AI-driven multi-center medical image collaborative diagnosis and data security sharing method

Abstract

The invention relates to the technical field of wisdom medical treatment, and discloses an AI-driven multi-center medical image collaborative diagnosis and data security sharing method, which comprises the steps of S1, data acquisition and preprocessing, S2, data security encryption and transmission, S3, AI-driven collaborative diagnosis and S4, data security storage and traceability. A data standard and a triggering mechanism are unified: a DICOM 3.0 standard unified acquisition format is adopted in the scheme, real-time monitoring of the equipment state every 10 seconds is realized through a special data acquisition terminal, meanwhile, a local cache (500GB) 80% capacity triggering uploading mechanism is set, manual intervention is avoided, and the data circulation triggering efficiency is improved by more than 90%; in the preprocessing link, through format verification, wavelet transform noise reduction (db4 wavelet basis, three-layer decomposition) and normalization ([0, 1] interval) standardized operation, cross-center data adaptation difference is eliminated, data adaptation time is shortened to 1 / 5 of the original time, and the rapid diagnosis requirement of emergency diseases is met.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, specifically to an AI-driven method for multi-center collaborative diagnosis and secure data sharing of medical images. Background Technology

[0002] In the field of smart healthcare, multi-center collaborative medical imaging diagnosis is a key direction for improving the diagnostic capabilities of complex diseases and optimizing the allocation of medical resources. However, the current multi-center medical imaging diagnostic system still faces three major pain points, which seriously restrict its development and application: (i) The problem of data silos is prominent and the foundation for collaboration is weak. Medical imaging data in various central hospitals are mostly stored in local systems. Due to technical barriers and data ownership divisions in hospital information systems (HIS) and picture archiving and communication systems (PACS), data is difficult to circulate across centers, forming "data silos." For example, although CT and MRI image data from different hospitals mostly follow the DICOM standard, differences in equipment models and data annotation rules require additional format adaptation when accessing data across centers. Adaptation time usually accounts for more than 30% of the total time spent on collaborative diagnosis. At the same time, there is a lack of a unified data sharing trigger mechanism among centers, and most rely on manual application and transmission, resulting in low data flow efficiency and an inability to meet the rapid diagnostic needs of emergency conditions (such as stroke and acute tumors).

[0003] (ii) High data security risks and great difficulty in protecting privacy. Medical imaging data contains patient privacy information (such as identity information and lesion details), and its transmission and sharing must strictly comply with regulations such as the Personal Information Protection Law and the Measures for the Administration of Network Security in Medical and Health Institutions. However, current technologies often employ a single encryption algorithm (such as RSA or AES), which suffers from insufficient encryption strength or low efficiency. When using a single asymmetric encryption (such as RSA-2048), encrypting imaging data (typically tens to hundreds of MB) takes more than 5 minutes, failing to meet real-time transmission requirements. When using a single symmetric encryption (such as AES-128), key management is difficult, and once the key is leaked, all encrypted data is at risk of being leaked. Furthermore, data transmission often relies on ordinary network channels and lacks dedicated encryption transmission protocols, resulting in a high risk of packet loss and tampering during transmission. Statistics show that the data leakage rate for existing unencrypted or weakly encrypted imaging data transmission is approximately 0.8%, far exceeding the 0.1% security threshold in the medical industry. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an AI-driven method for multi-center medical image collaborative diagnosis and secure data sharing, which has advantages such as breaking down data silos and improving data flow efficiency, thus solving the problems of prominent data silos and weak collaborative foundation.

[0005] (II) Technical Solution To achieve the above-mentioned goals of breaking down data silos and improving data flow efficiency, the present invention provides the following technical solution: an AI-driven multi-center medical image collaborative diagnosis and data security sharing method, including S1 data acquisition and preprocessing, S2 data security encryption and transmission, S3 AI-driven collaborative diagnosis and S4 data security storage and traceability, wherein S1 data acquisition and preprocessing includes S101 multi-center image data acquisition, S102 data preprocessing initialization and S103 preprocessing quality detection; Among them, S2 data security encryption and transmission includes S201 layered data encryption, S202 establishment of encrypted data transmission channel and S203 transmission data verification and feedback; Among them, S3AI-driven collaborative diagnosis includes S301 AI model loading and initialization, S302 multi-center AI parallel analysis, S303 multi-center doctor collaborative interaction and S304 diagnostic result integration and confirmation. Among them, S4 data security storage and traceability includes S401 distributed security storage and S402 full-process data traceability.

[0006] Preferably, in the S101 multi-center image data acquisition: each central hospital connects to the imaging equipment through a dedicated data acquisition terminal, acquires medical image data according to the DICOM 3.0 standard, and simultaneously acquires basic patient information (such as patient ID and examination type). The acquisition frequency is dynamically adjusted according to the patient flow of each central hospital, generally set to acquire device status information once every 10 seconds to ensure the real-time nature of data acquisition. The acquired data is temporarily stored in the local cache server of each center, with a cache capacity set to 500GB. When the cached data reaches 80%, the data upload preprocessing process is automatically triggered.

