Sugar net recognition system based on block chain and federal learning

Abstract

The invention discloses a sugar net recognition system based on a block chain and federal learning, and the system comprises a data collection module which is used for collecting fundus image data and related medical data; the federal learning module is connected with the data acquisition module, the federal learning module comprises a plurality of participant nodes, and each participant node performs model training based on local data; the block chain module is connected with the federated learning module and used for storing model parameters, training data abstracts and data access record key information in the federated learning process; the sugar network identification module is used for receiving the global sugar network identification model trained by the federal learning module; and the result output and feedback module is connected with the sugar net identification module. According to the method, the data collaborative training model is integrated through federal learning, so that the sugar net recognition accuracy is improved, misdiagnosis and missed diagnosis are reduced, the block chain ensures the data to be safe and credible, the innovative formula assists optimization, the system application is wide, clinical and remote medical treatment is facilitated, and the sugar net diagnosis service quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of diabetic retinopathy identification systems, and in particular to a diabetic retinopathy identification system based on blockchain and federated learning. Background Art

[0002] Diabetic retinopathy (DR) is a common and serious microvascular complication in diabetic patients and one of the main causes of blindness in adults. Early accurate diagnosis and timely intervention are crucial to delaying the progression of DR and protecting patients' vision. However, traditional DR diagnosis faces many challenges.

[0003] In terms of medical data, data related to diabetic retinopathy are scattered in different medical institutions, ophthalmology clinics and other places. These data are often difficult to share and integrate effectively due to issues such as privacy protection and data ownership. Each institution has its own patient data set, but these data sets are relatively isolated and limited in size, making it difficult to cover various types and stages of diabetic retinopathy cases, which limits the accuracy and generalization ability of diagnostic models trained based on a single data set. Moreover, there are differences in data collection standards and quality in different regions and medical institutions, which further increases the difficulty of data integration and utilization.

[0004] In terms of diagnostic technology, it currently mainly relies on ophthalmologists to make subjective judgments through funduscopy and other methods. This method not only relies on the doctor's experience and professional level, but also has low efficiency and is prone to misdiagnosis and missed diagnosis. With the development of artificial intelligence technology, although some diabetic retinopathy identification methods based on machine learning have been proposed, these methods usually require a large amount of labeled data for model training, and obtaining large-scale, high-quality labeled data is costly and difficult. In addition, the traditional centralized machine learning model training method requires all data to be concentrated on a central server for processing, which not only poses the risk of data privacy leakage, but may also face the pressure of data transmission and storage, especially for sensitive information such as medical data.

[0005] In terms of technology development trends, blockchain technology, with its decentralization, immutability, and traceability, provides new ideas for solving data security and sharing issues. It can ensure the authenticity and integrity of data, while protecting the rights and interests of data owners, allowing different institutions to interact and cooperate on data based on trust. Federated learning is an emerging distributed machine learning technology that allows multiple participants to jointly train a global model without sharing the original data, effectively solving data privacy and data island issues, and is very suitable for use in scenarios such as the medical field that have strict requirements on data privacy.

[0006] Therefore, it is of great practical significance to develop a diabetic retinopathy identification system by combining blockchain and federated learning technology. This system can integrate diabetic retinopathy data scattered across the country, realize collaborative training and optimization of models under the premise of protecting data privacy, improve the accuracy and efficiency of diabetic retinopathy identification, provide stronger support for the early diagnosis and treatment of diabetic retinopathy, and is expected to improve the prognosis of diabetic retinopathy patients and reduce the risk of blindness. At the same time, it also provides new solutions and technical means for the digital and intelligent development of the medical field. Summary of the invention

[0007] The present invention proposes a diabetic retinopathy identification system based on blockchain and federated learning to solve the problems mentioned in the above-mentioned prior art.

[0008] In order to achieve the above objectives, the present invention adopts the following technical solutions: a diabetic retinopathy identification system based on blockchain and federated learning, comprising:

[0009] A data acquisition module is used to collect fundus image data and related medical data, wherein the fundus image data includes fundus images of diabetic retinopathy patients at different stages and normal fundus images, and the related medical data includes basic information, medical history, and blood sugar test data of the patient;

[0010] A federated learning module is connected to the data acquisition module, and includes multiple participant nodes. Each participant node performs model training based on local data, and exchanges parameters and updates the model through a federated learning algorithm, so as to jointly train a global diabetic retinopathy recognition model without leaking the original data of each party.

[0011] A blockchain module, connected to the federated learning module, is used to store model parameters, training data summaries, and key information of data access records in the federated learning process to ensure that the data cannot be tampered with and is traceable;

[0012] The diabetic retinopathy recognition module receives the global diabetic retinopathy recognition model trained by the federated learning module, and uses the model to analyze and recognize the input fundus image to determine whether the patient has diabetic retinopathy and the severity of diabetic retinopathy;

[0013] The result output and feedback module is connected to the diabetic retinopathy identification module, outputs the result of diabetic retinopathy identification to the user in a visual manner, and optimizes and improves the system according to the user's feedback.

[0014] Introducing the formula for diabetic retinopathy identification accuracy: Among them, TP represents true positive (number of samples correctly identified as diabetic retinopathy patients), TN represents true negative (number of samples correctly identified as non-diabetic retinopathy patients), FP represents false positive (number of samples incorrectly identified as diabetic retinopathy patients), and FN represents false negative (number of samples incorrectly identified as non-diabetic retinopathy patients). This formula is used to measure the accuracy of the system's recognition of diabetic retinopathy and is one of the key indicators for evaluating system performance.

[0015] Introduce the data update frequency formula: Where f represents the data update frequency, and n represents the number of data updates within time T. This formula can be used to reasonably plan the data update strategy of the data acquisition module to ensure that the data set used by the system is timely and representative.

