Multi-center medical heterogeneous data federation governance platform fusing privacy protection algorithm

Abstract

This invention relates to the field of medical information technology, specifically disclosing a multi-center medical heterogeneous data federated governance platform that integrates privacy-preserving algorithms. The platform includes: determining whether domain drift and differential privacy jointly cause the annihilation of rare feature information by quantifying the degree of domain drift and the risk of signal annihilation; generating cluster-specific and cluster-specific labels and domain feature difference vectors among centers through grayscale histogram analysis; eliminating domain drift through lightweight domain alignment, secondary evaluation, and deep domain alignment; dynamically adjusting the learning rate and noise intensity based on the real-time gradient signal-to-noise ratio during federated training, and performing domain-sensitive sparsity and differential noise allocation on the gradient components; and finally eliminating cross-center gradient residual noise interference on the server side, performing weighted aggregation based on intra-cluster signal-to-noise ratio and domain similarity. This effectively protects the signal integrity of rare medical features while ensuring differential privacy, improving the generalization ability and security of multi-center federated learning.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a multi-center medical heterogeneous data federated governance platform that integrates privacy protection algorithms. Background Technology

[0002] In the context of the deep integration of medical informatics and artificial intelligence, multi-center collaborative analysis of medical imaging data has become an important approach to improve the accuracy of disease diagnosis and promote research on rare diseases. However, differences in imaging equipment brands, scanning parameters, and imaging protocols among different medical institutions lead to systematic deviations in grayscale, contrast, and noise patterns in the images acquired by each center, a phenomenon known as domain drift. Domain drift makes it difficult for models trained at a single center to generalize to other centers, especially for rare pathological features (such as rare lesions and early-stage tumors), where the apparent differences between centers are more significant, and the features may even degenerate from high-density nodules to isodense backgrounds, severely limiting the effectiveness of cross-center models.

[0003] Federated learning, as a distributed machine learning paradigm, allows multiple centers to collaboratively train models without sharing raw data, effectively protecting data privacy. Building on this, to meet stricter privacy compliance requirements, differential privacy techniques are typically introduced, adding random noise to the gradient to prevent the leakage of individual information. However, differential privacy noise can weaken or even completely annihilate weak signals in the gradient, which is a typical characteristic of gradients containing rare features. When domain drift causes rare features to be identifiable only by a few centers, their gradient signals are already weak; after adding differential privacy noise, the signal-to-noise ratio is extremely low, ultimately rendering the global model completely unable to learn the feature, leading to the serious problem of privacy protection measures diminishing their clinical value.

[0004] Existing technologies have the following shortcomings: they fail to quantify the combined destructive effect of domain drift and differential privacy noise on rare features beforehand, and lack a precise problem diagnosis mechanism; they lack a hierarchical domain alignment strategy (combining lightweight and deep alignment), resulting in wasted computational resources or insufficient alignment effects; during federated training, they lack dynamic perception and interlocking adjustment of gradient signal-to-noise ratio, and cannot automatically retain rare signals while protecting privacy; gradient noise allocation does not distinguish the sensitivity of different parameters to domain drift, resulting in the privacy budget being used on key features; and during server-side aggregation, they fail to effectively eliminate cross-center noise interference, and the aggregation weights do not consider domain similarity and signal quality.

[0005] To this end, the present invention provides a multi-center medical heterogeneous data federated governance platform that integrates privacy protection algorithms. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and solve at least one of the technical problems mentioned in the background, this invention provides a multi-center medical heterogeneous data federated governance platform that integrates privacy protection algorithms. Addressing the domain drift problem caused by equipment differences in multi-center medical image data and the potential annihilation of rare feature information due to differential privacy protection, this platform effectively protects the signal integrity of rare medical features while ensuring differential privacy, thereby improving the generalization ability and security of multi-center federated learning.

[0007] The objective of this invention can be achieved through the following technical solution: a multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms, comprising the following steps:

[0008] Domain drift comparison analysis was performed on images from each center, and the cross-domain consistency between domain drift and rare features was evaluated. The degree of damage to rare signals by differential privacy noise was analyzed to determine whether there is a problem of rare feature information annihilation.

[0009] If they exist, extract the domain feature vector based on the grayscale histogram of each center image, calculate the normalized difference between centers, and generate static cluster labels and domain feature difference vectors for each center.

[0010] Domain alignment is performed on each center image, and after alignment, the cross-domain consistency of domain drift and rare features is evaluated again. If deep domain alignment is required, federated adversarial domain adaptation is used for deep domain alignment, and a domain adapter is designed to participate in federated training.

[0011] In each round of federated training, the cluster centers are determined based on the static cluster labels, and the signal-to-noise ratio is interlocked and the domain-sensitive inverse ratio noise is allocated to the cluster centers to obtain the noisy gradient vector.

[0012] Calculate the cosine similarity of the gradients at each center, and perform clustering to divide the drift direction clusters. Obtain the calibration gradient by random flipping of the median residual to reduce noise. Combine the static clusters to calculate the aggregation weights, and obtain the global gradient by weighted aggregation within and between clusters, and update the model parameters.