[0007] Preferably, the S102 data preprocessing initialization involves starting the preprocessing module, firstly performing format verification on the acquired image data, and then allowing the image data that passes the verification to enter the noise reduction processing stage. A noise reduction algorithm based on wavelet transform is adopted, with the wavelet basis set to db4 and the decomposition level set to 3. The noise-reduced image data is then normalized to normalize the pixel values ​​to the [0,1] interval. The normalization formula is: x_{norm}=\frac{x - x_{min}}{x_{max} - x_{min}}, where x is the original pixel value, and x_{min} and x_{max} are the minimum and maximum pixel values ​​in the image data, respectively. At the same time, the patient's basic information is desensitized, removing sensitive information such as name and ID number, and retaining only necessary non-sensitive information such as patient ID and examination type.

[0008] Preferably, the S103 preprocessing quality detection involves constructing a preprocessing quality evaluation index system, including image clarity (assessed using edge strength, with an edge strength ≥ 50 being acceptable), data integrity (data missing rate ≤ 0.1% being acceptable), and desensitization integrity (sensitive information removal rate 100% being acceptable). An AI quality detection model is used to automatically detect the preprocessed data. If the detection pass rate is ≥ 95%, the process proceeds to the next stage. If the detection fails, the data is returned to the corresponding preprocessing stage for reprocessing, with a maximum of 3 reprocessing attempts. If the data still fails after 3 attempts, a warning message is sent to the administrator.

[0009] Preferably, the S201 layered data encryption employs a layered encryption strategy, encrypting the preprocessed image data and non-sensitive patient information separately. For the image data, a block encryption method based on the national cryptographic algorithm SM4 is used, with a block length of 128 bits, a key length of 128 bits, and a CBC encryption mode. The initialization vector (IV) is randomly generated, and a new IV is generated for each encryption. For the non-sensitive patient information, an asymmetric encryption method based on the RSA algorithm is used, with a public key length of 2048 bits. The private key is kept separately by each central hospital. A digital digest is generated for each encrypted data block, using the SHA-256 hash algorithm to ensure data integrity.

[0010] Preferably, the S202 encrypted data transmission channel establishment involves: building an encrypted transmission channel based on a Virtual Private Network (VPN), using the IPsec protocol for channel encryption, setting the encryption algorithm to AES-256, the authentication algorithm to HMAC-SHA256, the key exchange protocol to IKEv2, and dynamically allocating the transmission bandwidth according to the network conditions of each central hospital, with a minimum guaranteed bandwidth of 10Mbps. Simultaneously, a sliding window protocol is used for flow control during transmission, with the window size initially set to 10 and dynamically adjusted according to network congestion. When the network packet loss rate is >2%, the window size is halved; when the network packet loss rate is <0.5%, the window size is increased by 1, with a maximum window size of 30.

[0011] Preferably, in S203, data transmission verification and feedback, the data receiving end verifies the digital digest of each data block in real time during the process of receiving encrypted data. If the verification passes, it continues to receive subsequent data; if the verification fails, it immediately sends a retransmission request to the data sending end. If no response is received within 1 second, the retransmission request is sent again, up to a maximum of 3 retransmission requests. If the correct data is still not received after 3 retransmissions, the data block is marked as a transmission failure, and a transmission anomaly report is sent to the administrator. At the same time, the transmission log is recorded, including transmission time, data volume, and transmission status information, for easy traceability later.

[0012] Preferably, the S301AI model loading and initialization involves loading a pre-trained multimodal medical image diagnosis model into the AI ​​analysis server of the collaborative diagnosis platform. This model is built on the Transformer architecture, containing 12 encoder layers and 6 decoder layers, with a hidden layer dimension of 768 and 12 attention heads. The model training dataset covers 1 million annotated image data from 20 central hospitals. The model parameters are initialized with a learning rate of 1e-4 and a momentum parameter of 0.9. Simultaneously, the historical diagnostic case databases of each central hospital are loaded. The case databases use distributed storage with a storage capacity of 10TB, facilitating case matching and reference during the diagnostic process.

