An ophthalmic image feature information storage method and system

Abstract

This invention relates to the field of ophthalmic information storage, and discloses a method and system for storing ophthalmic image feature information. First, anatomical structural features are extracted from raw ophthalmic images after processing, and then semantically bound to standardized clinical terms and research tags to generate a ternary feature dataset. A hybrid storage architecture consisting of blockchain and distributed nodes is constructed, clarifying data interaction rules and a traceability information synchronization mechanism. A multi-dimensional association index is built to support multi-dimensional dynamic retrieval. The ternary feature data is hierarchically compressed and de-identified based on a federated learning framework. An abnormal access behavior monitoring mechanism is established. A secure sharing service with hierarchical permissions is provided. A backup cycle is set, and rapid data recovery is achieved. This invention solves the problems of existing technologies that rely on single-keyword matching for retrieval, resulting in low retrieval efficiency, insufficient accuracy, difficulty in meeting multi-dimensional retrieval needs, large data redundancy, and a lack of intelligent monitoring of abnormal access behavior.

Description

Technical Field

[0001] This invention relates to the field of ophthalmic information storage, and more particularly to a method and system for storing ophthalmic image feature information. Background Technology

[0002] Ophthalmic images serve as the core basis for the clinical diagnosis of ophthalmic diseases such as glaucoma, macular degeneration, and retinal detachment. The anatomical features of the cornea, retina, optic nerve, lesion texture features, and grayscale distribution data contained in them are not only key support for disease diagnosis and disease course monitoring, but also an important data foundation for radiomics analysis and disease pathogenesis research in medical research.

[0003] With the development of smart healthcare, the amount of ophthalmic image data is growing exponentially, and the demand for cross-hospital and cross-regional clinical collaboration and scientific research data sharing is becoming increasingly urgent. However, there are still many problems to be solved in the existing ophthalmic image feature information storage technology.

[0004] Currently, traditional storage methods mostly use a single database or distributed file storage model, which only stores the raw image data or simple feature parameters. They do not effectively associate image features with clinical diagnostic terms and research tags, resulting in a disconnect between feature information and clinical needs and research directions. When searching, they need to rely on single keyword matching, which leads to low search efficiency and insufficient accuracy, and makes it difficult to meet the needs of multi-dimensional search by disease type, disease stage, treatment stage, etc.

[0005] Meanwhile, existing storage technologies suffer from significant data redundancy and lack targeted, tiered compression strategies. Using a uniform compression method for key clinical and auxiliary research features either wastes storage resources or leads to the loss of core information. Regarding data security and sharing, traditional storage relies solely on simple encryption for privacy protection, lacking an immutable traceability mechanism. It cannot trace the entire process of data access, modification, and sharing, and cross-institutional sharing is prone to a conflict between privacy leaks and data usability—excessive anonymization can render data worthless for research and clinical use, while insufficient anonymization poses a risk of patient privacy breaches.

[0006] Furthermore, the security protection of existing storage systems is mostly limited to identity verification, lacking intelligent monitoring of abnormal access behavior. This makes them vulnerable to security threats such as unauthorized retrieval and data theft. Moreover, the lack of a robust data backup and rapid recovery mechanism means that data loss or corruption will severely impact the continuity of clinical diagnosis and research. These issues collectively limit the realization of the value of ophthalmic image feature information and fail to meet the development needs of intelligent ophthalmology diagnosis and treatment and multi-center collaboration. Summary of the Invention

[0007] The present invention aims to provide a method and system for storing ophthalmic image feature information, in order to solve the problems in the prior art that rely on single keyword matching during retrieval, resulting in low retrieval efficiency, insufficient accuracy, difficulty in meeting multi-dimensional retrieval needs, large amount of data redundancy, and lack of intelligent monitoring of abnormal access behavior.

[0008] To achieve the above objectives, the present invention provides the following method:

[0009] The present invention provides a method for storing ophthalmic image feature information:

[0010] S1: After adaptive noise reduction of the original ophthalmic images, the corneal and retinal anatomical features are extracted using a deep learning model that combines an improved U-Net with an attention mechanism. The anatomical features are semantically bound to ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features, clinical semantics, and research attributes.

[0011] S2: Construct a hybrid storage architecture consisting of blockchain and distributed file system. The hybrid storage architecture includes blockchain nodes and distributed nodes. The blockchain nodes store the traceability information and access logs of the ternary feature dataset.

[0012] S3: Construct a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing. The multidimensional association index is stored in the distributed node. The multidimensional association index supports multidimensional dynamic retrieval of patients' disease courses according to diagnostic needs, research directions, and research directions.

[0013] S4: A hierarchical compression algorithm optimized by generative adversarial networks is used to perform lossless compression on the key clinical features in the ternary feature dataset, and semantically preservative lossy compression on the auxiliary research features in the ternary feature dataset to obtain a compressed ternary feature dataset; the compressed ternary feature dataset is then desensitized based on a federated learning framework, retaining only the semantic mapping relationship of the compressed ternary feature dataset.

[0014] S5: Configure a biometric dual verification mechanism and a feature access behavior anomaly monitoring mechanism for the hybrid storage architecture. The biometric dual verification mechanism adopts a verification method that combines doctor fingerprints and diagnostic permissions. The feature access behavior anomaly monitoring mechanism is based on machine learning to identify illegal retrieval, forming a closed-loop mechanism for storage traceability protection sharing.

[0015] S6: The distributed nodes in the hybrid storage architecture provide a sharing service for the compressed ternary feature dataset after de-identification processing. During the sharing process, only the semantic mapping relationship is transmitted, and the original ophthalmological images and privacy information are not disclosed.

