Medical image segmentation model construction method, system, device, medium and product

Abstract

The invention relates to the technical field of medical image segmentation, and discloses a medical image segmentation model construction method, system and device, a medium and a product, and the method comprises the steps: obtaining the gradient information of a local medical image segmentation model which is trained by each client node according to local medical image data; on the basis of gradient information, global model parameters are updated; the global model parameters are used for updating the local medical image segmentation model. By means of the federal learning framework, in the cross-institution data cooperation process, the safety of patient data is strictly ensured, the worry of medical institutions in the aspect of data sharing is eliminated, data sharing and cooperation between the medical institutions are promoted, the medical image segmentation technology is powerfully promoted to be applied to wider medical scenes, and the medical image segmentation method has the advantages of being high in practicability and easy to popularize. The generalization ability of the medical image segmentation model is improved, and the accuracy of cross-mechanism segmentation of the medical image segmentation model is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical image segmentation technology, and in particular to a medical image segmentation model construction method, system, equipment, medium and product. Background Art

[0002] Currently, medical image segmentation methods based on deep learning have become a research hotspot. They can automatically learn image feature patterns and accurately segment target areas. However, the training of deep learning models requires a large amount of training data. Because medical institutions need to protect patient privacy, data sharing is difficult. As a result, deep learning models are not able to fully learn the diverse features of data from different medical institutions, resulting in limited generalization capabilities. When a model trained on data from one medical institution is applied to data from other medical institutions, the distribution of data features changes due to factors such as equipment differences and different scanning parameters, and the model performance decreases significantly. Summary of the Invention

[0003] In view of this, the present invention provides a method, system, device, medium and product for constructing a medical image segmentation model to solve the problem that the medical image segmentation model constructed in the prior art is difficult to achieve accurate segmentation across institutions.

[0004] In a first aspect, the present invention provides a method for constructing a medical image segmentation model, applicable to a central server, the method comprising:

[0005] Obtaining gradient information of the local medical image segmentation model, where the local medical image segmentation model is trained by each client node based on local medical image data;

[0006] Based on the gradient information, the global model parameters are updated; the global model parameters are used to update the local medical image segmentation model.

[0007] With the help of the federated learning framework, this invention not only strictly ensures the security of patient data in the process of cross-institutional data collaboration, but also eliminates the concerns of medical institutions in data sharing, promotes data sharing and cooperation among medical institutions, and effectively promotes the application of medical image segmentation technology in a wider range of medical scenarios, improves the generalization ability of medical image segmentation models, and achieves the accuracy of cross-institutional segmentation of medical image segmentation models.

[0008] In an optional embodiment, updating global model parameters based on gradient information includes:

[0009] Calculate the data distribution characteristics of each client node;

[0010] Based on the adaptive weight allocation strategy, the weight of the data distribution characteristics of each client node in the parameter aggregation is determined;

[0011] Update the global model parameters based on the weights and gradient information of each client node.

[0012] In this implementation, an adaptive weight allocation strategy is adopted to dynamically adjust the aggregation weight according to the node data distribution, which shows great advantages in actual cross-institutional medical image segmentation tasks, greatly shortens the model training time, and significantly reduces computing resource consumption, enabling the model to be applied to clinical practice more quickly, providing doctors with accurate image segmentation results in a timely manner, and improving the timeliness and efficiency of medical services.

[0013] In an optional implementation, the gradient information is obtained after being processed by a gradient obfuscation algorithm.

[0014] In this implementation, with the help of the gradient obfuscation algorithm and the federated learning framework, effective protection of data privacy can be achieved in the cross-institutional data collaboration process.

[0015] In an optional embodiment, an encryption protocol is used to send the updated global model parameters to each client node.

[0016] In this embodiment, the central server distributes the updated global model parameters to each medical institution node. The distribution process uses secure network connection and data encryption technology to ensure the secure transmission of model parameters.

[0017] In a second aspect, the present invention provides a method for constructing a medical image segmentation model, applicable to a client node, the method comprising:

[0018] A dynamic noise scheduling mechanism is used to generate synthetic image data based on local medical image data and a pre-built spatiotemporal conditional diffusion network.

[0019] Based on local medical image data and synthetic image data, an initial deep learning model is trained to obtain a local medical image segmentation model;

[0020] Calculate the gradient information of the local medical image segmentation model;

[0021] The gradient information is processed by the gradient obfuscation algorithm and sent to the central server;

[0022] After the central server updates the global model parameters based on the gradient information, receiving the global model parameters;

[0023] Update local medical image segmentation models based on global model parameters.

[0024] This paper utilizes a medical image segmentation framework that integrates diffusion models with federated learning. This architecture combines the diffusion model's superior ability to generate high-quality data with the powerful distributed collaboration of federated learning. By leveraging spatiotemporal conditional diffusion networks to generate synthetic images, this approach enables cross-institutional data collaboration while maintaining data privacy, creating an unprecedented solution for medical image segmentation.

[0025] In an optional embodiment, the spatiotemporal conditional diffusion network includes:

[0026] The input layer is used to receive local medical imaging data and perform data preprocessing on the local medical imaging data;

[0027] The convolutional layer is connected to the input layer and is used to extract multi-scale features from the pre-processed local medical imaging data;

[0028] The recurrent layer, connected to the convolutional layer, is used to capture dynamic features in the temporal and spatial dimensions;

[0029] The attention mechanism layer is connected to the recurrent layer to calculate the attention weight of each position;

[0030] The output layer, connected to the attention mechanism layer, is used to output synthetic image data.

[0031] In this implementation, a unique spatiotemporal conditional diffusion network architecture, comprised of convolutional, recurrent, attention, and deconvolutional layers, can deeply mine the spatiotemporal characteristics of medical images to generate high-quality synthetic images. Its dynamic noise scheduling mechanism flexibly adjusts noise intensity based on the number of iterations, helping the network accurately learn the complex mapping relationship from noisy images to real images.

