A medical image archive data management system

Abstract

The application discloses a medical image archive data management system, belongs to the zero-knowledge proof technical field, and comprises a verifiable zero-knowledge image fingerprint generation module, a multi-modal federal feature fusion training module, a dynamic contribution intelligent quantification incentive module, a dynamic privacy calculation sandbox module, an intelligent diagnosis routing network module and an archive evaluation optimization module. The application generates verifiable zero-knowledge image fingerprints through a blockchain and a zero-knowledge proof technology. Global features and local features of original medical images are mapped into polynomials and encrypted through a hash function to generate fingerprints containing only feature parameters, so that the image content is not directly exposed. In the verification stage, a bilinear pairing function and an elliptic curve base point are calculated to ensure the legality of the fingerprints and not to leak the original data. The blockchain notarization further guarantees the unforgeability of the fingerprints.

Description

Technical Field

[0001] This invention belongs to the field of zero-knowledge proof technology, specifically referring to a medical image archive data management system. Background Technology

[0002] With the rapid development of medical informatization, the amount of medical image data has exploded. Traditional medical image archiving and communication systems can no longer meet the needs of modern medicine for data security, privacy protection, multi-institutional collaboration and intelligent applications.

[0003] However, existing medical image archive data management still has certain shortcomings. Existing medical images are usually stored in plaintext or weakly encrypted form, lacking deep protection for sensitive information and vulnerable to attacks such as SQL injection and unauthorized access. The integrity and authenticity of image data are difficult to trace, and it is impossible to effectively prevent tampering or forgery, which leads to doubts about the credibility of medical records. There is a risk of privacy leakage when sharing and jointly analyzing data, especially in cross-institutional collaboration scenarios. The lack of dynamic adjustment of computing resources or data access permissions results in high-value nodes having limited resources and inefficient nodes consuming too many resources. To address this, a medical image archive data management system is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a medical image archive data management system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a medical image archive data management system, comprising a verifiable zero-knowledge image fingerprint generation module, a multimodal federated feature fusion training module, a dynamic contribution intelligent quantification incentive module, a dynamic privacy computing sandbox module, an intelligent diagnostic routing network module, and an archive evaluation and optimization module;

[0006] The verifiable zero-knowledge image fingerprint generation module uses blockchain and zero-knowledge proof technology to generate verifiable feature fingerprints from original medical images.

[0007] The multimodal federated feature fusion training module performs distributed model training of multimodal images locally based on the generated feature fingerprints.

[0008] The dynamic contribution intelligent quantification incentive module updates the contribution based on the locally trained distributed model, quantifies the collaborative value of each node through smart contracts, and constructs an incentive closed loop.

[0009] The dynamic privacy computation sandbox module, based on the contribution value quantification result, enables cross-institutional encrypted data interaction through secure multi-party computation and homomorphic encryption technology;

[0010] The intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability map based on a distributed model, and intelligently matches the optimal medical institution node according to diagnostic needs to complete precise task routing.

[0011] The archive evaluation and optimization module is used to monitor the module's operating data in real time and dynamically optimize storage strategies and resource allocation through adaptive algorithms.

[0012] Preferably, the verifiable zero-knowledge image fingerprint generation module is wirelessly connected to the multimodal federated feature fusion training module; the multimodal federated feature fusion training module is wirelessly connected to the dynamic contribution intelligent quantification incentive module and the intelligent diagnostic routing network module; the dynamic contribution intelligent quantification incentive module is wirelessly connected to the dynamic privacy computing sandbox module; the dynamic privacy computing sandbox module is wirelessly connected to the intelligent diagnostic routing network module; the intelligent diagnostic routing network module is wirelessly connected to the archive evaluation and optimization module; and the archive evaluation and optimization module is wirelessly connected to the verifiable zero-knowledge image fingerprint generation module and the multimodal federated feature fusion training module.

[0013] Preferably, the verifiable zero-knowledge image fingerprint generation module includes the following feature fingerprint generation steps: acquiring original medical images from a medical institution, standardizing the images, extracting global and local features of the images, and mapping the local features of the images to a polynomial P. i (x)=a i x 2 +b i x+d i a i ,b i ,d i The characteristic parameters are represented by the encrypted polynomial combination input into the hash function H;

[0014] Specifically, the feature vector is encrypted using a zero-knowledge proof algorithm to generate a verifiable feature fingerprint, as shown in the following formula:

[0015]

[0016] In the formula, F represents the zero-knowledge image feature fingerprint, H represents the collision-resistant hash function, n represents the number of local features, and c i P represents the randomly generated encryption coefficient. i (x) represents the local feature polynomial of the medical image, where x represents the variable of the polynomial.

