A method, device and medium for vision monitoring based on fundus OCT images
By employing a vision monitoring method based on fundus OCT images, utilizing an attention-enhanced dynamic U-shaped segmentation network and multi-dimensional biomarker extraction, combined with encryption algorithms and blockchain storage, this approach addresses the issues of insufficient segmentation accuracy and privacy security in existing vision monitoring technologies, achieving high-precision prediction of vision change trends and early intervention.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO PENGFENGCHENG MEDICAL TECHNOLOGY CO LTD
- Filing Date
- 2026-03-17
- Publication Date
- 2026-07-31
AI Technical Summary
Existing vision monitoring solutions suffer from complex data processing, poor equipment compatibility, insufficient segmentation accuracy, inability to adaptively focus on macular lesions, insufficient feature extraction accuracy, and failure to integrate multi-dimensional biomarkers, resulting in a lack of predictive ability for vision change trends and making it difficult to meet the needs of early detection and early intervention.
A vision monitoring method based on fundus OCT images is adopted. Preprocessing is performed by identifying device type and noise characteristics. Attention-enhanced dynamic U-shaped segmentation network is used to segment the retinal interlayer boundaries, extract multi-dimensional biomarkers, and data security and privacy protection are achieved through encryption algorithms and blockchain storage. Combined with a time-series prediction model, the future vision change trend is output.
It achieves subpixel-level segmentation of retinal layer boundaries, integrates personalized assessments of multi-dimensional biomarkers, improves monitoring accuracy, and ensures privacy and security through smart contracts and encryption strategies. It outputs future vision change trends and abnormal risks, enabling early detection and early intervention.
Smart Images
Figure CN122492545A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of healthcare informatics technology, and in particular to a vision monitoring method, device and medium based on fundus OCT images. Background Technology
[0002] Fundus OCT images can clearly show the microstructure of the retina and are the core data source for vision health monitoring. Their information and intelligent analysis has become an important support for ophthalmological clinical diagnosis.
[0003] In existing technologies, some vision monitoring solutions rely on multimodal data fusion to achieve detection, which has problems such as complex data processing, poor equipment compatibility, and high deployment costs. Another type of solution focuses on the analysis of single-modal fundus OCT images. Due to the use of fixed network structure and hole rate settings, it cannot adaptively focus on high-incidence areas of lesions such as the macula, resulting in insufficient feature extraction accuracy, poor thin-layer high-resolution segmentation effect, and poor boundary localization accuracy.
[0004] In addition, current monitoring systems mostly use traditional U-shaped networks for feature extraction, which has limited segmentation accuracy. Monitoring is based only on a single indicator of retinal thickness and does not integrate multi-dimensional biomarkers, resulting in a lack of ability to predict trends in vision changes and making it difficult to meet the needs of early clinical detection and intervention. Summary of the Invention
[0005] To address the technical problems mentioned above, this application provides a vision monitoring method, device, and medium based on fundus OCT images. The method includes: acquiring a single-modal three-dimensional image of the subject's fundus; identifying the device type and noise characteristics of the single-modal three-dimensional image; determining a denoising algorithm to preprocess the single-modal three-dimensional image based on the device type and noise characteristics; inputting the preprocessed single-modal three-dimensional image into a preset attention-enhanced dynamic U-shaped segmentation network to output retinal interlayer boundary segmentation results, and extracting multi-dimensional biomarkers based on the retinal interlayer boundary segmentation results; classifying the multi-dimensional biomarkers and the single-modal three-dimensional image according to preset classification rules, and encrypting and storing the classified multi-dimensional biomarkers and the single-modal three-dimensional image using a preset encryption algorithm; performing temporal registration of the retinal interlayer boundary segmentation results with preset historical time data according to a preset SIFT algorithm, and performing personalized evaluation based on preset multi-dimensional indicators and a preset temporal prediction model to output a preset future vision change trend.
[0006] In one example, the device type and noise features of the single-modal 3D image are identified. Based on the device type and noise features, a denoising algorithm is determined to preprocess the single-modal 3D image. Specifically, this includes: identifying the OCT device type corresponding to the single-modal 3D image using a preset device type identification algorithm; matching the corresponding noise pattern from a preset noise feature library based on the OCT device type; selecting a denoising algorithm that matches the noise pattern from a preset denoising algorithm pool; the denoising algorithm pool includes Gaussian filtering, median filtering, and wavelet denoising; and denoising the single-modal 3D image using the denoising algorithm to obtain a preprocessed single-modal 3D image.
