A blockchain-based artificial intelligence detection method for medical imaging data

Abstract

The present invention relates to an artificial intelligence detection method for medical imaging data based on blockchain, and belongs to the field of blockchain application technology. The method uses artificial intelligence target detection technology to assist doctors in diagnosis, realizes safe sharing of medical imaging data, and medical imaging data training parameters are anonymously stored on the blockchain. Except for the server, any medical structure cannot obtain the medical imaging data information of other medical structures. Since the data is stored on the blockchain, the information is prevented from being tampered with, and the authenticity of the data is guaranteed. At the same time, the method can save storage space and bandwidth on the chain. According to the method, each medical structure uploads the training parameters instead of the medical imaging data itself, realizes global parameter updates, and avoids uploading massive medical imaging data.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence detection method for medical imaging data based on blockchain, and belongs to the field of blockchain application technology. Background Art

[0002] Medical institutions are extremely busy, seeing numerous patients daily and generating vast amounts of medical data. Imaging data accounts for over 90% of all medical data and is rapidly increasing at an annual rate exceeding 30%. With the continuous improvement and upgrading of medical testing equipment, the rapid growth of medical imaging data stands in stark contrast to the slow growth in the number of physicians. The burden on physicians from tasks such as reading and analyzing CT scans is increasing, causing eye fatigue and compromising diagnostic accuracy. Therefore, utilizing computer-assisted diagnosis and AI-based object detection technologies to assist physicians in making informed decisions is crucial for reducing their burden, increasing diagnostic accuracy, and improving patient care efficiency.

[0003] Object detection technology has attracted considerable attention from both academia and industry due to its wide range of real-world applications. Object detection primarily involves extracting information such as the location and category of a target through detection and classification. With the continuous advancement of computing power, object detection technology has flourished, evolving into two main categories: two-stage and single-stage.

[0004] The two-stage algorithm generates candidate regions on the target object based on feature extraction, then classifies and regresses the regions, using rectangular boxes to determine the exact location information and determine the category. Two-stage algorithms have higher accuracy. Faster R-CNN, R-FCN, and FPN are all two-stage algorithms. Among the two-stage algorithms, R-CNN is the earliest deep learning target detection algorithm, which significantly improves the mean average precision (mAP) of target detection. He et al. proposed SPP-NET by introducing the spatial pyramid pooling layer (SPP), which greatly improves the detection efficiency compared to R-CNN. Fast R-CNN improves the pooling kernel of R-CNN, mapping the candidate regions into feature vectors of uniform scale, improving the efficiency of model training.

[0005] Single-stage algorithms achieve candidate region generation, classification, and regression in a single step. YOLO, SSD, and RetinaNet are all examples of single-stage object detection algorithms. YOLO, with its simple network structure, is fast and offers good real-time performance, making it a representative example of this type of algorithm. SSD optimizes multi-scale object detection by extracting features using pre-set anchor boxes for different convolutional layers. YOLOv3 continues the previous DarkNet architecture while introducing a residual structure to improve detection accuracy for small-scale objects. Therefore, overall, single-stage algorithms are highly efficient.

[0006] However, for object detection, ensuring accurate predictions requires a large amount of accurate data. If the training sample size is too small, poor learning results will result. Incorrect data can even lead to misleading results.

[0007] Sharing medical imaging data between medical institutions, research centers, and other organizations is a simple and efficient solution for obtaining large amounts of medical imaging data. However, current centralized medical data storage solutions are not conducive to data sharing, pose the risk of data resale, and are even less conducive to protecting patient privacy, such as their medical conditions. Summary of the Invention

[0008] The purpose of this invention is to address the defects and shortcomings of the existing technology, and to solve technical problems such as insufficient utilization of medical imaging data and difficulties in information sharing, and to innovatively propose an artificial intelligence detection method for medical imaging data based on blockchain.

[0009] First, the relevant contents of the present invention are explained.

[0010] Definition 1: Object Detection

[0011] It refers to extracting the required information such as the location and category of the predetermined target in the detection object through detection and classification.

[0012] Definition 2: Medical institution

[0013] It refers to an institution that provides necessary medical examinations, diagnosis, treatment and other services to patients. It is the source of medical imaging data. It also includes institutions such as research institutes that generate relevant medical imaging data.

[0014] Definition 3: Blockchain

[0015] It refers to a distributed ledger that is decentralized, tamper-proof, and traceable, achieved using technologies such as distributed storage and digital signatures.

[0016] Definition 4: Smart Contract

[0017] It refers to a computer protocol that disseminates, verifies or executes contracts in an information-based manner. Smart contracts in blockchain are written in scripting languages ​​and faithfully execute codes according to established rules without human intervention.

