the vitreous body is divided by its echo characteristics; and the retina is located based on the reflection characteristics of the interface between it and the choroid.

[0028] Establish an independent coordinate system for each divided eye tissue area, and then accurately calculate and analyze the features in different areas;

[0029] Thickness change rate calculation: For each ocular tissue area, the thickness change rate over time is calculated in the time dimension of the four-dimensional ultrasound data;

[0030] Echo intensity change analysis: statistics on the echo intensity distribution of each eye tissue area in the four-dimensional ultrasound data, dividing the echo intensity into several intervals, and calculating the number or proportion of pixels in each interval at different times;

[0031] Analyze the changing trend of the echo intensity distribution of ocular tissue during the process of intraocular pressure changes, and calculate the average echo intensity change and echo intensity change rate;

[0032] Texture feature extraction and analysis: Using the gray-level co-occurrence matrix analysis method, we extract the texture features of ocular tissue ultrasound images, calculate multiple characteristic parameters of the gray-level co-occurrence matrix, including but not limited to contrast, correlation, energy, and entropy, and compare the differences in ocular tissue texture features under different intraocular pressure conditions;

[0033] Feature correlation analysis and feature vector construction were performed to calculate the correlation coefficients between the extracted thickness change rate, echo intensity change-related parameters, texture feature parameters and intraocular pressure values. Based on the correlation analysis results, features with significant correlation with intraocular pressure were screened out and feature vectors were constructed.

[0034] Optionally, the correlation coefficient between the calculated and extracted thickness change rate, echo intensity change related parameters, texture feature parameters and intraocular pressure value is specifically calculated using the Pearson correlation coefficient method, and the formula is:

[0035]

[0036] Where n is the number of samples, a i and b i are the observed values ​​of variables A and B respectively, and are the means of variables A and B, respectively, where a i represents thickness change rate, echo intensity change related parameters, texture feature parameters, b i Indicates that i The corresponding intraocular pressure value.

[0037] The present invention proposes an intraocular pressure measurement system based on ultrasound and deep learning, comprising:

[0038] Data acquisition module, used to collect eye ultrasound data and intraocular pressure data;

[0039] A four-dimensional reconstruction module is used to perform four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye;

[0040] Ultrasound and intraocular pressure relationship building module, used to establish a relationship model between four-dimensional eye ultrasound data and corresponding intraocular pressure data;

[0041] The neural network model training module is used to build and train an LSTM neural network intraocular pressure recognition model using reconstructed four-dimensional ultrasound data of the internal structure of the eye, intraocular pressure data, and a relationship model;

[0042] The intraocular pressure analysis and recognition module is used to identify the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model.

[0043] Optionally, the data acquisition module includes an eye ultrasound acquisition submodule, an intraocular pressure data acquisition submodule, and a data transmission module, and the eye ultrasound acquisition submodule and the intraocular pressure data acquisition submodule are both communicatively connected to the data transmission module, wherein:

[0044] (1) The ocular ultrasound acquisition submodule includes:

[0045] Ultrasonic transducer, array-type surrounds the eye, operating frequency 2-10MHz, multi-angle transmission and reception of ultrasonic signals;

[0046] Signal transmitting and receiving module, controls the ultrasonic transducer to transmit pulses and receive reflected echo signals;

[0047] Echo acquisition module, converts analog echo signals into digital signals, with a sampling frequency of 15-45MHz;

[0048] A microprocessor for preprocessing collected data;

[0049] A communication interface module, used to transmit the collected ultrasonic signal to an external device, which is transmitted to the data transmission module in the present invention;

[0050] Power module, using rechargeable battery;

[0051] (4) an intraocular pressure data acquisition submodule, using a tonometer, including but not limited to an applanation tonometer, an indentation tonometer, a non-contact tonometer, and a dynamic contour tonometer;

[0052] (5) A data transmission module, which is used to send the eye data collected by the eye ultrasound acquisition submodule and the intraocular pressure data acquisition submodule to the server for data analysis and modeling.

[0053] Optionally, it also includes an intraocular pressure display module, which displays the intraocular pressure measurement results and related information in the form of charts and data lists, including but not limited to three-dimensional dynamic charts, line charts, and bar charts.

[0054] The beneficial effects of the above technical solution of the present invention are as follows:

[0055] 1. This invention utilizes a head-mounted ultrasound device to continuously collect ocular ultrasound data. Combined with four-dimensional ultrasound reconstruction technology, it can obtain three-dimensional dynamic changes in the internal structure of the eye over time, thereby enabling continuous dynamic monitoring of intraocular pressure. This continuous dynamic measurement method can provide more comprehensive and accurate information on intraocular pressure changes, helping to gain a deeper understanding of the patterns and trends of intraocular pressure changes. For patients with large intraocular pressure fluctuations or those at risk for potential eye diseases, it can promptly detect abnormalities, providing an important basis for early diagnosis and treatment.