[0007] In one example, a preprocessed single-modal 3D image is input into a pre-defined attention-enhanced dynamic U-shaped segmentation network, which outputs retinal interlayer boundary segmentation results. Specifically, this includes: pre-constructing an attention-enhanced dynamic U-shaped segmentation network; the network includes an encoding path, intermediate attention fusion layers, and a decoding path; each layer of the encoding path has a backbone and branch paths; the backbone is composed of cascaded residual modules to extract local detail features of the image; the branch paths are dilated convolutions, and the dilation rate is adjusted based on the retinal thickness distribution of the input image; the preprocessed single-modal 3D image is then input into a pre-defined attention-enhanced dynamic U-shaped segmentation network. The modal 3D image is input into the encoding path, and a deep feature map is output after passing through each layer of the encoding path. The deep feature map is input into the intermediate attention fusion layer, and the importance weight of each pixel in the deep feature map is calculated through a preset spatial attention mechanism to generate an attention-weighted feature map. The attention-weighted feature map is input into the decoding path to upsample the attention-weighted feature map through a deconvolution operation to obtain an upsampled weighted feature map. The upsampled weighted feature map is concatenated with the features from the attention-weighted feature map from the corresponding encoding layer to output the segmentation result of the retinal layer boundary.
[0008] In one example, multi-dimensional biomarkers are extracted based on the retinal interlayer boundary segmentation results. Specifically, this includes: calculating the distance between retinal layers based on their locations within the segmentation results to obtain retinal thickness distribution parameters; analyzing the texture features of the nerve fiber layer region within the segmentation results using a gray-level co-occurrence matrix to extract texture feature parameters including contrast, correlation, and energy; performing blood flow signal analysis on a preset region based on the original gray-level signal of the OCT image and the retinal interlayer boundary segmentation results to calculate retinal blood flow density parameters; calculating the foveal thickness and macular volume parameters based on the boundary location of the macular region within the segmentation results; and integrating the retinal thickness distribution parameters, nerve fiber layer texture feature parameters, retinal blood flow density parameters, foveal thickness, and macular volume parameters to obtain multi-dimensional biomarkers.
[0009] In one example, the multidimensional biomarkers and the single-modal 3D images are classified according to preset classification rules, and the classified multidimensional biomarkers and single-modal 3D images are encrypted and stored using a preset encryption algorithm. Specifically, this includes: dividing the multidimensional biomarkers and the single-modal 3D images into basic feature data, sensitive feature data, and original image data according to preset classification rules; the basic feature data includes the mean retinal thickness and overall structural integrity parameters; the sensitive feature data includes lesion area coordinates and texture feature parameters; the original image data is the acquired single-modal 3D image; encrypting the basic feature data using the AES symmetric encryption algorithm; encrypting the sensitive feature data using an algorithm combining SHDPPC-CP-ABE attribute-based encryption and homomorphic encryption; encrypting the original image data using the SHDPPC-CP-ABE attribute-based encryption algorithm; uploading the encrypted basic feature data, sensitive feature data, and original image data to the IPFS interplanetary file system for distributed storage to obtain the corresponding IPFS storage address; and writing the hash value of the IPFS storage address, the preset data access strategy, and the data classification tag into Hyperledger. The evidence record is obtained from the blocks of the Fabric consortium blockchain.
[0010] In one example, after writing the hash value of the IPFS storage address, access policy, and data classification label into a block of the Hyperledger Fabric consortium blockchain to obtain the evidence record, the method further includes: deploying a smart contract to monitor access behavior to encrypted basic feature data, sensitive feature data, and raw image data; when abnormal access behavior that meets preset warning rules is detected, a warning mechanism is triggered, and a warning notification is sent to preset contacts to suspend the permissions of the abnormal access account.