[0032] In a third aspect, the present invention provides a medical image segmentation model construction system, the system comprising:

[0033] The client node processing module is used to generate synthetic image data based on local medical image data and a pre-built spatiotemporal conditional diffusion network; train an initial deep learning model based on the local medical image data and the synthetic image data to obtain a local medical image segmentation model; calculate the gradient information of the local medical image segmentation model; and send the gradient information to the central server after processing it using a gradient obfuscation algorithm;

[0034] The central server processing module is used to obtain gradient information of the local medical image segmentation model; based on the gradient information, the global model parameters are updated; and the global model parameters are used to update the local medical image segmentation model.

[0035] This invention integrates diffusion models and federated learning technology to achieve accurate dynamic segmentation of medical images across medical institutions. It can output accurate image segmentation results for disease diagnosis, helping doctors to more accurately judge the condition and formulate treatment plans, which is of great significance to improving the efficiency and quality of medical diagnosis.

[0036] In a fourth aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to execute the medical image segmentation model construction method of the first aspect or any corresponding embodiment thereof.

[0037] In a fifth aspect, the present invention provides a computer-readable storage medium storing computer instructions, which are used to enable a computer to execute the medical image segmentation model construction method of the above-mentioned first aspect or any corresponding embodiment thereof.

[0038] In a sixth aspect, the present invention provides a computer program product comprising computer instructions for enabling a computer to execute the method for constructing a medical image segmentation model according to the first aspect or any corresponding embodiment thereof.

[0039] It should be noted that the medical image segmentation model construction system, computer device, computer-readable storage medium, and computer program product provided herein correspond to the aforementioned medical image segmentation model construction method. Therefore, for the beneficial effects of the medical image segmentation model construction system, computer device, computer-readable storage medium, and computer program product, please refer to the description of the corresponding beneficial effects of the medical image segmentation model construction method above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 is a flowchart of a method for constructing a medical image segmentation model according to an embodiment of the present invention;

[0042] Figure 2 is a flowchart of another method for constructing a medical image segmentation model according to an embodiment of the present invention;

[0043] Figure 3 is a structural block diagram of a medical image segmentation model construction system according to an embodiment of the present invention;

[0044] Figure 4 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0046] Medical data involves patient privacy. Due to data security and privacy considerations, it is difficult for medical institutions to achieve data sharing. This seriously restricts the training effect and generalization ability of deep learning models, making it difficult for the models to maintain good performance on diverse data from different medical institutions.

[0047] In view of this, according to an embodiment of the present invention, an embodiment of a method for constructing a medical image segmentation model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0048] In this embodiment, a method for constructing a medical image segmentation model is provided, which can be executed by a central server, a terminal, a mobile terminal, and other devices. Figure 1 FIG. 1 is a flow chart of a method for constructing a medical image segmentation model according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0049] Step S101 , obtaining gradient information of a local medical image segmentation model, where the local medical image segmentation model is trained by each client node based on local medical image data.

[0050] Step S102: Based on the gradient information, update the global model parameters; the global model parameters are used to update the local medical image segmentation model. The global model is the initial deep learning model generated at the beginning of federated learning training.

[0051] The specific process for building a medical image segmentation model is as follows: Each client node selects medical image data for training from the local data storage module, determining the selection range based on the training task and requirements. For example, for the tumor segmentation task, image data containing tumor cases is selected. The selected data is then input into the spatiotemporal conditional diffusion network (ST-DiffNet) to generate synthetic images. Before data input, the original images undergo necessary preprocessing, such as normalization and denoising, to improve data quality and ensure that ST-DiffNet generates high-quality synthetic images.

[0052] The client node then uses the local original medical image data and the generated synthetic images to train the local segmentation model in the model training module. During the training process, a suitable loss function is selected, such as the Dice loss function commonly used in segmentation tasks, to measure the degree of overlap between the predicted results and the true annotations. Stochastic gradient descent (SGD) or its variants, such as the Adam algorithm, is used for optimization. Appropriate hyperparameters such as the learning rate and number of iterations are set. For example, the initial value of the learning rate is set to 0.001, and as the training progresses, a learning rate decay strategy is adopted, reducing the learning rate to 0.9 times the original value every 10 rounds of training. In each round of training, the data is input into the model in batches, and the loss value between the predicted results and the true annotations is calculated. The model parameters are adjusted through the backpropagation algorithm to enable the model to accurately segment the target area in the image.

[0053] After local model training is complete, the client node calculates the local model's gradient information and obfuscates the gradient using a gradient obfuscation algorithm to generate an obfuscated gradient vector. This obfuscated gradient information is then uploaded to the central server via a secure network connection. Data encryption and compression techniques are used during the upload process to ensure the security and transmission efficiency of the gradient information.

[0054] The central server receives the obfuscated gradient information uploaded by the nodes of each medical institution, calculates the weight of each node according to the adaptive weight allocation strategy (FL-AdaW), takes the weighted sum of the obfuscated gradient vectors, and updates the global model parameters. The aggregation process takes into account issues such as network latency and node failure, and adopts a fault-tolerant mechanism, such as setting a retry count and retrying a certain number of nodes that fail to successfully upload gradient information to ensure the accuracy and completeness of parameter aggregation. Finally, the central server distributes the updated global model parameters to each client node. The distribution process also uses secure network connections and data encryption technology to ensure the secure transmission of model parameters. After receiving the global model parameters, each medical institution node updates its local copy of the global model.

[0055] In actual use, client nodes use the updated global model to segment new medical images, and the model outputs segmentation results. Doctors use these segmentation results to assist in disease diagnosis. For example, in tumor diagnosis, segmentation results can accurately determine the tumor's location, size, and shape, providing important evidence for developing treatment plans.

[0056] Experimental data shows that compared to traditional single-modality segmentation methods or those that do not employ the present invention's technology, segmentation accuracy, as measured by the Dice coefficient, has increased significantly from 0.82 to 0.92, a 12% improvement. This significant improvement enables doctors to more accurately determine the location, size, and shape of lesions, providing a highly reliable basis for disease diagnosis, helping to develop more targeted and accurate treatment plans, and effectively improving cure rates.