[0056] 2. The present invention establishes a relationship model between four-dimensional eye ultrasound data and corresponding intraocular pressure data, and performs correlation analysis on the four-dimensional eye ultrasound data and the corresponding intraocular pressure data. It can deeply explore the impact mechanism of intraocular pressure changes on the internal structure of the eye in spatial and temporal dimensions, and can mine ultrasound data features closely related to intraocular pressure. These features can then be applied to the subsequent construction of an intraocular pressure recognition model, which can more accurately reflect the true state of intraocular pressure.

[0057] 3. This invention leverages the deep learning algorithm's ability to process complex data and learn features by establishing a relationship model between four-dimensional ocular ultrasound data and intraocular pressure data, and using this data to construct and train an LSTM neural network intraocular pressure recognition model. As measurement data accumulates, the system can continuously update and optimize the LSTM neural network model using new data, enabling it to adapt to changes in different patient groups and various ocular conditions, thereby improving the model's generalization and predictive performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the intraocular pressure measurement method based on ultrasound and deep learning of the present invention;

[0059] Figure 2 This is a functional block diagram of the intraocular pressure measurement system based on ultrasound and deep learning of the present invention;

[0060] Figure 3 This is a functional block diagram of the data acquisition module of the intraocular pressure measurement system based on ultrasound and deep learning of the present invention;

[0061] Figure 4 This is a principle block diagram of the ocular ultrasound acquisition submodule of the intraocular pressure measurement system based on ultrasound and deep learning of the present invention. DETAILED DESCRIPTION

[0062] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0063] like Figure 1 As shown, a method for measuring intraocular pressure based on ultrasound and deep learning includes the following steps:

[0064] S1. Collect ocular ultrasound data and intraocular pressure data through a head-mounted ultrasound device and an intraocular pressure measurement device;

[0065] S2, performing four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye;

[0066] S3. Establishing a relationship model between four-dimensional ocular ultrasound data and corresponding intraocular pressure data;

[0067] S4. Construct and train an LSTM neural network intraocular pressure recognition model using the reconstructed four-dimensional eye internal structure ultrasound data, intraocular pressure data, and relationship model;

[0068] S5. Based on the trained LSTM neural network intraocular pressure recognition model, the patient's intraocular pressure is classified, identified and displayed.

[0069] In step S1, the patient puts on the head-mounted ultrasound device and adjusts the position to ensure that the ultrasound transducer fits well with the eye without obstruction or pressure. The medical staff enters the patient's basic information (name, age, gender, medical history, etc.) into the accompanying software and sets measurement parameters based on the patient's condition, such as the measurement interval (5-15 minutes) and measurement duration (1-24 hours).

[0070] Next, the head-mounted ultrasound device collects ocular ultrasound data and intraocular pressure data. The ultrasound transducer automatically activates the ultrasound signal transmitter at preset intervals. The transmission frequency is initially 5 MHz and automatically adjusts to between 2 and 10 MHz based on the strength of the reflected signal from the ocular tissue. The pulse width is 2 to 8 microseconds, and the transmission interval is 60 to 180 microseconds. The ultrasound transducer array receives the echo signal reflected from the ocular tissue, and the data acquisition module converts it into a digital signal with a sampling frequency of 15 to 45 MHz, generating the ocular ultrasound data sequence.

[0071] While collecting ocular ultrasound data, IOP is measured using a standard intraocular pressure measurement device (e.g., applanation tonometer, indentation tonometer, non-contact tonometer, or dynamic contour tonometer). The tonometer probe is gently placed on the surface of the patient's eyeball to measure the IOP.

[0072] In this embodiment, step S2 performs four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence, specifically including:

[0073] (1) Time dimension integration:

[0074] Assume that a 2D ultrasound image sequence has T frames in the time dimension. For each frame, mark the corresponding time point t (t = 0, 1, ..., T-1). By analyzing the time interval Δt between adjacent frames, the uniformity and accuracy of the time series are ensured. If the time interval is found to be uneven, it may be due to equipment failure or interference during the acquisition process. Appropriate time correction processing is required, such as linear interpolation or resampling to adjust the time series so that the image sequence accurately reflects the physiological changes of the eye in the time dimension.