[0016] Introduce the model performance improvement rate formula: Where R represents the model performance improvement rate, Accuracy new Indicates the accuracy of diabetic retinopathy recognition of the optimized model, Accuracy old Represents the accuracy of diabetic retinopathy recognition of the model before optimization. This formula is used to evaluate the effect of model optimization and reflects the degree of performance improvement of the system through continuous improvement and learning.

[0017] Furthermore, the data acquisition module includes:

[0018] Image acquisition unit, equipped with a professional fundus camera or equipment capable of taking fundus images, and corresponding image acquisition software, to ensure that the acquired fundus images have clear resolution and accurate color reproduction, and the image resolution is not less than 512×512 pixels;

[0019] A data preprocessing unit preprocesses the collected fundus images and related medical data, including image denoising, image enhancement, and data standardization operations to improve the quality and usability of the data. Image denoising uses mean filtering, median filtering, or Gaussian filtering algorithms, image enhancement uses contrast enhancement and brightness adjustment methods, and data standardization normalizes the data to a specific interval;

[0020] The data annotation unit is where professional medical personnel annotate fundus images. The annotation content includes whether the patient has diabetic retinopathy, the stage of diabetic retinopathy (such as mild, moderate, severe), and information on the location of the lesion. The annotation accuracy is no less than 95%.

[0021] Furthermore, the federated learning module includes:

[0022] The participant node management unit is responsible for managing and coordinating the joining, exiting and communication of each participant node, ensuring stable connection and data transmission between the participant nodes, and using encrypted communication protocols to ensure communication security;

[0023] Local model training unit: each participant node uses local data to conduct preliminary training of the diabetic retinopathy recognition model. The training algorithms include but are not limited to convolutional neural network (CNN) and recurrent neural network (RNN) deep learning algorithms. Appropriate algorithms and model structures are selected according to the characteristics of local data. Stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimization algorithms are used to optimize model parameters during training.

[0024] The model parameter aggregation unit aggregates the trained model parameters of each participant's nodes through a federated learning algorithm to obtain updated parameters of the global model. The aggregation algorithm can adopt a federated averaging algorithm (Federated Averaging) or other suitable aggregation methods. In the parameter aggregation process, the data volume and data quality factors of each participant's nodes are considered for weighted processing.

[0025] Introduce the parameter aggregation formula of the federated learning model: where θ global represents the global model parameters, n is the number of participating nodes, and w i is the weight of the i-th participant node, θ i is the local model parameter after training of the i-th participant node. Weight w i According to the data volume of the participating nodes d i and data quality assessment indicator q i Determine, for example The amount of data d i It can be measured by the number of data samples, and the data quality evaluation index q i The accuracy, completeness and labeling consistency of the data can be comprehensively considered for quantitative assessment.

[0026] The model update and distribution unit updates the global diabetic retinopathy recognition model according to the aggregated parameters, and distributes the updated model to each participating node, so that each participating node can continue the next round of local training and parameter upload to achieve continuous optimization of the model.

[0027] Furthermore, the blockchain module includes:

[0028] The blockchain network construction unit uses the alliance chain architecture to build the blockchain network. The nodes in the alliance chain are composed of participating medical institutions and scientific research institutions. The nodes reach a consensus through the practical Byzantine fault tolerance algorithm (PBFT) to ensure the consistency and reliability of the data on the blockchain.

[0029] The data storage unit stores the model parameters, training data summary, and key information of data access records in the process of federated learning in the form of blocks on the blockchain. Each block contains the hash value of the previous block to form a chain structure to ensure that the data cannot be tampered with and is traceable. The data storage adopts a distributed storage method to improve the storage security and reliability of the data.

[0030] Smart contract unit: write and deploy smart contracts to regulate data sharing, model updates, node incentive rules and operation procedures in the federated learning process. Smart contracts are automatically executed without third-party intervention to ensure that all parties collaborate as agreed. For example, when a participating node successfully completes a round of federated learning tasks, the smart contract automatically gives corresponding rewards according to the preset incentive mechanism. The rewards can be digital currency, points or other forms of rights and interests.

[0031] The data encryption and privacy protection unit uses encryption technology to encrypt data stored on the blockchain to ensure the security and privacy of the data. Before the data is uploaded to the blockchain, an asymmetric encryption algorithm (such as RSA) is used to encrypt the data. Only authorized nodes can decrypt and access the data. At the same time, zero-knowledge proof privacy protection technology is used to perform data verification and calculation without leaking the original data.

[0032] Furthermore, the diabetic retinopathy identification module includes:

[0033] The model loading unit obtains the trained global diabetic retinopathy recognition model from the federated learning module and loads it into the memory to prepare for diabetic retinopathy recognition;

[0034] The image analysis unit extracts and analyzes the input fundus images, and uses the convolutional neural network algorithm in deep learning to automatically extract the features of the fundus images, including vascular morphology and lesion features. In the feature extraction process, the pooling layer and convolution layer network structure are used to process the image, reduce the data dimension, and extract key features.

[0035] The classification and diagnosis unit uses the trained model to classify and diagnose fundus images based on the extracted image features to determine whether the patient has diabetic retinopathy and the severity of the diabetic retinopathy. The classification results can be divided into normal, mild diabetic retinopathy, moderate diabetic retinopathy, and severe diabetic retinopathy. The threshold judgment and probability estimation methods are used to determine the final diagnosis results during the diagnosis process.

[0036] The result evaluation unit evaluates and verifies the results of diabetic retinopathy identification, uses accuracy, recall rate, and F1 value evaluation indicators to quantitatively evaluate the identification results, and continuously adjusts and optimizes the model parameters by comparing them with the diagnosis results of professional doctors to improve the accuracy and reliability of diabetic retinopathy identification.