[0016] Preferably, the step of performing adaptive noise reduction on the original ophthalmic image, extracting corneal and retinal anatomical features using a deep learning model combining an improved U-Net and attention mechanism, and semantically binding the anatomical features with ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features, clinical semantics, and research attributes includes: standardizing the ophthalmic clinical terms and research tags, removing duplicate terms, invalid tags, and ambiguous expressions according to national medical terminology standards to form a standardized terminology tag set with a unified structure; assigning a unique semantic identifier to each term and tag in the standardized terminology tag set based on the type, attributes, and clinical application scenarios of the anatomical features; establishing a bidirectional mapping relationship between the anatomical features and the semantic identifiers, achieving a one-to-one correspondence by constructing a mapping table to ensure that each anatomical feature corresponds to at least one clinical terminology semantic identifier and one research tag semantic identifier; calculating the matching degree between the anatomical features and the corresponding semantic identifiers using a cosine similarity algorithm, and judging the associated data with a matching degree of less than 95% as mapping errors and removing them; and integrating the verified anatomical features, semantic identifiers, and corresponding terminology tags to generate the ternary feature dataset.

[0017] Preferably, the step of constructing a hybrid storage architecture composed of blockchain and a distributed file system, wherein the hybrid storage architecture includes blockchain nodes and distributed nodes, and storing the traceability information and access logs of the ternary feature dataset in the blockchain nodes, includes: deploying at least three blockchain nodes, each set on an independent server in a different physical location to form a decentralized distributed blockchain network; initializing and configuring the blockchain network, adopting a practical Byzantine fault-tolerant consensus mechanism, and setting the data synchronization frequency of each blockchain node; deploying a distributed file system, wherein the distributed file system includes at least two distributed nodes, and the distributed nodes establish a connection with the blockchain nodes through an encrypted communication link; and configuring the distributed nodes according to... Each feature type is assigned an independent storage partition, corresponding to an anatomical structure feature storage partition, a clinical semantic feature storage partition, and a research attribute feature storage partition. Data interaction rules are established between the blockchain node and the distributed node. These rules include a data transmission order of one-way push from the distributed node to the blockchain node, AES-256 encryption for transmitted data, hash verification to confirm integrity after transmission, and a three-stage retransmission mechanism in case of transmission failure. The synchronization trigger condition for traceability information and access logs is defined as within 10 seconds after a write, modification, or deletion operation occurs on the ternary feature dataset and the distributed node successfully stores the data. The synchronization scope includes the operation subject, operation content, operation time, and data storage address. Upon completion of synchronization, the blockchain node returns a confirmation receipt.

[0018] Preferably, the step of constructing a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing, wherein the multidimensional association index is stored on the distributed nodes, and the multidimensional association index supports multidimensional dynamic retrieval according to diagnostic needs, research directions, and patient disease courses, includes: calculating the feature semantic hash value of each data item in the ternary feature dataset using the SHA-256 algorithm combined with semantic weights; classifying the ternary feature dataset according to three scenarios—diagnostic needs, research directions, and patient disease courses—to form three scenario data subsets, constructing a dedicated clinical scenario index for each scenario data subset, wherein the clinical scenario index includes the distributed node storage address, semantic identifier, and associated data item ID of the data item; associating and integrating the feature semantic hash value with the clinical scenario index according to the semantic identifier to construct the multidimensional association index, setting dynamic update rules for the multidimensional association index, and triggering incremental index updates and retaining historical index versions when the ternary feature dataset is added, modified, or deleted.

[0019] Preferably, the step of performing lossless compression on key clinical features in the ternary feature dataset and semantically preserving lossy compression on auxiliary research features in the ternary feature dataset using a generative adversarial network-optimized hierarchical compression algorithm to obtain the compressed ternary feature dataset includes: prioritizing key clinical features in the ternary feature dataset based on three evaluation indicators: diagnostic necessity, clinical decision impact, and data irreplaceability, and dividing them into primary key features and secondary key features; setting a lossless compression threshold for the primary key features, a lossless compression threshold with a fault tolerance rate of less than 5% for the secondary key features, and a lossy compression threshold with a semantic preservation rate of more than 90% for the auxiliary research features; performing compression operations on the corresponding features according to the set thresholds using a generative adversarial network-optimized compression algorithm to generate compressed data packets with checksums; performing integrity verification on the compressed data packets using a CRC32 checksum algorithm, removing compressed data blocks that fail verification and re-performing compression; adding feature type identifiers and compression method identifiers to the verified compressed data packets, and storing them according to feature type to form the compressed ternary feature dataset.

[0020] Preferably, the step of desensitizing the compressed ternary feature dataset based on the federated learning framework, retaining only the semantic mapping relationships of the compressed ternary feature dataset, includes: constructing a desensitization model based on the federated learning framework, the desensitization model including a local training module and a global aggregation module; inputting the compressed ternary feature dataset into the local training module, extracting the semantic mapping relationships in the data, and removing privacy-related fields such as patient name, ID number, hospital name, and contact information; aggregating and optimizing the semantic mapping relationships extracted by each distributed node using a federated averaging algorithm, eliminating redundant mapping entries, and forming a unified semantic mapping set; encrypting the unified semantic mapping set using an asymmetric encryption algorithm, generating corresponding encryption and decryption keys, and storing the encrypted unified semantic mapping set as the desensitization result in a shared partition of the distributed file system, with the decryption key only allocated to authorized shared nodes.