[0075] The optical flow method is used to track the motion information of eye tissue between different time frames and calculate the displacement vector field (Δx t (x,y),Δy t (x,y)). The optical flow method assumes continuity of pixel intensity in an image and calculates the speed of pixel movement between frames to determine the direction and displacement of ocular tissue. For key ocular structures such as the cornea and lens, the method accurately tracks their positional changes over time, providing dynamic information for subsequent spatial integration.

[0076] (2) Spatial dimension integration:

[0077] The spatial coordinates (x, y) of the 2D ultrasound image are determined, and the spatial position of each pixel in each frame is clearly defined. Based on the characteristics of the ocular anatomy, pixels in different regions (such as the cornea, anterior chamber, and lens) are classified and labeled to accurately handle the spatial relationships between different tissues during the integration process.

[0078] Incorporating knowledge of ocular anatomy, two-dimensional ultrasound images from different time frames are spatially aligned and spliced. Taking into account the potential for subtle rotations and translations of the eye during physiological activity, image registration technology based on feature point matching is employed. By identifying key feature points in the image (such as the corneal vertex and lens edge), correspondences are established between frames. Based on these correspondences, geometric transformations (such as rotation, translation, and scaling) are then performed on the images, ensuring accurate spatial alignment of the images from different time frames and constructing a three-dimensional sequence of ocular structural images.

[0079] (3) 4D ultrasound data generation:

[0080] Using the integrated information of the time dimension and space dimension obtained above, the four-dimensional ultrasound data is calculated according to a specific reconstruction formula. Let the ultrasound signal intensity at the point (x, y) of the frame image be I t (x, y), then the reconstruction formula of four-dimensional ultrasound data P(x, y, t) is:

[0081]

[0082] Where ω(tt′) is the time interpolation weight function, according to the time interval Δ t The specific form is determined by the characteristics of the data. This formula fuses the ultrasound signal intensity information from different time frames and spatial locations to generate four-dimensional ultrasound data that can reflect the dynamic changes of the internal structure of the eye over time. Its spatial resolution can reach 0.1-0.5mm and its temporal resolution can reach 0.1-1 second. It can clearly show the detailed dynamic changes of eye tissue and intraocular pressure over time.

[0083] In this embodiment, step S3 establishes a relationship model between the four-dimensional ultrasound data and the corresponding intraocular pressure data, specifically:

[0084] Data annotation: annotate the four-dimensional ultrasound data and the corresponding intraocular pressure data, and divide the samples into a normal intraocular pressure group and a patient intraocular pressure group according to the intraocular pressure value range (for example, the normal intraocular pressure range is 10-21 mmHg, which can be adjusted according to actual clinical standards).

[0085] Feature extraction and analysis: Extract features from 4D ultrasound data and analyze the characteristic differences in ultrasound data of internal ocular structures under different intraocular pressure conditions. For example, characteristic parameters such as the rate of change of ocular tissue thickness and echo intensity are calculated and correlated with intraocular pressure values ​​to identify feature vectors that are closely related to intraocular pressure. Methods such as principal component analysis (PCA) can be used to reduce feature dimensionality and improve computational efficiency.

[0086] Construct a relationship model and use the multivariate linear regression method to construct a mathematical model between ultrasound data features and intraocular pressure values. The feature vector is used as the independent variable and the intraocular pressure value is the dependent variable. The model parameters are determined by training the model to establish a quantitative correspondence between ultrasound data and intraocular pressure. Assume that the intraocular pressure value is y and the ultrasound data feature vector is X = (x1, x2, ..., x n ) (where n is the dimension of the eigenvector), the multiple linear regression model can be expressed as y = β0 + β1x1 + β2x2 + ... + β n x n +ε(where β0, β1, β2,…β n is the regression coefficient to be determined, and ε is the error term). Using a large number of labeled data samples collected (X i ,y i (i is the sample size) and estimate the regression coefficients using the least squares method. After the model is built, the validation set data is used to evaluate the model and calculate metrics such as the mean squared error (MSE). Based on the evaluation results, if the model performance is poor, adjust the model parameters or adopt other improvement measures.

[0087] Among them, feature extraction is performed on the four-dimensional ultrasound data to analyze the characteristic differences of the ultrasound data of the internal structure of the eye under different intraocular pressure conditions, specifically:

[0088] (1) Division and positioning of eye tissue regions

[0089] Image segmentation algorithms are used to precisely delineate and locate different tissue regions within ocular ultrasound images. For example, by analyzing ultrasound image features such as grayscale and texture, the boundaries and extents of major ocular structures, including the cornea, anterior chamber, lens, vitreous body, and retina, are accurately identified. The cornea is located based on its unique curved structure and relatively high echogenicity; the anterior chamber is identified based on its relatively low echogenicity and its positional relationship to the cornea and lens; the lens is identified by its distinctive shape and high echogenicity; the vitreous body is delineated by its relatively uniform low echogenicity; and the retina is located based on the reflective characteristics of its interface with the choroid.