[0017] Preferably, the verifiable zero-knowledge image fingerprint generation module includes a zero-knowledge image feature fingerprint verification step that maps the encrypted zero-knowledge image feature fingerprint F to a point F = H(·)·G on an elliptic curve.

[0018] Specifically, zero-knowledge image feature fingerprints are verified using a bilinear pairing function, implemented as follows:

[0019]

[0020] In the formula, e(·,·) represents the bilinear pairing function, G represents the generator on the elliptic curve, and G i Represents the encryption coefficient c i The corresponding elliptic curve base point, G j This represents the challenge parameter d generated by the validator. j The corresponding base point, d j This represents the random challenge coefficient generated by the verifier, and m represents the number of verification challenge parameters;

[0021] It should be understood that the verifier determines the encryption coefficient c. i The random challenge coefficient d generated by the validator j ,calculate and And verify whether its pairing result is equal to the left side.

[0022] Preferably, the multimodal federated feature fusion training module includes the following local multimodal feature extraction and fusion steps: defining a globally shared model on the federated learning server, which is responsible for global model aggregation and parameter updates but does not access the original data; using the zero-knowledge image feature fingerprint F as the identity identifier for each client, which is bound to the local model during training; and each client holding a copy of the local model, which is consistent with the global model structure.

[0023] Specifically, each client extracts local features from its local multimodal data and maps them to a multimodal feature vector of a unified dimension. The zero-knowledge image feature fingerprint is then concatenated with the multimodal feature vector to generate a joint feature vector X. fused X fused =Concat(X) n ,F), Concat represents the vector concatenation operation, X n This represents a multimodal feature vector.

[0024] Preferably, the multimodal federated feature fusion training module includes the following distributed model training steps: training a local model using the global model input joint feature vector for each client, minimizing the task loss, as shown in the formula:

[0025]

[0026] In the formula, θ represents the local model parameters, and L tk This represents the task-related loss function. This represents the encrypted concatenated eigenvector. Indicates encrypted binding operation, L py λ represents the privacy protection constraint term, and λ represents the weight coefficient of the privacy protection constraint term.

[0027] Preferably, the multimodal federated feature fusion training module, the global model aggregation step includes: the client encrypting the local model parameters and uploading them to the federated learning server;

[0028] By aggregating global model parameters through federated averaging, the following is achieved:

[0029]

[0030] In the formula, θ gl The parameters θ represent the parameters of the global model. The federated learning server takes the current parameters θ from all clients. i Weighted averages are used to form a global model, where N represents the total number of clients participating in federated learning, and w i Represents the client weight, θ i The local model parameters of client i;

[0031] Specifically, all raw data is stored locally on the client, while the server only stores the encrypted fingerprint F and the aggregated global model parameters. After the global model is updated, the federated learning server sends the new parameters to the client, repeating the training and aggregation process until convergence.

[0032] Preferably, the dynamic contribution intelligent quantification incentive module includes contribution indicators such as model update quality, data value, and training efficiency. Model update quality refers to the improvement of global model performance based on local model parameter updates. Data value quantifies the data contribution of nodes through data diversity, data volume, or task relevance. Training efficiency measures the computational resource input of nodes or the model convergence speed.

[0033] Specifically, through multi-dimensional weighted fusion of indicators, C i C i C i =α·Q i +β·D i +γ·E i C i Q represents the contribution of node i. i D represents the quality of model updates. i E represents the value of data. i α, β, and γ represent training efficiency, and α, β, and γ represent dynamic weights.

[0034] Differential privacy is used to obfuscate the contribution metric, preventing nodes from inferring sensitive data from the contribution score. This is based on the contribution score C. i The formula for allocating rewards is as follows:

[0035]

[0036] In the formula, R i R represents the reward for node i. total Indicates the total reward pool;

[0037] Incentive closed-loop construction: Nodes construct the reward R allocated by the smart contract. i By generating revenue and creating positive incentives, nodes with low contribution are penalized to achieve reverse constraints. In subsequent training rounds, the smart contract dynamically adjusts the weights of nodes based on their historical performance, and optimizes the federated learning aggregation through contribution evaluation results.

