An AI-based collaborative management method for ophthalmic medical data

Abstract

This application discloses an AI-based collaborative management method for ophthalmic medical data, relating to the field of data processing technology. The method includes: acquiring encrypted pre-test images and calculating a texture interference coefficient; pre-setting a first threshold; if the texture interference coefficient is not less than the first threshold, extracting feature points from each pre-test image and acquiring identical images from different periods as an evolutionary group; establishing a tracer curve; determining the final mutation zone based on the tracer curve; calculating the disturbance intensity of each image in the evolutionary group, and obtaining a damage index based on the final mutation zone and the disturbance intensity of each image; accurately quantifying the impact of encryption operations on the geometric parameters of vascular bifurcation points in ophthalmic medical image data through the texture interference coefficient and disturbance intensity, accurately measuring the actual damage level of the image data, ensuring the security and diagnostic accuracy of the image data; and synchronizing ophthalmic equipment data from multiple institutions to the cloud in real time, fully improving cross-disciplinary resource integration capabilities.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an AI-based collaborative management method for ophthalmic medical data. Background Technology

[0002] Ophthalmic images contain a lot of private data. When sharing or collaborating on data management, encryption or obfuscation is usually performed to ensure data security. However, ophthalmic medical data contains many subtle data features. For example, there are bifurcation points where retinal vessels (arteries or veins) split into two or more branches. The data contained at each bifurcation point is an important feature for medical diagnosis.

[0003] However, in existing technologies, when performing encryption operations on private areas, if important features of blood vessels, such as bifurcation points, are covered, an edge diffusion effect will occur, causing the pixels at the edge of the blood vessel to mix with the pixels of the surrounding tissue. Furthermore, the sharp angles at the bifurcation points will be smoothed, resulting in measurement angle deviations. When performing data sharing and collaborative processing, the receiver cannot accurately restore the image, leading to subtle deviations that affect the judgment results. Summary of the Invention

[0004] This application provides an AI-based collaborative management method for ophthalmic medical data, which solves the problem of neglecting the influencing factors of subtle features in the prior art, leading to recognition bias. It achieves the technical effect of accurately identifying feature changes and improving data security and recognition accuracy.

[0005] This application provides an AI-based collaborative management method for ophthalmic medical data, the method comprising:

[0006] S1: Obtain the features of the set of blood vessel bifurcation points and extract geometric parameters to form a normal database; select several image data from the normal database for encryption to obtain encrypted pre-test images, and calculate the texture interference coefficient;

[0007] S2: A first threshold is preset. The texture interference coefficient is compared with the first threshold. If it is less than the threshold, the disturbance intensity of each pre-test image is directly determined based on the texture interference coefficient. If the texture interference coefficient is not less than the first threshold, several feature points of each pre-test image are extracted and the same images from different periods are obtained as an evolution group. Based on the evolution group and each feature point, a tracer curve is established. The final mutation zone is determined based on the tracer curve. The disturbance intensity of each image in the evolution group is calculated, and the damage index is obtained based on the final mutation zone and the disturbance intensity of each image.

[0008] S3: Set an identification code based on the damage index, pre-set a repair table, determine the repair level and compensation mechanism by comparing the identification code with the repair table, and repair the encrypted image data according to the compensation mechanism; each damage index corresponds to a unique identification code.

[0009] Furthermore, the set features of vascular bifurcation points are obtained, and geometric parameters are extracted to form a normal database. This includes: collecting several healthy ophthalmic medical image data, extracting vascular network features and vascular bifurcation point features, extracting geometric parameters of vascular bifurcation points, calculating the normal range of each parameter, and forming a normal database. The normal database contains set features, the position coordinates of each bifurcation point, and metadata of the image to which it belongs. The geometric parameters include bifurcation angle, bifurcation diameter ratio, radius of curvature, and bifurcation symmetry.

[0010] Further, the texture interference coefficient is calculated, including: taking the blood vessel bifurcation point as the center, cutting out a central region and several neighborhoods, determining the index set, calculating the change of each index before and after encryption, obtaining the absolute difference of each index, and weighting and summing the absolute differences of multiple indices to obtain the texture interference coefficient; the index set includes contrast, spatial frequency attenuation rate, sharpness loss rate and homogeneity.

[0011] The process involves calculating the standard deviation of the grayscale difference between each neighboring region and the central region to obtain the contrast ratio; calculating the amplitude spectrum and total energy of the central region, defining a high-frequency region, calculating the high-frequency energy of the high-frequency region, and obtaining the spatial frequency attenuation rate; the total energy refers to the sum of the squares of all amplitudes in the amplitude spectrum, and the high-frequency energy refers to the sum of the squares of all amplitudes in the high-frequency region; calculating the gradient amplitude image of the central region, extracting edge pixels, selecting pixels with gradient amplitudes greater than the gradient threshold, calculating the standard deviation of the gradient amplitudes of all edge pixels, and obtaining the sharpness loss rate; quantizing the central region at the grayscale level, setting the parameters of the co-occurrence matrix: distance is 1, angle is 0 degrees; counting the occurrence frequency of pixel pairs to form a co-occurrence matrix, and calculating homogeneity based on the co-occurrence matrix.

