A deep learning-based hippocampus three-dimensional segmentation method for brain MRI

Abstract

The application provides a kind of brain MRI hippocampus three-dimensional segmentation method based on deep learning, it is related to data processing technical field, through server node initialization and through blockchain encryption release public parameter combination;Through user node decryption public parameter combination and adopts MFSOA algorithm training to generate sub-target parameter combination, then upload through blockchain encryption;Finally, through the server node aggregation sub-target parameter obtains target public parameter combination, and carries out hippocampus segmentation based on target public parameter combination, combines blockchain technology with federal learning framework, makes original brain MRI data always remain in user node local, solves the trust and privacy problem of medical data sharing, breaks the data island;At the same time, MFSOA algorithm effectively improves the model optimization efficiency and segmentation accuracy, finally significantly enhances the generalization ability and robustness of model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a three-dimensional segmentation method for the hippocampus in brain MRI based on deep learning. Background Technology

[0002] The hippocampus is a key region of the brain responsible for learning and memory, and its morphological changes (such as atrophy) are important biomarkers for neurodegenerative diseases such as Alzheimer's. Therefore, accurate and automated 3D segmentation of the hippocampus in brain MRI images is of crucial clinical significance for early disease diagnosis, progression monitoring, and treatment evaluation. Currently, segmentation methods based on deep learning, especially 3D convolutional neural networks (3D-CNN), have achieved significant results in hippocampal segmentation. Among them, U-Net and its 3D variant (3D U-Net), due to their excellent encoder-decoder structure and skip connections, can effectively fuse multi-scale features and have become the mainstream model for medical image segmentation.

[0003] However, existing technologies still face the following two major challenges:

[0004] Data privacy and data silos: High-quality medical imaging data is often scattered across different hospitals and research institutions. Due to patient privacy regulations (such as HIPAA and GDPR), it is difficult to centralize this data for training a unified, high-generalization model. This results in models trained independently by each institution performing well on local data but experiencing a sharp decline in performance on external data, i.e., the data silo effect, which limits the overall performance of the model and its clinical application.

[0005] The inefficiency of hyperparameter optimization is a significant issue: the performance of deep learning models is highly dependent on the selection of hyperparameters, such as learning rate, batch size, and optimizer type. Traditional hyperparameter optimization methods, such as grid search and random search, are computationally expensive and inefficient. While some emerging intelligent optimization algorithms (such as genetic algorithms and particle swarm optimization) have made improvements, they are still prone to getting trapped in local optima when dealing with high-dimensional and complex hyperparameter spaces. This prevents the model from achieving its optimal performance and results in slow convergence, prolonging the model development cycle. Summary of the Invention

[0006] This application aims to provide a deep learning-based method for three-dimensional segmentation of the hippocampus in brain MRI, which addresses the problems of data silos and low efficiency of hyperparameter optimization in existing technologies.

[0007] This application provides a deep learning-based method for three-dimensional segmentation of the hippocampus in brain MRI, including:

[0008] A 3D segmentation model is generated using a deep learning model on a server node, and the hyperparameters of the 3D segmentation model are initialized to obtain a common parameter combination.

[0009] The public parameter combination is encrypted using the public key of the user node in the blockchain to obtain the encrypted public parameter combination, and the encrypted public parameter combination is published to the blockchain.

[0010] The user node collects the encrypted public parameter combination from the blockchain and decrypts the encrypted public parameter combination using its own private key to obtain the initial public parameter combination.

[0011] Based on the initial common parameter combination, the user node calls the MFSOA algorithm to train the initial common parameter combination and obtain the sub-target common parameter combination;

[0012] The user node uses the server node's public key to encrypt the sub-target public parameter combination, obtaining the encrypted sub-target public parameter combination, and publishes the encrypted sub-target public parameter combination in the blockchain;

[0013] The encrypted sub-target common parameter combination is aggregated by the server node to obtain the target common parameter combination. Based on the three-dimensional segmentation model and the target common parameter combination, the hippocampus of the three-dimensional brain MRI image is segmented to obtain the hippocampus segmentation result.

[0014] In one possible implementation, a 3D segmentation model is generated using a deep learning model on a server node, and the hyperparameters of the 3D segmentation model are initialized to obtain a common parameter combination, including:

[0015] A 3D U-net model is constructed using server nodes to obtain a 3D segmentation model;

[0016] The hyperparameters of the three-dimensional segmentation model are randomly initialized between the upper and lower bounds, and the randomly initialized hyperparameters are encoded into a vector to obtain a common parameter combination.

[0017] Repeatedly obtain multiple combinations of common parameters.

[0018] In one possible implementation, based on the initial common parameter combination, the user node calls the MFSOA algorithm to train the initial common parameter combination to obtain the sub-target common parameter combination, including:

[0019] The fitness of each initial common parameter combination is obtained based on the training data stored on the user node, and the optimal and worst parameter combinations are determined based on the fitness.

[0020] Based on the fitness corresponding to the initial common parameter combination, a fitness decision search strategy is used to perform a random information fusion search on the initial common parameter combination to obtain the common parameter combination after the random information fusion search.

[0021] Based on the optimal parameter combination, a solution space position transformation strategy is used to perform position transformation search on the common parameter combination after the random information fusion search, so as to obtain the common parameter combination after the position transformation search.