[0038] Preferably, the dynamic privacy computation sandbox module, based on the contribution value quantification result, achieves cross-institutional encrypted data interaction through secure multi-party computation and homomorphic encryption technology; each institution encrypts its local data using homomorphic encryption and dynamically adjusts the encryption parameters according to the contribution value quantification result.

[0039] High-contribution nodes employ lightweight encryption to reduce computational overhead;

[0040] Nodes with low contribution values ​​are subjected to strong encryption, and their data access permissions are restricted.

[0041] By distributing encryption keys through a distributed key generation protocol, it is ensured that each organization holds only a local share of the key, thus avoiding the risk of single-point leakage.

[0042] Secure Multi-Party Computation: Select a suitable MPC protocol based on the type of cross-organizational collaboration task. The protocol must support dynamic participant adjustment based on contribution value C. i Nodes can be dynamically added or removed, with each organization using encrypted data as input;

[0043] It should be understood that the encrypted shared value generated through the MPC protocol must meet the contribution value constraint: the shared value of nodes with high contribution values ​​has a higher weight. Joint computation is performed on the encrypted shared value, and the output is still in encrypted form E(R). During the computation process, the legality of the operations of each party is verified through zero-knowledge proof.

[0044] Homomorphic encryption: For model parameters that need to be updated frequently, partial homomorphic encryption is used to avoid the high communication overhead of MPC; for complex calculations, a hybrid mode of homomorphic encryption and MPC is combined: the linear part is accelerated by homomorphic encryption, and the nonlinear part is completed by MPC.

[0045] Based on contribution value C i Decryption permissions are determined dynamically:

[0046] Nodes with high contribution values ​​can directly participate in decryption;

[0047] Nodes with low contribution values ​​need to obtain results indirectly through a trusted institution;

[0048] Decryption is performed in stages: intermediate results are decrypted first, and then the final data is released gradually.

[0049] The decrypted result needs to be added with differential privacy noise to prevent the original data from being reversed, and all operation records are stored in the blockchain;

[0050] Dynamically adjust sandbox resource allocation based on historical contribution values:

[0051] Nodes with high contribution values ​​should be allocated more computing resources;

[0052] Low contribution value nodes limit the calculation frequency;

[0053] When a new organization joins, the key and shared value are renegotiated through the MPC protocol. Data from departing organizations is securely deleted after verification with zero-knowledge proofs. The sandbox computation results are fed back to the global model training through federated learning, and the encryption strategy and MPC protocol parameters of the next round of sandboxes are adjusted based on the quantification results of the contribution values.

[0054] Preferably, the intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability map based on a distributed model, and intelligently matches the optimal medical institution node according to diagnostic needs to complete accurate task routing;

[0055] Specifically, the steps for constructing a cross-institutional diagnostic capability map include: collecting the core diagnostic capabilities of various medical institutions, including: hardware equipment, expert resources, data quality and response capabilities; and constructing a graph database to establish the triplet relationship between institutions, capabilities and tasks.

[0056] By analyzing public reports, academic papers, and certifications of medical institutions through natural language processing, structured capability descriptions are extracted. Each institution uploads encrypted diagnostic results, and contribution values ​​are generated through federated learning. The confidence level of each institution under the task type is then output.

[0057] The institution's capability score is dynamically adjusted based on new case data, and changes in capability are recorded and documented via blockchain.

[0058] Each institution acts as a routing node, negotiating task allocation through a secure multi-party computation protocol. The requesting party broadcasts an encrypted task description, each institution calculates the matching degree locally and submits the results, and the coordination server aggregates the results to select the optimal institution.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. This invention generates verifiable zero-knowledge image fingerprints using blockchain and zero-knowledge proof technology. It maps the global and local features of the original medical image into a polynomial and encrypts them using a hash function to generate a fingerprint containing only feature parameters, thus avoiding direct exposure of the image content. During the verification stage, the fingerprint's legality is ensured without leaking the original data by using a bilinear pairing function and elliptic curve base point calculation. Blockchain storage further guarantees the fingerprint's immutability.