[0013] Preferably, the S302 multi-center AI parallel analysis involves decrypting the encrypted image data transmitted to the collaborative diagnostic platform (decryption is only possible by authorized personnel and requires multi-factor authentication, including password, USB key, and biometrics), and distributing it to multiple AI analysis nodes for parallel analysis. The amount of data processed by each AI analysis node is dynamically allocated based on node performance. Node performance evaluation indicators include CPU utilization (≤80%), GPU utilization (≤85%), and memory usage (≤75%). The AI ​​model performs lesion detection, segmentation, and qualitative analysis on the image data, and outputs preliminary diagnostic results, including lesion location, size, and nature probability information. The analysis time is ≤30 seconds per case. S303 Multi-center Doctor Collaborative Interaction: The collaborative diagnostic platform pushes the preliminary diagnostic results output by each AI analysis node to the doctor's workstation of the corresponding central hospital. Doctors can view image data, AI preliminary diagnostic results, and related historical cases through the platform's interactive interface. It also supports real-time text, voice, and video communication between doctors from multiple centers. The communication data adopts an end-to-end encryption method, and the encryption algorithm is the same as the image data encryption algorithm in step 4. Doctors can modify and supplement the preliminary diagnostic results and mark doubtful points. The platform records every operation of the doctor to form an operation log. The operation log is tamper-proof and is stored using blockchain technology. The blockchain nodes include the servers of each central hospital and third-party supervision servers to ensure the authenticity and traceability of the log. S304 Diagnostic Result Integration and Confirmation: When doctors from multiple centers reach a consensus on the diagnostic results or after sufficient discussion, the collaborative diagnostic platform activates the diagnostic result integration module to integrate the revised diagnostic results from each doctor with weights. The weights are determined based on the doctor's professional field, title, and historical diagnostic accuracy. For example, the weight for a chief physician is 0.3, for an associate chief physician it is 0.2, for an attending physician it is 0.15, and for a resident physician it is 0.1. For every 1% increase in historical diagnostic accuracy above the average level, the weight increases by 0.01, with a maximum weight not exceeding 0.4. After integration, a final diagnostic report is generated, which is confirmed by attending physicians and above from each central hospital. After confirmation, the diagnostic report is stamped with an electronic seal (using digital signature technology based on the national cryptographic algorithm SM2) and sent back to the patient and relevant departments.

[0014] Preferably, the S401 distributed secure storage stores the final diagnostic report, preprocessed image data (encrypted), and operation log data in a distributed storage system. The storage system employs a multi-replica mechanism, storing three copies of each data block on storage nodes in different regions. These storage nodes are connected via high-speed fiber optic cables, with a data transmission rate ≥1Gbps. Simultaneously, an access control list (ACL)-based permission management strategy is adopted, assigning different access permissions to different roles (e.g., doctors, nurses, administrators, patients). For example, patients can only view their own diagnostic reports and related image data, but cannot modify them; doctors can view all relevant data of the patients under their care and perform diagnostic operations, but cannot delete data; administrators have the highest privileges and can perform system configuration and user management operations, but these operations must leave audit logs. S402 Full-Process Data Traceability: Constructs a data lifecycle traceability system that uses a unique data identifier (generated using a UUID, 36 bits long) to link information at each stage of data collection, preprocessing, encrypted transmission, collaborative diagnosis, and storage, including the operator, operation time, operation content, and data status. Users can query the full-process traceability information of corresponding data by entering the data identifier or patient ID. The traceability system has a response time of ≤2 seconds and also supports regulatory queries from regulatory authorities. Regulatory authorities can access the traceability system through a dedicated interface to view the flow of all data and ensure the compliance of data use.

[0015] (III) Beneficial Effects Compared with existing technologies, this invention provides an AI-driven method for multi-center collaborative diagnosis and secure data sharing of medical images, which has the following beneficial effects: 1. This AI-driven multi-center medical image collaborative diagnosis and data security sharing method breaks down data silos, improves data flow efficiency, and unifies data standards and triggering mechanisms: The solution adopts the DICOM 3.0 standard to unify the acquisition format, and realizes real-time monitoring of device status every 10 seconds through a dedicated data acquisition terminal. At the same time, it sets a local cache (500GB) 80% capacity trigger upload mechanism to avoid manual intervention, and improves data flow triggering efficiency by more than 90%; The preprocessing stage eliminates cross-center data adaptation differences through format verification, wavelet transform noise reduction (db4 wavelet basis, 3-level decomposition), and normalization ([0,1] interval) standardization operations, and shortens the data adaptation time to 1 / 5 of the original time, meeting the rapid diagnosis needs of emergency diseases; Distributed case library support: Loading a 10TB multi-center historical diagnostic case library (covering 1 million cases from 20 centers) provides cross-center case references for AI models and doctors' diagnoses, especially significantly improving the diagnostic support for rare diseases. According to calculations, the AI ​​recognition accuracy for rare diseases can be increased from less than 70% to more than 85%.

[0016] 2. This AI-driven multi-center medical image collaborative diagnosis and data security sharing method strengthens data security and builds a strong privacy protection barrier through layered encryption and dedicated transmission channels: It adopts a layered strategy of "SM4 block encryption (image data) + RSA asymmetric encryption (non-sensitive information)". SM4 sets 128-bit blocks and keys, CBC mode dynamic IV, and the encryption time is controlled within 10 seconds / GB. RSA - 2048-bit public key ensures key security. At the same time, an IPsec protocol VPN channel is built (AES - 256 encryption, HMAC - SHA256 authentication), combined with dynamic bandwidth allocation (minimum 10Mbps) and sliding window traffic control (initial 10, packet loss rate adjusted accordingly), the data transmission latency is ≤50ms, the encryption and decryption success rate is ≥99.99%, and the data leakage risk is reduced to below 0.01%, far below the industry security threshold. Fine-grained permissions and immutable logs: ACL-based permission management policies assign differentiated permissions to doctors, patients, and administrators (e.g., patients can only view their own data). Operation logs are stored on the blockchain (multi-center servers + third-party monitoring nodes) to ensure that logs are immutable and that data usage complies with all medical industry regulations.