[0037] Furthermore, the result output and feedback module includes:

[0038] The result visualization unit outputs the results of diabetic retinopathy identification to users in the form of intuitive charts and reports, including the annotation results of fundus images, the diagnosis conclusion of diabetic retinopathy, and the severity assessment information. The charts can use bar charts and line charts to show the distribution of different types of diabetic retinopathy. The report contains detailed diagnostic instructions and suggestions.

[0039] The user feedback collection unit collects user feedback on the diabetic retinopathy identification results through the user interface, including evaluation of the accuracy of the diagnosis results, suggestions and requirements for the system. The feedback can be in the form of questionnaires and online comments to ensure that users can easily provide feedback information;

[0040] The system optimization unit optimizes and improves the system based on user feedback, including adjusting model parameters, improving data collection methods, optimizing the user interface, and continuously improving system performance and user experience. For example, if users report that the diagnosis results are inaccurate, the system can recheck the data collection and labeling process, or further optimize the model training algorithm to improve the accuracy of the diagnosis.

[0041] Furthermore, the system should meet the following performance and safety requirements:

[0042] The accuracy rate of diabetic retinopathy identification is no less than 90%, and the identification performance can be kept stable on diabetic retinopathy samples of different types and severity.

[0043] The system response time does not exceed 5 seconds, ensuring that users can obtain diabetic retinopathy identification results in a timely manner;

[0044] Data security complies with relevant medical data security standards and regulatory requirements, and uses multiple encryption technologies and access control mechanisms to ensure data confidentiality, integrity and availability;

[0045] The system has good scalability and compatibility, can be easily integrated with other medical information systems, and supports the access of new data sources and algorithms to adapt to the ever-evolving diabetic retinopathy identification technology and medical needs.

[0046] Furthermore, the system can be applied to the following scenarios and has corresponding expansion functions:

[0047] Clinical diagnosis assistance: Provide ophthalmologists with auxiliary reference for diabetic retinopathy diagnosis, improve the accuracy and efficiency of diagnosis, and reduce the rate of misdiagnosis and missed diagnosis. Doctors can refer to the system's diagnosis results during clinical diagnosis and make a comprehensive diagnosis based on their own professional judgment. The system can also provide case management functions to facilitate doctors to track and analyze patients' medical records and diagnosis results;

[0048] Telemedicine: Remote diabetic retinopathy diagnosis can be achieved through the Internet. Patients can take fundus images at primary medical institutions and upload them to the system for remote diagnosis by professional doctors, breaking geographical restrictions and improving the utilization efficiency of medical resources. The system should have remote data transmission and communication functions to ensure the safe and fast transmission of images and data, and support remote consultation and discussion functions to facilitate communication and collaboration between doctors.

[0049] Disease screening: used for large-scale diabetic retinopathy screening, rapid detection of diabetic retinopathy patients in the population, and early detection and treatment. The system can cooperate with community health service centers and physical examination institutions to carry out diabetic retinopathy screening projects. Through automated diabetic retinopathy identification and result analysis, it can improve screening efficiency and reduce screening costs. At the same time, it can collect and analyze screening data to provide decision-making support for disease prevention and control.

[0050] Scientific research cooperation: Provide data and technical support for medical research related to diabetic retinopathy, promote the development and innovation of diabetic retinopathy identification technology, and all parties involved can share data and models through blockchain and federated learning technology to carry out joint research projects. The system should provide a platform for data sharing and collaborative research, support scientific researchers in data analysis, model training and algorithm optimization, and protect the intellectual property rights and data rights of all parties.

[0051] Furthermore, the system has data updating and model optimization functions:

[0052] Data update mechanism: regularly (once a month) or irregularly update the fundus image data and related medical data in the data acquisition module to ensure the timeliness and diversity of the data. Data updates can be obtained through cooperation with medical institutions and scientific research institutions to obtain new case data, or through user feedback and follow-up to collect the latest information of patients. During the update process, strict quality control and annotation of new data are carried out to ensure the accuracy and availability of data;

[0053] Online model optimization: Based on new data and user feedback, the federated learning module is used to optimize and update the diabetic retinopathy identification model online to continuously improve the performance and accuracy of the model. During the online optimization process, the system automatically triggers the model update process. Each participating node performs local training based on the new data, and then uploads the updated model parameters to the blockchain for aggregation and update, ensuring that the model can adapt to new situations and needs in a timely manner.

[0054] Model evaluation and verification: Regularly (once every quarter) evaluate and verify the optimized model, and use an independent test data set to test the model's performance to ensure that the model's accuracy and reliability will not decline. Evaluation indicators include but are not limited to accuracy, recall rate, and F1 value. At the same time, invite professional doctors to evaluate and verify the model's diagnostic results, and further adjust and optimize the model based on the evaluation results.

[0055] Furthermore, the system has user management and authority control functions:

[0056] User registration and login: Users (including medical staff, researchers, and patients) need to register and log in to use the system. Registration information includes user name, contact information, and identity type. The system verifies and reviews user information to ensure the authenticity and legitimacy of the user's identity.

[0057] User rights management: Different rights are assigned according to the user's identity and role. For example, medical staff can perform data collection, model training, and result viewing operations, scientific researchers can perform data analysis and model optimization operations, and patients can view their own diagnosis results and related information. Rights management is implemented through access control lists (ACLs) or other rights management mechanisms to ensure that users can only operate within their own rights.