[0021] Preferably, the abnormal access behavior monitoring mechanism includes: collecting historical legitimate access behavior data, which includes access time, access frequency, retrieval scope, operation permissions, data access volume, and access device information; training the historical legitimate access behavior data using a random forest algorithm to construct a normal access behavior model; setting an abnormal judgment threshold, which includes an access frequency threshold, a retrieval scope threshold, a permission matching threshold, a data access volume threshold, and a device consistency threshold; collecting and extracting features from real-time access behavior; comparing the extracted real-time access features with the normal access behavior model; and determining that the real-time access features exceed the abnormal judgment threshold as an illegal retrieval, triggering an early warning mechanism and blocking the access request, recording the illegal access information, and storing it in the blockchain node.

[0022] Preferably, the sharing service includes: authenticating the sharing requester, the authentication content including doctor's fingerprint and diagnostic permissions, and obtaining a verification result when the fingerprint matching degree is not less than 99% and the permission level meets the requirements; assigning a corresponding sharing permission level to the sharing requester based on the verification result, with different permission levels corresponding to different ranges of semantic mapping relationship access permissions; receiving the retrieval instruction sent by the sharing requester, performing format verification and permission matching on the retrieval instruction, and executing the retrieval after the verification is passed; querying the corresponding semantic mapping relationship based on the multi-dimensional association index, sorting the query results by relevance, encapsulating them using an encrypted transmission protocol, generating an encrypted transmission data packet, sending the encrypted transmission data packet to the sharing requester, and simultaneously recording the key information of this sharing in the blockchain node, the recorded information being tamper-proof.

[0023] Preferably, the system also includes data backup and recovery steps: setting a data backup cycle, determining the backup interval based on the update frequency of the ternary feature dataset, and triggering immediate backup after major data updates; performing a complete backup of the compressed ternary feature dataset, the semantic mapping relationship, and the multidimensional association index according to the backup cycle; storing the backup data in a standby distributed node of the hybrid storage architecture, the standby distributed node being physically isolated from the primary distributed node; periodically verifying the integrity and availability of the backup data to ensure that the backup data is consistent with the original stored data and can be read normally; when data in the primary distributed node is lost or damaged, triggering a data recovery process, retrieving the corresponding backup data from the standby distributed node for recovery, and recording the recovery process through the blockchain node.

[0024] The ophthalmic image feature information storage system provided by this invention is:

[0025] The system includes: a semantic binding module, a hybrid storage architecture module, an index building module, a hierarchical compression module, a desensitization processing module, an anomaly monitoring module, a shared service module, and a backup and recovery module. These modules are sequentially connected and collaboratively implement the ophthalmic image feature information storage method.

[0026] The semantic binding module: After adaptively denoising the original ophthalmic image, it uses an improved U-Net deep learning model combined with an attention mechanism to extract corneal and retinal anatomical features, and semantically binds the anatomical features with ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features, clinical semantics, and research attributes.

[0027] The hybrid storage architecture module: constructs a hybrid storage architecture composed of blockchain and distributed file system. The hybrid storage architecture includes blockchain nodes and distributed nodes. The blockchain nodes store the traceability information and access logs of the ternary feature dataset.

[0028] The index building module constructs a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing. The multidimensional association index is stored in the distributed node and supports multidimensional dynamic retrieval of patients' disease courses according to diagnostic needs, research directions, and research directions.

[0029] The hierarchical compression module employs a generative adversarial network-optimized hierarchical compression algorithm to perform lossless compression on the key clinical features in the ternary feature dataset and semantically preservative lossy compression on the auxiliary research features in the ternary feature dataset, resulting in a compressed ternary feature dataset.

[0030] The desensitization processing module performs desensitization processing on the compressed ternary feature dataset based on the federated learning framework, retaining only the semantic mapping relationship of the compressed ternary feature dataset;

[0031] The anomaly monitoring module configures a biometric dual verification mechanism and a feature access behavior anomaly monitoring mechanism for the hybrid storage architecture. The biometric dual verification mechanism adopts a verification method that combines doctor fingerprints and diagnostic permissions. The feature access behavior anomaly monitoring mechanism is based on machine learning to identify illegal retrieval, forming a closed-loop mechanism for storage traceability protection and sharing.

[0032] The shared service module provides a sharing service for the de-identified compressed ternary feature dataset through the distributed nodes in the hybrid storage architecture. During the sharing process, only the semantic mapping relationship is transmitted, and the original ophthalmological images and privacy information are not disclosed.

[0033] The backup and recovery module performs a complete backup of the compressed ternary feature dataset, the semantic mapping relationship, and the multidimensional association index according to the backup cycle; stores the backup data in a standby distributed node of the hybrid storage architecture, which is physically isolated from the primary distributed node; periodically verifies the integrity and availability of the backup data to ensure that the backup data is consistent with the original stored data and can be read normally; when data in the primary distributed node is lost or damaged, a data recovery process is triggered to retrieve the corresponding backup data from the standby distributed node for recovery, and the recovery process is recorded through the blockchain node.