[0017] 3. This AI-driven multi-center medical image collaborative diagnosis and data security sharing method improves diagnostic efficiency and accuracy, enhances the collaborative mechanism, and enables deep collaboration between AI and doctors: Based on a Transformer architecture, a multimodal AI model (12 encoder layers, 6 decoder layers, and 768-dimensional hidden layers) achieves lesion detection, segmentation, and qualitative analysis, with an analysis time of ≤30 seconds per case and a preliminary diagnostic accuracy of ≥92%; it also supports real-time text, voice, and video interaction among doctors in multiple centers (end-to-end SM4 encryption), with an interaction latency of ≤50ms. Doctors can quickly optimize diagnostic opinions by combining AI results with historical cases, reducing the total collaborative diagnosis time to 1 / 3 of the original time. Dynamically weighted integrated diagnostic results: Weights are dynamically allocated based on the doctor's professional field, title (0.3 for chief physician, 0.2 for associate chief physician), and historical diagnostic accuracy (weight increases by 0.01 for every 1% increase), avoiding the subjectivity of "simple voting" and improving the accuracy of integrated diagnostic results by 10%~15%; at the same time, electronic signature (SM2 digital signature) and multi-center attending physician confirmation mechanism ensure the legal validity and reliability of diagnostic reports.

[0018] 4. This AI-driven multi-center medical image collaborative diagnosis and data security sharing method ensures data quality, achieves full-process traceability, and implements multi-dimensional quality assessment: It constructs an assessment system for "image clarity (edge ​​strength ≥50), data integrity (missing rate ≤0.1%), and desensitization integrity (100% removal of sensitive information)," and achieves automatic detection through an AI quality detection model with a pass rate ≥95%. The data quality compliance rate after preprocessing is increased to over 99%, providing a reliable data foundation for AI analysis and physician diagnosis, and reducing the AI ​​misjudgment rate by 8%~12%. Full lifecycle traceability: By using UUID (36 bits) to uniquely identify each stage of data collection, preprocessing, transmission, diagnosis, and storage, the traceability system has a response time of ≤2 seconds. Regulatory authorities can view the data flow in real time through a dedicated interface. The time to locate data problems has been shortened from 24 hours to less than 5 minutes, effectively reducing the difficulty of handling medical disputes. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a technical solution, specifically, an AI-driven multi-center medical image collaborative diagnosis and secure data sharing method, comprising the following methods: S1 Data acquisition and preprocessing: S101 Multi-center Imaging Data Acquisition: Each central hospital connects to imaging equipment through a dedicated data acquisition terminal to acquire medical imaging data according to the DICOM 3.0 standard. At the same time, it collects basic patient information (such as patient ID and examination type). The acquisition frequency is dynamically adjusted according to the patient flow of each central hospital, generally set to collect equipment status information once every 10 seconds to ensure the real-time nature of data acquisition. The acquired data is temporarily stored in the local cache server of each center, with a cache capacity set to 500GB. When the cache data reaches 80%, the data upload preprocessing process is automatically triggered. S102 Data Preprocessing Initialization: The preprocessing module is started. First, the format of the acquired image data is checked. The image data that passes the check enters the noise reduction process. A noise reduction algorithm based on wavelet transform is adopted, with the wavelet basis set to db4 and the decomposition level set to 3. The noise-reduced image data is normalized to normalize the pixel values ​​to the [0,1] interval. The normalization formula is: x_{norm}=\frac{x - x_{min}}{x_{max} - x_{min}}, where x is the original pixel value, and x_{min} and x_{max} are the minimum and maximum pixel values ​​in the image data, respectively. At the same time, the patient's basic information is desensitized, removing sensitive information such as name and ID number, and only retaining necessary non-sensitive information such as patient ID and examination type. S103 Preprocessing Quality Inspection: A preprocessing quality evaluation index system is constructed, including image sharpness (assessed using edge strength, with an edge strength ≥50 being acceptable), data integrity (data missing rate ≤0.1% being acceptable), and desensitization integrity (sensitive information removal rate 100% being acceptable). An AI quality inspection model automatically inspects the preprocessed data. If the inspection pass rate is ≥95%, the data proceeds to the next stage. If the inspection fails, the data is returned to the corresponding preprocessing stage for reprocessing, with a maximum of 3 reprocessing attempts. If the data still fails after 3 attempts, a warning message is sent to the administrator. S2 Data Security Encryption and Transmission: S201 Layered Data Encryption: A layered encryption strategy is adopted to encrypt preprocessed image data and non-sensitive patient information separately. For image data, a block encryption method based on the national cryptographic algorithm SM4 is used, with a block length of 128 bits, a key length of 128 bits, and CBC encryption mode. The initialization vector (IV) is randomly generated, and a new IV is generated for each encryption. For non-sensitive patient information, an asymmetric encryption method based on the RSA algorithm is used, with a public key length of 2048 bits. The private key is kept separately by each central hospital. A digital digest is generated for each encrypted data block, using the SHA-256 hash algorithm to ensure data integrity. S202 Encrypted Data Transmission Channel Establishment: An encrypted transmission