[0058] Data access control: Access control is performed on the data in the system. Different users have different access rights to data. For example, a patient's personal privacy data can only be viewed and processed by authorized medical institution staff, and researchers can only access some data under the premise of meeting research purposes and ethical requirements. Data access control is implemented through encryption technology and access authorization mechanism to ensure data security and privacy;

[0059] User behavior audit: Audit and record the user's operating behavior in the system, including login time, operation content, and data access records, for easy traceability and supervision. Audit records can be used to discover potential security issues and violations, and also provide a reference for system optimization and improvement.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention uses a federated learning module to enable multiple parties to collaboratively train the diabetic retinopathy recognition model without sharing the original data. This breaks the limitations of data silos and integrates diabetic retinopathy data resources scattered across different medical institutions, allowing the model to learn a wider range of richer case characteristics, thereby significantly improving the accuracy of diabetic retinopathy recognition. Compared with the model trained by a traditional single institution, the model of the present invention can better adapt to different types and stages of diabetic retinopathy lesions, effectively reduce the occurrence of misdiagnosis and missed diagnosis, provide patients with more reliable diagnostic results, assist in early detection and treatment, and reduce the risk of blindness caused by diabetic retinopathy.

[0062] With the help of blockchain modules, the present invention ensures the data is tamper-proof and traceable. This provides a strong guarantee for the security and authenticity of medical data and enhances the trust of all parties in the data. In the process of data sharing and model training, there is no need to worry about the data being maliciously tampered with or forged. At the same time, the source and flow of the data can be clearly traced, which meets the strict security and compliance requirements of medical data. This trust mechanism promotes broader cooperation between medical institutions and provides richer data support for diabetic retinopathy research and diagnosis.

[0063] The present invention achieves quantitative evaluation and precise optimization of system performance by introducing innovative formulas, such as the formula for diabetic retinopathy recognition accuracy, the formula for data update frequency, and the formula for model performance improvement rate. It can accurately measure the accuracy of diabetic retinopathy recognition, rationally plan data update strategies, and effectively evaluate the effect of model optimization, so that the system can continuously improve and perfect itself, and continuously improve the performance and quality of diabetic retinopathy recognition.

[0064] In practical applications, this system has wide applicability and scalability. It can be used to assist in clinical diagnosis, provide decision support for ophthalmologists, and improve diagnostic efficiency; support telemedicine, break geographical restrictions, and allow patients to enjoy professional diabetic retinopathy diagnosis services at the grassroots level; it is suitable for large-scale disease screening, improves screening efficiency, reduces costs, and assists in public health prevention and control; it can also provide a platform for scientific research cooperation and promote the innovation and development of diabetic retinopathy identification technology. In short, the present invention has important value in improving the accuracy of diabetic retinopathy diagnosis, ensuring data security, promoting medical cooperation and innovation, and has made positive contributions to improving the quality of medical services and prognosis of diabetic retinopathy patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a schematic block diagram of a diabetic retinopathy identification system based on blockchain and federated learning proposed in the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise" and "counterclockwise" indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0068] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0069] Reference Figure 1 :A diabetic retinopathy identification system based on blockchain and federated learning, including:

[0070] A data acquisition module is used to collect fundus image data and related medical data, wherein the fundus image data includes fundus images of diabetic retinopathy patients at different stages and normal fundus images, and the related medical data includes basic information, medical history, and blood sugar test data of the patient;

[0071] A federated learning module is connected to the data acquisition module, and includes multiple participant nodes. Each participant node performs model training based on local data, and exchanges parameters and updates the model through a federated learning algorithm, so as to jointly train a global diabetic retinopathy recognition model without leaking the original data of each party.

[0072] A blockchain module, connected to the federated learning module, is used to store model parameters, training data summaries, and key information of data access records in the federated learning process to ensure that the data cannot be tampered with and is traceable;

[0073] The diabetic retinopathy recognition module receives the global diabetic retinopathy recognition model trained by the federated learning module, and uses the model to analyze and recognize the input fundus image to determine whether the patient has diabetic retinopathy and the severity of diabetic retinopathy;

[0074] The result output and feedback module is connected to the diabetic retinopathy identification module, outputs the result of diabetic retinopathy identification to the user in a visual manner, and optimizes and improves the system according to the user's feedback.

[0075] Introducing the formula for diabetic retinopathy identification accuracy: Among them, TP represents true positive (number of samples correctly identified as diabetic retinopathy patients), TN represents true negative (number of samples correctly identified as non-diabetic retinopathy patients), FP represents false positive (number of samples incorrectly identified as diabetic retinopathy patients), and FN represents false negative (number of samples incorrectly identified as non-diabetic retinopathy patients). This formula is used to measure the accuracy of the system's recognition of diabetic retinopathy and is one of the key indicators for evaluating system performance.

[0076] Introduce the data update frequency formula: Where f represents the data update frequency, and n represents the number of data updates within time T. This formula can be used to reasonably plan the data update strategy of the data acquisition module to ensure that the data set used by the system is timely and representative.

[0077] Introduce the model performance improvement rate formula: Where R represents the model performance improvement rate, Accuracy new Indicates the accuracy of diabetic retinopathy recognition of the optimized model, Accuracy old Represents the accuracy of diabetic retinopathy recognition of the model before optimization. This formula is used to evaluate the effect of model optimization and reflects the degree of performance improvement of the system through continuous improvement and learning.

[0078] In the present invention, the data acquisition module includes:

[0079] Image acquisition unit, equipped with a professional fundus camera or equipment capable of taking fundus images, and corresponding image acquisition software, to ensure that the acquired fundus images have clear resolution and accurate color reproduction, and the image resolution is not less than 512×512 pixels;

[0080] A data preprocessing unit preprocesses the collected fundus images and related medical data, including image denoising, image enhancement, and data standardization operations to improve the quality and usability of the data. Image denoising uses mean filtering, median filtering, or Gaussian filtering algorithms, image enhancement uses contrast enhancement and brightness adjustment methods, and data standardization normalizes the data to a specific interval;

[0081] The data annotation unit is where professional medical personnel annotate fundus images. The annotation content includes whether the patient has diabetic retinopathy, the stage of diabetic retinopathy (such as mild, moderate, severe), and information on the location of the lesion. The annotation accuracy is no less than 95%.