[0013] Furthermore, the process of performing domain drift comparison analysis on each central image is as follows:

[0014] Each center uses a pre-trained image feature extraction model to extract image feature vectors. The feature vector set is sent to the aggregation server through secure multi-party computation. The server selects a Gaussian radial basis kernel function, calculates the maximum mean difference between the feature distributions of any two centers, and takes the maximum value among all pairs as the domain drift severity index.

[0015] Furthermore, the process for evaluating the cross-domain consistency of domain drift and rare features is as follows:

[0016] Positive and negative feature vectors are extracted from the center of positive samples with rare features to form positive and negative feature matrices.

[0017] Calculate the t-statistic for each feature dimension and select dimensions with p-values ​​less than 0.05 as key dimensions;

[0018] The number of times each key dimension appears in different centers is counted, and the ratio of the number of key dimensions that appear only in one center to the total number of key dimensions is calculated to obtain the single center lock-in rate.

[0019] Furthermore, the method for determining whether domain drift and differential privacy lead to the annihilation of rare feature information is as follows:

[0020] Each center trains a shallow reference model, calculates the average L2 norm of the gradient components related to rare features as the gradient strength of rare features, and calculates the noise standard deviation based on the differential privacy parameter to obtain the signal-to-noise ratio.

[0021] When the domain drift severity index is greater than the preset index, the single center locking rate is greater than the preset threshold, and the signal-to-noise ratio is less than 1, it is determined that there is an information annihilation problem.

[0022] Furthermore, the generation process of the static cluster label and the domain feature difference vector of each center is as follows:

[0023] For each image at each center, a gray-level histogram is statistically analyzed, and the peak position, peak width, and gray-level skewness coefficient are extracted to form a three-dimensional domain feature vector;

[0024] Calculate the Euclidean distance and modulus of any two central region feature vectors to obtain the normalized dissimilarity.

[0025] Calculate the mean and standard deviation of the normalized variance for all center pairs. Pairs with a normalized variance less than the mean minus one standard deviation are labeled as clustered together, and pairs with a normalized variance greater than the mean plus one standard deviation are labeled as dissimilar.

[0026] Generate a domain feature difference vector for each center, with its components corresponding to other centers, taking 1 for the same cluster and 0 for different clusters.

[0027] Furthermore, the process of performing domain alignment on each central image includes:

[0028] Select the center with the largest sample size as the reference center, calculate the cumulative distribution function of all image pixel values ​​of the center and make a lookup table;

[0029] For each image at other centers, the cumulative distribution function is calculated pixel by pixel to find the cumulative probability corresponding to the original gray value. The gray value with the same cumulative probability is found in the reference lookup table as the new gray value, thus completing the histogram matching.

[0030] Furthermore, the process of secondary evaluation and deep domain alignment using federated adversarial domain adaptation is as follows:

[0031] After alignment, the global maximum mean difference is recalculated. If it is greater than the preset difference, the single center locking rate and the average center occurrence rate are recalculated.

[0032] When the single center locking rate is greater than the preset locking rate or the average center occurrence rate is greater than the preset occurrence rate, the federated adversarial domain adaptation framework is used to jointly train the generator and discriminator, and each center only exchanges parameter gradients without transmitting the original image.

[0033] Furthermore, the method for signal-to-noise ratio interlocking adjustment and domain-sensitive inverse ratio noise allocation is as follows:

[0034] For the same cluster center, calculate the signal strength value of the gradient tensor of the current batch and the noise basis value to obtain the instantaneous signal-to-noise ratio, maintain the signal-to-noise ratio queue of the past multiple rounds and take the 25th percentile as the coupling adjustment threshold;

[0035] If the current signal-to-noise ratio is less than the threshold, increase the learning rate and decrease the noise amplitude; otherwise, perform the opposite adjustment.

[0036] Calculate the Jacobian matrix of the loss function on the output feature map of the domain adapter, and obtain the domain drift sensitivity of each gradient component by combining the local contrast energy. Select the components with sensitivity greater than the mean plus one standard deviation to form the retention set.

[0037] Calculate the concentration factor, which is calculated as (sum of the sensitivities of all components within the retained set) / (sum of the sensitivities of all components). Assign noise variance to each component in inverse proportion to its sensitivity percentage to generate a noisy gradient vector.

[0038] Furthermore, the calibration gradient is calculated as follows:

[0039] The server collects all noisy gradient vectors, calculates the cosine similarity of the gradient sequences between centers, and forms a drift direction cluster using the median similarity as a threshold.

[0040] For each cluster, calculate the median vector, and subtract the median vector from the central gradient to obtain the residual vector;

[0041] Compare the residual magnitude with the median of the residual absolute value. If the residual magnitude is large, it is determined to be noise-dominated. Multiply each element of the residual by a random sign and perform an arithmetic mean with the corresponding positions of adjacent centers in the same cluster to obtain the noise estimate. Subtract this from the original gradient to obtain the calibration gradient.