[0090] An independent coordinate system is established for each divided ocular tissue region to facilitate the subsequent precise calculation and analysis of features within each region. For example, a two-dimensional coordinate system is established with the corneal vertex as the origin, with the horizontal x-axis and the vertical y-axis to describe changes in features within the corneal region. For the lens, a three-dimensional coordinate system is established with its center as the origin. In addition to the x- and y-axes, an additional axis is added to represent the thickness of the lens, allowing for a more comprehensive analysis of changes in the lens across different dimensions.

[0091] (2) Calculation of thickness change rate

[0092] For each ocular tissue region, the rate of change of its thickness over time is precisely calculated within the time dimension of the 4D ultrasound data. For the cornea, for example, in each ultrasound image frame, the distance between the front and back surfaces of the cornea is measured (the front and back surface boundary is determined based on the grayscale changes in the corneal tissue in the ultrasound image), yielding corneal thickness values ​​T(t) at different moments (where t represents time).

[0093] Calculate thickness change rate The difference method is used for approximate calculation, that is, (Δt is the time interval between two consecutive image frames.) By calculating the differences in the rate of change of corneal thickness under different intraocular pressure conditions, it can be found that when intraocular pressure increases, the cornea may show a trend of increasing or thinning thickness due to pressure, and the rate of change may be related to the magnitude of the intraocular pressure change.

[0094] Similarly, the rate of change in thickness of other ocular tissues, such as the lens and retina, is calculated. Changes in lens thickness may affect its refractive power, while changes in retinal thickness may be related to the effects of intraocular pressure on the retinal nerve fiber layer. By comparing the differences in the rate of change in these tissue thicknesses under normal and elevated intraocular pressure conditions, important characteristic information related to intraocular pressure can be obtained.

[0095] (3) Analysis of echo intensity changes

[0096] Count the echo intensity distribution of each ocular tissue region in the 4D ultrasound data. Divide the echo intensity into several intervals (for example, low, medium, and high based on experience and data characteristics) and calculate the number or proportion of pixels in each interval at different times.

[0097] Analyze the changing trends in the echogenicity distribution of ocular tissues during changes in intraocular pressure. For example, when intraocular pressure rises, certain tissues may experience changes in their internal acoustic properties due to structural deformation or changes in blood circulation, leading to changes in echogenicity. This may manifest as either increased echogenicity (e.g., increased tissue density) or decreased echogenicity (e.g., tissue edema leading to echo attenuation) in certain areas.

[0098] Parameters related to echo intensity changes, such as the average echo intensity change and the echo intensity change rate, can be used as important features to reflect the impact of intraocular pressure changes on ocular tissues.

[0099] (4) Texture feature extraction and analysis

[0100] Texture analysis methods such as the gray-level co-occurrence matrix (GLCM) are used to extract texture features from ultrasound images of ocular tissue. Multiple characteristic parameters of the GLCM, such as contrast, correlation, energy, and entropy, are calculated to describe the uniformity, directionality, and complexity of the internal structure of ocular tissue.

[0101] Compare the differences in ocular tissue texture characteristics under different intraocular pressure conditions. For example, under normal intraocular pressure, the texture of ocular tissue may be relatively uniform and regular. However, when intraocular pressure increases, the tissue may deform, shift, or undergo microstructural changes, resulting in changes in texture characteristics such as roughening, increased directionality, or disorder. By analyzing these changes in texture characteristics, we can further explore information related to intraocular pressure, providing more information for intraocular pressure measurement and eye disease diagnosis.

[0102] (5) Feature correlation analysis and feature vector construction

[0103] Calculate the correlation coefficient between the various features extracted above (thickness change rate, echo intensity change related parameters, texture feature parameters, etc.) and the intraocular pressure value. Use the Pearson correlation coefficient method to determine the strength of the correlation between each feature and intraocular pressure. The Pearson correlation coefficient method is as follows:

[0104]

[0105] Where n is the number of samples, a i and b i are the observed values ​​of variables A and B respectively, and are the means of variables A and B, respectively, where a i represents thickness change rate, echo intensity change related parameters, texture feature parameters, b i Indicates that i The corresponding intraocular pressure value.