[0082] In the present invention, the federated learning module includes:

[0083] The participant node management unit is responsible for managing and coordinating the joining, exiting and communication of each participant node, ensuring stable connection and data transmission between the participant nodes, and using encrypted communication protocols to ensure communication security;

[0084] Local model training unit: each participant node uses local data to conduct preliminary training of the diabetic retinopathy recognition model. The training algorithms include but are not limited to convolutional neural network (CNN) and recurrent neural network (RNN) deep learning algorithms. Appropriate algorithms and model structures are selected according to the characteristics of local data. Stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimization algorithms are used to optimize model parameters during training.

[0085] The model parameter aggregation unit aggregates the trained model parameters of each participant's nodes through a federated learning algorithm to obtain updated parameters of the global model. The aggregation algorithm can adopt a federated averaging algorithm (Federated Averaging) or other suitable aggregation methods. In the parameter aggregation process, the data volume and data quality factors of each participant's nodes are considered for weighted processing.

[0086] Introduce the parameter aggregation formula of the federated learning model: where θ global represents the global model parameters, n is the number of participating nodes, and w i is the weight of the i-th participant node, θ i is the local model parameter after training of the i-th participant node. Weight w i According to the data volume of the participating nodes d i and data quality assessment index q i Determine, for example The amount of data d i It can be measured by the number of data samples, and the data quality evaluation index q i The accuracy, completeness and labeling consistency of the data can be comprehensively considered for quantitative assessment.

[0087] The model update and distribution unit updates the global diabetic retinopathy recognition model according to the aggregated parameters, and distributes the updated model to each participating node, so that each participating node can continue the next round of local training and parameter upload to achieve continuous optimization of the model.

[0088] In the present invention, the blockchain module includes:

[0089] The blockchain network construction unit uses the alliance chain architecture to build the blockchain network. The nodes in the alliance chain are composed of participating medical institutions and scientific research institutions. The nodes reach a consensus through the practical Byzantine fault tolerance algorithm (PBFT) to ensure the consistency and reliability of the data on the blockchain.

[0090] The data storage unit stores the model parameters, training data summary, and key information of data access records in the process of federated learning in the form of blocks on the blockchain. Each block contains the hash value of the previous block to form a chain structure to ensure that the data cannot be tampered with and is traceable. The data storage adopts a distributed storage method to improve the storage security and reliability of the data.

[0091] Smart contract unit: write and deploy smart contracts to regulate data sharing, model updates, node incentive rules and operation procedures in the federated learning process. Smart contracts are automatically executed without third-party intervention to ensure that all parties collaborate as agreed. For example, when a participating node successfully completes a round of federated learning tasks, the smart contract automatically gives corresponding rewards according to the preset incentive mechanism. The rewards can be digital currency, points or other forms of rights and interests.

[0092] The data encryption and privacy protection unit uses encryption technology to encrypt data stored on the blockchain to ensure the security and privacy of the data. Before the data is uploaded to the blockchain, an asymmetric encryption algorithm (such as RSA) is used to encrypt the data. Only authorized nodes can decrypt and access the data. At the same time, zero-knowledge proof privacy protection technology is used to perform data verification and calculation without leaking the original data.

[0093] In the present invention, the diabetic retinopathy identification module comprises:

[0094] The model loading unit obtains the trained global diabetic retinopathy recognition model from the federated learning module and loads it into the memory to prepare for diabetic retinopathy recognition;

[0095] The image analysis unit extracts and analyzes the input fundus images, and uses the convolutional neural network algorithm in deep learning to automatically extract the features of the fundus images, including vascular morphology and lesion features. In the feature extraction process, the pooling layer and convolution layer network structure are used to process the image, reduce the data dimension, and extract key features.

[0096] The classification and diagnosis unit uses the trained model to classify and diagnose fundus images based on the extracted image features to determine whether the patient has diabetic retinopathy and the severity of the diabetic retinopathy. The classification results can be divided into normal, mild diabetic retinopathy, moderate diabetic retinopathy, and severe diabetic retinopathy. The threshold judgment and probability estimation methods are used to determine the final diagnosis results during the diagnosis process.

[0097] The result evaluation unit evaluates and verifies the results of diabetic retinopathy identification, uses accuracy, recall rate, and F1 value evaluation indicators to quantitatively evaluate the identification results, and continuously adjusts and optimizes the model parameters by comparing them with the diagnosis results of professional doctors to improve the accuracy and reliability of diabetic retinopathy identification.

[0098] In the present invention, the result output and feedback module includes:

[0099] The result visualization unit outputs the results of diabetic retinopathy identification to users in the form of intuitive charts and reports, including the annotation results of fundus images, the diagnosis conclusion of diabetic retinopathy, and the severity assessment information. The charts can use bar charts and line charts to show the distribution of different types of diabetic retinopathy. The report contains detailed diagnostic instructions and suggestions.

[0100] The user feedback collection unit collects user feedback on the diabetic retinopathy identification results through the user interface, including evaluation of the accuracy of the diagnosis results, suggestions and requirements for the system. The feedback can be in the form of questionnaires and online comments to ensure that users can easily provide feedback information;

[0101] The system optimization unit optimizes and improves the system based on user feedback, including adjusting model parameters, improving data collection methods, optimizing the user interface, and continuously improving system performance and user experience. For example, if users report that the diagnosis results are inaccurate, the system can recheck the data collection and labeling process, or further optimize the model training algorithm to improve the accuracy of the diagnosis.

[0102] In the present invention, the system should meet the following performance and safety requirements:

[0103] The accuracy rate of diabetic retinopathy identification is no less than 90%, and the identification performance can be kept stable on diabetic retinopathy samples of different types and severity.