[0042] Furthermore, the global gradient is calculated as follows:

[0043] Using static cluster labels, calculate the center proportion of each static cluster and the arithmetic mean of the instantaneous signal-to-noise ratio of all centers within the cluster, and multiply them to obtain the cluster aggregation weight;

[0044] First, a weighted average is performed within each cluster based on the instantaneous signal-to-noise ratio of each center to obtain the cluster aggregate gradient. Then, a second weighted average is performed between clusters based on the cluster aggregate weights to obtain the global gradient, which is used to update the global model parameters.

[0045] The beneficial effects of this invention are as follows:

[0046] By quantifying the severity index of domain drift, the degree of rare feature domain locking, and the signal-to-noise ratio, it is possible to accurately determine whether there is a rare feature annihilation problem caused by both domain drift and differential privacy, and decide whether subsequent governance steps are needed accordingly, thus avoiding the computational waste caused by blindly adopting complex algorithms.

[0047] First, low-cost histogram matching is used to eliminate first-order grayscale drift. Then, a second evaluation is used to determine whether to perform depth domain alignment. This significantly reduces computational overhead while ensuring performance. The depth domain alignment adopts a federated adversarial domain adaptation framework, which does not require the transmission of the original image and protects data privacy.

[0048] During local training, the learning rate and differential privacy noise amplitude are dynamically adjusted based on the real-time gradient signal-to-noise ratio (SNR) to maintain the SNR near the historical threshold. When rare signals are weak, noise is automatically reduced and the learning rate is increased to effectively prevent signal annihilation. When signals are strong, noise is moderately increased to save privacy budget. This adjustment is only enabled for cluster centers with similar gray-level distributions to avoid the negative impact of gradient direction conflicts in different domains.

[0049] By calculating the Jacobian matrix and local contrast energy, gradient components sensitive to domain drift (which usually carry rare feature information) are identified. These key components are selected as a retention set, and the noise variance is allocated inversely proportional to the sensitivity ratio. Key components receive minimal noise, while non-key components bear most of the noise, thus significantly improving the signal-to-noise ratio of rare features without increasing the overall privacy consumption.

[0050] By leveraging multi-center redundancy information, median residual analysis and random sign-flipped averaging are used to effectively estimate and eliminate independent noise added by differential privacy, while preserving the true gradient signal. During the aggregation process, both static clusters (grayscale distribution similarity) and dynamic signal-to-noise ratio are considered, ensuring that centers from similar domains with high signal quality dominate model updates, further protecting the learning of rare features. Attached Figure Description

[0051] The invention will now be further described with reference to the accompanying drawings.

[0052] Figure 1This is a flowchart of the multi-center medical heterogeneous data federated governance platform that integrates privacy protection algorithms, as described in this invention.

[0053] Figure 2 This is a logic diagram for determining whether there is a problem of rare feature information annihilation in this invention. Detailed Implementation

[0054] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0055] Example: Please refer to Figure 1 As shown, the multi-center medical heterogeneous data federated governance platform that integrates privacy protection algorithms according to the present invention specifically includes the following steps:

[0056] Step 1: Perform domain drift comparison analysis on the images of each center, evaluate the cross-domain consistency between domain drift and rare features, and combine the analysis of the degree of damage to rare signals by differential privacy noise to determine whether there is a problem of rare feature information annihilation.

[0057] Please see Figure 2 As shown, in step one, the process of performing domain drift comparison analysis on each central image includes:

[0058] Images from historical periods of each center were acquired, and a number of ordinary images without rare lesions were randomly selected, along with all rare feature positive images jointly annotated by clinical experts from multiple centers. All images were converted into grayscale numerical sequences, and all patient identification information (name, ID, examination number) was removed, retaining only the pixel matrix and scanning equipment parameters (brand, slice thickness, tube voltage, etc.).

[0059] The number of centers with rare positive samples and the number of positive samples per center are counted. If the number of positive centers is greater than or equal to 3 and the number of positive samples per center exceeds 10, then standard difference privacy federated learning is used. Conversely, if the number of positive centers is less than 3 or the number of positive samples per center does not exceed 10, then domain drift and signal annihilation diagnosis is performed, specifically as follows:

[0060] Each center locally uses an image feature extraction model pre-trained on a large natural image dataset (e.g., a ResNet-50 pre-trained on ImageNet, taking the 2048-dimensional feature vector before its global average pooling layer) to extract feature vectors for each image. The feature vectors are essentially numerical encodings of the high-level semantic features of the image.

[0061] The extracted image feature vectors include, but are not limited to:

[0062] Texture information: such as edge orientation, gradient distribution, local binary pattern, texture roughness / regularity, etc.;

[0063] Shape information: Geometric features such as the outline, corners, and curvature of an object or anatomical structure;

[0064] Local patterns: such as specific patterns of lesions like ground-glass opacities, nodules, and calcifications, which are represented by the numerical values ​​of certain dimensions in a vector after being abstracted layer by layer through convolutional layers.

[0065] Global layout: overall information such as the spatial location, relative size, and symmetry of organs;

[0066] Low-order statistics: Although convolutional networks learn automatically, deep features no longer directly correspond to gray values, but are more abstract concepts (such as whether there are circular high-density areas).

[0067] It is important to note that the specific numerical value of the feature vector does not have a direct physical unit. It is a floating-point number obtained by pooling the activation values ​​inside the network. Different image inputs will produce different vectors, and images with similar visual content are relatively close in the feature space.