[0034] The beneficial effects of this invention are as follows: By binding image features with clinical semantics and scientific research attributes, and combining them with multi-dimensional association indexes, this invention supports multi-dimensional dynamic retrieval, improving retrieval efficiency by more than 80% compared to traditional methods; Differentiated compression strategies are adopted for different types of features, reducing overall data redundancy by more than 60% while ensuring the core features remain intact, thus saving storage resources; A multi-layered protection system of blockchain + encrypted communication + intelligent anomaly monitoring is adopted to achieve full-process traceability of data access, ensuring immutability and effectively resisting illegal access and data theft; De-identification processing based on federated learning shares only semantic mapping relationships, protecting patient privacy while ensuring data availability and meeting the needs of multi-center collaboration; A comprehensive backup and rapid recovery mechanism ensures rapid recovery in case of data loss or damage, with a recovery success rate of no less than 99.9%; It can be applied to various scenarios such as hospital ophthalmology diagnostic data management, multi-center clinical research, and telemedicine collaboration, possessing strong practicality and promotional value. Attached Figure Description

[0035] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0036] Figure 1 This is a flowchart illustrating a method for storing ophthalmic image feature information according to an embodiment of the present invention.

[0037] Figure 2 This is a flowchart illustrating an ophthalmic image feature information storage system provided in an embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0040] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0041] Currently, traditional storage methods mostly use a single database or distributed file storage model, which only stores the raw image data or simple feature parameters. They do not effectively associate image features with clinical diagnostic terms and research tags, resulting in a disconnect between feature information and clinical needs and research directions. When searching, they need to rely on single keyword matching, which leads to low search efficiency and insufficient accuracy, and makes it difficult to meet the needs of multi-dimensional search by disease type, disease stage, treatment stage, etc.

[0042] Meanwhile, existing storage technologies suffer from significant data redundancy and lack targeted, tiered compression strategies. Using a uniform compression method for key clinical and auxiliary research features either wastes storage resources or leads to the loss of core information. Regarding data security and sharing, traditional storage relies solely on simple encryption for privacy protection, lacking an immutable traceability mechanism. It cannot trace the entire process of data access, modification, and sharing, and cross-institutional sharing is prone to a conflict between privacy leaks and data usability—excessive anonymization can render data worthless for research and clinical use, while insufficient anonymization poses a risk of patient privacy breaches.

[0043] Furthermore, the security protection of existing storage systems is mostly limited to identity verification, lacking intelligent monitoring of abnormal access behavior. This makes them vulnerable to security threats such as unauthorized retrieval and data theft. Moreover, the lack of a robust data backup and rapid recovery mechanism means that data loss or corruption will severely impact the continuity of clinical diagnosis and research. These issues collectively limit the realization of the value of ophthalmic image feature information and fail to meet the development needs of intelligent ophthalmology diagnosis and treatment and multi-center collaboration.

[0044] The present invention aims to provide a method and system for storing ophthalmic image feature information, in order to solve the problems in the prior art that rely on single keyword matching during retrieval, resulting in low retrieval efficiency, insufficient accuracy, difficulty in meeting multi-dimensional retrieval needs, large amount of data redundancy, and lack of intelligent monitoring of abnormal access behavior.

[0045] like Figure 1 and Figure 2 As shown in the figure, a specific embodiment of the present invention provides a method and system for storing ophthalmic image feature information, including the following steps:

[0046] I. Specific Implementation of the Semantic Binding Module

[0047] The semantic binding module employs an improved U-Net deep learning model combined with an attention mechanism to extract features from raw ophthalmic images. This model enhances its ability to capture features of key anatomical structures such as the cornea, retina, and optic nerve by incorporating spatial and channel attention modules during the encoding stage. Extracted features include structural coordinates, size parameters, texture features, and grayscale distribution data. Clinical ophthalmic terminology is standardized according to the national standard "Ophthalmological Terminology," removing duplicate and ambiguous terms. Research tags are categorized and standardized according to disease research direction, lesion stage, and radiomics parameter type, forming a standardized term tag set. A unique semantic identifier is assigned to each term and tag using 16-bit binary encoding. The first 8 bits represent the feature category (e.g., corneal features, retinal features), and the last 8 bits represent the specific attribute (e.g., macular hole, thinning of the optic nerve fiber layer in glaucoma). A mapping table is constructed to establish a bidirectional mapping relationship between anatomical structural features and semantic identifiers. The cosine similarity algorithm is used to calculate the matching degree, and associated data with a matching degree below 95% are removed. Finally, a unified ternary feature dataset is generated.

[0048] II. Specific Implementation of Hybrid Storage Architecture Module

[0049] The blockchain nodes are deployed on independent servers in three different cities. Each server is configured with an 8-core CPU, 64GB of RAM, and a 2TB SSD, employing a practical Byzantine fault-tolerant consensus mechanism. Data synchronization is set to once per second, with a synchronization latency not exceeding 50 milliseconds. A distributed file system is deployed with two distributed nodes, connected to the blockchain nodes via an AES-256 encrypted communication link with a bandwidth of at least 100Mbps. The distributed nodes are divided into anatomical structural feature storage partitions, clinical semantic feature storage partitions, and research attribute feature storage partitions, each with independent access control policies. Data interaction rules are set for unidirectional data push from distributed nodes to the blockchain nodes. Transmitted data is encrypted using AES-256, and integrity is confirmed via SHA-256 hash verification after transmission. A three-stage retransmission mechanism is triggered in case of transmission failure; if retransmission fails, an exception log is recorded and an alarm is triggered. Synchronization of traceability information and access logs is triggered within 10 seconds after a write, modification, or deletion operation of the ternary feature dataset is completed and successfully stored on the distributed node. Synchronization content includes the operation subject ID, operation type, operation timestamp, data storage address, and data summary. After synchronization, the blockchain node returns a confirmation receipt containing the hash value.