[0022] A lens-back learning strategy is used to perform a global mutation search on the common parameter combination after the position transformation search, so as to obtain the common parameter combination after the global mutation search.

[0023] An association control search strategy is used to perform association control search on the optimal and worst parameter combinations to obtain the optimal and worst parameter combinations after the association control search.

[0024] The optimal parameter combination, the worst parameter combination, and the common parameter combination after the association control search are merged back into a population to obtain the target population.

[0025] Determine the current number of training iterations and whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, redetermine the optimal parameter combination based on the target population to obtain the sub-target common parameter combination. Otherwise, based on the target population, return to the step of determining the optimal parameter combination and the worst parameter combination based on the fitness, and proceed to the next training iteration.

[0026] In one possible implementation, based on the fitness corresponding to the initial common parameter combination, a fitness decision search strategy is used to perform a random information fusion search on the initial common parameter combination to obtain a common parameter combination after the random information fusion search, including:

[0027] Based on the fitness corresponding to the initial common parameter combinations, the probability of selection for each initial common parameter combination is determined as follows:

[0028]

[0029] in, Indicates the first i The probability of being selected corresponding to each initial common parameter combination. Indicates the first i The fitness corresponding to each initial common parameter combination, where N represents the total number of initial common parameter combinations;

[0030] For any initial common parameter combination, based on the selection probability corresponding to each initial common parameter combination, a roulette wheel strategy is used to randomly select N / 2 other common parameter combinations for the initial common parameter combination, thereby obtaining N / 2 first target parameter combinations corresponding to the initial common parameter combination.

[0031] Based on the fitness of the initial common parameter combination and the fitness of the first target parameter combination, the direction decision factor is determined as follows:

[0032]

[0033] in, Indicates the first i The first initial common parameter combination corresponds to the first j Each search direction decision factor Represents a symbolic function. Indicates the first i The first initial common parameter combination corresponds to the first j The fitness corresponding to each first objective parameter combination;

[0034] Based on the stated direction decision factor, a random information fusion search is performed on the initial common parameter combination to obtain the common parameter combination after the random information fusion search:

[0035]

[0036] in, Indicates the first t During the training process, the first i The first initial common parameter combination d dimensional hyperparameters, Indicates the first t During the training process, the first i The first common parameter combination after random information fusion search d dimensional hyperparameters, Indicates the first i The first initial common parameter combination corresponds to the first j The first combination of target parameters d dimensional hyperparameters, d =1,2,…,D, where D represents the total dimension of hyperparameters in the common parameter combination.

[0037] In one possible implementation, based on the optimal parameter combination, a solution space location transformation strategy is used to perform a location transformation search on the common parameter combination after the random information fusion search, resulting in the common parameter combination after the location transformation search, including:

[0038] Determine the current training iteration count, and obtain the position transition selection parameters based on the current training iteration count:

[0039]

[0040] in, This indicates the position conversion selection parameters. This represents the minimum value of the position selection parameter. This indicates the maximum value of the position transformation selection parameter; Represents the sine function. Represents pi (π). This indicates the preset maximum number of training iterations;

[0041] For any combination of common parameters after random information fusion search, the first position transformation coefficient and the second position transformation coefficient are obtained as follows:

[0042]

[0043]

[0044] in, Indicates the conversion factor for the first position. Represents a random number between (0,1). This represents the conversion factor for the second position. Indicates the total positional transformation order. Indicates the first t During the training process, the first k The first common parameter combination after random information fusion search d dimensional hyperparameters, d =1,2,…,D, where D represents the total dimension of hyperparameters in the common parameter combination. Represented as the first k After a random information fusion search, the common parameter combination is randomly matched with other common parameter combinations. d dimensional hyperparameters;

[0045] Based on the optimal parameter combination, the position transformation selection parameter, the first position transformation coefficient, and the second position transformation coefficient, a position transformation search is performed on the common parameter combination after the random information fusion search, resulting in the following common parameter combination after the position transformation search:

[0046]

[0047] in, Indicates the first t During the training process, the first k The first position transformation search followed by the common parameter combination d dimensional hyperparameters, The first parameter combination represents the optimal parameter combination. d dimensional hyperparameters, Indicates the position transition control factor. s Indicates the order of position transformation.

[0048] In one possible implementation, a lens back learning strategy is used to perform a global mutation search on the common parameter combination after the position transformation search, resulting in the common parameter combination after the global mutation search, including:

[0049] For any given combination of common parameters after a positional transformation search, the corresponding global mutation search position is obtained as follows:

[0050]

[0051] in, Indicates the first t During the training process, the first m The first position transformation search followed by the common parameter combination d dimensional hyperparameters, Indicates the first m The global mutation search position corresponding to the common parameter combination after the position transformation search is the nth position. d dimensional hyperparameters, Indicates the first d The upper limit of the hyperparameters, Indicates the first d The lower bound of the hyperparameter;

[0052] Determine whether the fitness of the common parameter combination after the position transformation search is less than the fitness of its corresponding global mutation search position. If so, use the global mutation search position as the common parameter combination after the global mutation search; otherwise, use the common parameter combination after the position transformation search as the common parameter combination after the global mutation search.