channel based on a Virtual Private Network (VPN) is established, using the IPsec protocol for channel encryption. The encryption algorithm is set to AES-256, the authentication algorithm to HMAC-SHA256, and the key exchange protocol to IKEv2. The transmission bandwidth is dynamically allocated according to the network conditions of each central hospital, with a minimum guaranteed bandwidth of 10Mbps. Simultaneously, a sliding window protocol is used for flow control during transmission. The initial window size is set to 10, and it is dynamically adjusted according to the network congestion level. When the network packet loss rate > 2%, the window size is halved; when the network packet loss rate < 0.5%, the window size is increased by 1, with a maximum window size of 30. S203 Data Transmission Verification and Feedback: During the process of receiving encrypted data, the data receiving end verifies the digital digest of each data block in real time. If the verification is successful, it continues to receive subsequent data; if the verification fails, it immediately sends a retransmission request to the data sending end. If no response is received within 1 second, the retransmission request is sent again, up to a maximum of 3 retransmission requests. If the correct data is still not received after 3 retransmissions, the data block is marked as a transmission failure, and a transmission anomaly report is sent to the administrator. At the same time, the transmission log is recorded, including transmission time, data volume, and transmission status information, for easy traceability later. S3AI-driven collaborative diagnostics: S301AI Model Loading and Initialization: The pre-trained multimodal medical image diagnosis model is loaded into the AI ​​analysis server of the collaborative diagnosis platform. This model is built on the Transformer architecture and contains 12 encoder layers and 6 decoder layers. The hidden layer dimension is 768, and the number of attention heads is 12. The model training dataset covers 1 million annotated image data from 20 central hospitals. The model parameters are initialized with a learning rate of 1e - 4 and a momentum parameter of 0.9. At the same time, the historical diagnosis case databases of each central hospital are loaded. The case database adopts a distributed storage method with a storage capacity of 10TB, which facilitates case matching and reference during the diagnosis process. S302 Multi-center AI Parallel Analysis: After decrypting the encrypted image data transmitted to the collaborative diagnostic platform (decryption is only possible by authorized personnel and requires multi-factor authentication, including password, USB key, and biometrics), the data is distributed to multiple AI analysis nodes for parallel analysis. The amount of data processed by each AI analysis node is dynamically allocated based on node performance. Node performance evaluation indicators include CPU utilization (≤80%), GPU utilization (≤85%), and memory usage (≤75%). The AI ​​model performs lesion detection, segmentation, and qualitative analysis on the image data, and outputs preliminary diagnostic results, including lesion location, size, and nature probability information. The analysis time is ≤30 seconds per case. S303 Multi-center Doctor Collaborative Interaction: The collaborative diagnostic platform pushes the preliminary diagnostic results output by each AI analysis node to the doctor's workstation of the corresponding central hospital. Doctors can view image data, AI preliminary diagnostic results, and related historical cases through the platform's interactive interface. It also supports real-time text, voice, and video communication between doctors from multiple centers. The communication data adopts an end-to-end encryption method, and the encryption algorithm is the same as the image data encryption algorithm in step 4. Doctors can modify and supplement the preliminary diagnostic results and mark doubtful points. The platform records every operation of the doctor to form an operation log. The operation log is tamper-proof and is stored using blockchain technology. The blockchain nodes include the servers of each central hospital and third-party supervision servers to ensure the authenticity and traceability of the log. S304 Diagnostic Result Integration and Confirmation: When doctors from multiple centers reach a consensus on the diagnostic results or after sufficient discussion, the collaborative diagnostic platform activates the diagnostic result integration module to integrate the revised diagnostic results of each doctor with weights. The weights are determined based on the doctor's professional field, title, and historical diagnostic accuracy. For example, the weight of a chief physician is 0.3, that of an associate chief physician is 0.2, that of an attending physician is 0.15, and that of a resident physician is 0.1. For every 1% increase in historical diagnostic accuracy above the average level, the weight increases by 0.01, with a maximum weight not exceeding 0.4. After integration, a final diagnostic report is generated, which is confirmed by attending physicians and above from each central hospital. After confirmation, the diagnostic report is stamped with an electronic seal (using digital signature technology based on the national cryptographic algorithm SM2) and sent back to the patient and relevant departments. S4 Data Security Storage and Traceability: S401 Distributed Secure Storage: The final diagnostic report, pre-processed image data (encrypted), and operation log data are stored in a distributed storage system. The system employs a multi-replica mechanism, with three copies of each data block stored on storage nodes in different regions. These nodes are connected via high-speed fiber optic cables, achieving a data transfer rate of ≥1Gbps. A Access Control List (ACL)-based permission management policy is used to assign different access permissions to different roles (e.g., doctors, nurses, administrators, patients). For example, patients can only view their own diagnostic reports and related image data, but