[0018] Definition 5: Hash function

[0019] A hash function is a function that can map an input of arbitrary length (also called a function preimage) to an output of fixed length (i.e., a hash value). Hash functions are one of the fundamental building blocks of modern cryptography, possessing properties such as one-way and collision resistance.

[0020] Definition 6: Digital Signature

[0021] It refers to the use of public key cryptography technology to sign digital information, generating an unforgeable digital string that can only be generated by the sender of the information. This digital string is also a valid proof of the authenticity of the information sent by the sender.

[0022] Definition 7: Proof-of-Work Consensus Mechanism

[0023] It means that users perform complex calculations that are time-consuming and the answers can be quickly verified by the service provider. The time, equipment and energy consumed are used as guarantee costs to ensure that services and resources are used for real needs. The contribution of medical institutions in sharing medical imaging data can be evaluated through proof-of-work consensus.

[0024] The present invention is implemented by adopting the following technical solutions.

[0025] A blockchain-based artificial intelligence detection method for medical imaging data includes the following steps:

[0026] Step 1: Initialization.

[0027] Medical institutions generate public-private key pairs and broadcast their public keys to other medical institutions and servers.

[0028] Step 2: Define the task.

[0029] The medical imaging data of various medical institutions is fully utilized to form a data set and train a global model. For a specific training task, the server defines and subdivides the task.

[0030] Step 3: Distribute training tasks.

[0031] After the server determines the training task, it distributes the task (such as CT diagnosis task).

[0032] First, a temporary session key is generated for each medical institution. The server session key and training task are encrypted with the public key of each medical institution and sent to the corresponding medical institution.

[0033] Step 4: Learn locally.

[0034] Each medical institution uses the target detection algorithm to train its model locally and updates the parameters locally.

[0035] Step 5: Upload local update parameters.

[0036] Each medical institution encrypts the locally updated parameters using a temporary session key and uploads the encrypted training parameters to the blockchain.

[0037] Step 6: Update global parameters.

[0038] The server decrypts the local updated parameters of each medical institution in the blockchain and aggregates them into global parameters.

[0039] Step 7: Auxiliary diagnosis.

[0040] Each medical institution sends the diagnostic tasks to be performed to the server through a smart contract. After the server makes a judgment, the smart contract estimates the accuracy of the auxiliary diagnosis results. When the accuracy exceeds the threshold, the correct result will be returned to the medical institution, otherwise FALSE will be returned.

[0041] Step 8: Request data.

[0042] A medical institution requests medical imaging data from other medical institutions, sends the request signature to the smart contract, and after verification by the contract, forwards the request to the corresponding medical institution and then returns the medical imaging data to the requester.

[0043] Beneficial effects

[0044] Compared with the prior art, the method of the present invention has the following beneficial effects:

[0045] 1. This method enables secure sharing of medical imaging data. Medical imaging data training parameters are anonymously stored on the blockchain. No medical institution, other than the server, can access the medical imaging data of another medical institution. Furthermore, because the data is stored on the blockchain, it prevents tampering and ensures data authenticity.

[0046] 2. This method can save on-chain storage space and bandwidth. Each medical institution uploads training parameters instead of the medical imaging data itself, achieving global parameter updates while avoiding the upload of massive medical imaging data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the method of the present invention;

[0048] Figure 2 This is a schematic diagram of the application server diagnosis process in the method of the present invention;

[0049] Figure 3 The figure shows the application data flow in the method of the present invention. DETAILED DESCRIPTION

[0050] The specific implementation process of the method of the present invention is further described in detail below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, a schematic diagram of a blockchain-based artificial intelligence medical imaging data detection model is elaborated in detail, which fully utilizes the data of various medical institutions to form a massive data set with the participation of many medical institutions, thereby obtaining a better global training model.

[0052] A blockchain-based artificial intelligence detection method for medical imaging data includes the following steps:

[0053] Step 1: Initialization.

[0054] Medical institutions generate public-private key pairs and broadcast their public keys to other medical institutions and servers.

[0055] Specifically, each medical institution generates a public-private key {PK j , SK j}{j=1,2,···,K},PK j Represents the public key of the jth medical institution, SK j represents the private key of the jth medical institution, K indicates that there are K medical institutions in total, and their public keys are broadcast to other medical institutions and servers.

[0056] Step 2: Define the task.

[0057] In order to make full use of the medical imaging data of various medical institutions, a massive data set is formed with the participation of many institutions, so as to train a better global model.

[0058] For a certain training task, the server defines and subdivides the task, which may include clarifying the learning object, related equipment, application type (such as CT image), training set format, training model (such as convolutional neural network), training requirements (such as learning rate), etc.

[0059] Step 3: Distribute training tasks.

[0060] The server determines the training task tx t , tx t Represents the tth task and performs task distribution (taking CT diagnosis as an example).