[0012] Furthermore, the disturbance intensity is obtained by measuring the geometric parameters before and after encryption and calculating the rate of change of each bifurcation point in the pre-test image, and then performing a weighted sum based on the rate of change of the geometric parameters.

[0013] Furthermore, several feature points are extracted from each pre-test image, and the same images from different periods are obtained as evolutionary groups. This includes: using a random forest model to predict the sensitivity score of each bifurcation point, with a value range of [0,1]. For each image, the sensitivity scores are sorted in descending order, and several bifurcation points are selected as feature points in sequence. For each feature point, the same images from different periods are collected according to the image data type to form an evolutionary group.

[0014] Furthermore, the horizontal axis of the tracer curve represents time, calculated from the earliest acquisition time in the evolutionary group; the vertical axis represents a specific set of indicators for the feature points, set according to the core parameter characteristics of different diseases; and a smooth curve is drawn using linear interpolation based on the indicator values ​​of each feature point in all images of the evolutionary group as the tracer curve.

[0015] Further, the final anomaly zone is determined based on the tracer curve, including: for each feature point, calculating the rate of change of the index in the tracer curve; if the rate of change exceeds the second threshold for the first time, it is marked as the starting point; monitoring continues until the rate of change does not exceed the second threshold, then it is marked as the ending point; the anomaly duration is determined based on the starting point and the ending point, and the initial anomaly zone is determined based on the change trajectory; image data within the anomaly duration is acquired, and supplementary regions associated with the initial anomaly zone are identified; the final anomaly zone is formed based on the initial anomaly zone and the supplementary regions.

[0016] Furthermore, the supplementary region is based on the edge features of the initial anomaly region, and a segmentation algorithm is used to identify texture distortion regions, edge jagged artifacts, and local contrast attenuation bands in each image as supplementary regions.

[0017] Furthermore, the repair level is determined jointly based on the damage index and the characteristic distribution, which includes edge distribution, central distribution, and core distribution; the repair level includes three levels: mild, moderate, and severe; if the damage index is less than 0.3 and belongs to the edge distribution, it is set as mild; if the damage index is in the range [0.3, 0.6) or belongs to the central distribution, it is set as moderate; if the damage index is not less than 0.6 or belongs to the core distribution, it is determined as severe.

[0018] The feature distribution is based on the center of the final mutation zone as the origin, with the region radius normalized to 1. For each feature point, its coordinates are calculated, and the distance from the origin is calculated based on the coordinates. If the distance is greater than 0.7, it is determined to be an edge point; if the distance is not greater than 0.7 and not less than 0.4, it is determined to be a middle point; if the distance is less than 0.4, it is determined to be a core point. The proportion of different points is counted, and the one with the highest proportion is set as the corresponding feature distribution attribute.

[0019] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0020] By using texture interference coefficients and disturbance intensity, the impact of encryption operations on the geometric parameters of vascular bifurcation points in ophthalmic medical imaging data is precisely quantified. Tracer curves are used to dynamically track changes in feature points over time, identifying anomalous interference introduced by the encryption operation and ensuring data consistency over time. A damage index is calculated based on the final anomaly zone and the intensity of image disturbance to accurately measure the actual degree of damage to the image data. Furthermore, different repair levels and compensation mechanisms are set according to the damage index and feature distribution, employing multiple repair strategies to repair the encrypted image data, ensuring the security of the image data and diagnostic accuracy. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of an AI-based collaborative management method for ophthalmic medical data in an embodiment of the present invention. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] In ophthalmic imaging, retinal vessels (arteries or veins) intersect at points where they branch into two or more branches. Each intersection contains geometric information. When an artery and vein intersect, the artery typically compresses the vein, creating specific morphological features including: vein thinning (the percentage reduction in vein diameter at the intersection); angular offset (the angle of path deviation caused by compression); and compression marks (specific morphological features formed around the intersection). When performing encryption operations on private areas, if the bifurcation point is covered, an edge diffusion effect occurs, causing the pixels at the vessel edge to blend with the surrounding tissue pixels. Sharp angles at the bifurcation point are smoothed, leading to measurement angle deviations. For example, with linear blurring, blurring along a specific direction causes apparent displacement of the vessel, blurring subtle thinning features at the compression point. When using block encryption algorithms such as AES or DES, converting pixel values ​​into encrypted data blocks disrupts the continuity of vessel edges, damaging geometric features, including bifurcation point position offsets, angle measurement errors, and vessel diameter measurement errors. Furthermore, the vessel edges are prone to jaggedness, causing the gradual changes in features at the compression point to be altered in a stepped manner.