[0106] Based on the correlation analysis results, features significantly correlated with IOP were selected and a feature vector was constructed. These features were then combined into a vector in a specific order and used as input variables for the subsequent construction of a model for the relationship between ultrasound data and IOP, accurately reflecting the intrinsic connection between ocular ultrasound data features and IOP values.

[0107] In this embodiment, step S4 constructs and trains an LSTM neural network intraocular pressure recognition model by reconstructing the four-dimensional intraocular structure ultrasound data, intraocular pressure data, and the relationship model, specifically:

[0108] The model was trained using ultrasound data and intraocular pressure data samples with established correspondence, and the sample data was divided into training set, validation set, and test set in a ratio of 7:2:1 or 8:1:1;

[0109] Construct an LSTM neural network model with 1-3 LSTM layers, 30-100 hidden units per layer, followed by a fully connected layer for classification output. The model is trained using a 4D ultrasound data sequence of the internal structure of the eye as input and the corresponding intraocular pressure data as output.

[0110] Using a stochastic gradient descent optimization algorithm, we adjust model weights and bias parameters through backpropagation based on the discrepancies between the model's predictions and the actual intraocular pressure data to minimize the loss function (e.g., mean squared error). We regularly evaluate model performance using a validation set. Training is stopped when the model's loss or accuracy on the validation set no longer decreases to prevent overfitting. Finally, we evaluate model performance using a test set to ensure high accuracy and reliability.

[0111] In this embodiment, step S5 classifies, identifies and displays the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model. Specifically, the ultrasonic intraocular pressure data collected and reconstructed in real time is input into the trained LSTM neural network intraocular pressure recognition model to determine whether the intraocular pressure is normal and whether there is any abnormal change trend, and at the same time calculate the dynamic change indicators of intraocular pressure, including the intraocular pressure fluctuation amplitude and the intraocular pressure change frequency.

[0112] like Figure 2 As shown, the present invention proposes an intraocular pressure measurement system based on ultrasound and deep learning, comprising:

[0113] The data acquisition module 101 is used to collect eye ultrasound data and intraocular pressure data;

[0114] A four-dimensional reconstruction module 102 is used to perform four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye;

[0115] Ultrasound and intraocular pressure relationship building module 103, used to establish a relationship model between four-dimensional eye ultrasound data and corresponding intraocular pressure data;

[0116] A neural network model training module 104 is used to construct and train an LSTM neural network intraocular pressure recognition model using the reconstructed four-dimensional intraocular structure ultrasound data, intraocular pressure data, and the relationship model;

[0117] The intraocular pressure analysis and recognition module 105 is used to identify the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model.

[0118] In this embodiment, the data acquisition module 101 includes an eye ultrasound acquisition submodule 1011, an intraocular pressure data acquisition submodule 1012, and a data transmission module 1013. The eye ultrasound acquisition submodule 1011 and the intraocular pressure data acquisition submodule 1012 are both communicatively connected to the data transmission module 1013. The data transmission module 1013 can be an RS485 module, an Ethernet module, or a similar module. Figure 3 As shown. Among them:

[0119] (1) Figure 4 As shown, the eye ultrasound acquisition submodule 1011 includes:

[0120] Ultrasonic transducer, array-type surrounds the eye, operating frequency 2-10MHz, multi-angle transmission and reception of ultrasonic signals;

[0121] Signal transmitting and receiving module, controls the ultrasonic transducer to transmit pulses and receive reflected echo signals;

[0122] Echo acquisition module, converts analog echo signals into digital signals, with a sampling frequency of 15-45MHz;

[0123] A microprocessor for preprocessing collected data;

[0124] A communication interface module, used to transmit the collected ultrasonic signal to an external device, which is transmitted to the data transmission module in the present invention;

[0125] The power module uses a rechargeable battery.

[0126] (6) an intraocular pressure data acquisition submodule, using a tonometer, including but not limited to an applanation tonometer, an indentation tonometer, a non-contact tonometer, and a dynamic contour tonometer;

[0127] (7) A data transmission module, which is used to send the eye data collected by the eye ultrasound acquisition submodule and the intraocular pressure data acquisition submodule to the server for data analysis and modeling.

[0128] In this embodiment, an intraocular pressure display module 106 is further included to display intraocular pressure measurement results and related information in the form of charts and data lists, including but not limited to three-dimensional dynamic charts, line charts, and bar charts.