[0061] 2. This invention utilizes zero-knowledge image fingerprints and a multimodal feature fusion training process within a federated learning framework. By aggregating local model parameters from distributed nodes through a federated learning server, it avoids centralizing the original dataset, reducing the risk of privacy leaks. The local model generates joint features by concatenating zero-knowledge fingerprints and multimodal feature vectors, preserving the complete description of the image while enhancing the model's ability to perceive multimodal correlations. The global parameter update mechanism of federated learning, combined with privacy protection constraints, ensures that model performance is improved while meeting compliance requirements. By dynamically binding fingerprints to model parameters, it enhances the fairness and accuracy of cross-institutional collaboration.

[0062] 3. This invention uses indicators such as model update quality, data value, and training efficiency, combined with differential privacy fuzzification processing, to dynamically evaluate the contribution of each node to the global model, avoid the leakage of sensitive information, and automatically allocate rewards and implement penalties through smart contracts, forming a closed loop of positive incentives and negative constraints. The dynamic weight adjustment mechanism, combined with historical performance, optimizes the federated learning aggregation strategy, thereby improving the overall training efficiency and model convergence speed.

[0063] 4. This invention constructs an intelligent diagnostic routing network based on a distributed model. It aggregates diagnostic capability data from various institutions through federated learning, generates a cross-institutional capability map, and extracts structured capability descriptions using natural language processing. After parsing diagnostic needs, it matches the optimal institution through a graph neural network. Combined with a secure multi-party computation protocol, it dynamically negotiates task allocation. Blockchain notarization and zero-knowledge proof verification mechanisms ensure the transparency and legitimacy of the routing process, avoid node interference, and improve the accurate matching of medical resources and the rapid response to emergency tasks. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the structure of a medical image archive data management system according to the present invention;

[0065] Figure 2 The present invention provides the operation flow of a medical image archive data management system. Figure 1 ;

[0066] Figure 3 The present invention provides the operation flow of a medical image archive data management system. Figure 2 ;

[0067] Figure 4 The present invention provides the operation flow of a medical image archive data management system. Figure 3 . Detailed Implementation

[0068] 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.

[0069] Example

[0070] Please see Figures 1-4 As shown, the present invention provides a technical solution including a verifiable zero-knowledge image fingerprint generation module, a multimodal federated feature fusion training module, a dynamic contribution intelligent quantification incentive module, a dynamic privacy computing sandbox module, an intelligent diagnostic routing network module, and an archive evaluation and optimization module;

[0071] The verifiable zero-knowledge image fingerprint generation module uses blockchain and zero-knowledge proof technology to generate verifiable feature fingerprints from original medical images.

[0072] The multimodal federated feature fusion training module performs distributed model training of multimodal images locally based on the generated feature fingerprints.

[0073] The dynamic contribution intelligent quantification incentive module updates the contribution based on the locally trained distributed model, quantifies the collaborative value of each node through smart contracts, and constructs an incentive closed loop.

[0074] The dynamic privacy computation sandbox module, based on the contribution value quantification result, enables cross-institutional encrypted data interaction through secure multi-party computation and homomorphic encryption technology;

[0075] The intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability map based on a distributed model, and intelligently matches the optimal medical institution node according to diagnostic needs to complete precise task routing.

[0076] The archive evaluation and optimization module is used to monitor the module's operating data in real time and dynamically optimize storage strategies and resource allocation through adaptive algorithms.

[0077] In this embodiment, the verifiable zero-knowledge image fingerprint generation module is wirelessly connected to the multimodal federated feature fusion training module; the multimodal federated feature fusion training module is wirelessly connected to the dynamic contribution intelligent quantification incentive module and the intelligent diagnostic routing network module; the dynamic contribution intelligent quantification incentive module is wirelessly connected to the dynamic privacy computing sandbox module; the dynamic privacy computing sandbox module is wirelessly connected to the intelligent diagnostic routing network module; the intelligent diagnostic routing network module is wirelessly connected to the archive evaluation and optimization module; and the archive evaluation and optimization module is wirelessly connected to the verifiable zero-knowledge image fingerprint generation module and the multimodal federated feature fusion training module.