[0053] In one possible implementation, an association control search strategy is employed to perform an association control search on the optimal and worst parameter combinations, resulting in the optimal and worst parameter combinations after the association control search, including:

[0054] Determine the current training iteration count, and obtain the association control search factor based on the current training iteration count:

[0055]

[0056] in, Indicates the search factors for association control;

[0057] Based on the worst-case parameter combination and the association control search factor, the basic association parameters are determined as follows:

[0058]

[0059] in, Indicates the first d The underlying correlation parameters corresponding to the dimensional hyperparameters The worst parameter combination is represented by the first... d dimensional hyperparameters, This represents the first common parameter combination after a random global mutation search. d The hyperparameters are defined such that the common parameter combination after random global mutation search is neither the worst nor the best parameter combination.

[0060] Based on the optimal and worst parameter combinations, the target correlation parameters are obtained as follows:

[0061]

[0062] in, Indicates the first d The target correlation parameter corresponding to the hyperparameter. This represents the fitness of the worst-case parameter combination. This represents the fitness of the optimal parameter combination; The first parameter combination represents the optimal parameter combination. d dimensional hyperparameters, This represents the Euclidean distance between the worst and best parameter combinations.

[0063] Based on the basic association parameters and the target association parameters, the optimal and worst parameter combinations after the association control search are obtained as follows:

[0064]

[0065] in, This represents the first optimal parameter combination after the association control search. d dimensional hyperparameters, This represents the worst parameter combination after the association control search. d dimensional hyperparameters, Represents the first random number between (0,1). This represents the second random number between (0,1).

[0066] In one possible implementation, the encrypted sub-target common parameter combinations are aggregated by the server node to obtain the target common parameter combination, including: the encrypted sub-target common parameter combinations are aggregated by the server node using the FedAvg algorithm to obtain the target common parameter combination.

[0067] In one possible implementation, hippocampal segmentation is performed on a three-dimensional brain MRI image based on the aforementioned three-dimensional segmentation model and a combination of target common parameters to obtain hippocampal segmentation results, including:

[0068] Based on the target common parameter combination, the three-dimensional segmentation model is deployed on the server node, and the user node is allowed to schedule the three-dimensional segmentation model deployed on the server node to perform hippocampal segmentation on the three-dimensional brain MRI image to obtain the hippocampal segmentation result.

[0069] In one possible implementation, hippocampal segmentation is performed on a three-dimensional brain MRI image based on the aforementioned three-dimensional segmentation model and a combination of target common parameters to obtain hippocampal segmentation results, including:

[0070] The target common parameter combination and the three-dimensional segmentation model are made public, enabling user nodes to obtain the target common parameter combination and the three-dimensional segmentation model, and deploy the three-dimensional segmentation model on the user nodes according to the target common parameter combination, so that the user nodes can use the three-dimensional segmentation model deployed locally to perform hippocampal segmentation on the three-dimensional brain MRI image to obtain the hippocampal segmentation result.

[0071] This application provides a deep learning-based method for 3D segmentation of the hippocampus in brain MRI. It combines blockchain technology with a federated learning framework, ensuring that the original brain MRI data remains locally on the user node. Only encrypted public parameter combinations, free of any privacy information, are transmitted on the blockchain. The decentralized, immutable, and traceable characteristics of the blockchain provide security for parameter publication, transmission, and aggregation, effectively solving the trust issue in medical data sharing. This allows different medical institutions to collaboratively train models without violating privacy regulations, breaking down data silos and fully utilizing multi-source heterogeneous data. This significantly improves the generalization ability and robustness of the final model. Furthermore, the MFSOA algorithm is used on the user node to train the initial public parameter combinations, effectively addressing the low optimization efficiency of existing technologies and improving the model's segmentation accuracy. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart illustrating a deep learning-based three-dimensional segmentation method for the hippocampus in brain MRI, as proposed in one embodiment of this application.

[0074] Figure 2 This is a schematic diagram of the process for obtaining the combination of common parameters of sub-targets according to an embodiment of this application. Detailed Implementation

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] like Figure 1 As shown, this application provides a method for three-dimensional segmentation of the hippocampus in brain MRI based on deep learning, including:

[0077] S101. Generate a 3D segmentation model using a deep learning model through the server node, and initialize the hyperparameters of the 3D segmentation model to obtain a common parameter combination.

[0078] For example, a 3D U-Net model can be built as the basis for a 3D segmentation model using server nodes (such as a large medical research center or cloud platform). The hyperparameters of this model include learning rate, batch size, optimizer momentum, weight decay coefficient, etc. Assuming there are a total of D dimensions, the hyperparameters of each dimension can be randomly initialized and encoded into a common parameter combination. Multiple common parameter combinations can be obtained for user nodes to train.

[0079] S102. The public parameter combination is encrypted using the public key of the user node in the blockchain to obtain the encrypted public parameter combination, and the encrypted public parameter combination is published to the blockchain.

[0080] The public keys of all registered user nodes (such as Hospital A and Hospital B) in the blockchain network can be obtained through server nodes. These public keys are then used to encrypt multiple combinations of public parameters to generate an encrypted data packet. Finally, this encrypted data packet is broadcast to the blockchain network as a transaction. Once on-chain, the initial parameters are public, transparent, and tamper-proof to all participants.

[0081] S103. The encrypted public parameter combination in the blockchain is collected by the user node, and the encrypted public parameter combination is decrypted using its own private key to obtain the initial public parameter combination.

[0082] S104. Based on the initial common parameter combination, the user node calls the MFSOA algorithm to train the initial common parameter combination and obtain the sub-target common parameter combination.