[0129] In summary, the ultrasound- and deep learning-based intraocular pressure measurement method and system of this invention utilizes a variety of advanced technologies to achieve accurate intraocular pressure measurement and analysis. Its core principle is to collect ocular ultrasound data and intraocular pressure data using a head-mounted ultrasound device, leveraging the non-invasive and penetrating nature of ultrasound technology to obtain information about the internal structure of the eye, while simultaneously using a tonometer to obtain accurate intraocular pressure values.

[0130] First, in terms of data processing, an image processing algorithm based on spatiotemporal characteristics is used to reconstruct the ultrasound data sequence into four-dimensional ultrasound. By integrating the temporal dimension and using optical flow to track ocular tissue movement, the uniformity and accuracy of the time series are ensured, and the tissue displacement vector field is obtained. In the spatial dimension, image registration is performed based on the ocular anatomy to achieve accurate alignment of different frames, thereby generating four-dimensional ultrasound data that clearly demonstrates the detailed dynamic changes of ocular tissue and intraocular pressure over time.

[0131] Secondly, a relationship model between four-dimensional ocular ultrasound data and intraocular pressure data was established. Samples were grouped through data annotation, and feature extraction and analysis were performed on the four-dimensional ultrasound data. This included calculating the thickness change rate of each ocular tissue region, parameters related to echo intensity change, and texture feature parameters. Correlation analysis was then performed with intraocular pressure values, and significant correlated features were screened to construct feature vectors. A mathematical model was then constructed using multiple linear regression to determine the quantitative correspondence between ultrasound data features and intraocular pressure values.

[0132] Then, an LSTM neural network IOP recognition model was constructed and trained based on the reconstructed 4D ultrasound data, IOP data, and their relationship model. Using data samples with established relationships, the model was divided into training, validation, and test sets. A specifically structured LSTM neural network was constructed, using the 4D ultrasound data sequence as input and IOP data as output. Using a stochastic gradient descent optimization algorithm, the model parameters were adjusted through backpropagation based on the discrepancy between the model predictions and the actual data, minimizing the loss function and ultimately training a highly accurate IOP recognition model.

[0133] Finally, the real-time collected and reconstructed ultrasound intraocular pressure data is fed into a trained LSTM neural network model to identify the patient's intraocular pressure and calculate dynamic changes. The intraocular pressure display module presents the results in intuitive charts and data tables, providing comprehensive, accurate, and dynamic intraocular pressure information for ophthalmic clinical diagnosis and research.

[0134] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for measuring intraocular pressure based on ultrasound and deep learning, characterized in that: The following steps are involved: Collect ocular ultrasound data and intraocular pressure data through head-mounted ultrasound equipment and intraocular pressure measurement equipment; The continuously collected ultrasound intraocular pressure data sequence is reconstructed by four-dimensional ultrasound to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye; Establish a relationship model between four-dimensional ocular ultrasound data and corresponding intraocular pressure data; By reconstructing four-dimensional ultrasound data of the internal structure of the eye, intraocular pressure data and relationship models, an LSTM neural network intraocular pressure recognition model is constructed and trained; Based on the trained LSTM neural network intraocular pressure recognition model, the patient's intraocular pressure is identified and displayed; The relationship model between the four-dimensional ultrasound data and the corresponding intraocular pressure data is established as follows: Data annotation: annotate the four-dimensional ultrasound data and the corresponding intraocular pressure data, and divide the samples into normal intraocular pressure group and patient intraocular pressure group according to the intraocular pressure value range; Feature extraction and analysis: extract features from four-dimensional ultrasound data, analyze the characteristic differences of ultrasound data of internal eye structures under different intraocular pressure conditions, and perform correlation analysis with intraocular pressure values ​​to determine the characteristic vector related to intraocular pressure; A relationship model was constructed using multiple linear regression to build a mathematical model between ultrasound data features and intraocular pressure values, with the feature vector as the independent variable and the intraocular pressure value as the dependent variable. The model parameters were determined by training the model, and a quantitative correspondence between ultrasound data and intraocular pressure was established. The feature extraction of the four-dimensional ultrasound data and the analysis of the feature differences of the ultrasound data of the internal structure of the eye under different intraocular pressure states are specifically as follows: Ocular tissue region division and positioning utilizes image segmentation algorithms to analyze ultrasound image features, not limited to grayscale and texture, to identify the boundaries and extents of ocular structures, including but not limited to the cornea, anterior chamber, lens, vitreous body, and retina. The cornea is positioned based on its arcuate structure and echo intensity; the anterior chamber is determined based on echo characteristics and its positional relationship with the cornea and lens; the lens is identified by its shape and echogenicity; the vitreous body is divided based on its echogenicity; and the retina is positioned based on the reflective characteristics of its interface with the choroid. Establish an independent coordinate system for each divided eye tissue area, and then accurately calculate and analyze the features in different areas; Thickness change rate calculation: For each ocular tissue area, the thickness change rate over time is calculated in the time dimension of the four-dimensional ultrasound data; Echo intensity change analysis: statistics on the echo intensity distribution of each eye tissue area in the four-dimensional ultrasound data, dividing the echo intensity into several intervals, and calculating the number or proportion of pixels in each interval at different times; Analyze the changing trend of the echo intensity distribution of ocular tissue during the process of intraocular pressure changes, and calculate the average echo intensity change and echo intensity change rate; Texture feature extraction and analysis: Using the gray-level co-occurrence matrix analysis method, we extract the texture features of ocular tissue ultrasound images, calculate multiple characteristic parameters of the gray-level co-occurrence matrix, including but not limited to contrast, correlation, energy, and entropy, and compare the differences in ocular tissue texture features under different intraocular pressure conditions; Feature correlation analysis and feature vector construction were performed to calculate the correlation coefficients between the extracted thickness change rate, echo intensity change-related parameters, texture feature parameters and intraocular pressure values. Based on the correlation analysis results, features with significant correlation with intraocular pressure were screened out and feature vectors were constructed.