[0078] Preferably, the verifiable zero-knowledge image fingerprint generation module includes the following feature fingerprint generation steps: acquiring original medical images from a medical institution, standardizing the images, extracting global and local features of the images, and mapping the local features of the images to a polynomial P. i (x)=a i x 2 +b i x+d i a i ,b i ,d i The characteristic parameters are represented by the encrypted polynomial combination input into the hash function H;

[0079] Specifically, the feature vector is encrypted using a zero-knowledge proof algorithm to generate a verifiable feature fingerprint, as shown in the following formula:

[0080]

[0081] In the formula, F represents the zero-knowledge image feature fingerprint, H represents the collision-resistant hash function, n represents the number of local features, and c i P represents the randomly generated encryption coefficient. i (x) represents the local feature polynomial of the medical image, where x represents the variable of the polynomial.

[0082] Preferably, the verifiable zero-knowledge image fingerprint generation module includes a zero-knowledge image feature fingerprint verification step that maps the encrypted zero-knowledge image feature fingerprint F to a point F = H(·)·G on an elliptic curve.

[0083] Specifically, zero-knowledge image feature fingerprints are verified using a bilinear pairing function, implemented as follows:

[0084]

[0085] In the formula, e(·,·) represents the bilinear pairing function, G represents the generator on the elliptic curve, and G i Represents the encryption coefficient ci The corresponding elliptic curve base point, G j This represents the challenge parameter d generated by the validator. j The corresponding base point, d j This represents the random challenge coefficient generated by the verifier, and m represents the number of verification challenge parameters;

[0086] It should be understood that the verifier determines the encryption coefficient c. i The random challenge coefficient d generated by the validator j ,calculate and And verify whether its pairing result is equal to the left side.

[0087] Preferably, the multimodal federated feature fusion training module includes the following local multimodal feature extraction and fusion steps: defining a globally shared model on the federated learning server, which is responsible for global model aggregation and parameter updates but does not access the original data; using the zero-knowledge image feature fingerprint F as the identity identifier for each client, which is bound to the local model during training; and each client holding a copy of the local model, which is consistent with the global model structure.

[0088] Specifically, each client extracts local features from its local multimodal data and maps them to a multimodal feature vector of a unified dimension. The zero-knowledge image feature fingerprint is then concatenated with the multimodal feature vector to generate a joint feature vector X. fused X fused =Concat(X) n ,F), Concat represents the vector concatenation operation, X n This represents a multimodal feature vector.

[0089] Preferably, the multimodal federated feature fusion training module includes the following distributed model training steps: training a local model using the global model input joint feature vector for each client, minimizing the task loss, as shown in the formula:

[0090]

[0091] In the formula, θ represents the local model parameters, and L tk This represents the task-related loss function. This represents the encrypted concatenated eigenvector. Indicates encrypted binding operation, L py λ represents the privacy protection constraint term, and λ represents the weight coefficient of the privacy protection constraint term.

[0092] Preferably, the multimodal federated feature fusion training module, the global model aggregation step includes: the client encrypting the local model parameters and uploading them to the federated learning server;

[0093] By aggregating global model parameters through federated averaging, the following is achieved:

[0094]

[0095] In the formula, θ gl The parameters θ represent the parameters of the global model. The federated learning server takes the current parameters θ from all clients. i Weighted averages are used to form a global model, where N represents the total number of clients participating in federated learning, and w i Represents the client weight, θ i The local model parameters of client i;

[0096] In this embodiment, all original data is kept locally on the client. The server only stores the encrypted fingerprint F and the aggregated global model parameters. After the global model is updated, the federated learning server sends the new parameters to the client and repeats the training and aggregation process until convergence.

[0097] Preferably, the dynamic contribution intelligent quantification incentive module includes contribution indicators such as model update quality, data value, and training efficiency. Model update quality refers to the improvement of global model performance based on local model parameter updates. Data value quantifies the data contribution of nodes through data diversity, data volume, or task relevance. Training efficiency measures the computational resource input of nodes or the model convergence speed.

[0098] Specifically, through multi-dimensional weighted fusion of indicators, C i C i C i =α·Q i +β·D i +γ·E i C i Q represents the contribution of node i. i D represents the quality of model updates. i E represents the value of data. i α, β, and γ represent training efficiency, and α, β, and γ represent dynamic weights.