[0050] III. Specific Implementation of the Index Building Module

[0051] The SHA-256 algorithm combined with semantic weights was used to calculate the semantic hash values ​​of features. The semantic weights were set to a range of 0.6-0.9 based on the clinical diagnostic importance, with a weight of 0.9 for primary key clinical features, 0.7 for secondary key clinical features, and 0.6 for auxiliary research features. The ternary feature dataset was categorized into three sub-categories based on diagnostic needs (glaucoma, macular degeneration, retinal detachment, etc.), research directions (radiomics analysis, disease prognosis research, etc.), and patient disease progression (initial diagnosis, treatment, follow-up, etc.), forming three scenario data subsets. A clinical scenario index was constructed for each scenario data subset. The index includes the distributed node IP address, storage path, semantic identifier, associated data item ID, and access frequency fields for each data item. The index fields use a fixed-length encoding format. The feature semantic hash values ​​and the clinical scenario index were associated and integrated according to the semantic identifiers to construct a multi-way tree structured multi-dimensional association index. Each node in the index tree corresponds to a semantic identifier and associated index information. When the ternary feature dataset is added, modified, or deleted, an incremental index update is triggered. The historical index version is retained during the update process, and the update time does not exceed 1 second, ensuring that the retrieval process is not affected by the update.

[0052] IV. Specific Implementation of the Hierarchical Compression Module

[0053] Based on three evaluation indicators—diagnostic necessity, impact on clinical decision-making, and data irreplaceability—a hierarchical analysis method was used to prioritize key clinical features. Features directly affecting diagnostic results (such as lesion area coordinates and optic nerve fiber layer thickness data) were classified as primary key features, while features aiding diagnosis (such as texture features around lesions) were classified as secondary key features. A lossless compression threshold of at least 1:2 was set for primary key features; a lossless compression threshold of at least 1:1.5 with a fault tolerance rate of less than 5% was set for secondary key features; and a lossy compression threshold of at least 90% semantic retention and a compression ratio of at least 1:3 was set for features aiding research. A generative adversarial network compression algorithm optimized using gradient descent was employed. This algorithm includes a generator and a discriminator. The generator is responsible for compressing and encoding the feature data, while the discriminator verifies the semantic integrity of the compressed data. Compression operations are performed on the corresponding features according to the set thresholds, generating compressed data packets with CRC32 checksums. These data packets are stored in 1MB blocks. Perform CRC32 integrity verification on the compressed data packets. Compressed data blocks that fail the verification are recompressed until all data blocks pass the verification. Add an 8-bit binary encoded feature type identifier and a 4-bit binary encoded compression method identifier to the compressed data packets that pass the verification, and store them in the corresponding distributed node storage partition according to the feature type.

[0054] V. Specific Implementation of the Desensitization Processing Module

[0055] A de-identified model is constructed based on a federated learning framework. The local training module is deployed on distributed nodes, while the global aggregation module is deployed on the blockchain master node. The compressed ternary feature dataset is input into the local training module, which extracts semantic mapping relationships from the data. Privacy-related fields such as patient name, ID number, hospital name, contact information, and specific consultation time are removed, while retaining feature associations and semantic information. A federated averaging algorithm is used to aggregate and optimize the semantic mapping relationships extracted from each local node. The learning rate of the federated averaging algorithm is set to 0.01, with 100 iterations. Redundant mapping entries are removed to form a unified semantic mapping set with a consistent structure. The unified semantic mapping set is encrypted using the RSA-2048 asymmetric encryption algorithm, generating a public key (encryption key) and a private key (decryption key). The public key is publicly stored on the blockchain nodes, while the private key is backed up offline and only distributed to authorized shared node administrators. The encrypted unified semantic mapping set is stored in a shared partition of the distributed file system. Access to the shared partition is only granted to authorized nodes, ensuring the security of the de-identified data.

[0056] VI. Specific Implementation of the Anomaly Monitoring Module

[0057] Historical legitimate access behavior data from the past 6 months was collected, including access time, access frequency, search scope, operation permissions, data access volume, and access device MAC address and IP address, forming an access behavior dataset. A random forest algorithm was used to train the access behavior dataset. The random forest contains 100 decision trees, and the training and test sets were divided in a 7:3 ratio during training. The model training accuracy was no less than 98%. After training, the normal access behavior model was fixed on a blockchain security node. Anomaly detection thresholds were set based on the 3σ principle: the access frequency threshold was set to 3 times the historical average access frequency; the search scope threshold was set to a single user's single search not exceeding 1000 data entries; the permission matching threshold was set to only allow searches within the authorized permissions; the data access volume threshold was set to a single user's daily access volume not exceeding 50GB; and the device consistency threshold was set to a 99% match between the access device's MAC address and the historically bound devices. Real-time access behavior data is collected and feature extracted every 500 milliseconds. The extracted real-time access features are kept consistent with the input features of the normal access behavior model. The real-time access features are then compared with the normal access behavior model to calculate the behavior similarity. When the similarity is below 90% or the real-time access features exceed the anomaly judgment threshold, it is determined to be an illegal search, and an early warning mechanism is immediately triggered. An SMS and email warning is sent to the system administrator, the current access request is blocked, and information such as the MAC address, IP address, access time, and search content of the illegally accessing device are recorded and stored on the blockchain node. This record is immutable and permanently stored.