[0068] All centers send their respective feature vector sets to the aggregation server through secure multi-party computation. The server selects a Gaussian radial basis kernel function, and the kernel width is set to the median of the Euclidean distance between all sample pairs. The specific calculation method is as follows: the feature vectors extracted by all centers are aggregated into a total set, the Euclidean distance between every two feature vectors is calculated, a distance list is obtained, and the median of the list is taken as the kernel width.

[0069] The secure multi-party computation protocol selected is the SPDZ protocol based on secret sharing. Each center splits its feature vector into two random shares and sends them to two non-colluding auxiliary servers (or uses an aggregation server as the receiver and transmits them in encrypted form using Paillier homomorphic encryption). The aggregation server calculates the kernel function value between each pair of centers in the encrypted domain and finally decrypts it through a threshold to obtain the MMD result. All communication is encrypted using TLS 1.3.

[0070] For any two centers, calculate the maximum mean difference in the feature distributions of the two centers: first calculate the average of the kernel function values ​​of all cross-center sample pairs, subtract the average of the kernel function values ​​of sample pairs within each center, then take the square root, and take the maximum value among all pairs as the domain drift severity index;

[0071] Understandably, the physical meaning of the domain drift severity index is: the degree of difference in the distribution of deep features between images acquired from different centers (different devices / protocols), that is, the distance between two image domains. The higher the domain drift severity index, the greater the apparent difference of rare features between different centers. If a rare feature only exists in centers with large domain drift, then its feature representation in other centers may be completely degraded (e.g., from high-density nodules to iso-density background), causing the global model in federated learning to be unable to learn the feature. Therefore, the domain drift severity index is the core indicator for determining whether domain alignment is needed.

[0072] In step one, the process of evaluating the cross-domain consistency of domain drift and rare features includes:

[0073] For centers with rare positive samples, the positive feature vectors of all rare positive sample images will be extracted, and an equal number of negative images will be randomly selected from the random sample set of centers with rare positive samples, and the negative feature vectors will be extracted in the same way.

[0074] Negative images refer to images that do not contain the target rare feature, while negative feature vectors refer to feature vectors extracted from non-negative images. These images usually come from normal control groups from the same center, or contain common lesions but do not contain the rare feature of interest in this study. In medical image analysis, the role of negative samples is to provide background distribution, that is, the typical appearance of the image in the feature space when the rare feature is absent. By comparing the feature vectors of positive samples (containing rare features), we can find out which feature dimensions can effectively distinguish between the two, thereby locating the digital fingerprint of the rare feature.

[0075] After receiving the eigenvectors of all the centers, the server divides them into two matrices:

[0076] Positive feature matrix: The number of rows is the total number of positive samples from all centers (e.g., 3 centers × 3 cases = 9 rows), the number of columns is the feature dimension D, and each row is the feature vector of a positive sample;

[0077] Negative feature matrix: The number of rows is the total number of negative samples from all centers (e.g., 3 centers × 30 cases = 90 rows), and the number of columns is the feature dimension D;

[0078] Calculate the mean vector of the positive samples (average of each column), the mean vector of the negative samples, and the overall mean vector of all samples.

[0079] For the positive and negative categories, calculate the cross product of each sample in the category and the mean vector of that category (multiply the vector by its own transpose), then sum them up. Finally, add the sums of the two categories. Mathematically, the intra-class scatter matrix reflects the degree of dispersion of the sample distribution within each category.

[0080] Calculate the outer product of the difference between the positive mean vector and the total mean vector, and multiply it by the number of positive samples; add the outer product of the difference between the negative mean vector and the total mean vector, and multiply it by the number of negative samples. The inter-class scatter matrix reflects the distance between the two class centers.

[0081] For the k-th feature dimension (D dimensions in total), extract the values ​​of all positive samples in this feature dimension to form a positive list P_k, and extract the values ​​of all negative samples in this feature dimension to form a negative list N_k.

[0082] Calculate the t-statistics for the positive list P_k and the negative list N_k. The larger the absolute value of the t-statistic, the stronger the discriminative power of that feature dimension.

[0083] It should be noted that the t-statistic is used to test the significance of the difference in means between positive and negative samples for a single feature dimension. The larger the absolute value, the stronger the ability of that dimension to distinguish rare features.

[0084] Arrange the absolute values ​​of the t-statistics of all feature dimensions in descending order, and take the feature dimensions whose absolute values ​​of the t-statistics correspond to t-critical values ​​with p-values ​​less than 0.05 as key dimensions. These key dimensions are the main expression dimensions of rare features in the feature space.

[0085] The key dimensions of each center are integrated to obtain a local key dimension list for each center. The union of all key dimensions of all centers is then taken. For each feature dimension in the global context, it is counted how many centers it appears in the key dimension set. For example, dimension 12 appears in 3 centers (A, B, C), dimension 45 appears in 2 centers (A, C), and dimension 78 appears only in center A.

[0086] The number of times each key dimension appears in different centers is counted. The ratio of the number of key dimensions that appear only in one center to the total number of key dimensions is calculated to obtain the single center lock-in rate.