[0099] Differential privacy is used to obfuscate the contribution metric, preventing nodes from inferring sensitive data from the contribution score. This is based on the contribution score C. i The formula for allocating rewards is as follows:

[0100]

[0101] In the formula, R i R represents the reward for node i. total Indicates the total reward pool;

[0102] In this embodiment, the incentive closed loop is constructed as follows: nodes are rewarded R by the smart contract. iBy generating revenue and creating positive incentives, nodes with low contribution are penalized to achieve reverse constraints. In subsequent training rounds, the smart contract dynamically adjusts the weights of nodes based on their historical performance, and optimizes the federated learning aggregation through contribution evaluation results.

[0103] Preferably, the dynamic privacy computing sandbox module, based on the contribution value quantification result, achieves cross-institutional encrypted data interaction through secure multi-party computation and homomorphic encryption technology;

[0104] In this embodiment, each institution encrypts its local data using homomorphic encryption and dynamically adjusts the encryption parameters based on the contribution value quantification results:

[0105] High-contribution nodes employ lightweight encryption to reduce computational overhead;

[0106] Nodes with low contribution values ​​are subjected to strong encryption, and their data access permissions are restricted.

[0107] By distributing encryption keys through a distributed key generation protocol, it is ensured that each organization holds only a local share of the key, thus avoiding the risk of single-point leakage.

[0108] Secure Multi-Party Computation: Select a suitable MPC protocol based on the type of cross-organizational collaboration task. The protocol must support dynamic participant adjustment based on contribution value C. i Nodes can be dynamically added or removed, with each organization using encrypted data as input;

[0109] It should be understood that the encrypted shared value generated through the MPC protocol must meet the contribution value constraint: the shared value of nodes with high contribution values ​​has a higher weight. Joint computation is performed on the encrypted shared value, and the output is still in encrypted form E(R). During the computation process, the legality of the operations of each party is verified through zero-knowledge proof.

[0110] Homomorphic encryption: For model parameters that need to be updated frequently, partial homomorphic encryption is used to avoid the high communication overhead of MPC; for complex calculations, a hybrid mode of homomorphic encryption and MPC is combined: the linear part is accelerated by homomorphic encryption, and the nonlinear part is completed by MPC.

[0111] Based on contribution value C i Decryption permissions are determined dynamically:

[0112] Nodes with high contribution values ​​can directly participate in decryption;

[0113] Nodes with low contribution values ​​need to obtain results indirectly through a trusted institution;

[0114] Decryption is performed in stages: intermediate results are decrypted first, and then the final data is released gradually.

[0115] The decrypted result needs to be added with differential privacy noise to prevent the original data from being reversed, and all operation records are stored in the blockchain;

[0116] Dynamically adjust sandbox resource allocation based on historical contribution values:

[0117] Nodes with high contribution values ​​should be allocated more computing resources;

[0118] Low contribution value nodes limit the calculation frequency;

[0119] In this embodiment, when a new organization joins, the key and shared value are renegotiated through the MPC protocol, and the data of the departing organization is securely deleted after verification by zero-knowledge proof; the sandbox calculation results are fed back to the global model training through federated learning, and the encryption strategy and MPC protocol parameters of the next round of sandbox are adjusted according to the contribution value quantification results.

[0120] In this embodiment, the intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability map based on a distributed model, and intelligently matches the optimal medical institution node according to diagnostic needs to complete precise task routing.

[0121] Specifically, the steps for constructing a cross-institutional diagnostic capability map include: collecting the core diagnostic capabilities of various medical institutions, including: hardware equipment, expert resources, data quality and response capabilities; and constructing a graph database to establish the triplet relationship between institutions, capabilities and tasks.

[0122] By analyzing public reports, academic papers, and certifications of medical institutions through natural language processing, structured capability descriptions are extracted. Each institution uploads encrypted diagnostic results, and contribution values ​​are generated through federated learning. The confidence level of each institution under the task type is then output.

[0123] The institution's capability score is dynamically adjusted based on new case data, and changes in capability are recorded and documented via blockchain.

[0124] Each institution acts as a routing node, negotiating task allocation through a secure multi-party computation protocol. The requesting party broadcasts an encrypted task description, each institution calculates the matching degree locally and submits the results, and the coordination server aggregates the results to select the optimal institution.