[0083] For example, a user node in Hospital A continuously listens to the blockchain. When it detects a new block containing initial parameters published by the server, it downloads the encrypted data packet. Hospital A uses its private key to decrypt the data packet, recovering N initial public parameter combinations. Next, Hospital A initiates a local training process on its local dataset containing hundreds of labeled brain MRI (Magnetic Resonance Imaging) samples. The core of this process is to call the MFESOA algorithm to optimize these multiple public parameter combinations, ultimately obtaining the optimal public parameter combination, thus acquiring the initial public parameter combination.

[0084] S105. The user node uses the server node's public key to encrypt the sub-target public parameter combination to obtain the encrypted sub-target public parameter combination, and publishes the encrypted sub-target public parameter combination in the blockchain.

[0085] S106. The encrypted sub-target common parameter combination is aggregated through the server node to obtain the target common parameter combination. Based on the three-dimensional segmentation model and the target common parameter combination, the hippocampus of the three-dimensional brain MRI image is segmented to obtain the hippocampus segmentation result.

[0086] For example, encrypted sub-target parameters from hospitals A and B can be collected from the blockchain via server nodes. After decryption, and assuming the FedAvg algorithm is used for aggregation, a combination of target common parameters is obtained. Finally, the server loads this combination of target common parameters into its 3D U-Net model, forming the final global segmentation model. This global segmentation model can be deployed as a cloud service. When new brain MRI images need analysis, hospital A or hospital B can upload the images to the server. The server calls the model to perform segmentation and returns the results (such as a 3D mask of the hippocampus) to the hospital. The entire process involves the original data remaining locally, protecting patient privacy.

[0087] It should be noted that the examples given in the embodiments of this application are merely to enable those skilled in the art to more easily understand the technical solutions shown in the embodiments of this application, and are not intended to limit the technical solutions.

[0088] This application provides a deep learning-based three-dimensional segmentation method for the hippocampus in brain MRI, combining blockchain technology with a federated learning framework. This ensures that the original brain MRI data remains locally on the user node, with only encrypted public parameter combinations, free of any privacy information, transmitted on the blockchain. The decentralized, immutable, and traceable characteristics of the blockchain provide security for parameter publication, transmission, and aggregation, effectively solving the trust issue in medical data sharing. This allows different medical institutions to collaboratively train models without violating privacy regulations, breaking down data silos and fully utilizing multi-source heterogeneous data. This significantly improves the generalization ability and robustness of the final model. Furthermore, the MFSOA algorithm is used on the user node to train the initial public parameter combinations, effectively addressing the low optimization efficiency of existing technologies and improving the model's segmentation accuracy.

[0089] In one possible implementation, a 3D segmentation model is generated using a deep learning model on a server node, and the hyperparameters of the 3D segmentation model are initialized to obtain a common parameter combination, including:

[0090] A 3D U-net model is constructed using server nodes to obtain a 3D segmentation model;

[0091] The hyperparameters of the three-dimensional segmentation model are randomly initialized between the upper and lower bounds, and the randomly initialized hyperparameters are encoded into a vector to obtain a common parameter combination.

[0092] Repeatedly obtain multiple combinations of common parameters.

[0093] For the hyperparameter initialization of the 3D U-net model, various initialization methods exist in the existing technology, so existing techniques can also be used for initialization. Furthermore, obtaining multiple common parameter combinations allows all user nodes to train the model on the same basis.

[0094] like Figure 2 As shown, based on the initial common parameter combination, the user node calls the MFSOA algorithm (Multimodal Fusion Search Optimization Algorithm) to train the initial common parameter combination and obtain the sub-target common parameter combination, including:

[0095] S201. Obtain the fitness corresponding to each initial common parameter combination based on the training data stored on the user node, and determine the optimal parameter combination and the worst parameter combination based on the fitness.

[0096] For example, we can first obtain the loss function value corresponding to the initial common parameter combination, and then take the reciprocal of the loss function value to obtain the fitness corresponding to the initial common parameter combination. In the process of taking the reciprocal, in order to avoid the denominator being zero, we can also add the loss function value to a very small constant term (such as 0.0001) before taking the reciprocal.

[0097] S202. Based on the fitness corresponding to the initial common parameter combination, a fitness decision search strategy is used to perform a random information fusion search on the initial common parameter combination to obtain the common parameter combination after the random information fusion search.

[0098] S203. Based on the optimal parameter combination, a solution space position transformation strategy is used to perform a position transformation search on the common parameter combination after the random information fusion search, so as to obtain the common parameter combination after the position transformation search.

[0099] S204. A lens-back learning strategy is used to perform a global mutation search on the common parameter combination after the position transformation search, so as to obtain the common parameter combination after the global mutation search.

[0100] S205. An association control search strategy is used to perform an association control search on the optimal and worst parameter combinations to obtain the optimal and worst parameter combinations after the association control search.

[0101] S206. The optimal parameter combination, the worst parameter combination, and the common parameter combination after the association control search are re-integrated into a population to obtain the target population.

[0102] S207. Determine the current number of training iterations and determine whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, redetermine the optimal parameter combination based on the target population to obtain the sub-target common parameter combination. Otherwise, based on the target population, return to the step of determining the optimal parameter combination and the worst parameter combination based on the fitness, and proceed to the next training iteration.