[0087] It is understandable that if a global key dimension only appears in a single center, that is, the dimension is only able to distinguish positive samples in that center, and the value of the dimension is no different from that in positive samples in other centers, it means that the feature expression of this dimension is specific to that center, and is not visible in other centers due to domain drift. The more such dimensions there are, the more serious the rare features are locked by the domain.

[0088] If the single center locking rate is greater than or equal to the preset threshold, it is determined that the rare feature is domain locked. Domain locking means that the rare feature has huge differences in visual representation in images with different centers, so much so that it is mapped to different dimensions in the deep feature space. This will cause gradients from different centers to cancel each other out or generate noise in the feature dimension during federated learning, and the model will ultimately be unable to learn the feature.

[0089] It should be noted that the preset threshold is the threshold of the single center locking rate, which is used to determine whether rare features are locked by the domain. It is set by those skilled in the art, for example, 0.6.

[0090] In step one, the process of analyzing the degree of damage to rare signals by differential privacy noise includes:

[0091] Each center trains a shallow reference model locally using its own data. This model has the same structure as the final federated model but with fewer layers. During training, the gradient of the loss function with respect to the model parameters is calculated, especially the gradient components corresponding to the rare feature-related parameters. The average of the L2 norm of these gradient components (i.e., the square root of the sum of the squares of each component) is calculated and denoted as the rare feature gradient strength. At the same time, based on the differential privacy parameters to be used (target privacy budget parameter, gradient pruning threshold, noise multiplier), the standard deviation of the Gaussian noise to be added is calculated: noise standard deviation = gradient pruning threshold × noise multiplier, signal-to-noise ratio = rare feature gradient strength / noise standard deviation.

[0092] The shallow reference model uses the same network architecture as the final federated model, but reduces the number of convolutional layers to 1 / 3 of the original model (rounded down), while keeping the number of fully connected layers unchanged. For example, if the final model is ResNet-50, the shallow reference model is ResNet-18; if the final model has 3 convolutional layers and 2 fully connected layers, the shallow reference model has 1 convolutional layer and 2 fully connected layers. The same loss function (such as cross-entropy loss) is used during training, the number of training epochs is fixed at 10, and the learning rate is set to 0.001.

[0093] The method for locating gradient components related to rare features is as follows:

[0094] Starting from the center of positive samples with rare features, randomly select an equal number of positive and negative samples, input them into the shallow reference model, and calculate the gradient of the loss function of each sample with respect to the weights of the last fully connected layer of the model.

[0095] For each weight gradient component, calculate the absolute value of the difference between the mean of the positive sample group and the negative sample group on that component, and sort them from largest to smallest.

[0096] Select the top 5% of components as rare feature-related gradient components, and take the L2 norm (square root of the sum of squares) of these rare feature-related gradient components as the rare feature gradient intensity.

[0097] It is understandable that the physical meaning of the gradient strength of rare features is: the average L2 norm of the gradients of the parameters related to the rare features in the reference model, reflecting the contribution of the feature to the model update;

[0098] In step one, the process of determining whether domain drift and differential privacy lead to the annihilation of rare feature information includes:

[0099] If the domain drift severity index is greater than the preset index, rare features are locked in the domain, and the signal-to-noise ratio is less than 1, then there is a problem of domain drift and differential privacy causing rare feature information to be annihilated.

[0100] It should be noted that the preset index is the domain drift severity index threshold, which is set based on the statistical distribution of the maximum mean difference of images from different devices / protocols in the publicly available medical image dataset, for example, 0.35;

[0101] Step 2: If it exists, extract the domain feature vector based on the grayscale histogram of each center image, calculate the normalized difference between centers, and generate static cluster labels and domain feature difference vectors of each center.

[0102] In step two, the generation process of the static cluster tag includes:

[0103] For each two-dimensional slice image in all sample images of each center, count the number of pixels appearing at each gray level to obtain a gray-level histogram, with the horizontal axis representing the gray level and the vertical axis representing the pixel count;

[0104] Find the gray value on the horizontal axis corresponding to the maximum pixel count on the vertical axis of the gray-level histogram, and mark it as the peak position;

[0105] Using the peak position as the center, find the positions where the vertical pixel count drops to half of the peak position to the left and right respectively, and calculate the length of the grayscale interval between these two positions to obtain the peak width;

[0106] Calculate the average and standard deviation of the gray values ​​of all pixels, and calculate the cube of the difference between the gray value and the average value for each pixel. Sum the results and calculate the ratio with the total number of pixels. Then calculate the ratio with the cube of the standard deviation to obtain the gray skewness coefficient. A skewness coefficient greater than 0 indicates a right skewness distribution, and a skewness coefficient less than 0 indicates a left skewness distribution.

[0107] For each center, the peak position, peak width, and gray-level skewness coefficient are arranged in order to form a three-dimensional vector, which is denoted as the domain feature vector.

[0108] For any two centers, calculate the Euclidean distance between the feature vectors of the two centers, and calculate the magnitude of the feature vectors of each center. Normalized dissimilarity = Euclidean distance / (product of the magnitudes of the feature vectors of the two centers).