[0011] In one example, based on a preset SIFT algorithm, the retinal interlayer boundary segmentation result is temporally registered with preset historical time data. Then, based on preset multi-dimensional indicators and a preset temporal prediction model, a personalized assessment is performed, outputting a preset future vision change trend. Specifically, this includes: rigidly registering the retinal interlayer boundary segmentation result with the retinal interlayer boundary segmentation result at a preset historical time using the preset SIFT algorithm; calculating the displacement change matrix between the two images at the two time points to align the image data from multiple time points in spatial location; and calculating the temporal changes of multi-dimensional indicators based on the rigidly registered image data. These multi-dimensional indicators include the difference in retinal thickness and blood flow density between the retinal interlayer boundary segmentation result at the time of the segmentation and the preset historical time. The process involves: obtaining the rate of change and the coefficient of variation of the texture features of the nerve fiber layer; acquiring the subject's personal attribute information, including age, eye habits, and past medical history; retrieving a personalized normal threshold range matching the subject from a pre-constructed multi-dimensional indicator database based on the personal attribute information; comparing the temporal change of the multi-dimensional indicators with the personalized normal threshold range, and marking the indicator as an abnormal indicator when the temporal change of any indicator exceeds the corresponding threshold range; inputting multi-dimensional biomarker data at at least three time points, including the time of the retinal interlayer boundary segmentation result, into a pre-trained LSTM-based temporal prediction model; and outputting the subject's vision change trend within a preset future time period through the temporal prediction model, wherein the vision change trend includes the rate of change of the multi-dimensional indicators and the probability of abnormal risk.
[0012] In one example, the method further includes: receiving treatment suggestions entered by doctors according to preset doctor permission levels and authorization scopes; and displaying to subjects the retinal interlayer boundary segmentation results, the multidimensional biomarkers, the labeled abnormal indicators, the personalized assessment report, the vision change trend, and the treatment suggestions according to preset subject permission levels and authorization scopes.
[0013] On the other hand, embodiments of this application provide a vision monitoring device based on fundus OCT images, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the above-mentioned vision monitoring methods based on fundus OCT images.
[0014] On the other hand, embodiments of this application provide a non-volatile computer storage medium for vision monitoring based on fundus OCT images, which stores computer-executable instructions that can execute any of the above-mentioned vision monitoring methods based on fundus OCT images.
[0015] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This solution utilizes an attention-enhanced dynamic U-shaped segmentation network, incorporating dynamic hole rate adjustment and spatial attention mechanisms to achieve sub-pixel-level segmentation of retinal layer boundaries. It also integrates multi-dimensional biomarkers such as retinal thickness, blood flow density, and texture features to construct a personalized assessment model, improving monitoring accuracy. Simultaneously, a three-level data grading system is built, employing differentiated encryption strategies such as AES and SHDPPC-CP-ABE+ homomorphic encryption for basic features, sensitive features, and original images, respectively. This is combined with IPFS distributed storage and Hyperledger Fabric consortium blockchain for evidence storage, and real-time monitoring of abnormal access behavior triggers early warnings via smart contracts, balancing privacy and diagnostic efficiency. Furthermore, high-precision registration of multi-time-point images is achieved based on the SIFT algorithm, and an LSTM temporal prediction model integrates personalized features of the subject to output future vision change trends and abnormal risk probabilities, truly enabling early detection and early intervention. Attached Figure Description
[0016] To more clearly illustrate the technical solution of this application, some embodiments of this application will be described in detail below with reference to the accompanying drawings, in which: Figure 1 A schematic flowchart illustrating a vision monitoring method based on fundus OCT images provided in this application embodiment; Figure 2 The system of a vision monitoring method based on fundus OCT images provided in the embodiments of this application adopts a layered architecture design diagram; Figure 3 A system architecture diagram of a vision monitoring method based on fundus OCT images provided in this application embodiment; Figure 4 A schematic diagram of an attention-enhanced dynamic U-shaped segmentation network structure for a vision monitoring method based on fundus OCT images provided in this application embodiment; Figure 5 A flowchart illustrating the hierarchical privacy protection module of a vision monitoring method based on fundus OCT images, provided for embodiments of this application; Figure 6 This is a schematic diagram of the structure of a vision monitoring device based on fundus OCT images provided in an embodiment of this application. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0018] Some embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0019] Figure 1 This is a schematic flowchart illustrating a vision monitoring method based on fundus OCT images, provided as an embodiment of this application. This method can be applied to various business areas. Certain input parameters or intermediate results in this process can be manually adjusted to help improve accuracy.
[0020] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.
[0021] Based on this Figure 1 The process may include the following steps: S101: Acquire single-modal three-dimensional images of the subject's fundus.