[0103] For example, the combination in the target population can be used as the initial common parameter combination, and the steps of determining the optimal and worst parameter combinations based on the fitness can be returned to proceed to the next training.

[0104] The MFSOA algorithm provided in this application constructs a multi-level, multi-dimensional composite search strategy, significantly improving the efficiency and accuracy of parameter optimization. First, fitness evaluation clarifies the optimal and worst solutions, providing precise direction for subsequent searches. Then, a random information fusion strategy enhances population diversity, effectively avoiding premature convergence; a solution space location transformation strategy refines the neighborhood of high-quality solutions, accelerating local convergence. More importantly, a lens-based back-learning strategy introduces global mutation, giving the algorithm a strong ability to escape local optima, while an association control strategy dynamically balances global exploration and local development by regulating the optimal and worst solutions. Finally, through iterative fusion and selection, a virtuous cycle of optimization is formed. This scheme not only overcomes the limitations of single optimization algorithms and avoids the problem of existing optimization algorithms (such as particle swarm optimization) easily getting trapped in local optima, but also significantly improves convergence speed and solution quality, making it particularly suitable for solving complex optimization problems with high dimensions and multiple modalities.

[0105] Optionally, after each search of common parameter combinations, out-of-bounds handling can be performed to ensure training effectiveness.

[0106] In one possible implementation, based on the fitness corresponding to the initial common parameter combination, a fitness decision search strategy is used to perform a random information fusion search on the initial common parameter combination to obtain a common parameter combination after the random information fusion search, including:

[0107] Based on the fitness corresponding to the initial common parameter combinations, the probability of selection for each initial common parameter combination is determined as follows:

[0108]

[0109] in, Indicates the first i The probability of being selected corresponding to each initial common parameter combination. Indicates the first i The fitness corresponding to each initial common parameter combination, where N represents the total number of initial common parameter combinations;

[0110] For any initial common parameter combination, based on the selection probability corresponding to each initial common parameter combination, a roulette wheel strategy is used to randomly select N / 2 other common parameter combinations for the initial common parameter combination, thereby obtaining N / 2 first target parameter combinations corresponding to the initial common parameter combination.

[0111] Based on the fitness of the initial common parameter combination and the fitness of the first target parameter combination, the direction decision factor is determined as follows:

[0112]

[0113] in, Indicates the first i The first initial common parameter combination corresponds to the first j Each search direction decision factor Represents a symbolic function. Indicates the first i The first initial common parameter combination corresponds to the first j The fitness corresponding to each first objective parameter combination;

[0114] Based on the stated direction decision factor, a random information fusion search is performed on the initial common parameter combination to obtain the common parameter combination after the random information fusion search:

[0115]

[0116] in, Indicates the first t During the training process, the first i The first initial common parameter combination d dimensional hyperparameters, Indicates the first t During the training process, the first i The first common parameter combination after random information fusion search d dimensional hyperparameters, Indicates the first i The first initial common parameter combination corresponds to the first j The first combination of target parameters d dimensional hyperparameters, d =1,2,…,D, where D represents the total dimension of hyperparameters in the common parameter combination.

[0117] Traditional random search is blind, while this strategy introduces a fitness-weighted roulette wheel selection mechanism, giving the search process a clear survival-of-the-fittest orientation. High-performance initial common parameter combinations (high fitness) have a higher probability of being selected as information sources, meaning the algorithm will naturally focus its search on regions that have been proven to be better. This positive feedback mechanism allows the population to converge to the optimal solution region more quickly, significantly accelerating the algorithm's convergence speed and avoiding wasting time in a large number of invalid solution spaces.

[0118] In one possible implementation, based on the optimal parameter combination, a solution space location transformation strategy is used to perform a location transformation search on the common parameter combination after the random information fusion search, resulting in the common parameter combination after the location transformation search, including:

[0119] Determine the current training iteration count, and obtain the position transition selection parameters based on the current training iteration count:

[0120]

[0121] in, This indicates the position conversion selection parameters. This indicates the minimum value of the position conversion selection parameter, which can be set to 0.001; This indicates the maximum value of the position conversion selection parameter, which can be set to 1; Represents the sine function. Represents pi (π). This indicates the preset maximum number of training iterations;

[0122] For any combination of common parameters after random information fusion search, the first position transformation coefficient and the second position transformation coefficient are obtained as follows:

[0123] in, Indicates the conversion factor for the first position. Represents a random number between (0,1). This represents the conversion factor for the second position. This indicates the total positional transformation order and can be set to 3 or 4; Indicates the first t During the training process, the first k The first common parameter combination after random information fusion search d dimensional hyperparameters, d =1,2,…,D, where D represents the total dimension of hyperparameters in the common parameter combination. Represented as the first k After a random information fusion search, the common parameter combination is randomly matched with other common parameter combinations. d dimensional hyperparameters;

[0124] Based on the optimal parameter combination, the position transformation selection parameter, the first position transformation coefficient, and the second position transformation coefficient, a position transformation search is performed on the common parameter combination after the random information fusion search, resulting in the following common parameter combination after the position transformation search:

[0125]

[0126] in, Indicates the first t During the training process, the first k The first position transformation search followed by the common parameter combination d dimensional hyperparameters, The first parameter combination represents the optimal parameter combination. d dimensional hyperparameters, This represents the position transition control factor, which can be set to 0.4; sIndicates the order of position transformation.