cannot modify them; doctors can view all relevant data for their assigned patients and perform diagnostic operations, but cannot delete data; administrators have the highest privileges, enabling system configuration and user management operations, but all operations must be logged. S402 Full-Process Data Traceability: Constructs a data lifecycle traceability system that uses a unique data identifier (generated using a UUID, 36 bits long) to link information at each stage of data collection, preprocessing, encrypted transmission, collaborative diagnosis, and storage, including the operator, operation time, operation content, and data status. Users can query the full-process traceability information of corresponding data by entering the data identifier or patient ID. The traceability system has a response time of ≤2 seconds and also supports regulatory queries from regulatory authorities. Regulatory authorities can access the traceability system through a dedicated interface to view the flow of all data and ensure compliance in data use. Furthermore, this method breaks down data silos, improves data flow efficiency, and unifies data standards and triggering mechanisms: The solution adopts the DICOM 3.0 standard to unify the acquisition format, and realizes real-time monitoring of device status every 10 seconds through a dedicated data acquisition terminal. At the same time, it sets up a local cache (500GB) with 80% capacity to trigger the upload mechanism, avoiding manual intervention and improving data flow triggering efficiency by more than 90%. In the preprocessing stage, through format verification, wavelet transform noise reduction (db4 wavelet basis, 3-level decomposition), and normalization ([0,1] interval) standardization operations, the differences in cross-center data adaptation are eliminated, and the data adaptation time is shortened to 1 / 5 of the original time, meeting the needs of rapid diagnosis of emergency diseases. Distributed case library support: Loading a 10TB multi-center historical diagnostic case library (covering 1 million cases from 20 centers) provides cross-center case references for AI models and doctors' diagnoses, especially significantly improving the diagnostic support for rare diseases. According to calculations, the AI ​​recognition accuracy for rare diseases can be increased from less than 70% to more than 85%. Furthermore, this method strengthens data security and builds a solid privacy protection barrier through layered encryption and dedicated transmission channels: It adopts a layered strategy of "SM4 block encryption (image data) + RSA asymmetric encryption (non-sensitive information)". SM4 uses 128-bit blocks and keys, CBC mode dynamic IVs, and encryption time is controlled within 10 seconds / GB. RSA - 2048-bit public key ensures key security. Simultaneously, an IPsec protocol VPN channel (AES - 256 encryption, HMAC - SHA256 authentication) is established, combined with dynamic bandwidth allocation (minimum 10Mbps) and sliding window flow control (initial 10, adjusted according to packet loss rate). Data transmission latency is ≤50ms, encryption / decryption success rate is ≥99.99%, and the risk of data leakage is reduced to below 0.01%, far below the industry security threshold. Fine-grained permissions and immutable logs: ACL-based permission management policies assign differentiated permissions to doctors, patients, and administrators (e.g., patients can only view their own data). Operation logs are stored on the blockchain (multi-center servers + third-party monitoring nodes) to ensure that logs are immutable and that data usage complies with all medical industry regulations. Furthermore, this method improves diagnostic efficiency and accuracy, enhances the collaborative mechanism, and enables deep collaboration between AI and doctors: a multimodal AI model based on the Transformer architecture (12 encoder layers, 6 decoder layers, and 768-dimensional hidden layers) achieves lesion detection, segmentation, and qualitative analysis, with an analysis time of ≤30 seconds per case and a preliminary diagnostic accuracy of ≥92%; it also supports real-time text, voice, and video interaction among doctors in multiple centers (end-to-end SM4 encryption), with an interaction latency of ≤50ms. Doctors can quickly optimize diagnostic opinions by combining AI results with historical cases, reducing the total time for collaborative diagnosis to one-third of the original time; Dynamically weighted integrated diagnostic results: Weights are dynamically allocated based on the doctor's professional field, title (0.3 for chief physician, 0.2 for associate chief physician), and historical diagnostic accuracy (weight increases by 0.01 for every 1% increase), avoiding the subjectivity of "simple voting" and improving the accuracy of integrated diagnostic results by 10%~15%; at the same time, electronic signature (SM2 digital signature) and multi-center attending physician confirmation mechanism ensure the legal validity and reliability of diagnostic reports; Furthermore, this method ensures data quality, achieves full-process traceability, and enables multi-dimensional quality assessment: it constructs an assessment system for "image clarity (edge ​​strength ≥50), data integrity (missing rate ≤0.1%), and desensitization integrity (100% removal of sensitive information)," and achieves automatic detection through an AI quality detection model, with a detection pass rate ≥95%. The data quality compliance rate after preprocessing is improved to over 99%, providing a reliable data foundation for AI analysis and doctor diagnosis, and reducing the AI ​​misjudgment rate by 8%~12%. Full lifecycle traceability: By using UUID (36 bits) to uniquely identify each stage of data collection, preprocessing, transmission, diagnosis, and storage, the traceability system has a response time of ≤2 seconds. Regulatory authorities can view the data flow in real time through a dedicated interface. The time to locate data problems has been shortened from 24 hours to less than 5 minutes, effectively reducing the difficulty of handling medical disputes.