[0057] In this embodiment, through the federated learning framework, diverse data from multiple medical institutions can be integrated for model training. There are many differences in patient groups, equipment types, scanning parameters, etc. in different medical institutions, resulting in a high diversity of data. The model can learn the rich features of data from these different sources, thereby significantly enhancing the generalization ability. When the trained model is applied to new medical institution data, it can better adapt to changes in data and maintain a high level of segmentation accuracy. Compared with the traditional model, which has a 40% performance drop when applied across institutions, the model of the present invention can control the drop in segmentation accuracy within 10% when applied across institutions. This effectively solves the problem of insufficient generalization of traditional models, improves the practicality of medical image segmentation models in different medical environments, and enables the model to provide reliable support for medical diagnosis in a wider range.

[0058] In this embodiment, with the help of the federated learning framework, in the process of cross-institutional data collaboration, not only the security of patient data is strictly ensured, but also the concerns of medical institutions in data sharing are eliminated, data sharing and cooperation between medical institutions are promoted, and the application of medical image segmentation technology in a wider range of medical scenarios is strongly promoted. The generalization ability of the medical image segmentation model is improved, and the accuracy of cross-institutional segmentation of the medical image segmentation model is achieved.

[0059] In some optional implementations, updating global model parameters based on gradient information includes:

[0060] Calculate the data distribution characteristics of each client node.

[0061] Based on the adaptive weight allocation strategy, the weight of the data distribution characteristics of each client node in parameter aggregation is determined.

[0062] Update the global model parameters based on the weights and gradient information of each client node.

[0063] In response to the heterogeneity problem of data distribution of different client nodes in federated learning, an adaptive weight allocation strategy (FL-AdaW) is adopted in this embodiment. Each time the parameters are aggregated, the central server first calculates the distribution characteristics of the data of each node. For the distribution of data features, taking CT image data as an example, each pixel value in the image is regarded as a data point, and the mean and variance of all pixel values ​​are calculated. The mean reflects the average grayscale level of the data, and the variance reflects the degree of discreteness of the data. For the distribution of data categories, in the segmentation task, if the goal is to segment tumors, normal tissues and background, the proportion of each category in the node data is counted.

[0064] Then, based on these distribution characteristics, the weight of each node in parameter aggregation is determined using an adaptive weight calculation function. Assuming the data distribution characteristics of node i are fi and the adaptive weight calculation function is W(fi), the weight of node i in parameter aggregation is wi = W(fi). For example, nodes with data distributions that are more similar to the global data distribution are given a higher weight; nodes with significantly different data distributions are given a lower weight. In actual calculations, the adaptive weight calculation function can use methods based on distance metrics, such as calculating the Euclidean distance or Mahalanobis distance between the node data distribution characteristics and the global data distribution characteristics. The smaller the distance, the more similar the node data distribution is to the global distribution, and the higher the weight; the larger the distance, the lower the weight.

[0065] Through this adaptive weight allocation strategy, the model can better adapt to the data distribution of different nodes during the aggregation process, improving convergence speed and performance. Experimental tests show that compared with the traditional fixed weight allocation strategy, the model convergence speed is increased by 3 times after adopting FL-AdaW. In a brain MRI image segmentation task involving 5 client nodes, the traditional fixed weight allocation strategy may require 100 rounds of training for the model to converge to a certain accuracy. After adopting FL-AdaW, only about 33 rounds of training are required to achieve the same accuracy, greatly improving the model training efficiency and enabling the model to be applied to actual medical image segmentation tasks more quickly.

[0066] The central server then distributes the updated global model parameters to each client node. This distribution process also uses secure network connections and data encryption technology to ensure the secure transmission of model parameters. After receiving the global model parameters, each client node updates its local copy of the global model.

[0067] There are often significant differences in the data distribution of different medical institutions. The traditional fixed weight allocation strategy is difficult to fully adapt to this heterogeneity, while the adaptive weight allocation strategy can calculate the distribution characteristics of each node data in real time and accurately. The adaptive weight allocation strategy can calculate the distribution characteristics of each node data including key information such as mean, variance, and data category distribution, and dynamically allocate weights based on these characteristics. In a chest CT image segmentation task involving multiple medical institution nodes, the traditional strategy may require a large number of rounds of training before the model converges, while after adopting the adaptive weight allocation strategy, the model convergence speed is increased by 3 times. In this embodiment, an adaptive weight allocation strategy is adopted to dynamically adjust the aggregation weight according to the node data distribution, which shows great advantages in the actual cross-institutional medical image segmentation task, greatly shortens the model training time, significantly reduces the consumption of computing resources, enables the model to be applied to clinical practice more quickly, and provides doctors with accurate image segmentation results in a timely manner, thereby improving the timeliness and efficiency of medical services.

[0068] In some optional implementations, the gradient information is obtained after being processed by a gradient obfuscation algorithm.

[0069] Before uploading the gradient information to the central server, the gradient is obfuscated. Specifically, a random noise vector with the same dimension as the gradient vector is generated, and the elements of the noise vector obey a specific distribution, such as a Gaussian distribution. The mean of the Gaussian distribution is set to 0, and the variance is adjusted according to the pre-set noise intensity parameter α, which is α2. For example, when α = 0.01, the elements of the noise vector obey a Gaussian distribution with a mean of 0 and a variance of 0.0001. Then, according to the pre-set noise intensity parameter α, the noise vector and the gradient vector are weighted and summed. Assuming that the gradient vector is g and the noise vector is n, the obfuscated gradient vector g' = g + α × n. When the central server performs parameter aggregation, it uses the obfuscated gradient vector for calculation. In this way, even if an attacker obtains the gradient information, it is difficult to restore the original data through model inversion due to noise interference.

[0070] During the parameter aggregation phase of federated learning, the gradient obfuscation algorithm significantly interferes with gradient information by adding directional noise. Even if an attacker were to obtain this obfuscated gradient information, it would be extremely difficult to recover the original data. IEEETMI 2024 verified that the implementation of this algorithm reduced the risk of privacy leakage by up to 78%. In practical applications, medical institution nodes do not need to share original medical imaging data; they only upload obfuscated gradient information. This not only strictly ensures the security of patient data and fully complies with strict medical data privacy regulations, but also eliminates concerns about data sharing among medical institutions, promotes data sharing and collaboration among them, and strongly promotes the application of medical image segmentation technology in a wider range of medical scenarios, opening up a new, secure and reliable path for data collaboration in the medical industry.

[0071] In this embodiment, with the help of the gradient obfuscation algorithm and the federated learning framework, effective protection of data privacy can be achieved in the cross-institutional data collaboration process.