[0022] In some embodiments of this application, this step is implemented through an image acquisition module. This module is compatible with mainstream OCT device types such as spectral-domain and swapped-source, supports a scan rate of ≥80kHz, an axial resolution of ≥5μm, and an output image resolution of not less than 512×512 pixels, which can clearly present key structural information such as the retinal nerve fiber layer and the macula.
[0023] During the data acquisition process, the device model, scanning parameters, and basic information of the subjects were automatically recorded, and the raw image data was uploaded to the cloud or local server in DICOM or raw format. To ensure the generalization ability and training quality of the model, this application pre-constructed a database containing 8,000 fundus OCT three-dimensional images. The samples cover various visual health states, including normal vision, myopia, hyperopia, glaucoma, and macular degeneration, and are stratified and labeled according to age, eye habits, and past medical history, providing a rich data foundation for subsequent feature extraction, model training, and personalized evaluation.
[0024] S102: Identify the device type and noise characteristics of the single-modal 3D image, and determine a denoising algorithm based on the device type and noise characteristics to preprocess the single-modal 3D image.
[0025] In some embodiments of this application, this step is implemented using an adaptive preprocessing module.
[0026] First, the system uses a device type recognition algorithm to automatically identify the OCT device type corresponding to the input image. Then, based on the identified device type, it matches the corresponding typical noise pattern from a pre-set noise feature library. This noise feature library is created by pre-collecting 500 images of each of 10 mainstream OCT devices, extracting the noise spectrum features of different devices (such as salt-and-pepper noise, Gaussian noise, complex noise, etc.), and establishing a mapping table between noise type and denoising algorithm.
[0027] Furthermore, based on the matched noise pattern, the system adaptively selects the optimal denoising scheme from a preset denoising algorithm pool: for Gaussian noise, Gaussian filtering is selected; for salt-and-pepper noise, median filtering is selected; and for complex noise or scenarios that require preservation of details, wavelet denoising is selected.
[0028] The original image is denoised using the selected algorithm, which improves the signal-to-noise ratio of the processed image, effectively eliminating device differences and noise interference, and providing high-quality data input for subsequent high-precision segmentation.
[0029] S103: Input the preprocessed single-modal 3D image into a preset attention-enhanced dynamic U-shaped segmentation network, output the retinal interlayer boundary segmentation result, and extract multi-dimensional biomarkers based on the retinal interlayer boundary segmentation result.
[0030] In some embodiments of this application, this step is implemented by a feature extraction module, which has a built-in attention-enhanced dynamic U-shaped segmentation network.
[0031] The network structure comprises four encoding paths, an intermediate attention fusion layer, and four decoding paths. Each encoding path has a backbone and branch paths: the backbone consists of cascaded residual modules used to capture local detail features of the image. Specifically, the first backbone layer is connected to three 32-channel 3×3 residual modules after passing through a 32-channel 7×7 convolutional layer, batch normalization, and ReLU activation; the second to fourth backbone layers have 3, 5, and 7 cascaded residual modules, respectively; the branch paths use dilated convolutions, and the dilation rate is dynamically adjusted based on the retinal thickness distribution of the input image, ranging from 3 to 8. The optimal dilation rate is determined by calculating the thickness variance of each region of the retina in real time to adaptively expand the receptive field and adapt to the differences in retinal structure among different subjects.
[0032] The deep feature map output from the encoding path is fed into the intermediate attention fusion layer. This layer adopts a spatial attention mechanism, which automatically focuses on high-incidence areas of lesions such as the macula and optic nerve fiber layer by calculating the importance weight of each pixel in the feature map, suppressing background interference, and generating attention-weighted feature maps.
[0033] Subsequently, the attention-weighted feature map enters the decoding path, and is upsampled layer by layer through deconvolution operation. The upsampled feature map is then concatenated with the attention-weighted feature map from the corresponding coding layer to gradually restore the image resolution and finally output a high-precision retinal layer boundary segmentation result.
[0034] During the model training phase, a transfer learning strategy was employed: the network was first pre-trained on the publicly available OCT dataset, and then fine-tuned using a self-developed dataset of 8000 examples. Training parameters were set to a learning rate of 0.0008, a batch size of 16, and 300 iterations. The loss function was a hybrid of cross-entropy loss and Dice loss. Test results show that the network achieves an average intersection-over-union ratio of 0.94, a retinal layer segmentation error ≤0.8μm, and a boundary localization error less than 0.9 pixels, significantly outperforming the traditional U-Net.