[0127] By performing positional transformation searches based on the optimal parameter combination within the solution space, local exploitation capabilities can be effectively improved. This allows for refined and intensive searches within identified high-quality regions to approximate the global optimum. Furthermore, multi-stage positional transformation strategies enable large-scale global exploitation, quickly locating promising regions. By designing positional transformation selection parameters, more global exploitation can be performed in the early stages of the algorithm, while more local exploitation can be conducted in the later stages, thus more accurately locating the optimal hyperparameters.

[0128] In one possible implementation, a lens back learning strategy is used to perform a global mutation search on the common parameter combination after the position transformation search, resulting in the common parameter combination after the global mutation search, including:

[0129] For any given combination of common parameters after a positional transformation search, the corresponding global mutation search position is obtained as follows:

[0130]

[0131] in, Indicates the first t During the training process, the first m The first position transformation search followed by the common parameter combination d dimensional hyperparameters, Indicates the first m The global mutation search position corresponding to the common parameter combination after the position transformation search is the nth position. d dimensional hyperparameters, Indicates the first d The upper limit of the hyperparameters, Indicates the first d The lower bound of the hyperparameter;

[0132] Determine whether the fitness of the common parameter combination after the position transformation search is less than the fitness of its corresponding global mutation search position. If so, use the global mutation search position as the common parameter combination after the global mutation search; otherwise, use the common parameter combination after the position transformation search as the common parameter combination after the global mutation search.

[0133] By continuously replacing inferior solutions with inverse solutions, the population can maintain a wider distribution throughout the solution space, avoiding the clustering of all individuals in a small area. This effectively improves training diversity and ensures global expansiveness. The introduction of an adaptive selection strategy guarantees the algorithm's training speed. This diversity makes the algorithm insensitive to initial values ​​and gives it stronger adaptability and robustness to different types of optimization problems.

[0134] In one possible implementation, an association control search strategy is employed to perform an association control search on the optimal and worst parameter combinations, resulting in the optimal and worst parameter combinations after the association control search, including:

[0135] Determine the current training iteration count, and obtain the association control search factor based on the current training iteration count:

[0136]

[0137] in, Indicates the search factors for association control;

[0138] Based on the worst-case parameter combination and the association control search factor, the basic association parameters are determined as follows:

[0139]

[0140] in, Indicates the first d The underlying correlation parameters corresponding to the dimensional hyperparameters The worst parameter combination is represented by the first... d dimensional hyperparameters, This represents the first common parameter combination after a random global mutation search. d The hyperparameters are defined such that the common parameter combination after random global mutation search is neither the worst nor the best parameter combination.

[0141] Based on the optimal and worst parameter combinations, the target correlation parameters are obtained as follows:

[0142]

[0143] in, Indicates the first d The target correlation parameter corresponding to the hyperparameter. This represents the fitness of the worst-case parameter combination. This represents the fitness of the optimal parameter combination; The first parameter combination represents the optimal parameter combination. d dimensional hyperparameters, This represents the Euclidean distance between the worst and best parameter combinations.

[0144] Based on the basic association parameters and the target association parameters, the optimal and worst parameter combinations after the association control search are obtained as follows:

[0145]

[0146] in, This represents the first optimal parameter combination after the association control search. ddimensional hyperparameters, This represents the worst parameter combination after the association control search. d dimensional hyperparameters, Represents the first random number between (0,1). This represents the second random number between (0,1).

[0147] By performing correlation control search on the optimal and worst parameter combinations, the probability of collisions is increased, which in turn increases the probability of the algorithm escaping local optima. The gradually increasing speed also helps the algorithm reduce the probability of getting trapped in local optima in the later stages.

[0148] In one possible implementation, the encrypted sub-target common parameter combinations are aggregated by the server node to obtain the target common parameter combination, including: the encrypted sub-target common parameter combinations are aggregated by the server node using the FedAvg algorithm to obtain the target common parameter combination.

[0149] In one possible implementation, hippocampal segmentation is performed on a three-dimensional brain MRI image based on the aforementioned three-dimensional segmentation model and a combination of target common parameters to obtain hippocampal segmentation results, including:

[0150] Based on the target common parameter combination, the three-dimensional segmentation model is deployed on the server node, and the user node is allowed to schedule the three-dimensional segmentation model deployed on the server node to perform hippocampal segmentation on the three-dimensional brain MRI image to obtain the hippocampal segmentation result.

[0151] In one possible implementation, hippocampal segmentation is performed on a three-dimensional brain MRI image based on the aforementioned three-dimensional segmentation model and a combination of target common parameters to obtain hippocampal segmentation results, including:

[0152] The target common parameter combination and the three-dimensional segmentation model are made public, enabling user nodes to obtain the target common parameter combination and the three-dimensional segmentation model, and deploy the three-dimensional segmentation model on the user nodes according to the target common parameter combination, so that the user nodes can use the three-dimensional segmentation model deployed locally to perform hippocampal segmentation on the three-dimensional brain MRI image to obtain the hippocampal segmentation result.

[0153] The model obtained in this application supports two deployment modes: one is to deploy it as a cloud service on a server node, which users can call through a network API (Application Program Interface), reducing the hardware requirements of the user end; the other is to make the model and parameters completely public, which users can download to their local deployment, meeting the needs of scenarios with higher requirements for data sovereignty and offline use. This flexibility enables the present invention to adapt to different clinical application environments and needs.