[0072] In some optional implementations, an encrypted protocol is used to send the updated global model parameters to each client node.

[0073] During the training process of traditional deep learning models, all data needs to be concentrated in a central node for training. This not only increases the cost of data transmission, especially for large amounts of medical imaging data, which takes a long time to transmit and occupies a large amount of network bandwidth, but also brings serious risks of data privacy leakage. Once the central node is attacked, all data is at risk of leakage.

[0074] In this embodiment, the central server distributes the updated global model parameters to each medical institution node. The distribution process uses secure network connection and data encryption technology to ensure the secure transmission of the model parameters.

[0075] In this embodiment, a method for constructing a medical image segmentation model is provided, which can be executed by client nodes, terminals, mobile terminals and other devices of medical institutions, etc. Figure 2 FIG. 1 is a flow chart of a method for constructing a medical image segmentation model according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0076] Step S201 : Using a dynamic noise scheduling mechanism, synthetic image data is generated based on local medical image data and a pre-built spatiotemporal conditional diffusion network.

[0077] Step S202: Based on the local medical image data and the synthetic image data, an initial deep learning model is trained to obtain a local medical image segmentation model.

[0078] Step S203: Calculate the gradient information of the local medical image segmentation model.

[0079] Step S204: The gradient information is processed by the gradient obfuscation algorithm and then sent to the central server.

[0080] Step S205 , receiving the global model parameters after the central server updates the global model parameters based on the gradient information;

[0081] Step S206: updating the local medical image segmentation model based on the global model parameters.

[0082] Specifically, each client node is equipped with a local medical image data storage module, utilizing the advanced Ceph distributed file system. This system leverages the storage space of multiple servers, ensuring data reliability through redundant storage and data verification mechanisms such as erasure coding. Even if some nodes fail, the original data can be restored based on data from other nodes. Data is stored by patient ID, examination time, imaging modality, and other categories, facilitating rapid retrieval and improving data management efficiency.

[0083] The spatiotemporal conditional diffusion network (ST-DiffNet) runs within the client node and is responsible for generating high-quality synthetic images. Its network structure can fully exploit the spatiotemporal characteristics of medical images. During the model training phase, local segmentation model training is carried out with the help of locally stored original medical image data and the synthetic images generated by ST-DiffNet. Efficient training algorithms such as stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, RMSProp, and Adam can be used, and they can be flexibly selected according to different data characteristics and model requirements to accelerate the training process and improve model convergence. During the training process, by calculating the loss function value between the predicted results and the true annotations, the backpropagation algorithm is used to continuously adjust the model parameters, so that the model can accurately learn the characteristics of the target area in the medical image.

[0084] To generate high-quality synthetic images, ST-DiffNet also uses a dynamic noise scheduling mechanism. In the diffusion model, the addition and removal of noise are core steps. In the initial stage, high-intensity noise is added to the input image. As the network iterates, the noise intensity gradually decreases. Specifically, a carefully designed noise scheduling function β(t) is used to control the change in noise intensity, where t represents the number of iterations. Initially, β(0) is set to a large value, such as 0.1. At this time, the image is almost completely submerged in noise, making it difficult to discern the original structure. As the number of iterations increases, β(t) gradually decreases, such as β(T) = 0.001, where T is the total number of iterations. In each iteration, Gaussian noise is added to the image according to the value of β(t). The mean of the Gaussian noise is 0, and the variance is adjusted according to β(t), for example, the variance is β(t). In the process of generating synthetic images, the network gradually removes noise and restores a clear image by continuously learning the mapping relationship from noisy images to real images.

[0085] In practice, for the input lung CT image, in the first iteration, according to β(0) = 0.1, a Gaussian noise matrix with a mean of 0 and a variance of 0.1 is generated, which is the same size as the CT image. This noise matrix is ​​added to the CT image pixel by pixel to obtain the initial noise image. As the iteration proceeds, β(t) gradually decreases. For example, at the 10th step, β(10) = 0.05. At this time, the variance of the generated Gaussian noise becomes 0.05, the noise intensity decreases, and the original structure in the image gradually emerges. By continuously adjusting the noise intensity, the network can learn the complex mapping relationship from the noise image to the real image, thereby generating high-quality synthetic images. Experimental results show that the synthetic images generated by this dynamic noise scheduling mechanism have a peak signal-to-noise ratio (PSNR) ≥ 42dB, which can effectively replace the sharing of original data in federated learning, protecting data privacy and providing high-quality data for model training.

[0086] During the federated learning parameter aggregation phase, this invention also employs an innovative gradient obfuscation algorithm to defend against model inversion attacks. After the client node completes local model training, it calculates the local model's gradient information. This gradient information reflects the direction and magnitude of changes in model parameters on local data, and its dimension corresponds to the number of model parameters. For example, for a model with one million parameters, the calculated gradient vector dimension is also one million.

[0087] Before uploading the gradient information to the central server, the gradient is obfuscated. Specifically, a random noise vector with the same dimension as the gradient vector is generated, and the elements of the noise vector obey a specific distribution, such as a Gaussian distribution. The mean of the Gaussian distribution is set to 0, and the variance is adjusted according to the pre-set noise intensity parameter α, which is α2. For example, when α = 0.01, the elements of the noise vector obey a Gaussian distribution with a mean of 0 and a variance of 0.0001. Then, according to the pre-set noise intensity parameter α, the noise vector and the gradient vector are weighted and summed. Assuming that the gradient vector is g and the noise vector is n, the obfuscated gradient vector g' = g + α × n. When the central server performs parameter aggregation, it uses the obfuscated gradient vector for calculation. In this way, even if an attacker obtains the gradient information, it is difficult to restore the original data through model inversion due to noise interference.

[0088] In practical applications, for the gradient vector g calculated by a client node, the generated noise vector n has the same dimension as g. Each element of n is randomly generated according to a Gaussian distribution (mean 0, variance α2). Then g and α×n are added element by element to obtain the obfuscated gradient vector g'. After the central server receives the obfuscated gradient vectors uploaded by multiple client nodes, it will determine the weight of each node according to the adaptive weight allocation strategy (FL-AdaW) when performing parameter aggregation, sum the weighted obfuscated gradient vectors, and then update the global model parameters. Experimental verification (IEEETMI2024 verification) shows that the risk of privacy leakage is reduced by 78% after adopting this gradient obfuscation algorithm, effectively ensuring data security in the federated learning process.