[0125] Working Principle: Using blockchain and zero-knowledge proof technology, raw medical images are transformed into verifiable encrypted feature fingerprints. The images are standardized, global and local features are extracted, and local features are mapped to a polynomial form. The polynomial is then encrypted with randomly generated encryption coefficients and input into a collision-resistant hash function to generate the fingerprint. During verification, the fingerprint is mapped to points on an elliptic curve and mathematically verified using a bilinear pairing function and challenge parameters generated by the verifier. The fingerprint is stored on the blockchain, achieving full lifecycle traceability and anti-counterfeiting capabilities. The multimodal federated feature fusion training module, based on the generated zero-knowledge image fingerprints, performs distributed model training on multimodal images locally. The federated learning server defines the entire... The shared model uses fingerprints as identifiers to bind local model copies. Clients extract local features from multimodal data, map them into feature vectors of a unified dimension, and concatenate them with fingerprints to generate joint feature vectors. During training, clients optimize the task loss function on encrypted features and introduce constraints such as differential privacy to protect data privacy. The federated server aggregates the encrypted local model parameters and updates the global model through weighted averaging, ensuring that data does not leave the local machine. This enables cross-agency collaboration and model performance improvement. The dynamic contribution intelligent quantification incentive module quantifies the contribution of nodes in federated learning, constructs an incentive loop to improve collaboration efficiency, and the contribution indicators include model update quality, data value, and training efficiency.

[0126] By dynamically weighting and fusing multi-dimensional indicators, and combining differential privacy fuzzing, sensitive information leakage is prevented. Smart contracts allocate rewards based on contribution, with high-contribution nodes receiving positive incentives and low-contribution nodes being penalized. Historical performance dynamically adjusts weights, optimizing the federated learning aggregation strategy to achieve sustainable and fair collaboration. Based on contribution value quantification, secure multi-party computation and homomorphic encryption technology enable cross-institutional encrypted data interaction. High-contribution nodes use lightweight encryption to reduce overhead, while low-contribution nodes use strong encryption to restrict access. A distributed key generation protocol ensures that each institution holds only a local key share, avoiding single-point leakage risks. The MPC protocol supports dynamic participant adjustment, and the joint computation output remains encrypted, with zero-knowledge proofs verifying the legitimacy of operations. Homomorphic encryption accelerates linear computation, and a hybrid mode handles complex tasks. Decryption permissions are dynamically allocated based on contribution value, and the results are stored on the blockchain after adding differential privacy noise. The intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability graph through a distributed model, achieving precise task routing. It first collects data from medical institutions... The system leverages the core capabilities of hardware, expert resources, and data quality. It utilizes natural language processing to parse publicly available reports and certifications, constructing structured capability descriptions. Federated learning aggregates encrypted diagnostic results, generating confidence scores for institutions under different tasks. Requesters broadcast encrypted task descriptions, each institution calculates its matching degree locally and submits results. A coordination server aggregates these results to select the optimal institution. The routing process incorporates a secure multi-party computation protocol to ensure privacy and efficiency in task allocation, dynamically updating the capability map to adapt to new case requirements. The record evaluation and optimization module monitors system operation data in real time, dynamically optimizing storage strategies and resource allocation through adaptive algorithms. It monitors storage load and resource consumption, tiering high-frequency data for high-speed media and migrating low-frequency data to low-cost storage. Based on contribution value quantification, resources are allocated to high-contribution nodes while limiting resource consumption by low-contribution nodes. Federated learning provides feedback to iteratively improve storage and resource allocation logic. All operation records are stored on the blockchain for auditing, ensuring transparency and compliance, while privacy protection technologies prevent sensitive data leakage.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0128] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A medical image archive data management system, characterized in that: It includes a verifiable zero-knowledge image fingerprint generation module, a multimodal federated feature fusion training module, a dynamic contribution intelligent quantification incentive module, a dynamic privacy computing sandbox module, an intelligent diagnostic routing network module, and an archive evaluation and optimization module; The verifiable zero-knowledge image fingerprint generation module uses blockchain and zero-knowledge proof technology to generate verifiable feature fingerprints from original medical images. The multimodal federated feature fusion training module performs distributed model training of multimodal images locally based on the generated feature fingerprints. The dynamic contribution intelligent quantification incentive module updates the contribution based on the locally trained distributed model, quantifies the collaborative value of each node through smart contracts, and constructs an incentive closed loop. The dynamic privacy computation sandbox module, based on the contribution value quantification result, enables cross-institutional encrypted data interaction through secure multi-party computation and homomorphic encryption technology; The intelligent diagnostic routing network module constructs a cross-institutional diagnostic capability map based on a distributed model, and intelligently matches the optimal medical institution node according to diagnostic needs to complete precise task routing. The archive evaluation and optimization module is used to monitor the module's operating data in real time and dynamically optimize storage strategies and resource allocation through adaptive algorithms. The steps for constructing a cross-institutional diagnostic capability map include: collecting the core diagnostic capabilities of each medical institution, including: hardware equipment, expert resources, data quality and response capabilities; and constructing a graph database to establish the triplet relationship between institutions, capabilities and tasks. By analyzing public reports, academic papers, and certifications of medical institutions through natural language processing, structured capability descriptions are extracted. Each institution uploads encrypted diagnostic results, and contribution values ​​are generated through federated learning. The confidence level of each institution under the task type is then output. The institution's capability score is dynamically adjusted based on new case data, and changes in capability are recorded and documented via blockchain. Each institution acts as a routing node, negotiates task allocation through a secure multi-party computation protocol, the requester broadcasts an encrypted task description, each institution calculates the matching degree locally and submits the results, and the coordination server aggregates the results to select the optimal institution. The dynamic contribution intelligent quantification incentive module uses contribution indicators including model update quality, data value, and training efficiency, which are weighted and fused through multi-dimensional indicators. According to contribution The formula for allocating rewards is as follows: ， In the formula, This represents the reward for node i. Indicates the total reward pool; Incentive loop construction: Nodes construct rewards based on smart contract allocations. By generating benefits and creating positive incentives, and by penalizing nodes with low contribution levels, reverse constraints are implemented, and the federated learning aggregation is optimized based on the contribution evaluation results.