[0035] Based on the segmentation results, the system further extracts multi-dimensional biomarkers, calculates the distance between each layer according to the location of the boundaries of each retinal layer, and obtains retinal thickness distribution parameters; analyzes the texture features of the nerve fiber layer region through gray-level co-occurrence matrix analysis, and extracts texture feature parameters such as contrast, correlation, and energy; combines the original gray-level signal of the OCT image with the segmentation results to perform blood flow signal analysis on specific regions and calculate retinal blood flow density parameters; and calculates the foveal thickness and macular volume parameters based on the boundary localization of the macular region.
[0036] These parameters are integrated to form a multi-dimensional set of biomarkers including retinal thickness, blood flow density, texture features, and macular morphology, providing rich data dimensions for subsequent personalized assessments.
[0037] S104: According to the preset classification rules, the multidimensional biomarkers and the single-modal three-dimensional images are classified, and the classified multidimensional biomarkers and single-modal three-dimensional images are encrypted and stored using a preset encryption algorithm.
[0038] In some embodiments of this application, this step is implemented through a graded privacy protection module.
[0039] First, the system divides the data to be stored into three levels according to the preset classification rules: basic feature data (such as the average retinal thickness and overall structural integrity parameters), sensitive feature data (such as the coordinates of the lesion area and texture feature parameters), and raw image data (i.e., the acquired single-modal three-dimensional images).
[0040] Different encryption strategies are adopted for different levels of data. Basic feature data is encrypted using a lightweight AES symmetric encryption algorithm with an encryption time of ≤0.1s, and authorized general ophthalmologists can directly decrypt and access it. Sensitive feature data is encrypted using an algorithm that combines SHDPPC-CP-ABE attribute-based encryption and homomorphic encryption, which supports limited computation in the encrypted state. Only authorized ophthalmologists with the title of chief physician or above can decrypt the complete data. The original image data is encrypted using the SHDPPC-CP-ABE attribute-based encryption algorithm. The encryption process is associated with a dual access strategy of "subject authorization + doctor's title + diagnosis and treatment scenario", and access is only allowed when the subject explicitly authorizes it and it is used for a specified diagnosis and treatment purpose.
[0041] Furthermore, after encryption, the encrypted data at each level is uploaded to the IPFS interplanetary file system for distributed storage, obtaining the corresponding IPFS storage address. Subsequently, the hash value of the IPFS storage address, the preset data access policy, and the data classification label are written into the blocks of the Hyperledger Fabric 2.4 consortium blockchain, forming an immutable evidence record.
[0042] Simultaneously, smart contracts are deployed to monitor data access behavior in real time and set early warning rules (such as high-frequency access on a single day, cross-regional access, access outside of working hours, etc.). When abnormal behavior is detected, an SMS / email alert is immediately triggered to the subject and the hospital's information security department, and the permissions of the abnormal access account are suspended until manual verification.
[0043] Tests showed that the total encryption time for a single image was ≤0.4s, the decryption time was ≤0.3s, the data access latency was less than 400ms, and all 1000 simulated unauthorized access attempts were blocked without any data leakage, achieving a balance between privacy and diagnostic efficiency.
[0044] S105: Based on the preset SIFT algorithm, the retinal layer boundary segmentation results are time-series registered with preset historical time data, and personalized evaluation is performed based on preset multi-dimensional indicators and preset time-series prediction models to output the vision change trend in the preset future time period.
[0045] In some embodiments of this application, this step is implemented through a monitoring and analysis module.
[0046] First, the pre-set SIFT algorithm is used to rigidly register the current retinal layer boundary segmentation result with the segmentation result of at least one historical time. The displacement change matrix between the two time points is calculated, so that the image data at multiple time points can be aligned with high precision in spatial position, with a registration error of less than 0.8 pixels.
[0047] Furthermore, based on the registered image data, the temporal changes of multi-dimensional indicators are calculated, including the retinal thickness difference Δd=|d2-d1|, the rate of change of blood flow density, the coefficient of variation (CV) of the texture features of the nerve fiber layer, and the changes in morphological parameters of the macular region.