[0021] 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 equivalents.

Claims

1. An AI-driven multi-center medical image collaborative diagnosis and secure data sharing method, comprising S1 data acquisition and preprocessing, S2 secure data encryption and transmission, S3 AI-driven collaborative diagnosis, and S4 secure data storage and traceability, characterized in that: The S1 data acquisition and preprocessing includes S101 multi-center image data acquisition, S102 data preprocessing initialization, and S103 preprocessing quality detection. Among them, S2 data security encryption and transmission includes S201 layered data encryption, S202 establishment of encrypted data transmission channel and S203 transmission data verification and feedback; Among them, S3AI-driven collaborative diagnosis includes S301 AI model loading and initialization, S302 multi-center AI parallel analysis, S303 multi-center doctor collaborative interaction and S304 diagnostic result integration and confirmation. Among them, S4 data security storage and traceability includes S401 distributed security storage and S402 full-process data traceability.

2. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S101 multi-center image data acquisition: Each central hospital connects to the imaging equipment through a dedicated data acquisition terminal to acquire medical image data according to the DICOM 3.0 standard, and simultaneously acquires the patient's basic information (patient ID, examination type). The acquisition frequency is dynamically adjusted according to the patient flow of each central hospital, generally set to acquire equipment status information once every 10 seconds to ensure the real-time nature of data acquisition. The acquired data is temporarily stored in the local cache server of each center, with a cache capacity set to 500GB. When the cache data reaches 80%, the data upload preprocessing process is automatically triggered.

3. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S102 data preprocessing initialization: The preprocessing module is started. First, the format of the acquired image data is checked. The image data that passes the check enters the noise reduction processing stage. A noise reduction algorithm based on wavelet transform is adopted, with the wavelet basis set to db4 and the decomposition level set to 3. The noise-reduced image data is normalized to normalize the pixel values ​​to the [0,1] interval. The normalization formula is: x_{norm}=\frac{x - x_{min}}{x_{max} - x_{min}}, where x is the original pixel value, and x_{min} and x_{max} are the minimum and maximum pixel values ​​in the image data, respectively. At the same time, the patient's basic information is desensitized, removing sensitive information such as name and ID number, and only retaining necessary non-sensitive information such as patient ID and examination type.

4. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S103 preprocessing quality inspection involves constructing a preprocessing quality evaluation index system, including image clarity (assessed using edge strength, with an edge strength ≥50 being acceptable), data integrity (data missing rate ≤0.1% being acceptable), and desensitization integrity (sensitive information removal rate 100% being acceptable). An AI quality inspection model is used to automatically inspect the preprocessed data; if the inspection pass rate is ≥95%, the process proceeds to the next stage. If the test fails, the data will be returned to the corresponding preprocessing stage for reprocessing. The maximum number of reprocessing attempts is 3. If the test still fails after 3 attempts, an alert will be sent to the administrator.

5. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S201 layered data encryption employs a layered encryption strategy, encrypting preprocessed image data and non-sensitive patient information separately. For image data, a block encryption method based on the national cryptographic algorithm SM4 is used, with a block length of 128 bits, a key length of 128 bits, and a CBC encryption mode. The initialization vector (IV) is randomly generated, and a new IV is generated for each encryption. For non-sensitive patient information, an asymmetric encryption method based on the RSA algorithm is used, with a public key length of 2048 bits. The private key is kept separately by each central hospital. A digital digest is generated for each encrypted data block, using the SHA-256 hash algorithm to ensure data integrity.

6. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S202 encrypted data transmission channel is established as follows: An encrypted transmission channel based on a Virtual Private Network (VPN) is built, using the IPsec protocol for channel encryption. The encryption algorithm is set to AES-256, the authentication algorithm to HMAC-SHA256, and the key exchange protocol to IKEv2. The transmission bandwidth is dynamically allocated according to the network conditions of each central hospital, with a minimum guaranteed bandwidth of 10Mbps. At the same time, a sliding window protocol is used for flow control during transmission. The window size is initially set to 10 and dynamically adjusted according to the network congestion level. When the network packet loss rate is >2%, the window size is halved; when the network packet loss rate is <0.5%, the window size is increased by 1, with a maximum window size of 30.

7. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S203 data transmission verification and feedback: During the process of receiving encrypted data, the data receiving end verifies the digital digest of each data block in real time. If the verification is successful, it continues to receive subsequent data; if the verification fails, it immediately sends a retransmission request to the data sending end. If no response is received within 1 second, the retransmission request is sent again, up to a maximum of 3 retransmission requests. If the correct data is still not received after 3 retransmissions, the data block is marked as a transmission failure, and a transmission anomaly report is sent to the administrator. At the same time, the transmission log is recorded, including transmission time, data volume, and transmission status information, for subsequent traceability.

8. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S301AI model loading and initialization involves loading a pre-trained multimodal medical image diagnosis model onto the AI ​​analysis server of the collaborative diagnosis platform. This model is built on the Transformer architecture, containing 12 encoder layers and 6 decoder layers, with a hidden layer dimension of 768 and 12 attention heads. The model training dataset covers 1 million annotated image data from 20 central hospitals. The model parameters are initialized with a learning rate of 1e-4 and a momentum parameter of 0.

9. Simultaneously, historical diagnostic case libraries from each central hospital are loaded. These case libraries use distributed storage with a capacity of 10TB, facilitating case matching and reference during the diagnostic process.

9. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S302 multi-center AI parallel analysis involves decrypting the encrypted image data transmitted to the collaborative diagnostic platform (decryption is only possible by authorized personnel and requires multi-factor authentication, including password, USB key, and biometrics), and distributing it to multiple AI analysis nodes for parallel analysis. The amount of data processed by each AI analysis node is dynamically allocated based on node performance. Node performance evaluation indicators include CPU utilization (≤80%), GPU utilization (≤85%), and memory usage (≤75%). The AI ​​model performs lesion detection, segmentation, and qualitative analysis on the image data, outputting preliminary diagnostic results, including lesion location, size, and nature probability information. The analysis time is ≤30 seconds per case. S303 Multi-center Doctor Collaborative Interaction: The collaborative diagnostic platform pushes the preliminary diagnostic results output by each AI analysis node to the doctor's workstation of the corresponding central hospital. Doctors can view image data, AI preliminary diagnostic results, and related historical cases through the platform's interactive interface. It also supports real-time text, voice, and video communication between doctors from multiple centers. The communication data adopts an end-to-end encryption method, and the encryption algorithm is the same as the image data encryption algorithm in step 4. Doctors can modify and supplement the preliminary diagnostic results and mark doubtful points. The platform records every operation of the doctor to form an operation log. The operation log is tamper-proof and is stored using blockchain technology. The blockchain nodes include the servers of each central hospital and third-party supervision servers to ensure the authenticity and traceability of the log. S304 Diagnostic Result Integration and Confirmation: When doctors from multiple centers reach a consensus on the diagnostic results or after sufficient discussion, the collaborative diagnostic platform activates the diagnostic result integration module to integrate the revised diagnostic results from each doctor with weights. The weights are determined based on the doctor's professional field, title, and historical diagnostic accuracy. For example, the weight for a chief physician is 0.3, for an associate chief physician it is 0.2, for an attending physician it is 0.15, and for a resident physician it is 0.

1. For every 1% increase in historical diagnostic accuracy above the average level, the weight increases by 0.01, with a maximum weight not exceeding 0.

4. After integration, a final diagnostic report is generated, which is confirmed by attending physicians and above from each central hospital. After confirmation, the diagnostic report is stamped with an electronic seal (using digital signature technology based on the national cryptographic algorithm SM2) and sent back to the patient and relevant departments.

10. The AI-driven multi-center medical image collaborative diagnosis and secure data sharing method according to claim 1, characterized in that: The S401 distributed secure storage system stores the final diagnostic report, preprocessed image data (encrypted), and operation log data. The system employs a multi-replica mechanism, storing three copies of each data block on storage nodes in different regions. These nodes are connected via high-speed fiber optic cables, achieving a data transmission rate of ≥1Gbps. Furthermore, it utilizes an access control list (ACL)-based permission management strategy, assigning different access permissions to different roles (doctors, nurses, administrators, and patients). For example, patients can only view their own diagnostic reports and related image data, but cannot modify them; doctors can view all relevant data for their assigned patients and perform diagnostic operations, but cannot delete data. Administrators have the highest privileges and can perform system configuration and user management operations, but these operations must be logged. S402 Full-Process Data Traceability: Constructs a data lifecycle traceability system that uses a unique data identifier (generated using a UUID, 36 bits long) to link information at each stage of data collection, preprocessing, encrypted transmission, collaborative diagnosis, and storage, including the operator, operation time, operation content, and data status. Users can query the full-process traceability information of corresponding data by entering the data identifier or patient ID. The traceability system has a response time of ≤2 seconds and also supports regulatory queries from regulatory authorities. Regulatory authorities can access the traceability system through a dedicated interface to view the flow of all data and ensure the compliance of data use.