2. The intraocular pressure measurement method based on ultrasound and deep learning according to claim 1, characterized in that The method of reconstructing the continuously collected ultrasound intraocular pressure data sequence by four-dimensional ultrasound is as follows: using an image processing algorithm based on spatiotemporal features to reconstruct the data sequence by four-dimensional ultrasound, assuming that the two-dimensional ultrasound image sequence has T frames, spatial coordinates (x, y), and the ultrasound signal intensity at point (x, y) of the t-th frame image is , obtain the eye tissue displacement vector field through the optical flow method , the reconstruction formula is: ； in is the time interpolation weight function.

3. The intraocular pressure measurement method based on ultrasound and deep learning according to claim 1, characterized in that The LSTM neural network intraocular pressure recognition model is constructed and trained by reconstructing the four-dimensional intraocular structure ultrasound data, intraocular pressure data and relationship model, specifically: The model was trained using ultrasound data and intraocular pressure data samples with established correspondence, and the sample data was divided into training set, validation set, and test set in a ratio of 7:2:1 or 8:1:1; Construct an LSTM neural network model with 1-3 LSTM layers, 30-100 hidden units per layer, followed by a fully connected layer for classification output. The model is trained using a 4D ultrasound data sequence of the internal structure of the eye as input and the corresponding intraocular pressure data as output. The stochastic gradient descent optimization algorithm was used to adjust the model weights and bias parameters through the back propagation algorithm according to the difference between the model prediction results and the actual intraocular pressure data to minimize the loss function.

4. The intraocular pressure measurement method based on ultrasound and deep learning according to claim 1, characterized in that The method for identifying the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model is as follows: ultrasonic intraocular pressure data collected and reconstructed in real time is input into the trained LSTM neural network intraocular pressure recognition model to determine whether the intraocular pressure is normal and whether there is any abnormal change trend, and at the same time calculate the dynamic change indicators of intraocular pressure, including the intraocular pressure fluctuation amplitude and the intraocular pressure change frequency.

5. The intraocular pressure measurement method based on ultrasound and deep learning according to claim 1, characterized in that The correlation coefficient between the thickness change rate, echo intensity change related parameters, texture feature parameters and intraocular pressure values ​​is calculated and extracted, specifically using the Pearson correlation coefficient method, and the formula is: ； Where n is the number of samples, and are the observed values ​​of variables A and B respectively, and are the means of variables A and B, respectively. Indicates thickness change rate, echo intensity change related parameters, texture feature parameters, Represents The corresponding intraocular pressure value.