[0048] Then, the personal attribute information of the subjects is obtained, including age, eye habits (such as prolonged desk work, outdoor work, etc.), and past medical history (such as hypertension, diabetes, etc.). Based on this information, personalized normal threshold ranges matching the subjects are retrieved from a pre-constructed multi-dimensional indicator database. The calculated time-series changes of the multi-dimensional indicators are compared with the personalized threshold ranges. When any indicator exceeds the threshold, that indicator is marked as abnormal.
[0049] Next, multidimensional biomarker data from at least three time points, including the current moment, are input into a pre-trained LSTM-based temporal prediction model (this model incorporates an attention mechanism, making it more attentive to indicators that significantly affect visual acuity changes, such as macular foveal thickness and nerve fiber layer thickness). The model outputs a prediction of the subject's visual acuity change trend over a preset future time period, including the rate of change of each indicator and the probability of abnormal risk. When the probability of abnormal risk is ≥60%, the system triggers a tiered warning based on the risk level: low risk triggers intervention suggestions, medium risk triggers reminders for regular follow-up examinations, and high risk triggers immediate medical attention.
[0050] Ultimately, the generated health assessment report and prediction results are pushed to the subject's terminal through the user interaction module. With the subject's authorization, doctors can decrypt and access the corresponding level of data and enter treatment suggestions, achieving the integration of accurate monitoring, secure sharing, and trend prediction.
[0051] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S101 to S105 will be described sequentially, but this does not mean that steps S101 and S105 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S101 to S105 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S101 to S105 can be appropriately adjusted according to actual needs.
[0052] pass Figure 1This approach utilizes an attention-enhanced dynamic U-shaped segmentation network, incorporating dynamic hole rate adjustment and spatial attention mechanisms to achieve sub-pixel-level segmentation of retinal layer boundaries. It also integrates multi-dimensional biomarkers such as retinal thickness, blood flow density, and texture features to construct a personalized assessment model, improving monitoring accuracy. Simultaneously, a three-level data grading system is constructed, employing differentiated encryption strategies such as AES and SHDPPC-CP-ABE+ homomorphic encryption for basic features, sensitive features, and original images, respectively. This is combined with IPFS distributed storage and Hyperledger Fabric consortium blockchain for evidence storage, and real-time monitoring of abnormal access behavior triggers early warnings via smart contracts, balancing privacy and diagnostic efficiency. Furthermore, high-precision registration of multi-time-point images is achieved based on the SIFT algorithm, and an LSTM temporal prediction model is used to integrate personalized features of the subject to output future vision change trends and abnormal risk probabilities, truly achieving early detection and early intervention.
[0053] Figure 2 The system of a vision monitoring method based on fundus OCT images provided in this application adopts a layered architecture design diagram.
[0054] exist Figure 2 The system of the present invention is shown in the figure, which adopts a layered architecture design and includes: Perception layer (image acquisition module): Acquires 3D fundus images output by different types of OCT devices, providing a data foundation for subsequent processing; Preprocessing layer (adaptive preprocessing module): automatically selects denoising algorithms based on image noise characteristics to improve image quality; Processing layer (feature extraction module, monitoring and analysis module): achieves accurate segmentation of retinal features, multi-dimensional index calculation and time-series predictive analysis; Security layer (tiered privacy protection module): Ensures data security through tiered encryption, blockchain storage, and anomaly alerts; Application layer (user interaction module): Provides human-computer interaction interface, supports permission configuration, report viewing and data sharing.
[0055] Figure 3 This is a system architecture diagram of a vision monitoring method based on fundus OCT images provided in an embodiment of this application.
[0056] exist Figure 3 The diagram illustrates the layered architecture of the perception layer, preprocessing layer, processing layer, security layer, and application layer, with each layer's core modules clearly labeled.
[0057] Figure 4 This is a schematic diagram of an attention-enhanced dynamic U-shaped segmentation network structure for a vision monitoring method based on fundus OCT images provided in an embodiment of this application.
[0058] exist Figure 4 The code includes annotations for the encoding path, intermediate attention fusion layer, decoding path, main route and branch routes, and dynamic adjustment logic for hole rate.
[0059] Figure 5 This document presents a flowchart illustrating the hierarchical privacy protection module of a vision monitoring method based on fundus OCT images, as provided in an embodiment of this application.
[0060] exist Figure 5 The document demonstrates the complete process of data classification → hierarchical encryption → IPFS storage → blockchain recording → access verification → anomaly alert.