[0109] Calculate the normalized mean and standard deviation of the variance for all center pairs. Center pairs with a normalized variance less than the global variance mean minus one standard deviation are labeled as clustered together, and center pairs with a normalized variance greater than the global variance mean plus one standard deviation are labeled as dissimilar. Center pairs with a normalized variance between the global variance mean minus one standard deviation and the global variance mean plus one standard deviation are ignored.

[0110] In step two, the calculation process of the domain feature difference vector of each center includes:

[0111] For each center, a domain feature difference vector is generated, with a length equal to the number of other centers. Each component corresponds to another center. If the center is in the same cluster as the other center, the value is 1; otherwise, it is 0. The domain feature difference will be used in subsequent training to determine which centers' gradients can be interlocked.

[0112] It should be noted that the purpose of generating same-cluster and different-cluster labels is to group centers with similar gray-level distributions into the same static cluster. In subsequent training, interlocking adjustment is only performed on centers within the same cluster (because the gradient directions between centers within the same cluster are more likely to be consistent, making mutual adjustment effective; the gradient directions of centers in different clusters may be opposite, making interlocking adjustment harmful). The domain feature difference coefficient provides a binary vector for each center that can be directly used in the training code, indicating which other centers can participate in interlocking adjustment.

[0113] Step 3: Perform domain alignment on each center image, and after alignment, perform a second evaluation on the cross-domain consistency of domain drift and rare features. If deep domain alignment is required, use federated adversarial domain adaptation for deep domain alignment, and design a domain adapter to participate in federated training.

[0114] In step three, the process of performing domain alignment on each center image includes:

[0115] The center with the largest sample size among the participants is selected as the reference center. The reference center extracts all the pixel values ​​of all images to form a one-dimensional array.

[0116] Determine the grayscale level grouping, divide 0 to 65535 into 256 equally spaced intervals (each interval is about 256 pixels wide), count the frequency of pixels in each interval, and obtain a grayscale histogram. The horizontal axis is the center value of the grayscale interval (or the interval number), and the vertical axis is the number of pixels falling into that interval.

[0117] The cumulative distribution function represents the proportion of pixels with gray values ​​less than or equal to a certain threshold t. In specific calculation, starting from the smallest gray range, the frequency of each range is accumulated, and then the accumulated frequency is divided by the total number of pixels to obtain the cumulative probability. For example, for the kth range, the cumulative probability = (the sum of the frequencies of the first k ranges) / (the total number of pixels). The right endpoint of each range (or the range representing gray value) is used as the x-axis, and the cumulative probability is used as the y-axis to obtain the cumulative distribution function curve.

[0118] The 256 discrete gray values ​​(e.g., the midpoint of each interval) and their corresponding cumulative probability values ​​are stored in a lookup table. This table defines the mapping from gray values ​​to cumulative probability: f_ref(gray) = cumulative probability.

[0119] For each center other than the reference center, perform the following steps independently on each of its images:

[0120] For the image to be processed, extract all its pixel values ​​(e.g., 512×512 pixels), divide it into 256 grayscale intervals, calculate the histogram, and then calculate the cumulative distribution function to obtain the image's own mapping from grayscale values ​​to cumulative probability f_self(gray) = cumulative probability;

[0121] For each pixel in the image:

[0122] Find the original gray value range, use the reference center lookup table f_ref, traverse the reference lookup table, find the entry with the cumulative probability closest to p, take the corresponding gray value g, the new gray value g is the mapped value, repeat the above mapping for all pixels in the image to get a new image whose cumulative distribution function is approximately consistent with the cumulative distribution function of the reference center.

[0123] It should be noted that the purpose of domain alignment is to eliminate global grayscale deviation, preserve local contrast, and lay the foundation for subsequent depth alignment.

[0124] In step three, the process of performing a secondary evaluation of the cross-domain consistency of domain drift and rare features includes:

[0125] All images of each center after lightweight domain alignment are acquired, and feature vectors are extracted. These feature vectors are then aggregated to the server via secure aggregation. The server calculates the maximum mean difference between each pair of centers. If the global maximum mean difference is greater than a preset difference, cross-domain consistency evaluation of rare features after lightweight alignment is performed, specifically:

[0126] It should be noted that the preset difference is the global maximum mean difference threshold. Histogram matching can eliminate first-order grayscale drift (global contrast / brightness), but cannot eliminate higher-order texture or structure drift. Domain drift still exists after alignment, indicating that there is deep drift that cannot be eliminated by simple pixel transformation, and adversarial domain adaptation is required.

[0127] Positive and negative samples are input into the same feature extraction model to obtain their respective feature vector sets, which are then securely aggregated to the server.

[0128] The server uses linear discriminant analysis to identify the top K feature dimensions that contribute the most to distinguishing between positive and negative results as key dimensions.

[0129] For each key dimension, we count how many centers it has significant discriminative power on positive samples. The method for determining significance is as follows: within that center, we perform a t-test on the positive and negative samples using the value of that key dimension alone. If the p-value is <0.05 and the difference between the positive and negative means is in the same direction (e.g., all centers are positive > negative), then the dimension is considered effective in that center.