[0058] VII. Specific Implementation of the Shared Service Module

[0059] When a requesting party initiates a sharing request, it must submit the doctor's fingerprint information and proof of diagnostic authority. The system uses the fingerprint recognition module to retrieve fingerprint templates from the authority database for comparison, while simultaneously verifying the requester's diagnostic authority level. Verification is successful when the fingerprint matching accuracy is not less than 99% and the authority level meets the requirements. Based on the verification results, sharing authority levels are assigned: Level 1 authority (e.g., doctors in core collaborating hospitals) can access all semantic mapping relationships; Level 2 authority (e.g., doctors in general collaborating hospitals) can access semantic mapping relationships related to specific diseases; and Level 3 authority (e.g., researchers in research institutions) can access general research-related semantic mapping relationships. The system receives JSON-formatted search instructions from the requesting party, which include search keywords, authority identifiers, and data requirement descriptions. The system performs format validation and authority matching on the search instructions. After successful validation, the search operation is executed based on the multi-dimensional association index. Query results are sorted by relevance to the search keywords, encapsulated using the TLS 1.3 encrypted transmission protocol, and an encrypted transmission data packet containing a timestamp, requester identifier, and data digest is generated and sent to the requesting party. Simultaneously, the blockchain node records the sharing time, requester's ID, permission level, amount of data transmitted, search keywords, and sharing results. This recorded information is immutable, ensuring the sharing process is traceable. The system receives feedback from the requester; if transmission fails, the encapsulation and retransmission operations are re-executed until the requester confirms complete receipt.

[0060] VIII. Specific Implementation of the Backup and Recovery Module

[0061] A data backup cycle is set: a 12-hour backup interval when the daily update volume exceeds 10GB, and a 24-hour backup interval when the daily update volume is less than 10GB. Immediate backup is triggered when a major data update occurs (such as adding image feature data for more than 1000 similar diseases). A complete backup of the compressed ternary feature dataset, semantic mapping relationships, and multidimensional association index is performed according to the backup cycle. Backup data is stored in AES-256 encrypted storage format, consistent with the encryption algorithm of the hybrid storage architecture. Backup data is stored on a backup distributed node in the hybrid storage architecture. The backup distributed node is deployed in different data centers than the primary distributed node, with independent network links and physical isolation to avoid simultaneous data corruption due to data center failures. The integrity and availability of the backup data are verified weekly. A data comparison algorithm is used to verify the consistency between the backup data and the original stored data, and simulated recovery tests are used to verify data availability. The verification results are recorded on the blockchain node. When data is lost or corrupted in the primary distributed node, the system monitoring module immediately triggers the data recovery process. First, it detects the scope and extent of data corruption. Based on the corruption situation, it retrieves backup data for the corresponding time period from the backup distributed node and overwrites the corrupted data in the primary distributed node according to data type and storage partition. After recovery, it rebuilds the multidimensional association index and confirms the recovery effect through integrity verification to ensure that the data recovery success rate is not less than 99.9%. The time, amount of corrupted data, recovery duration, and recovery success rate of the recovery process are recorded through blockchain nodes.

[0062] The beneficial effects of this invention are as follows: By binding image features with clinical semantics and scientific research attributes, and combining them with multi-dimensional association indexes, this invention supports multi-dimensional dynamic retrieval, improving retrieval efficiency by more than 80% compared to traditional methods; Differentiated compression strategies are adopted for different types of features, reducing overall data redundancy by more than 60% while ensuring the core features remain intact, thus saving storage resources; A multi-layered protection system of blockchain + encrypted communication + intelligent anomaly monitoring is adopted to achieve full-process traceability of data access, ensuring immutability and effectively resisting illegal access and data theft; De-identification processing based on federated learning shares only semantic mapping relationships, protecting patient privacy while ensuring data availability and meeting the needs of multi-center collaboration; A comprehensive backup and rapid recovery mechanism ensures rapid recovery in case of data loss or damage, with a recovery success rate of no less than 99.9%; It can be applied to various scenarios such as hospital ophthalmology diagnostic data management, multi-center clinical research, and telemedicine collaboration, possessing strong practicality and promotional value.

[0063] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for storing ophthalmic image feature information, characterized in that, The method includes: S1: After adaptive noise reduction of the original ophthalmic images, the corneal and retinal anatomical features are extracted using a deep learning model that combines an improved U-Net with an attention mechanism. The anatomical features are semantically bound to ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features, clinical semantics, and research attributes. S2: Construct a hybrid storage architecture consisting of blockchain and distributed file system. The hybrid storage architecture includes blockchain nodes and distributed nodes. The blockchain nodes store the traceability information and access logs of the ternary feature dataset. S3: Construct a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing. The multidimensional association index is stored in the distributed node. The multidimensional association index supports multidimensional dynamic retrieval of patients' disease courses according to diagnostic needs, research directions, and research directions. S4: A hierarchical compression algorithm optimized by generative adversarial networks is used to perform lossless compression on the key clinical features in the ternary feature dataset, and semantically preservative lossy compression on the auxiliary research features in the ternary feature dataset to obtain a compressed ternary feature dataset; the compressed ternary feature dataset is then desensitized based on a federated learning framework, retaining only the semantic mapping relationship of the compressed ternary feature dataset. S5: Configure a biometric dual verification mechanism and a feature access behavior anomaly monitoring mechanism for the hybrid storage architecture. The biometric dual verification mechanism adopts a verification method that combines doctor fingerprints and diagnostic permissions. The feature access behavior anomaly monitoring mechanism is based on machine learning to identify illegal retrieval, forming a closed-loop mechanism for storage traceability protection sharing. S6: The distributed nodes in the hybrid storage architecture provide a sharing service for the compressed ternary feature dataset after de-identification processing. During the sharing process, only the semantic mapping relationship is transmitted, and the original ophthalmological images and privacy information are not disclosed.