[0089] The gradient information is processed by the gradient obfuscation algorithm and sent to the central server. The central server plays a core coordination role in the entire system. The parameter aggregation module is responsible for receiving the model gradient information after gradient obfuscation processing uploaded by the nodes of each medical institution. During the aggregation process, according to the adaptive weight distribution strategy (FL-AdaW), the gradient information of different nodes is weighted and summed to generate updated global model parameters. The global model storage module adopts a high-performance database system, such as MongoDB or Redis, and combines it with efficient data indexing and storage structure to achieve fast storage and reading of model parameters, ensuring the efficiency of model updates and distribution. At the same time, the central server has strong computing power and stable network connection to cope with the processing and transmission needs of large amounts of data.

[0090] Multiple client nodes connect to a central server via a secure and stable network. This network connection utilizes SSL / TLS encryption to prevent data theft or tampering during transmission. Data compression techniques, such as the gzip algorithm, can also be used during data transmission to reduce data volume, improve transmission efficiency, and reduce network bandwidth consumption.

[0091] Client nodes use the updated global model to segment new medical images. During application, the input medical image undergoes the same preprocessing as during training, then is fed into the global model, which outputs the segmentation results. Doctors use these segmentation results to assist in disease diagnosis. For example, in tumor diagnosis, segmentation results can accurately determine the tumor's location, size, and shape, providing important evidence for treatment planning.

[0092] This paper utilizes a medical image segmentation framework that integrates diffusion models with federated learning. This architecture combines the diffusion model's superior ability to generate high-quality data with the powerful distributed collaboration of federated learning. By leveraging spatiotemporal conditional diffusion networks to generate synthetic images, this approach enables cross-institutional data collaboration while maintaining data privacy, creating an unprecedented solution for medical image segmentation.

[0093] In some optional embodiments, the spatiotemporal conditional diffusion network includes:

[0094] The input layer is used to receive local medical imaging data and perform data preprocessing on the local medical imaging data;

[0095] The convolutional layer is connected to the input layer and is used to extract multi-scale features from the pre-processed local medical imaging data;

[0096] The recurrent layer, connected to the convolutional layer, is used to capture dynamic features in the temporal and spatial dimensions;

[0097] The attention mechanism layer is connected to the recurrent layer to calculate the attention weight of each position;

[0098] The output layer, connected to the attention mechanism layer, is used to output synthetic image data.

[0099] Specifically, the input layer of the spatiotemporal conditional diffusion network can accept a variety of medical imaging data, such as common DICOM formats. For different modal images, such as CT, MRI, and PET, the input layer will perform preprocessing based on the image characteristics. CT images are grayscale normalized to unify the grayscale value range to [0, 1] to eliminate the impact of differences in scanning parameters between different devices. For MRI images, image registration may be performed to ensure that different sequence images are spatially aligned. The dimension and number of channels of the input data vary depending on the image type. For example, the common input dimension of a two-dimensional CT image is (height, width), and the number of channels may be 1 (grayscale image) or 3 (color image, such as an image that integrates multiple information); the input dimension of a three-dimensional MRI image is (height, width, depth), and the number of channels also depends on the specific situation.

[0100] The input image then enters a series of convolutional layers. These layers can use kernels of varying sizes, such as 3×3 and 5×5, to capture features at different scales. The kernel stride is adjusted as needed. When extracting low-level detail features, the stride is set to 1 to fully preserve image information; when performing downsampling, the stride is set to 2 to reduce image resolution. The number of kernels in each convolutional layer gradually increases from an initial 32 to 512 or even more. As the network deepens, more abstract and advanced features can be extracted. After processing through the convolutional layers, the size of the image feature map gradually decreases. For example, the input size of (256, 256) may become (32, 32) after several convolutional and pooling layers. However, the number of feature map channels continues to increase, from an initial 1 or 3 channels to hundreds of channels. This rich channel information covers a wide range of image features.

[0101] Next, the feature map processed by the convolutional layer enters the recurrent layer. In the present invention, the recurrent layer can use a long short-term memory network (LSTM) or a gated recurrent unit (GRU). Taking LSTM as an example, its input is the feature map sequence output by the convolutional layer, and the input dimension of each time step is the same as the number of feature map channels. LSTM can effectively capture the dynamic characteristics of the image in the time and space dimensions. When processing dynamic cardiac MRI images, the morphological change information of the heart at different times is input into the LSTM in sequence through time steps. The forget gate, input gate and output gate inside the LSTM work together to determine the retention, update and output of information. The forget gate controls the proportion of memory information retained at the previous moment to the current moment, the input gate determines the amount of new information added to the memory in the current input feature map, and the output gate generates output according to the current memory state. Through this mechanism, LSTM can learn the morphological change patterns of the heart at different times, such as the changing characteristics of myocardial thickness and ventricular volume during cardiac contraction and relaxation.

[0102] After the recurrent layer, the attention mechanism layer is introduced. The attention mechanism layer takes the feature vector output by the recurrent layer as input and calculates the attention weight for each position through a series of linear transformations and activation functions. For example, the calculated attention score is normalized using the Softmax function to obtain a weight value for each position. The weight value ranges from [0 to 1], and the sum of the weight values ​​for all positions is 1. In the task of segmenting lung tumors, the attention mechanism enables the network to focus more on the characteristics of the tumor region. When calculating the attention weight, the feature vector considers the shape, texture, grayscale value, and other features related to the tumor. Regions with obvious features are given higher weights, while irrelevant background regions are given lower weights. The attention weight is then weighted and summed with the feature vector to highlight the features of important regions, suppress background noise, and improve segmentation accuracy.