[0061] Figure 6 A schematic diagram of the structure of a vision monitoring device based on fundus OCT images provided in this application embodiment includes: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, which enable the at least one processor to perform any of the above-mentioned vision monitoring methods based on fundus OCT images.
[0062] Some embodiments of this application provide a non-volatile computer storage medium for vision monitoring based on fundus OCT images, which stores computer-executable instructions capable of executing any of the above-mentioned vision monitoring methods based on fundus OCT images.
[0063] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so they are described more simply; relevant parts can be referred to the descriptions of the method embodiments.
[0064] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0065] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0069] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0070] Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and non-volatile memory such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0071] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0072] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0073] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the technical principles of this application should fall within the protection scope of this application.
Claims
1. A method of vision monitoring based on OCT images of the fundus, characterized in that, The method includes: Acquire single-modal three-dimensional images of the fundus of the subject; Identify the device type and noise characteristics of the single-modal 3D image, and determine a denoising algorithm based on the device type and noise characteristics to preprocess the single-modal 3D image; The preprocessed single-modal 3D image is input into a preset attention-enhanced dynamic U-shaped segmentation network, which outputs the retinal interlayer boundary segmentation results and extracts multi-dimensional biomarkers based on the retinal interlayer boundary segmentation results. According to the preset classification rules, the multidimensional biomarkers and the single-modal three-dimensional images are classified, and the classified multidimensional biomarkers and single-modal three-dimensional images are encrypted and stored using a preset encryption algorithm. Based on the preset SIFT algorithm, the retinal layer boundary segmentation results are time-series registered with preset historical time data. Based on preset multi-dimensional indicators and preset time-series prediction models, a personalized evaluation is performed, and the vision change trend within a preset future time period is output.
2. The method according to claim 1, characterized in that, The process of identifying the device type and noise characteristics of the single-modal 3D image, and determining a denoising algorithm based on the device type and noise characteristics to preprocess the single-modal 3D image, specifically includes: The OCT device type corresponding to the single-modal 3D image is identified by a preset device type recognition algorithm; Based on the OCT device type, a corresponding noise pattern is matched from a preset noise feature library; Based on the noise pattern, a denoising algorithm matching the noise pattern is selected from a preset denoising algorithm pool; the denoising algorithm pool includes Gaussian filtering, median filtering, and wavelet denoising. The denoising algorithm is used to denoise the single-modal 3D image to obtain a preprocessed single-modal 3D image.
3. The method according to claim 1, characterized in that, The process of inputting the preprocessed single-modal 3D image into a preset attention-enhanced dynamic U-shaped segmentation network and outputting retinal interlayer boundary segmentation results specifically includes: A pre-constructed attention-enhanced dynamic U-shaped segmentation network is provided. The attention-enhanced dynamic U-shaped segmentation network includes an encoding path, an intermediate attention fusion layer, and a decoding path. Each layer of the encoding path is provided with a backbone path and branch paths. The backbone path is composed of cascaded residual modules to extract local detail features of the image. The branch paths are dilated convolutions, and the dilation rate is adjusted based on the retinal thickness distribution of the input image. The preprocessed single-modal 3D image is input into the encoding path, and after passing through each layer of the encoding path, a deep feature map is output. The deep feature map is input into the intermediate attention fusion layer, and the importance weight of each pixel in the deep feature map is calculated through a preset spatial attention mechanism to generate an attention-weighted feature map. The attention-weighted feature map is input into the decoding path to upsample the attention-weighted feature map through a deconvolution operation, thereby obtaining an upsampled weighted feature map. The upsampled weighted feature map is concatenated with the features from the attention-weighted feature map of the corresponding coding layer to output the segmentation result of the retinal layer boundary.
4. The method according to claim 1, characterized in that, Based on the retinal interlaminar boundary segmentation results, multidimensional biomarkers are extracted, specifically including: Based on the position of the boundaries of each retinal layer in the retinal layer boundary segmentation results, the distance between each layer is calculated to obtain the retinal thickness distribution parameters; The texture features of the nerve fiber layer region in the retinal interlayer boundary segmentation results were analyzed by gray-level co-occurrence matrix analysis, and texture feature parameters of the nerve fiber layer, including contrast, correlation, and energy, were extracted. Based on the original grayscale signal of the OCT image and the segmentation results of the retinal layer boundary, blood flow signal analysis is performed on the preset area to calculate the retinal blood flow density parameter; Based on the boundary localization of the macular region in the retinal interlayer boundary segmentation results, the foveal thickness and macular region volume parameters are calculated. By integrating the retinal thickness distribution parameters, nerve fiber layer texture feature parameters, retinal blood flow density parameters, macular foveal thickness, and macular region volume parameters, multidimensional biomarkers are obtained.