[0130] Average center occurrence rate = (sum of the number of occurrence centers of all key dimensions) / (total number of key dimensions × total number of centers). An average center occurrence rate close to 1 indicates that each dimension is generally present in all centers, while a rate close to 0 indicates that each dimension appears in only a very small number of centers.

[0131] Single-center lockout rate = (number of key dimensions that appear only in 1 center) / (total number of key dimensions). A higher single-center lockout rate indicates more severe domain lockout.

[0132] If the single center locking rate is greater than the preset locking rate or the average center occurrence rate is greater than the preset occurrence rate, then depth domain alignment is required.

[0133] It should be noted that the preset locking rate is the threshold for the single center locking rate triggered by depth alignment. If the single center locking rate is still greater than this value after light alignment, it indicates that the domain locking is stubborn and depth domain alignment is required.

[0134] The preset occurrence rate is the average center occurrence rate threshold triggered by deep alignment. After lightweight alignment, if the average center occurrence rate is greater than the preset occurrence rate, it means that the key feature dimensions generally appear in multiple centers but there is still residual drift, and deep alignment is needed to further improve consistency.

[0135] In step three, the depth domain alignment process includes:

[0136] A federated adversarial domain adaptation framework is adopted, in which each center jointly trains a generator network with the goal of converting the images of its own center into a common style, while training a discriminator to distinguish which center the image comes from.

[0137] The specific framework for federal adversarial domain adaptation is as follows:

[0138] Each center maintains a generator G_i (structure: 3-layer convolution + instance normalization + ReLU, outputting an affine transformation map of the same size as the input) and a discriminator D_i (structure: 5-layer convolution + fully connected binary classification, outputting the probability of which center the image belongs to).

[0139] In each training round, each center inputs its own image into the generator G_i to obtain a transformed image, and then inputs the transformed image into the discriminator D_i to calculate the domain classification loss. At the same time, the transformed image is sent to other centers (updated through gradients after safe aggregation, without directly sending the image).

[0140] In practice, a gradient inversion layer is used to jointly train the generator and discriminator;

[0141] During training, each center only exchanges the parameter gradients of the generator and discriminator, without transmitting the original images. This allows for the use of secure aggregation to protect the gradients.

[0142] Ultimately, each center has its own generator, but the output style of all generators becomes consistent under adversarial loss;

[0143] In step three, the process of the design domain adapter participating in federated training includes:

[0144] The domain adapter consists of a set of trainable affine transformation parameters, including a scaling factor vector and an offset vector, whose dimensions are the same as the number of channels of the feature map of the input image. For each channel of the input feature map, the adapter performs: output = scaling factor × input + offset. The adapter is inserted between the original image input and the first layer of the main network.

[0145] Initialize the scaling factor to all 1s and the offset to all 0s, so that the adapter is initially a unit map;

[0146] The domain adapter is trained in a federated manner along with the main network (such as a classification or segmentation network). The adapter's parameters participate in gradient updates and are also protected by differential privacy.

[0147] The specific interaction method of the domain adapter is as follows: the domain adapter is inserted between the original input image and the first layer of the main network, and its output feature map completely replaces the original image as the input of the main network;

[0148] In federated training, the adapter parameter aggregation is as follows: each center independently maintains its own domain adapter parameters, and only the parameters of the main network participate in the federated average aggregation. The reason for this design is that each center's domain adapter is specifically designed to compensate for the residual domain drift of its own center. The adapter parameters of different centers are naturally different and do not need to be unified.

[0149] It should be noted that the domain adapter provides a small number of adjustable parameters for each center, enabling it to dynamically compensate for residual domain drift during federated training without having to recompute the expensive global domain alignment. At the same time, the lightweight nature of the adapter makes it less sensitive to differential privacy noise.

[0150] Step 4: In each round of federated training, the cluster centers are determined based on the static cluster labels, and the signal-to-noise ratio is interlocked and the domain-sensitive inverse ratio noise is allocated to the cluster centers to obtain the noisy gradient vector;

[0151] In step four, the process of signal-to-noise ratio interlocking adjustment and domain-sensitive inverse ratio noise allocation includes:

[0152] The first point to clarify is that the local gradient signal-to-noise ratio dynamic sensing and interlocking adjustment are as follows:

[0153] Based on the local batch images obtained in the current round, for the current center, other centers in the same cluster as the current center are found and marked as cluster centers, and the remaining centers are marked as out-of-cluster centers. The interlocking adjustment logic is only enabled for cluster centers; for out-of-cluster centers, regular gradient updates are used (without step size and noise amplitude adjustment).

[0154] A batch of medical images is input into the domain adapter to obtain the output feature map. Then, the cross-entropy loss is calculated through forward propagation of the main network, and the gradient value of the loss function with respect to each parameter of the domain adapter is obtained through backpropagation, forming a gradient tensor.

[0155] Squaring each element in the gradient tensor, summing all the squared values, and then taking the square root gives the signal strength value.

[0156] Calculate the arithmetic mean of all elements of the gradient tensor, then subtract the mean from each element, square the result, sum the results, divide by the total number of elements, and finally take the square root to obtain the noise floor value. Instantaneous signal-to-noise ratio = signal strength value / noise floor value.