[0103] Finally, after a series of deconvolution layers and upsampling operations, the low-resolution feature map is restored to the original image resolution, and a synthetic image is output. The deconvolution layer uses a transposed convolution kernel whose size and stride correspond to those of the previous convolution layer to achieve feature map upsampling. For example, the previous convolution layer used a 3×3 convolution kernel with a stride of 2 for downsampling, while the deconvolution layer might use a 3×3 transposed convolution kernel with a stride of 2 for upsampling. The upsampling operation can also be combined with methods such as nearest neighbor interpolation and bilinear interpolation to further refine image details. After deconvolution and upsampling operations, the output synthetic image size and format are consistent with the input image. For example, the output two-dimensional synthetic image size is (256, 256) and the format is DICOM. The synthetic image can also retain the key features of the original image, remove noise and artifacts, and generate high-quality images for subsequent model training.

[0104] In this embodiment, a unique spatiotemporal conditional diffusion network structure, comprising convolutional layers, recurrent layers, attention mechanisms, and deconvolutional layers, is constructed to deeply mine the spatiotemporal characteristics of medical images and generate high-quality synthetic images. Its dynamic noise scheduling mechanism flexibly and dynamically adjusts noise intensity based on the number of iterations, helping the network accurately learn the complex mapping relationship from noisy images to real images. The resulting synthetic images have a PSNR of ≥42dB and can effectively replace original data for sharing.

[0105] In this embodiment, a medical image segmentation model construction system is also provided. The system is used to implement the above-mentioned embodiments and preferred embodiments. The details that have been described will not be repeated here. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, implementation using hardware, or a combination of software and hardware, is also possible and contemplated.

[0106] This embodiment provides a medical image segmentation model construction system. Figure 3 As shown, the system includes:

[0107] The client node processing module is used to generate synthetic image data based on local medical image data and a pre-built spatiotemporal conditional diffusion network; train an initial deep learning model based on the local medical image data and the synthetic image data to obtain a local medical image segmentation model; calculate the gradient information of the local medical image segmentation model; and send the gradient information to the central server after processing it using a gradient obfuscation algorithm;

[0108] The central server processing module 301 is used to obtain gradient information of the local medical image segmentation model; based on the gradient information, update the global model parameters; the global model parameters are used to update the local medical image segmentation model.

[0109] The central server processing module 302 is also used to calculate the data distribution characteristics of each client node; determine the weight of the data distribution characteristics of each client node in parameter aggregation based on the adaptive weight allocation strategy; and update the global model parameters based on the weight of each client node and the gradient information.

[0110] Specifically, a local data storage environment is established within each medical institution, namely on client nodes. High-capacity enterprise-class hard disk arrays are used, paired with the Ceph distributed file system for data storage. Ceph utilizes erasure coding technology to partition data and store it across different server nodes. Redundant storage and data validation ensure data reliability, enabling recovery of the original data even in the event of node failure. Data is stored by patient ID, examination time, and imaging modality. For example, CT images are stored in a directory structure of "patient ID / examination time / CT / specific image file" for easy and fast retrieval. When installing and configuring the spatiotemporal conditional diffusion network module developed using a deep learning framework (such as TensorFlow or PyTorch) on each client node, the number of GPUs and memory are innovatively allocated based on the hardware resources of the medical institution. For example, if four NVIDIA A100 GPUs are available, three can be allocated to ST-DiffNet for synthetic image generation, while one can be used for the model training module. Furthermore, innovative GPU optimization settings are implemented, such as adjusting the video memory allocation strategy and enabling GPU acceleration, to improve model efficiency. The model training module is also built based on the selected deep learning framework, and installs related dependent libraries such as the Scikit-image library for calculating loss functions and the Optuna library for optimization algorithms.

[0111] The central server needs to have strong computing and storage capabilities. It can use high-performance server hardware equipped with multi-core CPUs (such as Intel Xeon Platinum series), large-capacity memory (64GB or more) and high-speed storage devices (such as NVMe SSD). Install the CentOS or Ubuntu Server version of Linux operating system and deploy a parameter aggregation module developed based on a distributed computing framework (such as Apache Spark). Utilize the advantages of Apache Spark in quickly processing large-scale data in a cluster environment, and innovatively improve parameter aggregation efficiency by optimizing data partitioning and parallel computing strategies. The global model storage module can flexibly select relational databases (such as MySQL) or non-relational databases (such as MongoDB) according to the characteristics of the model and the amount of data. At the same time, innovatively configure secure network connections, use network security devices such as firewalls and intrusion detection systems, set firewall rules, and only allow specific medical institution node IP addresses to access relevant ports of the central server to prevent external illegal access.

[0112] Data preparation then begins. Client nodes collect medical imaging data, such as CT, MRI, and PET, from routine medical examinations. Medical ethics standards are strictly adhered to, patient privacy is fully protected, and informed consent is obtained from patients during data collection, clearly informing them of the purpose, method, and scope of data use. A professional medical image annotation team is organized to annotate the data, employing a multi-person cross-annotation and review mechanism to enhance annotation accuracy. For example, for brain MRI tumor segmentation annotation, three experienced radiologists independently annotate the data, which is then reviewed by a senior expert. Discrepancies are discussed and corrected to ensure reliable annotation. An annotation database is established, recording detailed information such as the annotating physician, time of day, and review results to facilitate quality control and traceability. Data preprocessing is performed before input into the spatiotemporal conditional diffusion network. For CT images, a unique grayscale normalization formula, \(I_{norm} = \frac{I - I_{min}}{I_{max} - I_{min}}\), is used to normalize the grayscale value range to [0, 1]. For MRI images, image registration is performed using the SIFT algorithm based on feature point matching, ensuring spatial alignment of images from different sequences. At the same time, all images are denoised using methods such as Gaussian filtering or median filtering, and the filtering parameters, such as the Gaussian filter standard deviation, are flexibly adjusted according to the characteristics of image noise to achieve the best denoising effect.