[0025] On the one hand, existing methods cannot accurately evaluate encrypted images and identify anomalous areas; on the other hand, during data sharing and collaborative processing, the recipient cannot accurately restore the image, leading to subtle deviations that affect the judgment results. Therefore, this application provides the following solution to the above problems.

[0026] Example 1: As Figure 1 As shown, an AI-based collaborative management method for ophthalmic medical data includes:

[0027] S1: Obtain the features of the set of blood vessel bifurcation points and extract geometric parameters to form a normal database; select several image data from the normal database for encryption to obtain encrypted pre-test images, and calculate the texture interference coefficient.

[0028] In some embodiments, a number of healthy ophthalmic medical image data (e.g., retinal images) are collected, covering different age groups, genders, and ethnicities. A U-Net-based vascular segmentation algorithm is used to extract vascular network features, and a bifurcation point detection algorithm (based on an improved Harris corner detection algorithm) is applied to identify vascular bifurcation point features. Geometric parameters of the vascular bifurcation points are extracted, and the normal range for each parameter is calculated (e.g., the normal range for bifurcation angle is 65°-110°), forming a normal database. This normal database contains ensemble features (including vascular network features and vascular bifurcation point features), the location coordinates of each bifurcation point, and metadata of the associated image. The geometric parameters include, but are not limited to, bifurcation angle (the angle between the main vessel and the branch vessel), bifurcation diameter ratio, radius of curvature, and bifurcation symmetry.

[0029] In the normal database, based on the geometric parameters of the blood vessel bifurcation points, the most characteristic image data (such as images that can identify the corresponding disease through the bifurcation angle) are selected and encrypted (by default, all image data use the same encryption method). The encrypted pre-test images are obtained, the texture interference coefficient is calculated, and the blurring degree and change status of key points are identified.

[0030] In some embodiments, the texture interference coefficient is calculated as follows: taking the blood vessel bifurcation point as the center, a central region and several neighborhoods are extracted to determine the index set. The changes of each index before and after encryption are calculated respectively to obtain the absolute difference of each index. The absolute differences of multiple indices are weighted and summed to obtain the texture interference coefficient. The index set includes contrast, spatial frequency attenuation rate, sharpness loss rate and homogeneity.

[0031] Specifically, taking the bifurcation point of the blood vessel as the center, a central region and several neighboring regions are extracted. When selecting the region, the area of ​​the central region should be larger than the area of ​​the neighboring regions. For example, a central region of 16×16 pixels and eight neighboring regions of 3×3 pixels are selected. The central region is then encrypted (using AES encryption, with a block size of 16x16 pixels) to obtain the encrypted central region image.

[0032] In some embodiments, the original central region image is used before encryption, and the following feature values ​​of the central region before and after encryption are calculated respectively: the contrast calculation includes: calculating the average gray value of the central region, and calculating the average gray value of each neighboring region; calculating the standard deviation of the gray value difference between each neighboring region and the central region, calculating before and after encryption respectively, and taking the absolute value of the difference between the standard deviation of the gray value difference after encryption and the standard deviation of the gray value difference before encryption as the absolute difference of contrast.

[0033] The spatial frequency attenuation rate calculation includes: performing a two-dimensional Fourier transform on the central region to obtain the amplitude spectrum, calculating the amplitude spectrum and total energy of the central region; defining a high-frequency region, which refers to the portion of the amplitude spectrum with a radius greater than 1 / 4 of the image size (e.g., if the image size is 32 pixels, the radius is > 8 pixels), calculating the high-frequency energy of the high-frequency region, and obtaining the spatial frequency attenuation rate; the spatial frequency attenuation rate is equal to 1 minus the ratio of high-frequency energy to total energy, calculated before and after encryption, and the absolute value of the difference is taken as the absolute difference of the spatial frequency attenuation rate. The total energy refers to the sum of the squares of all amplitudes in the amplitude spectrum, and the high-frequency energy refers to the sum of the squares of all amplitudes in the high-frequency region. In the Fourier transform, the amplitude represents the intensity of the frequency component. For an image, the amplitude of each frequency component reflects the salience of that frequency texture. For example, high high-frequency amplitude indicates rich detail (such as the edge of a blood vessel), while high low-frequency amplitude indicates a smooth region. After the image undergoes a two-dimensional Fourier transform, the set of amplitudes of all frequency components is called the amplitude spectrum. The amplitude spectrum is a matrix, where each element corresponds to the amplitude of a frequency point. For example, if the image size is 32×32 pixels, the frequency domain size will be the same after the Fourier transform. Define the high-frequency region, calculate the total energy (the sum of squares of all amplitudes in the amplitude spectrum) and the high-frequency energy (the sum of squares of the amplitudes of all frequency points within the high-frequency region), and the spatial frequency attenuation rate = 1 - (high-frequency energy / total energy). The closer this value is to 1, the more severe the loss of high-frequency details.