[0061] First, generate a temporary session key for each medical institution sk represents the temporary session key, subscript j represents the jth medical institution, and superscript txt represents the tth task, K represents a total of K medical institutions; the server session key and training task use the public key PK of each medical institution j After encryption, it is sent to the corresponding medical institution.

[0062] Step 4: Learn locally.

[0063] Each medical institution uses the target detection algorithm to train its model locally and updates the parameters locally. data represents the training parameters, subscript j is the medical institution number, and superscript tx t Number the task.

[0064] Step 5: Upload local update parameters.

[0065] Each medical institution will use the temporary session key to update the parameters locally Encrypt and train the parameters Encryption to obtain the encrypted ciphertext and will Upload to the blockchain. || represents cascade, Representatives The result of the hash operation.

[0066] Step 6: Update global parameters.

[0067] The server uses a temporary session key Decrypt the local updated parameters of each medical institution in the blockchain and aggregate them into global parameters.

[0068] Step 7: Perform auxiliary diagnosis.

[0069] like Figure 2 As shown in Figure 1, assume that medical institution H needs a server to diagnose a medical image (e.g., a CT image) x0 and obtain a diagnosis result y0. x represents the medical image, with the subscript being the image's serial number, and a serial number of 0 indicating that no diagnosis has been made. y represents the diagnosis result, with the subscript being the diagnostic result's serial number, and a serial number of 0 indicating that no diagnosis has been made.

[0070] In order not to expose the patient's privacy, the medical institution H does not want the server to know which specific medical image needs to be diagnosed. Therefore, the medical image graph {x1, x2, ···, x m} as the confusion input, non-zero sequence number represents the diagnosis. {x1,x2,···,x m} corresponding to the diagnosis results {y1,y2,···,y m};{x0,{x1,y1},{x2,y2},···,{x m ,y m}} is submitted to the smart contract by the medical institution H, and the smart contract then converts {x1,x2,···,x m}Send server.

[0071] Since {y1,y2,···,y m}, the server cannot identify the medical image x0 that needs to be judged, and will judge all inputs indiscriminately and send all judgment results {y′1,y′2,···,y′ m}Return to smart contract, y′ m Represents the server to the medical image x m The prediction result. At this time, let y′ i =y i The number of is n, then y′ i ≠y i The number of is mt, and the prediction accuracy is assumed to be n / m. Let the smart contract judgment threshold be M. When n / m>M, the smart contract returns y0 to the medical institution H, otherwise it returns false.

[0072] Each medical institution sends the medical images that need to be diagnosed to the server through the smart contract. After the server makes a judgment, the smart contract estimates the accuracy of the diagnosis result. When the accuracy exceeds the threshold, the correct result will be returned to the medical institution, otherwise false will be returned.

[0073] Step 8: Request data.

[0074] A medical institution requests medical imaging data from another medical institution and sends the request signature to the smart contract. After verification, the contract forwards the request to the corresponding medical institution and then returns the medical imaging data to the requester.

[0075] like Figure 3 As shown. Each block corresponds to a learning task, block tx i Corresponding to the i-th task tx i Suppose medical institution A wants to obtain block tx i Data of medical institution B B , send a request SK to the smart contract A (request). Request indicates the request of medical institution A, SK A (request) means using the private key SK of medical institution A A Digitally sign the request.

[0076] After the smart contract verifies the signature of medical institution A, it forwards the request to medical institution B. Medical institution B decrypts the corresponding data on the blockchain and uses the public key PK of the medical institution. A Encrypted data B, and use your own private key SK B After signing, it is sent to the smart contract.

[0077] Subsequently, the smart contract verifies the signature and forwards it to medical institution A, which uses its own private key SK A Decrypted to get data′ B , data′ B It represents the data obtained by medical institution A after decrypting the data of medical institution B. Medical institution A decrypts the data data′ B Perform hash operation.

[0078] Comparison result h(data′ B ) is consistent with the data on the chain. If they are consistent, it means that medical institution B has delivered real data, otherwise it means that the data has been tampered with. B ) represents the pair data′ B Perform hash operation.

Claims

1. A blockchain-based artificial intelligence detection method for medical imaging data, characterized in that: The following steps are involved: Step 1: Initialization: The medical institution generates a public-private key pair and broadcasts its public key to other medical institutions and servers; Step 2: Define the task; make full use of the medical imaging data of various medical institutions to form a data set and train a global model; for a certain training task, the server defines and subdivides the task; Step 3: Distribute training tasks; first generate a temporary session key for each medical institution; encrypt the server session key and training tasks with the public key of each medical institution and send them to the corresponding medical institution; Step 4: Local learning: Each medical institution uses the target detection algorithm to train its model locally and updates the parameters locally; Step 5: Upload the locally updated parameters; each medical institution encrypts the locally updated parameters using a temporary session key and uploads the encrypted training parameters to the blockchain; Step 6: Update global parameters; the server decrypts the local update parameters of each medical institution in the blockchain and aggregates them into global parameters; Step 7: Perform auxiliary diagnosis; each medical institution sends the diagnostic task to be performed to the server through the smart contract. After the server makes a judgment, the smart contract estimates the accuracy of the auxiliary diagnosis result. When the accuracy exceeds the threshold, the correct result is returned to the medical institution, otherwise FALSE is returned; Step 8: Request data; A medical institution requests medical imaging data from other medical institutions, sends the request signature to the smart contract, and after verification by the contract, forwards the request to the corresponding medical institution, and then returns the medical imaging data to the requester.