[0113] Subsequently, using a large amount of medical imaging data as a training set, an adversarial training approach was employed to enable the spatiotemporal conditional diffusion network to learn the mapping relationship from noisy images to real images. During training, innovative hyperparameter settings were used, such as an initial learning rate of 0.0002 and the Adam optimizer with β1 = 0.9 and β2 = 0.999. The noise scheduling function β(t) parameters were continuously adjusted, and based on experimental results, appropriate decay methods such as exponential decay were innovatively selected. After multiple rounds of training (e.g., 500 rounds), the spatiotemporal conditional diffusion network was able to stably generate high-quality synthetic images. Model parameters were periodically saved during training to facilitate resumption of training if interrupted. Simultaneously, the model was evaluated using a validation set, monitoring performance metrics such as PSNR and the structural similarity index (SSIM). Hyperparameters were adjusted based on the evaluation results to ensure effective model training. The local segmentation model was then trained using the synthetic images generated by the spatiotemporal conditional diffusion network and local raw medical imaging data. We innovatively selected a suitable segmentation model architecture, such as an improved U-Net, and adopted the Dice loss function as the optimization objective, with the formula \(DiceLoss = 1-\frac{2|A\capB|}{|A|+|B|}\). Stochastic gradient descent (SGD) was used to update parameters, with a learning rate of 0.01 and a momentum of 0.9. The data was divided into small batches (e.g., 32 images per batch) for training. Model performance was evaluated on the validation set after every 10 training rounds. Hyperparameters were adjusted based on the results. After approximately 200 rounds of training, the local segmentation model achieved good performance. During training, the model training process can be visualized, with loss function and accuracy curves plotted. By analyzing the changing trends of these curves, problems such as overfitting and underfitting can be promptly identified, and appropriate adjustments can be made.

[0114] After local model training is complete, the client node calculates the local model gradient information. Before uploading the gradient, a random noise vector with the same dimension as the gradient vector and elements following a Gaussian distribution \(N(0,\alpha^2)\)(α=0.01) is innovatively generated based on the gradient obfuscation algorithm. This noise vector is weighted and summed with the gradient vector to obtain the obfuscated gradient vector, which is then uploaded to the central server via a secure network connection using the SSL / TLS encryption protocol. During upload, the gradient information is compressed using the gzip compression algorithm, with an appropriate compression level set to minimize data size and improve transmission efficiency while ensuring data integrity.

[0115] The central server receives the obfuscated gradient information uploaded by nodes from each medical institution and calculates the data distribution characteristics of each node based on an adaptive weight allocation strategy. For each node's uploaded image data, taking CT images as an example, statistical features such as the mean \(\overline{x}=\frac{1}{n}\sum_{i=1}^{n}x_{i}\) and variance \(\sigma^{2}=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\overline{x})^{2}\) are extracted. For segmentation tasks, the proportion of tumor, normal tissue, and background in the node data is calculated.

[0116] Then, the weight of each node in the parameter aggregation is determined by the adaptive weight calculation function. Assuming that the data distribution characteristics of node \(i\) are \(f_{i}\) and the adaptive weight calculation function is \(W(f_{i})\), the weight of node \(i\) is \(w_{i}=W(f_{i})\). For example, using the weight calculation method based on Euclidean distance, the Euclidean distance between the data distribution characteristics of each node and the global data distribution characteristics is first calculated, \(d_{i}=\sqrt{\sum_{j=1}^{m}(f_{ij}-\overline{f}_{j})^{2}}\). The smaller the distance, the higher the weight. The specific weight calculation formula is \(w_{i}=\frac{1 / d_{i}}{\sum_{k=1}^{N}1 / d_{k}}\) (\(N\) is the total number of nodes).

[0117] After obtaining the weight of each node, the central server weights and sums the obfuscated gradient vectors. Let the obfuscated gradient vector uploaded by node \(i\) be \(g_{i}'\), then the aggregated gradient vector \(G=\sum_{i=1}^{N}w_{i}g_{i}'\). Using the aggregated gradient vector \(G\), the global model parameters are updated according to the selected optimization algorithm (such as a variant of the stochastic gradient descent algorithm). For example, if the Adam optimization algorithm is used, according to its update rule, combined with the learning rate \(\eta\), the gradient vector \(G\) and the previous parameters Calculate updated parameters _{t-1}-\eta\frac{m_{t}}{\sqrt{v_{t}}+\epsilon}\)(\(m_{t}\) and \(v_{t}\) are the bias-corrected first-order moment estimate and second-order moment estimate, respectively, and \(\epsilon\) is a small constant set for numerical stability.

[0118] After updating the global model parameters, the central server distributes the updated global model parameters to each medical institution node via SSL / TLS encryption. Upon receiving the updated global model parameters, each medical institution node updates its local copy of the global model. To ensure the accuracy and completeness of the model update, the central server and medical institution nodes innovatively implement a data verification mechanism, such as calculating the hash value of the parameter file. Upon receiving the parameter file, the node compares the hash values ​​and re-downloads any inconsistencies, preventing parameter transmission errors.

[0119] Finally, the model is applied and evaluated. Medical institution nodes use the updated global model to segment new medical images. In actual clinical applications, after a patient undergoes a CT, MRI, or PET scan, the image data is automatically transmitted to the local system. The system first performs the same innovative preprocessing as used during training, including grayscale normalization, image registration, and denoising. The preprocessed image is then input into the global model, which outputs the segmentation results. For example, in lung CT image segmentation, the model accurately segments lesions such as lung tissue, trachea, blood vessels, and possible tumors. Doctors use the segmentation results to assist in disease diagnosis. By observing the tumor's location, size, shape, and relationship to surrounding tissue, combined with clinical symptoms and other examination results, they can develop personalized treatment plans. For example, in patients with early-stage lung cancer, if the segmentation results indicate a small, localized tumor, doctors may consider surgical resection. However, if the tumor is larger or has metastasized, radiation therapy, chemotherapy, or targeted therapy may be prescribed.

[0120] Regular model innovation evaluations can also be conducted to ensure stable and accurate performance. Evaluation metrics include the Dice coefficient, sensitivity, and specificity. The Dice coefficient measures the degree of overlap between the predicted segmentation result and the true annotation, using the formula \(Dice = \frac{2|A\capB|}{|A|+|B|}\); sensitivity reflects the model's ability to correctly identify positive samples (such as lesions), using the formula \(Sensitivity = \frac{TP}{TP+FN}\); and specificity measures the model's ability to correctly identify negative samples (such as normal tissue), using the formula \(Specificity = \frac{TN}{TN+FP}\).