[0034] The sharpness loss rate calculation includes: using the Sobel operator to calculate the gradient magnitude image of the central region (obtaining the gradient magnitude of each pixel); extracting edge pixels and selecting pixels with gradient magnitudes greater than the gradient threshold (the gradient threshold is dynamically set based on all gradient magnitudes and historical experimental data, for example, set to 10% of the maximum gradient magnitude); calculating the standard deviation of the gradient magnitudes of all edge pixels, calculated before and after encryption, and the absolute value of the difference in sharpness index before and after encryption is used as the absolute difference in sharpness loss rate, representing the degree of sharpness change.

[0035] Homogeneity calculation includes: For the central region, homogeneity is calculated using a co-occurrence matrix (GLCM). This involves quantizing the central region to 8 bits, setting the parameters of the GLCM to a distance of 1 and an angle of 0 degrees, and counting the frequency of occurrence of pixel pairs (i, j) to form a GLCM matrix, where i and j are gray levels, both ranging from 0 to 255. The homogeneity formula is as follows: sum_{i,j}[GLCM(i,j) / (1+|ij|)], where |ij| is the gray level difference. The smaller the difference, the smaller the denominator, and the larger this term. High homogeneity indicates that there are many pixel pairs with similar gray levels in the image, resulting in uniform texture. Homogeneity is calculated before and after encryption, and the absolute value of the difference between the two is taken as the absolute difference of homogeneity.

[0036] In some embodiments, the contrast ratio measures the degree of brightness difference in a local area of ​​an image. High contrast indicates significant brightness variations within the area, which helps in identifying details such as vascular structures; low contrast makes the image appear flat and loses detail. Encryption operations may reduce contrast, thereby affecting the visibility of vascular bifurcation points. The spatial frequency attenuation rate measures the degree of loss of high-frequency components (corresponding to detail information) in the image. A high attenuation rate indicates severe loss of detail information, leading to image blurring. Encryption operations may weaken high-frequency information, affecting the accuracy of lesion detection. The sharpness loss rate quantifies the clarity of image edges. A high loss rate indicates blurred edges, affecting the identification of vascular boundaries. Encryption operations may blur edges, reducing the accuracy of geometric measurements. The homogeneity measures the uniformity of image texture. High homogeneity indicates smooth and uniform texture; low homogeneity indicates coarse and highly variable texture. Encryption operations may alter texture uniformity, affecting the consistency of texture analysis.

[0037] In some embodiments, the texture interference coefficient is a weighted sum of the four absolute differences mentioned above. ,in, These are the absolute differences in contrast, spatial frequency attenuation rate, sharpness loss rate, and homogeneity (i.e., the absolute values ​​of the differences before and after encryption). These are the corresponding weight coefficients, and the sum of the four weight coefficients is 1. The specific weight coefficients are adaptively adjusted according to the image resolution and blood vessel density. For example, for high-resolution images (such as 4K), contrast is more important, so the weights are set as: w1=0.4, w2=0.2, w3=0.2, w4=0.2; for low-resolution images, the weights are evenly distributed: w1=w2=w3=w4=0.25; if the blood vessel density is high (calculated by the blood vessel pixel ratio through a blood vessel segmentation algorithm), the sharpness loss rate is more important, so w3 is increased (for example, when the blood vessel density is >30%, w3=0.3, and other weights are adjusted accordingly). The weight adjustment rules are determined experimentally to ensure that TIC has the highest correlation with visual interference.

[0038] S2: A first threshold is preset. The texture interference coefficient is compared with the first threshold. If it is less than the threshold, the perturbation intensity of each pre-test image is directly determined based on the texture interference coefficient. The perturbation intensity refers to the degree of influence of the encryption operation on the geometric parameters of the bifurcation point. For example, it can cause the pixels at the edge of the blood vessel to be mixed with the pixels of the surrounding tissue, and the sharp angles at the bifurcation point are smoothed, leading to problems such as measurement angle deviation.

[0039] Specifically, for each bifurcation point in the pre-trial image, the geometric parameters before and after encryption are measured and their rate of change is calculated. The perturbation intensity is obtained by weighted summation based on the rate of change of the geometric parameters. For example, the values ​​of bifurcation angle, diameter ratio, radius of curvature, and symmetry are measured before and after encryption, and the relative rate of change of each parameter is calculated. The relative rate of change is the ratio of the absolute value of the difference between the values ​​before and after encryption to the value after encryption. This application does not limit the specific types of geometric parameters; they need to be set according to the core parameter characteristics corresponding to the relevant disease. When calculating the weighted summation, the corresponding data is pre-normalized to unify the data units, and the weight coefficient of each parameter is allocated based on clinical importance (for example, the bifurcation angle is most important for diagnosis, so its corresponding weight can be appropriately increased).