[0104] The system response time does not exceed 5 seconds, ensuring that users can obtain diabetic retinopathy identification results in a timely manner;

[0105] Data security complies with relevant medical data security standards and regulatory requirements, and uses multiple encryption technologies and access control mechanisms to ensure data confidentiality, integrity and availability;

[0106] The system has good scalability and compatibility, can be easily integrated with other medical information systems, and supports the access of new data sources and algorithms to adapt to the ever-evolving diabetic retinopathy identification technology and medical needs.

[0107] In the present invention, the system can be applied to the following scenarios and has corresponding expansion functions:

[0108] Clinical diagnosis assistance: Provide ophthalmologists with auxiliary reference for diabetic retinopathy diagnosis, improve the accuracy and efficiency of diagnosis, and reduce the rate of misdiagnosis and missed diagnosis. Doctors can refer to the system's diagnosis results during clinical diagnosis and make a comprehensive diagnosis based on their own professional judgment. The system can also provide case management functions to facilitate doctors to track and analyze patients' medical records and diagnosis results;

[0109] Telemedicine: Remote diabetic retinopathy diagnosis can be achieved through the Internet. Patients can take fundus images at primary medical institutions and upload them to the system for remote diagnosis by professional doctors, breaking geographical restrictions and improving the utilization efficiency of medical resources. The system should have remote data transmission and communication functions to ensure the safe and fast transmission of images and data, and support remote consultation and discussion functions to facilitate communication and collaboration between doctors.

[0110] Disease screening: used for large-scale diabetic retinopathy screening, rapid detection of diabetic retinopathy patients in the population, and early detection and treatment. The system can cooperate with community health service centers and physical examination institutions to carry out diabetic retinopathy screening projects. Through automated diabetic retinopathy identification and result analysis, it can improve screening efficiency and reduce screening costs. At the same time, it can collect and analyze screening data to provide decision-making support for disease prevention and control.

[0111] Scientific research cooperation: Provide data and technical support for medical research related to diabetic retinopathy, promote the development and innovation of diabetic retinopathy identification technology, and all parties involved can share data and models through blockchain and federated learning technology to carry out joint research projects. The system should provide a platform for data sharing and collaborative research, support scientific researchers in data analysis, model training and algorithm optimization, and protect the intellectual property rights and data rights of all parties.

[0112] In the present invention, the system has data updating and model optimization functions:

[0113] Data update mechanism: regularly (once a month) or irregularly update the fundus image data and related medical data in the data acquisition module to ensure the timeliness and diversity of the data. Data updates can be obtained through cooperation with medical institutions and scientific research institutions to obtain new case data, or through user feedback and follow-up to collect the latest information of patients. During the update process, strict quality control and annotation of new data are carried out to ensure the accuracy and availability of data;

[0114] Online model optimization: Based on new data and user feedback, the federated learning module is used to optimize and update the diabetic retinopathy identification model online to continuously improve the performance and accuracy of the model. During the online optimization process, the system automatically triggers the model update process. Each participating node performs local training based on the new data, and then uploads the updated model parameters to the blockchain for aggregation and update, ensuring that the model can adapt to new situations and needs in a timely manner.

[0115] Model evaluation and verification: Regularly (once every quarter) evaluate and verify the optimized model, and use an independent test data set to test the model's performance to ensure that the model's accuracy and reliability will not decline. Evaluation indicators include but are not limited to accuracy, recall rate, and F1 value. At the same time, invite professional doctors to evaluate and verify the model's diagnostic results, and further adjust and optimize the model based on the evaluation results.

[0116] In the present invention, the system has user management and authority control functions:

[0117] User registration and login: Users (including medical staff, researchers, and patients) need to register and log in to use the system. Registration information includes user name, contact information, and identity type. The system verifies and reviews user information to ensure the authenticity and legitimacy of the user's identity.

[0118] User rights management: Different rights are assigned according to the user's identity and role. For example, medical staff can perform data collection, model training, and result viewing operations, scientific researchers can perform data analysis and model optimization operations, and patients can view their own diagnosis results and related information. Rights management is implemented through access control lists (ACLs) or other rights management mechanisms to ensure that users can only operate within their own rights.

[0119] Data access control: Access control is performed on the data in the system. Different users have different access rights to data. For example, a patient's personal privacy data can only be viewed and processed by authorized medical institution staff, and researchers can only access some data under the premise of meeting research purposes and ethical requirements. Data access control is implemented through encryption technology and access authorization mechanism to ensure data security and privacy;

[0120] User behavior audit: Audit and record the user's operating behavior in the system, including login time, operation content, and data access records, for easy traceability and supervision. Audit records can be used to discover potential security issues and violations, and also provide a reference for system optimization and improvement.

[0121] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can replace or change the technical solution and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A diabetic retinopathy identification system based on blockchain and federated learning, characterized in that: include: A data acquisition module is used to collect fundus image data and related medical data, wherein the fundus image data includes fundus images of diabetic retinopathy patients at different stages and normal fundus images, and the related medical data includes basic information, medical history, and blood sugar test data of the patient; A federated learning module, connected to the data acquisition module, includes multiple participant nodes, each participant node performs model training based on local data, and exchanges parameters and updates the model through a federated learning algorithm, and jointly trains a global diabetic retinopathy recognition model without leaking the original data of each party; A blockchain module, connected to the federated learning module, for storing model parameters, training data summaries, and key information of data access records in the federated learning process; The diabetic retinopathy recognition module receives the global diabetic retinopathy recognition model trained by the federated learning module, and uses the model to analyze and recognize the input fundus image to determine whether the patient has diabetic retinopathy and the severity of diabetic retinopathy; A result output and feedback module is connected to the diabetic retinopathy identification module, outputs the result of diabetic retinopathy identification to the user in a visual manner, and optimizes and improves the system according to the user's feedback; Introducing the formula for diabetic retinopathy identification accuracy: Where TP means true positive, TN means true negative, FP means false positive, and FN means false negative; Introduce the data update frequency formula: Where f represents the data update frequency, and n represents the number of data updates within time T; Introduce the model performance improvement rate formula: R represents the model performance improvement rate, Accuracynew represents the diabetic retinopathy recognition accuracy of the optimized model, A ccuracyold represents the accuracy of diabetic retinopathy recognition of the model before optimization.