[0121] Evaluation is performed using an independent test dataset that contains multi-modal medical images from different medical institutions and different patients, as well as accurately labeled results. The test dataset is regularly (e.g., monthly) fed into the model to calculate various evaluation indicators. If model performance deteriorates, such as when the Dice coefficient falls below a set threshold (e.g., 0.9), innovative analysis of the causes may include changes in data distribution, overfitting or underfitting of the model, and insufficient learning of new lesion types. Appropriate improvement measures are taken for different reasons, such as re-collecting and labeling data, adjusting model architecture or hyperparameters, and retraining, to ensure that the model always provides accurate and reliable segmentation results for clinical diagnosis.

[0122] In actual applications, we can collect feedback from doctors to understand the convenience and practicality of the model's clinical use. For example, doctors may report that the visualization of the model's segmentation results is not intuitive or that the segmentation speed cannot meet urgent clinical needs. Based on this feedback, we can optimize the model output display method, such as using clearer color coding to display different segmentation areas to improve the visualization of segmentation results. If hardware allows, we can further optimize the model's inference speed through technical means such as model compression and hardware acceleration, reduce the model's image processing time, and improve clinical application efficiency.

[0123] This invention integrates diffusion models and federated learning technology to achieve accurate dynamic segmentation of medical images across medical institutions. It can output accurate image segmentation results for disease diagnosis, helping doctors to more accurately judge the condition and formulate treatment plans, which is of great significance to improving the efficiency and quality of medical diagnosis.

[0124] The present invention, with the help of the spatiotemporal conditional diffusion network in the dynamic diffusion architecture, can deeply analyze the spatiotemporal characteristics of medical images and lay the foundation for accurate segmentation. In the multimodal image segmentation task, the spatiotemporal conditional diffusion network can simultaneously process different modal images such as CT, MRI, and PET, giving full play to the unique advantages of each modality. Among them, CT images show a clear view of the skull structure, which can provide key information for judging the relationship between brain lesions and the skull, and help doctors accurately judge the location of lesions. MRI images are good at showing the soft tissue details of brain tissue, which helps doctors accurately identify subtle changes in brain tissue and improve the accuracy of lesion judgment. PET images can help discover early lesions or metabolic abnormalities by reflecting the metabolic status of brain tissue, providing a basis for early diagnosis of the disease. Through the fusion processing of multimodal images by the spatiotemporal conditional diffusion network, the characteristics of brain lesions can be captured comprehensively and accurately.

[0125] The system of the present invention can collaboratively process these multimodal images, extracting features from each modality through a spatiotemporal conditional diffusion network and performing fusion analysis. In practical applications, for complex tumor cases, the combined analysis of multimodal images can more accurately determine the tumor's benign or malignant nature, stage, and presence of metastases, providing more comprehensive and accurate information for clinical treatment and improving the accuracy and reliability of medical diagnoses. This overcomes the limitations of previous single-modality or traditional multimodal processing methods, providing stronger technical support for the diagnosis and treatment of diseases such as tumors.

[0126] The medical image segmentation model construction system in this embodiment is presented in the form of functional units, where the units refer to ASIC circuits, processors and memories that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0127] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0128] The embodiment of the present invention also provides a computer device having the above Figure 3 The medical image segmentation model building system shown.

[0129] See also Figure 4 , Figure 4 is a structural diagram of a computer device provided by an optional embodiment of the present invention, such as Figure 4 As shown, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in the memory or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 4 A processor 10 is taken as an example.

[0130] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0131] The memory 20 stores instructions that can be executed by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0132] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0133] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0134] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0135] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0136] A portion of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the form in which the computer program instruction exists in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc. Accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium that can be accessed by the computer.

[0137] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for constructing a medical image segmentation model, characterized in that: Applicable to a central server, the method includes: Obtaining gradient information of a local medical image segmentation model, where the local medical image segmentation model is trained by each client node based on local medical image data; Based on the gradient information, global model parameters are updated; the global model parameters are used to update the local medical image segmentation model.

2. The method according to claim 1, characterized in that The updating of global model parameters based on the gradient information includes: Calculate the data distribution characteristics of each client node; Determining the weight of the data distribution feature of each client node in parameter aggregation based on an adaptive weight allocation strategy; Based on the weight of each client node and the gradient information, global model parameters are updated.

3. The method according to claim 1, characterized in that The gradient information is obtained after being processed by a gradient obfuscation algorithm.

4. The method according to claim 1, wherein The updated global model parameters are sent to each client node using an encryption protocol.

5. A method for constructing a medical image segmentation model, characterized in that: Applicable to a client node, the method includes: A dynamic noise scheduling mechanism is used to generate synthetic image data based on local medical image data and a pre-built spatiotemporal conditional diffusion network. Based on the local medical image data and the synthetic image data, an initial deep learning model is trained to obtain a local medical image segmentation model; Calculating gradient information of the local medical image segmentation model; Processing the gradient information with a gradient obfuscation algorithm and sending it to a central server; After the central server updates the global model parameters based on the gradient information, receiving the global model parameters; The local medical image segmentation model is updated based on the global model parameters.

6. The method according to claim 5, characterized in that The spatiotemporal conditional diffusion network includes: An input layer, configured to receive the local medical imaging data and perform data preprocessing on the local medical imaging data; A convolutional layer, connected to the input layer, for performing multi-scale feature extraction on the pre-processed local medical image data; A recurrent layer, connected to the convolutional layer, is used to capture dynamic features in the temporal and spatial dimensions; An attention mechanism layer, connected to the recurrent layer, for calculating the attention weight of each position; An output layer, connected to the attention mechanism layer, is used to output the synthetic image data.

7. A medical image segmentation model construction system, characterized in that: The system comprises: A client node processing module is configured to generate synthetic image data based on local medical image data and a pre-built spatiotemporal conditional diffusion network; train an initial deep learning model based on the local medical image data and the synthetic image data to obtain a local medical image segmentation model; calculate gradient information of the local medical image segmentation model; and process the gradient information using a gradient obfuscation algorithm and send it to a central server; The central server processing module is used to obtain the gradient information of the local medical image segmentation model; based on the gradient information, update the global model parameters; the global model parameters are used to update the local medical image segmentation model.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the medical image segmentation model construction method according to any one of claims 1 to 6 by executing the computer instructions.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the medical image segmentation model construction method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The method comprises computer instructions for causing a computer to execute the medical image segmentation model construction method according to any one of claims 1 to 6.

Images