[0040] In some embodiments, the first threshold is preset based on historical experimental data. For example, 1000 healthy retinal images are collected, covering different resolutions (720p to 4K) and vascular densities. Each image is AES encrypted (block size 16×16), and the texture interference coefficient before and after encryption is calculated. Experts are invited to visually evaluate the encrypted images and give scores. The higher the score, the greater the interference. Images with expert scores greater than 1.5 are marked as "significant interference". A curve showing the relationship between the texture interference coefficient and the expert score is plotted. It is found that when the texture interference coefficient is greater than 0.1, 90% of the images are rated as "significant interference" by experts. Through curve analysis, the optimal threshold for the texture interference coefficient is determined to be 0.1, indicating that the texture interference exceeds the acceptable range for the human eye and affects the accuracy of diagnosis. In this application, the first threshold is set to 0.1. If it is greater than 0.1, it indicates significant texture interference, requiring further evaluation.

[0041] In some embodiments, if the texture interference coefficient is not less than a first threshold, several feature points of each pre-test image are extracted and the same images from different periods are obtained as an evolution group; based on the evolution group and each feature point, a tracer curve is established; the final mutation region is determined according to the tracer curve; the disturbance intensity of each image in the evolution group is calculated, and the damage index is obtained according to the final mutation region and the disturbance intensity of each image.

[0042] In some embodiments, a random forest model is used to predict the sensitivity score of each bifurcation point, with a value range of [0,1]. The training data comes from a large number of pre-test images (100 images), and the features include the geometric parameters, texture interference coefficient, and location information of the bifurcation point. The sensitivity is used to measure the ease with which each bifurcation point is subject to interference; the higher the sensitivity score, the more susceptible it is to interference and abnormal changes. For each image, the sensitivity scores are sorted in descending order, and several bifurcation points are selected sequentially as feature points; feature points are usually located in areas with dense blood vessels or near lesions. The specific number can be set according to the comprehensive sensitivity score, for example, the top 10% of bifurcation points, but no specific limitation is made here. For each feature point, the same images from different periods are collected to form an evolutionary group. The same images refer to images of the same disease or with the same features, with priority given to images of the same patient from different periods; the period range is determined according to the disease type in the image data: diabetic retinopathy is collected every 3 months, with 6 images collected (covering 18 months); glaucoma is collected every 6 months, with 4 images collected (covering 24 months).

[0043] In some embodiments, a tracer curve is used to track changes in feature points over time and identify the dynamic impact of encryption operations. The horizontal axis of the tracer curve represents time, calculated from the earliest acquisition time in the evolutionary group. The vertical axis represents a specific set of indicators for the feature points, such as bifurcation angle and texture interference coefficient, specifically set according to the core parameter characteristics of different diseases. Based on the indicator values ​​of each feature point in all images of the evolutionary group, a smooth curve is plotted using linear interpolation as the tracer curve; the curve contains data points from all time points. The tracer curve displays the trend of indicator changes over time. If the curve shows abrupt changes or a sustained deviation from the baseline, it indicates that the encryption operation has introduced abnormal interference. Encryption operations may disrupt the dynamic evolution tracking of lesions, leading to misdiagnosis. Through the tracer curve, the dynamic evolution process and the tracking of texture changes in feature points at different times can be identified, interference can be detected and corrected in a timely manner, and data consistency across the time dimension can be ensured.

[0044] In some embodiments, for each feature point, the rate of change of the index in the tracer curve (e.g., the monthly rate of change of the bifurcation angle) is calculated. If the rate of change exceeds a second threshold (e.g., 10%) for the first time, it is marked as the starting point. Monitoring continues until the rate of change does not exceed the second threshold, at which point it is marked as the ending point. The duration of the anomaly is determined based on the starting point and the ending point, and the initial anomaly zone is determined based on the change trajectory. Image data within the duration of the anomaly is acquired, and supplementary regions associated with the initial anomaly zone are identified. The final anomaly zone is formed based on the initial anomaly zone and the supplementary regions.

[0045] In some embodiments, identifying supplementary regions associated with the initial aberration region includes: using the U-Net segmentation algorithm to identify texture distortion regions, edge jagged artifacts, and local contrast attenuation bands in each image as supplementary regions based on the edge features of the initial aberration region; the segmentation model training data includes 1000 manually annotated distortion images; merging the initial aberration region with the segmentation-identified supplementary regions to form the final aberration region. During the merging process, connections are made based on the outermost data points to ensure coverage of all aberration features. Anatomical structures (such as vascular pathways) are considered during merging to ensure that the final aberration region does not disrupt key medical features. For example, in retinal images, the final aberration region often surrounds vascular bifurcation points or lesion points; therefore, vascular continuity is preferentially preserved during merging.

[0046] In some embodiments, the perturbation intensity of each image in the evolutionary group is calculated, and a damage index is obtained based on the final mutation region and the perturbation intensity of each image, including: the calculation formula for the damage index is as follows:

[0047]

[0048] in, It is a damage index used to measure the actual degree of damage to image data. The higher the damage index, the greater the damage caused to the image data of that feature by the encryption operation. It is the area of ​​the final mutation region; The total area of ​​the image is the same for all images. The total area of ​​the image refers to the overall area of ​​any one of the images. Both the area of ​​the final mutation zone and the total area of ​​the image are calculated by pixel counting. It is the average value of the disturbance intensity of the images in the evolutionary group. The disturbance intensity of each image is calculated and summed. The average value is obtained based on the sum of the disturbance intensities and the total number of images in the evolutionary group.