2. The method for artificial intelligence detection of medical imaging data based on blockchain as claimed in claim 1, characterized in that: In step 2, for a certain training task, the server defines and subdivides the task, including clarifying the learning object, related equipment, application type, training set format, training model and training requirements.

3. The method for artificial intelligence detection of medical imaging data based on blockchain as claimed in claim 1, characterized in that: The server determines the training task tx t , tx t Represents the tth task and distributes the tasks; first, a temporary session key is generated for each medical institution sk represents the temporary session key, the subscript j represents the jth medical institution, and the superscript tx t represents the tth task, K represents a total of K medical institutions; the server session key and training task use the public key PK of each medical institution j After encryption, it is sent to the corresponding medical institution; Each medical institution uses the target detection algorithm to train its model locally and updates the parameters locally data represents the training parameters, the subscript j is the medical institution number, and the superscript tx t Number the task; Each medical institution will use the temporary session key to update the parameters locally Encrypt and train the parameters Encryption to get the encrypted ciphertext and will Upload to blockchain; || represents cascade, Representatives The result after hash operation; The server uses the temporary session key Decrypt the local update parameters of each medical institution in the blockchain and aggregate them into global parameters.

4. The artificial intelligence detection method for medical imaging data based on blockchain as claimed in claim 1, characterized in that: In step 7, it is assumed that the medical institution H needs the server to diagnose the medical image x0 and obtain the diagnosis result y0; x represents the medical image, the subscript is the serial number of the medical image, and the serial number 0 represents that the diagnosis has not been made; y represents the diagnosis result, the subscript is the serial number of the diagnosis result, and the serial number 0 represents that the diagnosis has not been made; In order not to expose the patient's privacy, the medical institution H does not want the server to know which medical image needs to be diagnosed. Therefore, the medical image graph {x1, x2, ..., x m } as the obfuscated input, non-zero serial numbers represent diagnosed; {x1, x2, …, x m }Corresponding diagnosis results {y1, y2, ..., y m };{x0,{x1,y1},{x2,y2},…,{x m ,y m }} is submitted to the smart contract by the medical institution H, and the smart contract then converts {x1, x2, …, x m }SendServer; Since we do not receive {y1, y2, …, y m }, the server cannot identify the medical image x0 that needs to be judged, and will judge all inputs indiscriminately and send all judgment results {y′1, y′2, …, y′ m }Return to smart contract, y′ m Represents the server to the medical image x m The prediction result; at this time, let y′ i =y i If the number of is n, then y′ i ≠y i The number of is mt, and the prediction accuracy is assumed to be n / m; let the smart contract judgment threshold be M. When n / m>M, the smart contract returns y0 to the medical institution H, otherwise it returns false; Each medical institution sends the medical images that need to be diagnosed to the server through the smart contract. After the server makes a judgment, the smart contract estimates the accuracy of the diagnosis result. When the accuracy exceeds the threshold, the correct result is returned to the medical institution, otherwise false is returned.

5. The method for artificial intelligence detection of medical imaging data based on blockchain as claimed in claim 1, characterized in that: In step 8, each block corresponds to a learning task, block tx i Corresponding to the i-th task tx i ; Suppose medical institution A wants to get block tx i Medical institution B's data B , send a request SK to the smart contract A (request); request indicates the request from medical institution A, SK A (request) means using the private key SK of medical institution A A Digitally sign the request; After the smart contract verifies the signature of medical institution A, it forwards the request to medical institution B; medical institution B decrypts the corresponding data on the blockchain and uses the public key PK of the medical institution. A Encrypted data B , and use your own private key SK B After signing, send it to the smart contract; Subsequently, the smart contract verifies the signature and forwards it to medical institution A, which uses its own private key SK A Decrypted to get data′ B , data′ B It represents the data obtained by medical institution A after decrypting the data of medical institution B; medical institution A decrypts the data data′ B Perform hash operations; Comparison result h(data′ B ) is consistent with the data on the chain. If they are consistent, it means that medical institution B has delivered real data, otherwise it means that the data has been tampered with. B ) represents the pair data′ B ) to perform hash operation.

Images