6. An intraocular pressure measurement system based on ultrasound and deep learning, characterized by: include: Data acquisition module, used to collect eye ultrasound data and intraocular pressure data; The four-dimensional reconstruction module is used to perform four-dimensional ultrasound reconstruction on the continuously collected ultrasound intraocular pressure data sequence to obtain four-dimensional ultrasound data reflecting the three-dimensional dynamic changes of the internal structure of the eye; Ultrasound and intraocular pressure relationship building module, used to establish a relationship model between four-dimensional eye ultrasound data and corresponding intraocular pressure data; The neural network model training module is used to build and train an LSTM neural network intraocular pressure recognition model using reconstructed four-dimensional ultrasound data of the internal structure of the eye, intraocular pressure data, and a relationship model; The intraocular pressure analysis and recognition module is used to identify the patient's intraocular pressure based on the trained LSTM neural network intraocular pressure recognition model; The ultrasound and intraocular pressure relationship building module establishes a relationship model between four-dimensional eye ultrasound data and corresponding intraocular pressure data, specifically: Data annotation: annotate the four-dimensional ultrasound data and the corresponding intraocular pressure data, and divide the samples into normal intraocular pressure group and patient intraocular pressure group according to the intraocular pressure value range; Feature extraction and analysis: extract features from four-dimensional ultrasound data, analyze the characteristic differences of ultrasound data of internal eye structures under different intraocular pressure conditions, and perform correlation analysis with intraocular pressure values ​​to determine the characteristic vector related to intraocular pressure; A relationship model was constructed using multiple linear regression to build a mathematical model between ultrasound data features and intraocular pressure values, with the feature vector as the independent variable and the intraocular pressure value as the dependent variable. The model parameters were determined by training the model, and a quantitative correspondence between ultrasound data and intraocular pressure was established. The feature extraction of the four-dimensional ultrasound data and the analysis of the feature differences of the ultrasound data of the internal structure of the eye under different intraocular pressure states are specifically as follows: Ocular tissue region division and positioning utilizes image segmentation algorithms to analyze ultrasound image features, not limited to grayscale and texture, to identify the boundaries and extents of ocular structures, including but not limited to the cornea, anterior chamber, lens, vitreous body, and retina. The cornea is positioned based on its arcuate structure and echo intensity; the anterior chamber is determined based on echo characteristics and its positional relationship with the cornea and lens; the lens is identified by its shape and echogenicity; the vitreous body is divided based on its echogenicity; and the retina is positioned based on the reflective characteristics of its interface with the choroid. Establish an independent coordinate system for each divided eye tissue area, and then accurately calculate and analyze the features in different areas; Thickness change rate calculation: For each ocular tissue area, the thickness change rate over time is calculated in the time dimension of the four-dimensional ultrasound data; Echo intensity change analysis: statistics on the echo intensity distribution of each eye tissue area in the four-dimensional ultrasound data, dividing the echo intensity into several intervals, and calculating the number or proportion of pixels in each interval at different times; Analyze the changing trend of the echo intensity distribution of ocular tissue during the process of intraocular pressure changes, and calculate the average echo intensity change and echo intensity change rate; Texture feature extraction and analysis: Using the gray-level co-occurrence matrix analysis method, we extract the texture features of ocular tissue ultrasound images, calculate multiple characteristic parameters of the gray-level co-occurrence matrix, including but not limited to contrast, correlation, energy, and entropy, and compare the differences in ocular tissue texture features under different intraocular pressure conditions; Feature correlation analysis and feature vector construction were performed to calculate the correlation coefficients between the extracted thickness change rate, echo intensity change-related parameters, texture feature parameters and intraocular pressure values. Based on the correlation analysis results, features with significant correlation with intraocular pressure were screened out and feature vectors were constructed.

7. The intraocular pressure measurement system based on ultrasound and deep learning according to claim 6, characterized in that The data acquisition module includes an eye ultrasound acquisition submodule, an intraocular pressure data acquisition submodule, and a data transmission module. The eye ultrasound acquisition submodule and the intraocular pressure data acquisition submodule are both communicatively connected to the data transmission module, wherein: (1) The eye ultrasound acquisition submodule includes: Ultrasonic transducer, array-type surrounds the eye, operating frequency 2-10MHz, multi-angle transmission and reception of ultrasonic signals; Signal transmitting and receiving module, controls the ultrasonic transducer to transmit pulses and receive reflected echo signals; Echo acquisition module, converts analog echo signals into digital signals, with a sampling frequency of 15-45MHz; A microprocessor for preprocessing the collected data; A communication interface module, used to send the collected ultrasonic signals to an external device; Power module, using rechargeable battery; (2) Intraocular pressure data acquisition submodule, using tonometers, including but not limited to applanation tonometer, indentation tonometer, non-contact tonometer and dynamic contour tonometer; (3) Data transmission module, used to send the eye data collected by the eye ultrasound acquisition submodule and the intraocular pressure data acquisition submodule to the server for data analysis and modeling.

8. The intraocular pressure measurement system based on ultrasound and deep learning according to claim 6, characterized in that It also includes an intraocular pressure display module, which displays intraocular pressure measurement results and related information in the form of charts and data lists, including but not limited to three-dimensional dynamic charts, line charts, and bar charts.

Images