2. The medical image archive data management system according to claim 1, characterized in that: The verifiable zero-knowledge image fingerprint generation module includes the following feature fingerprint generation steps: acquiring original medical images from a medical institution, standardizing the images, extracting global and local features of the images, and mapping the global and local features of the images to a polynomial. The feature vector is encrypted using a zero-knowledge proof algorithm to generate a verifiable feature fingerprint. The formula is as follows: ， In the formula, F represents the zero-knowledge image feature fingerprint, H represents the collision-resistant hash function, and n represents the number of local features. This represents the randomly generated encryption coefficient. Let x represent the global and local feature polynomials of a medical image, where x represents the variables of the polynomial.

3. A medical image archive data management system according to claim 2, characterized in that: The verifiable zero-knowledge image fingerprint generation module includes a zero-knowledge image feature fingerprint verification step: mapping the encrypted zero-knowledge image feature fingerprint F to points on an elliptic curve, and verifying the zero-knowledge image feature fingerprint using a bilinear pairing function, as shown in the formula: ， In the formula, Let G denote the bilinear pairing function, and G denote the generator on the elliptic curve. Representation and encryption coefficient The corresponding elliptic curve base point, This represents the challenge parameters generated by the validator. The corresponding base point, represents the random challenge coefficient generated by the verifier, and m represents the number of verification challenge parameters.

4. A medical image archive data management system according to claim 1, characterized in that: The multimodal federated feature fusion training module includes the following local multimodal feature extraction and fusion steps: defining a globally shared model on the federated learning server, using the zero-knowledge image feature fingerprint F as the identity identifier for each client, with each client holding a local copy of the model consistent with the global model structure, each client extracting local features from the local multimodal data and mapping them to a multimodal feature vector of uniform dimension, and concatenating the zero-knowledge image feature fingerprint with the multimodal feature vector to generate a joint feature vector. .

5. A medical image archive data management system according to claim 4, characterized in that: The multimodal federated feature fusion training module includes the following distributed model training steps: each client trains its local model using the global model input joint feature vector, minimizing the task loss, as shown in the formula: ， In the formula, Indicates local model parameters, This represents the task-related loss function. This represents the encrypted joint feature vector. This indicates an encrypted binding operation. Indicates privacy protection constraints. This represents the weighting coefficient of the privacy protection constraint.

6. A medical image archive data management system according to claim 5, characterized in that: The multimodal federated feature fusion training module aggregating the global model includes the following steps: the client encrypts the local model parameters and uploads them to the federated learning server; the global model parameters are aggregated through federated averaging, as follows: ， In the formula, This represents the parameters of the global model, and N represents the total number of clients participating in federated learning. Indicates client weight. Local model parameters for client i.

Images