2. According to claim 1, a diabetic retinopathy identification system based on blockchain and federated learning is characterized in that: The data acquisition module comprises: Image acquisition unit, equipped with a fundus camera or equipment for taking fundus images, and image acquisition software, to ensure that the acquired fundus images have clear resolution and accurate color reproduction, and the image resolution is not less than 512×512 pixels; A data preprocessing unit preprocesses the collected fundus images and related medical data, including image denoising, image enhancement, and data standardization operations, to improve the quality and availability of data. Image denoising uses mean filtering, median filtering, or Gaussian filtering algorithms, image enhancement uses contrast enhancement and brightness adjustment methods, and data standardization normalizes the data to a specific interval; The data annotation unit is where medical personnel annotate fundus images. The annotation content includes whether the patient has diabetic retinopathy, the stage of diabetic retinopathy, and information on the location of the lesion. The accuracy of the annotation is no less than 95%.

3. According to claim 1, a diabetic retinopathy identification system based on blockchain and federated learning is characterized in that: The federated learning module includes: The participant node management unit is responsible for managing and coordinating the joining, exiting and communication of each participant node, ensuring stable connection and data transmission between the participant nodes, and using encrypted communication protocols to ensure communication security; Local model training unit: each participant node uses local data to conduct preliminary training of the diabetic retinopathy recognition model. The training algorithms include but are not limited to convolutional neural network (CNN) and recurrent neural network (RNN) deep learning algorithms. Appropriate algorithms and model structures are selected based on the characteristics of local data. Stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimization algorithms are used to optimize model parameters during training. The model parameter aggregation unit aggregates the trained model parameters of each participant's node through the federated learning algorithm to obtain the updated parameters of the global model. The aggregation algorithm adopts the federated averaging algorithm. In the parameter aggregation process, the data volume and data quality factors of each participant's node are considered for weighted processing; Introduce the parameter aggregation formula of the federated learning model: where θ global represents the global model parameters, n is the number of participating nodes, and w i is the weight of the i-th participant node, θ i is the local model parameter after training of the i-th participant node, weight w i According to the data volume of the participating nodes d i and data quality assessment indicator q i Sure, The amount of data d i The data quality evaluation index q is measured by the number of data samples. i Comprehensively consider the accuracy, completeness and consistency of data for quantitative evaluation; The model update and distribution unit updates the global diabetic retinopathy recognition model according to the aggregated parameters, and distributes the updated model to each participating node, so that each participating node can continue the next round of local training and parameter upload to achieve continuous optimization of the model.

4. According to claim 1, a diabetic retinopathy identification system based on blockchain and federated learning is characterized in that: The blockchain module includes: The blockchain network construction unit uses the alliance chain architecture to build the blockchain network. The nodes in the alliance chain are composed of participating medical institutions and scientific research institutions. The nodes reach a consensus through the practical Byzantine fault tolerance algorithm PBFT. The data storage unit stores the model parameters, training data summary, and key information of data access records in the process of federated learning in the form of blocks on the blockchain. Each block contains the hash value of the previous block to form a chain structure to ensure that the data cannot be tampered with and is traceable. The data storage adopts a distributed storage method to improve the storage security and reliability of the data. Smart contract unit, write and deploy smart contracts to regulate data sharing, model updates, node incentive rules and operation procedures in the federated learning process. Smart contracts are automatically executed without third-party intervention to ensure that all parties collaborate as agreed. When the participating nodes successfully complete a round of federated learning tasks, the smart contract automatically gives corresponding rewards according to the preset incentive mechanism. The rewards can be digital currency, points or other forms of rights and interests. The data encryption and privacy protection unit uses encryption technology to encrypt data stored on the blockchain to ensure the security and privacy of the data. Before uploading the data to the blockchain, the asymmetric encryption algorithm RSA is used to encrypt the data, and authorized nodes are set to decrypt and access the data. At the same time, zero-knowledge proof privacy protection technology is used to perform data verification and calculation without leaking the original data.

5. According to claim 1, a diabetic retinopathy identification system based on blockchain and federated learning is characterized in that: The diabetic retinopathy identification module comprises: The model loading unit obtains the trained global diabetic retinopathy recognition model from the federated learning module and loads it into the memory to prepare for diabetic retinopathy recognition; The image analysis unit extracts and analyzes the input fundus images, and uses the convolutional neural network algorithm in deep learning to automatically extract the features of the fundus images, including vascular morphology and lesion features. In the feature extraction process, the pooling layer and convolution layer network structure are used to process the image, reduce the data dimension, and extract key features. The classification and diagnosis unit uses the trained model to classify and diagnose fundus images based on the extracted image features to determine whether the patient has diabetic retinopathy and the severity of the diabetic retinopathy. The classification results are divided into normal, mild diabetic retinopathy, moderate diabetic retinopathy, and severe diabetic retinopathy. The threshold judgment and probability estimation methods are used to determine the final diagnosis results during the diagnosis process. The result evaluation unit evaluates and verifies the results of diabetic retinopathy identification, uses accuracy, recall rate, and F1 value evaluation indicators to quantitatively evaluate the identification results, and adjusts and optimizes model parameters by comparing with the diagnosis results of professional doctors to improve the accuracy and reliability of diabetic retinopathy identification.