[0154] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0155] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0158] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0159] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0160] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A deep learning-based method for three-dimensional segmentation of the hippocampus in brain MRI, characterized in that, The application relates to a method for generating a three-dimensional segmentation model of a hippocampus in a three-dimensional brain MRI image, comprising the following steps: a server node generates a three-dimensional segmentation model by adopting a deep learning model, and initializes hyperparameters of the three-dimensional segmentation model to obtain a common parameter combination; the common parameter combination is encrypted by a public key of a user node in a blockchain to obtain an encrypted common parameter combination, and the encrypted common parameter combination is published in the blockchain; the user node collects the encrypted common parameter combination in the blockchain, and decrypts the encrypted common parameter combination by adopting a private key of the user node to obtain an initial common parameter combination; the user node calls a multimodal fusion search optimization algorithm to train the initial common parameter combination on the basis of the initial common parameter combination to obtain a sub-target common parameter combination; the user node encrypts the sub-target common parameter combination by adopting a public key of the server node to obtain an encrypted sub-target common parameter combination, and publishes the encrypted sub-target common parameter combination in the blockchain; the server node aggregates the encrypted sub-target common parameter combination to obtain a target common parameter combination, and performs hippocampus segmentation on the three-dimensional brain MRI image on the basis of the three-dimensional segmentation model and the target common parameter combination to obtain a hippocampus segmentation result; the user node calls the multimodal fusion search optimization algorithm to train the initial common parameter combination on the basis of the initial common parameter combination to obtain a sub-target common parameter combination, comprising the following steps: based on training data stored on the user node, the fitness of each initial common parameter combination is obtained, and the optimal parameter combination and the worst parameter combination are determined according to the fitness; according to the fitness of the initial common parameter combination, a random information fusion search of the initial common parameter combination is performed by adopting a fitness decision search strategy to obtain a common parameter combination after random information fusion search; according to the optimal parameter combination, a position conversion search of the common parameter combination after random information fusion search is performed by adopting a solution space position conversion strategy to obtain a common parameter combination after position conversion search; a global mutation search of the common parameter combination after position conversion search is performed by adopting a lens reverse learning strategy to obtain a common parameter combination after global mutation search; an association control search of the optimal parameter combination and the worst parameter combination is performed by adopting an association control search strategy to obtain the optimal parameter combination and the worst parameter combination after association control search; the optimal parameter combination, the worst parameter combination and the common parameter combination after global mutation search after the association control search are re-fused into a population to obtain a target population; a current training number is determined, and it is determined whether the current training number is greater than or equal to a preset maximum training number; if yes, the optimal parameter combination is re-determined according to the target population to obtain a sub-target common parameter combination; if not, the step of determining the optimal parameter combination and the worst parameter combination according to the fitness is returned on the basis of the target population, and the next training is entered. According to the optimal parameter combination, a position conversion search strategy is used for position conversion search on the common parameter combination after the random information fusion search, to obtain a common parameter combination after position conversion search, including: Determine the current training number, and obtain a position conversion selection parameter according to the current training number: ； wherein, denotes a position conversion selection parameter, denotes a minimum value of the position conversion selection parameter, denotes a maximum value of the position conversion selection parameter; denotes a sine function, denotes a circle constant, denotes a preset maximum training number; For any one common parameter combination after the random information fusion search, obtain a first position conversion coefficient and a second position conversion coefficient: ； ； wherein, represents a first position conversion coefficient, represents a random number between (0, 1), represents a second position conversion coefficient, represents a total position conversion order, represents a d-dimensional hyperparameter of the common parameter combination after the kth random information fusion search in the tth training process, d = 1, 2, …, D, D represents the total dimension of the hyperparameter in the common parameter combination, represents a d-dimensional hyperparameter of the other common parameter combination randomly matched with the common parameter combination after the kth random information fusion search. According to the optimal parameter combination, a position conversion search strategy is used for position conversion search on the common parameter combination after the random information fusion search, to obtain a common parameter combination after position conversion search, including: ； wherein, represents the d-th dimensional hyperparameter of the common parameter combination after the k-th position conversion search in the t-th training process, represents the d-th dimensional hyperparameter of the optimal parameter combination, represents the position conversion control factor, and s represents the position conversion order.