[0049] S3: During data sharing, the sender evaluates the encrypted image data to obtain a damage index; sets an identification code based on the damage index and pre-sets a repair table; the receiver receives the image data, compares the identification code with the pre-received repair table, determines the repair level and compensation mechanism, and repairs the encrypted image data according to the compensation mechanism.

[0050] In some embodiments, the identification code is a unique identifier used to identify the damage index of the current image data, and each damage index corresponds to a unique identification code. Based on the damage index (DI) and the distribution of feature point locations, the restoration is divided into three levels, and different restoration strategies are adopted to form a restoration table. The restoration table is pre-sent to the recipient, who compares the restoration table with the received image data to determine the damage index and the corresponding restoration level and compensation mechanism, and then performs the restoration to ensure the security of the image data and the accuracy of the diagnosis.

[0051] In some embodiments, the repair level is determined jointly based on the damage index and the characteristic distribution, wherein the characteristic distribution includes a peripheral distribution, a central distribution, and a core distribution; the repair level includes three levels: mild, moderate, and severe; if the damage index is less than 0.3 and belongs to the peripheral distribution, it is set to mild; if the damage index is in the range [0.3, 0.6) or belongs to the central distribution, it is set to moderate; if the damage index is not less than 0.6 or belongs to the core distribution, it is determined to be severe.

[0052] In some embodiments, the feature distribution is based on the center of the final mutation region as the origin, and the region radius is normalized to 1. For each feature point, its coordinates are calculated, and the distance from the origin is calculated based on the coordinates. If the distance is greater than 0.7, it is determined to be an edge point; if the distance is not greater than 0.7 and not less than 0.4, it is determined to be a middle point; if the distance is less than 0.4, it is determined to be a core point. The proportion of different points is counted, and the one with the highest proportion is set as the corresponding feature distribution attribute.

[0053] In some embodiments, for mild interpolation-based local repair, specifically, a 5×5 pixel healthy region around the final mutation region is extracted; a bicubic interpolation algorithm is used to reconstruct the pixels in the final mutation region based on the surrounding pixel values; after repair, a similarity index is calculated, and if it is greater than 97%, the repair is considered successful.

[0054] In some embodiments, for semantic inpainting based on generative adversarial networks with moderate settings, specifically, a GAN generator is trained to take the encrypted image as input and output the inpainted image; the discriminator takes the original image as a reference, sets the loss function and parameter values, performs local inpainting on the final mutated area, and keeps the non-mutated area unchanged; the similarity index and the rate of change of vascular geometric parameters are calculated, and if the similarity index is greater than 95% and the rate of change of vascular geometric parameters is less than 5%, then the inpainting is considered qualified.

[0055] In some embodiments, for heavily configured multi-model collaborative repair, specifically, combining GAN and physical models, firstly, GAN is used to repair the overall texture, then a vascular geometric constraint model is used to correct the bifurcation point angles. Based on a normal database of bifurcation points, an optimization algorithm is used to minimize the angle error, and the objective function is set as follows. ,in It is the corrected bifurcation angle. It is the median of the normal angle range (assuming the normal range is (65°, 110°), with a median of 87.5 degrees); calculate the similarity index and the bifurcation angle error. If the similarity index is greater than 93% and the bifurcation angle error is less than 3%, then the repair is considered qualified.

[0056] In some embodiments, if the repair is unsatisfactory, it is repeated until the maximum number of repairs is reached (e.g., five times). If it is still unsatisfactory, a request is made to re-share the data. After the repair is completed, the damage index reduction rate is calculated, and clinical validation can also be performed, with multiple physicians assessing the diagnostic consistency of the repaired images.

[0057] In this embodiment, the impact of encryption operations on the geometric parameters of vascular bifurcation points in ophthalmic medical image data is precisely quantified by using texture interference coefficients and disturbance intensity. Tracer curves are used to dynamically track changes in feature points over time to identify abnormal interference introduced by the encryption operation, ensuring data consistency over time. A damage index is calculated based on the final anomaly zone and image disturbance intensity to accurately measure the actual degree of damage to the image data. Furthermore, different repair levels and compensation mechanisms are set according to the damage index and feature distribution, employing multiple repair strategies to repair the encrypted image data, ensuring the security of the image data and diagnostic accuracy.

[0058] This invention addresses the challenge of accurately assessing the impact of encryption operations on the quality of ophthalmic medical imaging data, providing a basis for secure data sharing. It further overcomes the difficulty in tracking the dynamic evolution of image data caused by encryption operations, ensuring data reliability over time. It resolves the issue of not being able to effectively measure the actual degree of damage to image data after encryption, providing a reference for data restoration. Multiple restoration strategies are offered for different degrees of damage, addressing the difficulties in restoring encrypted image data.