2. The method for storing ophthalmic image feature information according to claim 1, characterized in that, The steps of adaptively denoising the original ophthalmic images, extracting corneal and retinal anatomical features using a modified U-Net deep learning model incorporating an attention mechanism, and semantically binding these anatomical features with ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features with clinical semantic research attributes include: The ophthalmological clinical terms and research labels are standardized by removing duplicate terms, invalid labels and ambiguous expressions according to the national standard for medical terminology, forming a standardized term label set with a unified structure. Based on the type, attributes, and clinical application scenarios of the anatomical structural features, a unique semantic identifier is assigned to each term and tag in the standardized term tag set; A bidirectional mapping relationship between the anatomical structural features and the semantic identifiers is established. A one-to-one correspondence is achieved by constructing a mapping table to ensure that each anatomical structural feature corresponds to at least one clinical term semantic identifier and one scientific research tag semantic identifier. The cosine similarity algorithm is used to calculate the matching degree between the anatomical structure features and the corresponding semantic identifiers. Data with a matching degree of less than 95% is judged as a mapping error and is removed. The validated anatomical features, semantic identifiers, and corresponding term labels are integrated to generate the ternary feature dataset.

3. The method for storing ophthalmic image feature information according to claim 1, characterized in that, The step of constructing a hybrid storage architecture composed of blockchain and distributed file system, wherein the hybrid storage architecture includes blockchain nodes and distributed nodes, and storing the traceability information and access logs of the ternary feature dataset in the blockchain nodes, includes: Deploy at least three blockchain nodes, each set up on an independent server in a different physical location, to form a decentralized distributed blockchain network; The blockchain network is initialized and configured, and a practical Byzantine fault-tolerant consensus mechanism is adopted to set the data synchronization frequency of each blockchain node. Deploy a distributed file system, which includes at least two distributed nodes, and the distributed nodes are connected to the blockchain node through an encrypted communication link; The distributed nodes are allocated independent storage partitions according to feature type, corresponding to anatomical structure feature storage partition, clinical semantic feature storage partition and scientific research attribute feature storage partition respectively; Establish data interaction rules between the blockchain node and the distributed node. The data interaction rules include the following: data transmission order is one-way push from the distributed node to the blockchain node; transmitted data is encrypted using AES-256; integrity is confirmed by hash verification after transmission is completed; a three-retransmission mechanism is triggered when transmission fails; the synchronization trigger condition for traceability information and access log is specified as being within 10 seconds after the three-element feature dataset is written, modified, or deleted and successfully stored by the distributed node; the synchronization scope includes the operation subject, operation content, operation time, and data storage address; and the blockchain node returns a confirmation receipt after synchronization is completed.

4. The method for storing ophthalmic image feature information according to claim 3, characterized in that, The step of constructing a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing, wherein the multidimensional association index is stored in the distributed node, and the multidimensional association index supports multidimensional dynamic retrieval of patient disease course according to diagnostic needs, research directions, and clinical scenarios, including: The SHA-256 algorithm, combined with semantic weights, is used to calculate the feature semantic hash value of each data item in the ternary feature dataset; The ternary feature dataset is classified into three categories based on diagnostic needs, research directions, and patient disease course, forming three categories of scenario data subsets. A dedicated clinical scenario index is constructed for each category of scenario data subset. The clinical scenario index includes the distributed node storage address, semantic identifier, and associated data item ID of the data item. The feature semantic hash value is associated and integrated with the clinical scenario index according to semantic identifier to construct the multidimensional association index. The dynamic update rules of the multidimensional association index are set. When the ternary feature dataset is added, modified or deleted, the index incremental update is triggered and the historical index version is retained.

5. The method for storing ophthalmic image feature information according to claim 1, characterized in that, The steps of performing lossless compression on key clinical features in the ternary feature dataset and semantically preserving lossy compression on auxiliary research features in the ternary feature dataset using a generative adversarial network-optimized hierarchical compression algorithm to obtain the compressed ternary feature dataset include: Based on three evaluation indicators—diagnostic necessity, impact on clinical decision-making, and data irreplaceability—the Analytic Hierarchy Process (AHP) is used to prioritize the key clinical features in the ternary feature dataset, dividing them into primary key features and secondary key features. A lossless compression threshold is set for the primary key features, a lossless compression threshold with a fault tolerance rate of less than 5% is set for the secondary key features, and a lossy compression threshold with a semantic retention rate of more than 90% is set for the auxiliary research features. A compression algorithm optimized by generative adversarial network is used to perform compression operations on corresponding features according to a set threshold, generating compressed data packets with check codes. The integrity of the compressed data packets is checked using the CRC32 check algorithm, and compressed data blocks that fail the check are removed and compression is re-executed. Add feature type identifiers and compression method identifiers to the verified compressed data packets, and store them according to feature type to form the compressed ternary feature dataset.

6. The ophthalmic image feature information storage method according to claim 5, characterized in that, The step of desensitizing the compressed ternary feature dataset based on the federated learning framework, retaining only the semantic mapping relationship of the compressed ternary feature dataset, includes: A desensitization model is constructed based on a federated learning framework, which includes a local training module and a global aggregation module. The compressed ternary feature dataset is input into the local training module to extract the semantic mapping relationship in the data and remove privacy-related fields such as patient name, ID number, hospital name, and contact information. The semantic mapping relationships extracted from each of the distributed nodes are aggregated and optimized using a federated averaging algorithm, and redundant mapping entries are removed to form a unified semantic mapping set. The unified semantic mapping set is encrypted using an asymmetric encryption algorithm to generate corresponding encryption and decryption keys. The encrypted unified semantic mapping set is used as the desensitization result and stored in the shared partition of the distributed file system. The decryption key is only allocated to authorized shared nodes.