2. The deep learning-based brain MRI hippocampus three-dimensional segmentation method according to claim 1, wherein, A three-dimensional segmentation model is generated by a server node using a deep learning model, and the hyperparameters of the three-dimensional segmentation model are initialized to obtain a common parameter combination, including: A 3D U-net model is constructed by a server node to obtain a three-dimensional segmentation model; Randomly initialize between the upper limit and the lower limit of the hyperparameters of the three-dimensional segmentation model, and encode the hyperparameters after random initialization into a vector to obtain a common parameter combination; Repeat to obtain multiple common parameter combinations. 3.The deep learning-based brain MRI hippocampus three-dimensional segmentation method according to claim 1, characterized in that, According to the fitness corresponding to the initial common parameter combination, a fitness decision search strategy is used for random information fusion search on the initial common parameter combination to obtain a common parameter combination after random information fusion search, including: According to the fitness corresponding to the initial common parameter combination, determine the selection probability corresponding to each initial common parameter combination: ； wherein, represents the selected probability corresponding to the i-th initial common parameter combination, represents the fitness corresponding to the i-th initial common parameter combination, and N represents the total number of initial common parameter combinations; For any one initial common parameter combination, use a roulette strategy to randomly select N / 2 other common parameter combinations for the initial common parameter combination according to the selection probability corresponding to each initial common parameter combination, to obtain N / 2 first target parameter combinations corresponding to the initial common parameter combination; According to the fitness of the initial common parameter combination and the fitness corresponding to the first target parameter combination, determine a direction decision factor: ； wherein, denotes the jth search direction decision factor corresponding to the ith initial common parameter combination, denotes a sign function, denotes the fitness of the jth first target parameter combination corresponding to the ith initial common parameter combination; According to the direction decision factor, the initial common parameter combination is randomly information fusion searched to obtain a common parameter combination after random information fusion search: ； wherein, represents the d-th dimension hyper-parameter of the i-th initial common parameter combination in the t-th training process, represents the d-th dimension hyper-parameter of the common parameter combination after the i-th random information fusion search in the t-th training process, represents the d-th dimension hyper-parameter of the j-th first target parameter combination corresponding to the i-th initial common parameter combination, d = 1, 2, …, D, and D represents the total dimension of the hyper-parameters in the common parameter combination.

4. The deep learning-based brain MRI hippocampus three-dimensional segmentation method according to claim 3, characterized in that, A lens reverse learning strategy is used for global mutation search on the common parameter combination after position conversion search to obtain a common parameter combination after global mutation search, including: For any one common parameter combination after position conversion search, obtain its corresponding global mutation search position: ； wherein, represents the d-th dimensional hyper-parameter of the common parameter combination after the m-th position conversion search in the t-th training process, represents the d-th dimensional hyper-parameter of the global variation search position corresponding to the common parameter combination after the m-th position conversion search, represents the upper limit of the d-th dimensional hyper-parameter, represents the lower limit of the d-th dimensional hyper-parameter; Determine whether the fitness of the common parameter combination after position conversion search is less than the fitness of the corresponding global mutation search position, if yes, the global mutation search position is taken as the common parameter combination after global mutation search, otherwise the common parameter combination after position conversion search is taken as the common parameter combination after global mutation search.

5. The deep learning-based brain MRI hippocampus three-dimensional segmentation method according to claim 4, characterized in that, An association control search strategy is used for association control search on the optimal parameter combination and the worst parameter combination to obtain the optimal parameter combination and the worst parameter combination after association control search, including: Determine the current training number, and obtain an association control search factor according to the current training number: ； wherein represents the correlation control search factor; According to the worst parameter combination and the associated control search factor, the basic associated parameter is determined as: ； wherein, represents the dth dimension hyper-parameter corresponding to the base correlation parameter, represents the dth dimension hyper-parameter of the worst parameter combination, represents the dth dimension hyper-parameter of the common parameter combination after the random global variation search, and the common parameter combination after the random global variation search is neither the worst parameter combination nor the optimal parameter combination; According to the optimal parameter combination and the worst parameter combination, the target associated parameter is obtained as: ； wherein, represents the target correlation parameter corresponding to the d-th dimensional hyper-parameter, represents the fitness of the worst parameter combination, represents the fitness of the optimal parameter combination; represents the d-th dimensional hyper-parameter of the optimal parameter combination, represents the Euclidean distance between the worst parameter combination and the optimal parameter combination; According to the basic associated parameter and the target associated parameter, the optimal parameter combination and the worst parameter combination after the associated control search are obtained as: ； ； wherein, denotes the d-th hyperparameter of the best parameter combination after the associated control search, denotes the d-th hyperparameter of the worst parameter combination after the associated control search, denotes a first random number between (0, 1), denotes a second random number between (0, 1). 6.The deep learning based brain MRI hippocampus three-dimensional segmentation method according to claim 1, characterized in that, The encrypted sub-target public parameter combination is aggregated by the server node to obtain the target public parameter combination, including: the encrypted sub-target public parameter combination is aggregated by the server node using the FedAvg algorithm to obtain the target public parameter combination.

7. The deep learning-based brain MRI hippocampus three-dimensional segmentation method according to claim 1, wherein, Based on the three-dimensional segmentation model and the target public parameter combination, hippocampus segmentation is performed on the three-dimensional brain MRI image to obtain a hippocampus segmentation result, including: According to the target public parameter combination, the three-dimensional segmentation model is deployed on the server node, and the user node is allowed to schedule the three-dimensional segmentation model deployed on the server node to perform hippocampus segmentation on the three-dimensional brain MRI image to obtain a hippocampus segmentation result. 8.The deep learning based brain MRI hippocampus three-dimensional segmentation method according to claim 1, wherein, Based on the three-dimensional segmentation model and the target public parameter combination, hippocampus segmentation is performed on the three-dimensional brain MRI image to obtain a hippocampus segmentation result, including: The target public parameter combination and the three-dimensional segmentation model are disclosed, so that the user node obtains the target public parameter combination and the three-dimensional segmentation model, and deploys the three-dimensional segmentation model on the user node according to the target public parameter combination, so that the user node uses the three-dimensional segmentation model deployed locally to perform hippocampus segmentation on the three-dimensional brain MRI image to obtain a hippocampus segmentation result.

Images