6. According to claim 1, a diabetic retinopathy identification system based on blockchain and federated learning is characterized in that: The result output and feedback module includes: The result visualization unit outputs the results of diabetic retinopathy identification to users in the form of intuitive charts and reports, including the annotation results of fundus images, the diagnosis conclusion of diabetic retinopathy, and the severity assessment information. The charts use bar charts and line graphs to show the distribution of different types of diabetic retinopathy. The report contains detailed diagnostic instructions and suggestions. The user feedback collection unit collects user feedback on the diabetic retinopathy identification results through the user interface, including evaluation of the accuracy of the diagnosis results, suggestions and requirements for the system. The feedback can be in the form of questionnaires and online comments to ensure that users can easily provide feedback information; The system optimization unit optimizes and improves the system based on user feedback, including adjusting model parameters, improving data collection methods, optimizing the user interface, and improving system performance and user experience. If users report that the diagnosis results are inaccurate, the system will recheck the data collection and labeling process, or further optimize the model training algorithm to improve the accuracy of the diagnosis.

7. A diabetic retinopathy identification system based on blockchain and federated learning according to any one of claims 1 to 6, characterized in that: The system meets the following performance and safety requirements: The accuracy rate of diabetic retinopathy identification is no less than 90%, and it can maintain stable identification performance on diabetic retinopathy samples of different types and severity; The system response time does not exceed 5 seconds, ensuring that users can obtain diabetic retinopathy identification results in a timely manner; Data security complies with relevant medical data security standards and regulatory requirements, and uses multiple encryption technologies and access control mechanisms to ensure data confidentiality, integrity and availability; The system supports the access of new data sources and algorithms, and the evolving diabetic retinopathy identification technologies and medical needs.

8. A diabetic retinopathy identification system based on blockchain and federated learning according to any one of claims 1 to 6, characterized in that: The system is applied to the following scenarios and has corresponding expansion functions: Clinical diagnosis assistance: Provides auxiliary reference for ophthalmologists in diabetic retinopathy diagnosis, improves the accuracy and efficiency of diagnosis, and reduces the rate of misdiagnosis and missed diagnosis. Doctors refer to the system's diagnostic results during clinical diagnosis and make comprehensive diagnoses based on their own professional judgment. The system also provides case management functions to facilitate doctors to track and analyze patients' medical records and diagnostic results. Telemedicine: Remote diabetic retinopathy diagnosis is achieved through the Internet. Patients take fundus images at primary medical institutions, upload them to the system, and professional doctors conduct remote diagnosis, thereby improving the utilization efficiency of medical resources. The system has remote data transmission and communication functions to ensure the safe and fast transmission of images and data. It also supports remote consultation and discussion functions to facilitate communication and collaboration between doctors. Disease screening: used for large-scale diabetic retinopathy screening, rapid detection of diabetic retinopathy patients in the population, early detection and early treatment. The system cooperates with community health service centers and physical examination institutions to carry out diabetic retinopathy screening projects. Through automated diabetic retinopathy identification and result analysis, the screening efficiency is improved and the screening cost is reduced. At the same time, the screening data is counted and analyzed to provide decision-making support for disease prevention and control. Scientific research cooperation: Provide data and technical support for medical research related to diabetic retinopathy, promote the development and innovation of diabetic retinopathy identification technology, and all parties involved can share data and models through blockchain and federated learning technology to carry out joint research projects. The system provides a platform for data sharing and collaborative research, supports scientific researchers in data analysis, model training and algorithm optimization, and protects the intellectual property rights and data rights of all parties.

9. A diabetic retinopathy identification system based on blockchain and federated learning according to any one of claims 1 to 6, characterized in that: The system has data update and model optimization functions: Data update mechanism: Update the fundus image data and related medical data in the data collection module once a month to ensure the timeliness and diversity of the data. Data updates can be obtained through cooperation with medical institutions and scientific research institutions to obtain new case data, and collect the latest information of patients through user feedback and follow-up. During the update process, strict quality control and annotation of new data are carried out to ensure the accuracy and availability of data; Model online optimization: Based on new data and user feedback, the federated learning module is used to optimize and update the diabetic retinopathy identification model online to improve the performance and accuracy of the model. During the online optimization process, the system automatically triggers the model update process. Each participating node performs local training based on the new data, and then uploads the updated model parameters to the blockchain for aggregation and update. Model evaluation and verification: The optimized model is evaluated and verified once a quarter, and the performance of the model is tested using an independent test data set to ensure that the accuracy and reliability of the model do not decrease. Evaluation indicators include accuracy, recall rate, and F1 value.

10. A diabetic retinopathy identification system based on blockchain and federated learning according to any one of claims 1 to 6, characterized in that: The system has user management and authority control functions: User registration and login: Users include medical staff, researchers, and patients, who need to register and log in to use the system. Registration information includes user name, contact information, and identity type. The system verifies and reviews user information to ensure the authenticity and legitimacy of the user's identity. User rights management: Different rights are assigned according to the user's identity and role. Medical staff can perform data collection, model training and result viewing operations, scientific researchers can perform data analysis and model optimization operations, and patients can view their own diagnosis results and related information. Rights management is implemented through access control lists (ACLs) or other rights management mechanisms to ensure that users can only operate within their own rights. Data access control: Access control is performed on the data in the system. Different users have different access rights to data. Among them, the patient's personal privacy data is viewed and processed by authorized medical institution staff. Researchers can access some data under the premise of meeting the research purpose and ethical requirements. Data access control is implemented through encryption technology and access authorization mechanism to ensure data security and privacy; User behavior audit: Audit and record the user's operating behavior in the system, including login time, operation content, and data access records, for easy traceability and supervision. Audit records can be used to discover potential security issues and violations, and provide a reference for system optimization and improvement.

Images