[0059] In this embodiment, ophthalmic equipment data (such as fundus cameras, OCT, tonometers, etc.) from primary clinics, regional hospitals, and tertiary institutions are further synchronized to the cloud in real time, forming an eye health database covering the entire life cycle. Combining image data such as optical coherence tomography (OCT), fundus photography, and fluorescein angiography (FFA), a deep learning model is used to automatically classify and predict the risks of more than 10 diseases, including diabetic retinopathy and glaucoma. Based on patients' historical data (such as myopia progression, family history, and systemic diseases), a personalized eye health risk map is generated. For example, for adolescent myopia patients, AI automatically pushes personalized intervention plans such as orthokeratology lens fitting suggestions and reminders for outdoor activity time. This further enables secure data sharing across institutions within the ophthalmic medical consortium, and encrypts and shares ophthalmic medical data images according to the technical solutions of the above embodiments. The multilateral network of the medical consortium platform improves work efficiency, fully enhances cross-sectoral resource integration capabilities, ensures data security, guarantees medical safety and quality, and fully leverages the value of the data.

[0060] To further enhance security during data sharing, this application selects the optimal encryption operation for different image data characteristics based on the execution effects of the above embodiments, considering different encryption methods and corresponding damage indices, thereby reducing image damage while ensuring data security. Preferably, the CKKS homomorphic encryption scheme is used to generate a public key (PK) and a private key (SK). The public key is used to encrypt data, while the private key is securely stored by the data owner and not shared. The key length is set to 128 bits to balance security and computational efficiency. Image pixel values ​​are normalized to the [0,1] range and quantized into 16-bit floating-point numbers. The pixel matrix is ​​encrypted using the public key PK to obtain ciphertext data. The encryption process is executed in parallel on the GPU, with a single image encryption time ≤50ms (4K resolution). A 32×32 pixel region (ciphertext state) is extracted centered on the blood vessel bifurcation point, requiring no decryption. Neighboring regions are also extracted in ciphertext form.

[0061] Homomorphic texture feature calculation includes: Contrast calculation: In the encrypted state, the average gray value of the central region is calculated (achieved through homomorphic addition and multiplication); the gray value difference between the neighborhood and the central region is calculated, and the standard deviation is obtained (square root of the variance after homomorphic calculation). Spatial frequency attenuation rate calculation: The amplitude spectrum is calculated through homomorphic Fourier transform, supporting frequency domain transformation in the encrypted state; the proportion of high-frequency energy is calculated (homomorphic multiplication and summation) to obtain the attenuation rate. Sharpness loss rate and homogeneity calculation: The gradient magnitude is calculated using the homomorphic Sobel operator (approximate calculation), and edge detection is completed in the encrypted state; homogeneity is calculated through the homomorphic version of the gray-level co-occurrence matrix, and the pixel pair frequency is statistically analyzed. Texture interference coefficient calculation: The absolute difference of the above feature values ​​is calculated in the encrypted state, and the texture interference coefficient in encrypted form is obtained by weighted summation. The texture interference coefficient is decrypted using the private key SK to obtain the plaintext value (only the final result is decrypted, and the intermediate data is always encrypted).

[0062] Feature point selection and evolutionary group construction are both performed in encrypted form, and the index calculation of the tracer curve is completed through homomorphic operations. During final anomaly region identification and damage index calculation, only the final result is decrypted, ensuring that the original data is encrypted throughout the process. The receiver obtains the encrypted image and damage index (plaintext) and selects a strategy based on the repair table. Repair operations (such as interpolation or GAN) are performed on the decrypted data, but the repair parameters are optimized through homomorphic computation, reducing manual intervention and further ensuring data security.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based collaborative management method for ophthalmic medical data, characterized in that, include: S1: Obtain the features of the set of blood vessel bifurcation points and extract geometric parameters to construct a normal database; Several image data are selected from the normal database and encrypted to obtain the encrypted test image. The texture interference coefficient is then calculated. The process of calculating the texture interference coefficient includes: taking the blood vessel bifurcation point as the center, cutting out a central region and several neighborhoods, determining the index set, calculating the change of each index before and after encryption, obtaining the absolute difference of each index, and weighting and summing the absolute differences of multiple indices to obtain the texture interference coefficient. The set of metrics includes contrast, spatial frequency attenuation rate, sharpness loss rate, and homogeneity; In this process, the standard deviation of the grayscale difference between each neighboring region and the central region is calculated to obtain the contrast. Calculate the amplitude spectrum and total energy of the central region, define the high-frequency region, calculate the high-frequency energy of the high-frequency region, and obtain the spatial frequency attenuation rate; the total energy refers to the sum of the squares of all amplitudes in the amplitude spectrum, and the high-frequency energy refers to the sum of the squares of all amplitudes in the high-frequency region. Calculate the gradient magnitude image of the central region, extract edge pixels, select pixels with gradient magnitude greater than the gradient threshold, calculate the standard deviation of the gradient magnitude of all edge pixels, and obtain the sharpness loss rate. The central region is quantized to grayscale, and the parameters of the co-occurrence matrix are set as follows: distance is 1 and angle is 0 degrees; the occurrence frequency of pixel pairs is counted to form a co-occurrence matrix, and homogeneity is calculated based on the co-occurrence matrix. S2: A first threshold is preset. The texture interference coefficient is compared with the first threshold. If it is less than the threshold, the disturbance intensity of each pre-test image is directly determined based on the texture interference coefficient. If the texture interference coefficient is not less than the first threshold, several feature points of each pre-test image are extracted and the same images from different periods are obtained as an evolution group. Based on the evolution group and each feature point, a tracer curve is established. The final mutation zone is determined based on the tracer curve. The disturbance intensity of each image in the evolution group is calculated, and the damage index is obtained based on the final mutation zone and the disturbance intensity of each image. S3: Set an identification code based on the damage index, pre-set a repair table, determine the repair level and compensation mechanism by comparing the identification code with the repair table, and repair the encrypted image data according to the compensation mechanism; each damage index corresponds to a unique identification code.

2. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The process involves acquiring features of a set of vascular bifurcation points and extracting geometric parameters to construct a normal database. This includes collecting several healthy ophthalmological medical images, extracting vascular network features and vascular bifurcation point features, extracting geometric parameters of the vascular bifurcation points, calculating the normal range of each parameter, and forming a normal database. The normal database contains set features, the location coordinates of each bifurcation point, and metadata of the image to which it belongs; The geometric parameters include bifurcation angle, bifurcation diameter ratio, radius of curvature, and bifurcation symmetry.

3. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The disturbance intensity is obtained by measuring the geometric parameters of each bifurcation point in the pre-test image before and after encryption, calculating their rate of change, and then performing a weighted sum based on the rate of change of the geometric parameters.

4. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, Extract several feature points from each pre-test image and obtain the same images from different periods as evolutionary groups, including: using a random forest model to predict the sensitivity score of each bifurcation point, with a value range of [0,1]. For each image, the sensitivity scores are sorted in descending order, and several bifurcation points are selected as feature points in sequence. For each feature point, the same images from different periods are collected according to the image data type to form an evolutionary group.

5. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The horizontal axis of the tracer curve represents time, calculated from the earliest acquisition time in the evolutionary group; the vertical axis represents a specific set of indicators for the feature points, set according to the core parameter characteristics of different diseases; and a smooth curve is plotted using linear interpolation based on the indicator values ​​of each feature point in all images of the evolutionary group as the tracer curve.

6. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The process of determining the final mutation zone based on the tracer curve includes: for each feature point, calculating the rate of change of the index in the tracer curve; if the rate of change exceeds the second threshold for the first time, it is marked as the starting point; monitoring continues until the rate of change does not exceed the second threshold, then it is marked as the ending point; the mutation duration is determined based on the starting point and the ending point, and the initial mutation zone is determined based on the change trajectory. Acquire image data within the duration of the anomaly and identify supplementary regions associated with the initial anomaly zone; The final mutation region is formed based on the initial mutation region and the supplementary region.

7. The AI-based collaborative management method for ophthalmic medical data as described in claim 6, characterized in that, The supplementary region is based on the edge features of the initial anomaly region. A segmentation algorithm is used to identify texture distortion areas, edge jagged artifacts, and local contrast attenuation bands in each image, which are then used as supplementary regions.

8. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The formula for calculating the damage index is as follows: in, It is a damage index used to measure the actual degree of damage to image data. The higher the damage index, the greater the damage caused to the image data by the encryption operation. It is the area of ​​the final mutation region; It is the total area of ​​the image; It is the average value of the perturbation intensity of the images in the evolutionary group. The perturbation intensity of each image is calculated and summed. The average value is obtained based on the sum of the perturbation intensities and the total number of images in the evolutionary group. The area of ​​the final mutation zone and the total area of ​​the image are both calculated by pixel counting.

9. The AI-based collaborative management method for ophthalmic medical data as described in claim 1, characterized in that, The repair level is determined by both the damage index and the characteristic distribution, which includes edge distribution, central distribution, and core distribution; the repair level includes three levels: mild, moderate, and severe. If the damage index is less than 0.3 and belongs to the marginal distribution, it is set as mild; If the damage index is in the range [0.3, 0.6) or belongs to the middle distribution, it is set to moderate. If the damage index is not less than 0.6 or belongs to the core distribution, it is judged as severe; The characteristic distribution is based on the center of the final mutation region as the origin, and the region radius is normalized to 1. For each feature point, calculate its coordinates and the distance from the origin based on the coordinates. If the distance is greater than 0.7, it is determined to be an edge point; if the distance is not greater than 0.7 and not less than 0.4, it is determined to be a middle point; if the distance is less than 0.4, it is determined to be a core point. Calculate the proportion of different points and set the one with the highest proportion as the corresponding feature distribution attribute.

Images