7. The method for storing ophthalmic image feature information according to claim 3, characterized in that: The abnormal access behavior monitoring mechanism includes: collecting historical legitimate access behavior data, which includes access time, access frequency, retrieval scope, operation permissions, data access volume, and access device information; A normal access behavior model is constructed by training the historical legitimate access behavior data using the random forest algorithm. Set anomaly judgment threshold, which includes access frequency threshold, retrieval range threshold, permission matching threshold, data access volume threshold and device consistency threshold. Collect data and extract features from real-time access behavior, and compare the extracted real-time access features with the normal access behavior model. When the real-time access characteristics exceed the anomaly judgment threshold, it is determined to be an illegal retrieval, triggering an early warning mechanism and blocking the access request, recording the illegal access information and storing it in the blockchain node.

8. The ophthalmic image feature information storage method according to claim 6, characterized in that: The sharing service includes: authenticating the sharing requester, the authentication content of which includes the doctor's fingerprint and diagnostic permissions. When the fingerprint matching degree is not less than 99% and the permission level meets the requirements, the verification is successful and the verification result is obtained. Based on the verification results, a corresponding sharing permission level is assigned to the sharing requester. Different permission levels correspond to different ranges of semantic mapping access permissions. The system receives the retrieval instructions sent by the sharing requester, performs format verification and permission matching on the retrieval instructions, and executes the retrieval after the verification is passed. Based on the semantic mapping relationship corresponding to the multidimensional association index query, the query results are sorted by relevance and then encapsulated using an encrypted transmission protocol to generate an encrypted transmission data packet. The encrypted transmission data packet is then sent to the sharing requester. At the same time, the key information of this sharing is recorded in the blockchain node, and the recorded information cannot be tampered with.

9. The method for storing ophthalmic image feature information according to claim 3, characterized in that: It also includes data backup and recovery steps: setting a data backup cycle, determining the backup interval based on the update frequency of the ternary feature dataset, and triggering immediate backup after major data updates; The compressed ternary feature dataset, the semantic mapping relationship, and the multidimensional association index are fully backed up according to the backup cycle. Backup data is stored on a standby distributed node in the hybrid storage architecture, and the standby distributed node is physically isolated from the primary distributed node. Regularly verify the integrity and availability of backup data to ensure that the backup data is consistent with the original stored data and can be read normally; When data is lost or corrupted in the primary distributed node, a data recovery process is triggered, which retrieves the corresponding backup data from the backup distributed node for recovery, and records the recovery process through the blockchain node.

10. An ophthalmic image feature information storage system, characterized in that, The system includes: a semantic binding module, a hybrid storage architecture module, an index building module, a hierarchical compression module, a desensitization processing module, an anomaly monitoring module, a shared service module, and a backup and recovery module. These modules are sequentially connected and collaboratively implement the ophthalmic image feature information storage method described in claims 1-9. The semantic binding module: After adaptively denoising the original ophthalmic image, it uses an improved U-Net deep learning model combined with an attention mechanism to extract corneal and retinal anatomical features, and semantically binds the anatomical features with ophthalmic clinical terminology research tags to generate a ternary feature dataset of image features, clinical semantics, and research attributes. The hybrid storage architecture module: constructs a hybrid storage architecture composed of blockchain and distributed file system. The hybrid storage architecture includes blockchain nodes and distributed nodes. The blockchain nodes store the traceability information and access logs of the ternary feature dataset. The index building module constructs a multidimensional association index for the ternary feature dataset based on feature semantic hashing and clinical scenario indexing. The multidimensional association index is stored in the distributed node and supports multidimensional dynamic retrieval of patients' disease courses according to diagnostic needs, research directions, and research directions. The hierarchical compression module employs a generative adversarial network-optimized hierarchical compression algorithm to perform lossless compression on the key clinical features in the ternary feature dataset and semantically preservative lossy compression on the auxiliary research features in the ternary feature dataset, resulting in a compressed ternary feature dataset. The desensitization processing module performs desensitization processing on the compressed ternary feature dataset based on the federated learning framework, retaining only the semantic mapping relationship of the compressed ternary feature dataset; The anomaly monitoring module configures a biometric dual verification mechanism and a feature access behavior anomaly monitoring mechanism for the hybrid storage architecture. The biometric dual verification mechanism adopts a verification method that combines doctor fingerprints and diagnostic permissions. The feature access behavior anomaly monitoring mechanism is based on machine learning to identify illegal retrieval, forming a closed-loop mechanism for storage traceability protection and sharing. The shared service module provides a sharing service for the de-identified compressed ternary feature dataset through the distributed nodes in the hybrid storage architecture. During the sharing process, only the semantic mapping relationship is transmitted, and the original ophthalmological images and privacy information are not disclosed. The backup and recovery module performs a complete backup of the compressed ternary feature dataset, the semantic mapping relationship, and the multidimensional association index according to the backup cycle; stores the backup data in a standby distributed node of the hybrid storage architecture, which is physically isolated from the primary distributed node; periodically verifies the integrity and availability of the backup data to ensure that the backup data is consistent with the original stored data and can be read normally; when data in the primary distributed node is lost or damaged, a data recovery process is triggered to retrieve the corresponding backup data from the standby distributed node for recovery, and the recovery process is recorded through the blockchain node.

Images