5. The method according to claim 1, characterized in that, The step involves classifying the multidimensional biomarkers and the single-modal 3D image according to preset classification rules, and encrypting and storing the classified multidimensional biomarkers and single-modal 3D image using a preset encryption algorithm. Specifically, this includes: According to preset grading rules, the multidimensional biomarkers and the single-modal three-dimensional images are divided into basic feature data, sensitive feature data, and raw image data; the basic feature data includes the mean retinal thickness and overall structural integrity parameters; the sensitive feature data includes the lesion area coordinates and texture feature parameters; the raw image data is the acquired single-modal three-dimensional image. The basic feature data is encrypted using the AES symmetric encryption algorithm; The sensitive feature data is encrypted using an algorithm that combines SHDPPC-CP-ABE attribute-based encryption with homomorphic encryption. The original image data is encrypted using the SHDPPC-CP-ABE attribute-based encryption algorithm; The encrypted basic feature data, sensitive feature data and original image data are uploaded to the IPFS interplanetary file system for distributed storage, and the corresponding IPFS storage address is obtained. The hash value of the IPFS storage address, the preset data access policy, and the data classification label are written into the block of the Hyperledger Fabric consortium chain to obtain the evidence record.
6. The method according to claim 5, characterized in that, After writing the hash value, access policy, and data classification tag of the IPFS storage address into a block of the Hyperledger Fabric consortium blockchain to obtain the evidence record, the method further includes: Deploy smart contracts to monitor access to encrypted basic feature data, sensitive feature data, and raw image data; When abnormal access behavior that meets the preset warning rules is detected, the warning mechanism is triggered, and a warning notification is sent to the preset contacts, suspending the permissions of the abnormal access account.
7. The method according to claim 1, characterized in that, The process involves temporally registering the retinal slice boundary segmentation results with preset historical time data using a preset SIFT algorithm, and then performing a personalized assessment based on preset multi-dimensional indicators and a preset temporal prediction model to output a preset future vision change trend. Specifically, this includes: Using a preset SIFT algorithm, the retinal interlayer boundary segmentation result is rigidly registered with the retinal interlayer boundary segmentation result at a preset historical time. The displacement change matrix between the retinal interlayer boundary segmentation result and the retinal interlayer boundary segmentation result at the preset historical time is calculated so that the image data at multiple time points are aligned in spatial position. Based on the image data after rigid registration, the temporal changes of multi-dimensional indicators are calculated; the multi-dimensional indicators include the difference in retinal thickness between the time of the retinal interlayer boundary segmentation result and the preset historical time, the rate of change of blood flow density, and the coefficient of variation of the texture features of the nerve fiber layer. Obtain the personal attribute information of the subjects; the personal attribute information includes age, eye habits and past medical history; Based on the personal attribute information, retrieve the personalized normal threshold range that matches the subject from a pre-constructed multi-dimensional indicator database; The time-series changes of the multi-dimensional indicators are compared with the personalized normal threshold range. When the time-series change of any indicator exceeds the corresponding threshold range, the indicator is marked as an abnormal indicator. Multidimensional biomarker data at at least three time points, including the time of the retinal interlayer boundary segmentation result, are input into a pre-trained LSTM-based temporal prediction model. The time-series prediction model outputs the visual acuity change trend of the subject within a preset future time period; the visual acuity change trend includes the rate of change of multi-dimensional indicators and the probability of abnormal risk.
8. The method according to claim 1, characterized in that, The method further includes: Receive treatment suggestions entered by doctors based on preset doctor access levels and authorization scope; Based on the preset subject access level and authorization scope, the subject is shown the results of retinal layer boundary segmentation, the multidimensional biomarkers, the marked abnormal indicators, the personalized assessment report, the vision change trend, and the treatment recommendations.
9. A vision monitoring device based on fundus OCT images, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a vision monitoring method based on fundus OCT images as described in any one of claims 1-8.
10. A vision monitoring storage medium based on fundus OCT images, storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of executing the vision monitoring method based on fundus OCT images as described in any one of claims 1-8.