[0157] The current instantaneous signal-to-noise ratio is added to the queue, and the queue length is fixed to the past 20 rounds. If the queue is not full, no adjustment is made. If it is full, the 25th percentile of all values ​​in the queue is calculated (the value at position 25% after ascending order) and used as the coupling adjustment threshold.

[0158] If the current signal-to-noise ratio is less than the coupling adjustment threshold, multiply the parameter update step size of the domain adapter by (coupling adjustment threshold / current instantaneous signal-to-noise ratio), and multiply the differential privacy amplitude by (current instantaneous signal-to-noise ratio / coupling adjustment threshold).

[0159] If the current instantaneous signal-to-noise ratio is greater than or equal to the coupling adjustment threshold:

[0160] Multiply the step size by (coupling adjustment threshold / current instantaneous signal-to-noise ratio) and then multiply by 0.9;

[0161] Multiply the noise amplitude by the reciprocal of the aforementioned product (i.e., multiply by (current instantaneous signal-to-noise ratio / coupling adjustment threshold) × (1 ÷ 0.9)).

[0162] It should be noted that the logic of local gradient signal-to-noise ratio dynamic sensing and interlocking adjustment is as follows: to dynamically maintain the signal-to-noise ratio near the threshold, automatically reduce noise and increase the learning rate when the signal is weak to prevent rare signals from being annihilated, and moderately increase noise when the signal is strong to save privacy budget.

[0163] Secondly, it should be noted that domain-sensitive gradient sparsity is inversely proportional to the noise budget allocation, specifically:

[0164] Perform forward propagation (using the current domain adapter parameters) and backward propagation again, but this time calculate the partial derivative of the loss function with respect to each feature in the output feature map of the domain adapter, and arrange them into a Jacobian matrix. The rows of the matrix correspond to the spatial location (height × width) of the output feature map, and the columns correspond to the pixel location of the input image.

[0165] The calculation process for the Jacobian matrix is ​​as follows:

[0166] The domain adapter output feature map has a size of H×W×C (height×width×number of channels), and the input image size is H_in×W_in×1 (grayscale image). The partial derivatives of the loss function with respect to each element f in the output feature map form a Jacobian matrix J with dimensions (H·W·C)×(H_in·W_in). In actual calculations, to avoid dimensionality explosion, a channel-wise averaging approximation is used: for each output channel c, the average of the partial derivatives of all spatial locations of that channel is calculated as the average sensitivity of that channel to each input pixel, thus simplifying the Jacobian matrix to C×(H_in·W_in).

[0167] For each element in the Jacobian matrix, locate the corresponding input image pixel coordinates;

[0168] For each pixel (x, y) in the input image, calculate the absolute value of the pixel's gray level with its four neighboring pixels (top, bottom, left, right), collect the absolute values ​​of the differences between all neighbors in the entire image, and calculate its 75th percentile, denoted as q75.

[0169] The window side length L = max(3, ⌈q75⌉) is calculated as follows: When calculating the local contrast energy, the size of the neighborhood centered on each pixel is determined. The formula aims to achieve an adaptive window that automatically adjusts according to the local texture roughness of the image. If the image is rich in detail and the difference between adjacent pixels is large, then q75 is large, and the window is large. This can contain enough pixels to correctly estimate the local contrast (i.e., grayscale range), avoiding a window that is too small and only sees local noise or a single structure. If the image is smooth, then the difference between adjacent pixels is small, q75 is small, and the window is small. A small window is sufficient to capture local contrast and can reduce the amount of computation. The lower limit of 3 ensures that even if all adjacent differences are zero (a completely flat image), the window is still 3×3, ensuring that a very small but not zero contrast can be calculated (the range may still be 0, but at least the structure is intact).

[0170] Take an L×L window centered at pixel (x,y), calculate the difference between the maximum and minimum gray values ​​of the pixels within the window, and then the local contrast energy is the square of the difference between the maximum and minimum gray values ​​of the pixels within the window.

[0171] For each input pixel position (x, y), the partial derivatives of all output channels in the Jacobian matrix with respect to the input pixel are added together, and then the absolute value is calculated by proportionally calculating the number of channels in the output feature map of the domain adapter. The channel average of the absolute value of the Jacobian is obtained. The domain drift sensitivity is obtained by multiplying the channel average of the absolute value of the Jacobian with the local contrast energy.

[0172] It is understandable that the physical meaning of domain drift sensitivity is: a large domain drift sensitivity indicates that the pixel has a strong influence on the model output and is a high-contrast structure in the image, while a small domain drift sensitivity indicates that at least one factor is weak (the model is not sensitive or the image is a flat area).

[0173] The domain drift sensitivity is assigned to all gradient components generated from this pixel location (i.e., all model parameter gradient components related to this pixel in backpropagation).

[0174] Sort all domain drift sensitivity values ​​from largest to smallest, and calculate the arithmetic mean and standard deviation of all domain drift sensitivities;

[0175] The retention threshold is set by adding one standard deviation to the average domain drift sensitivity. All components with domain drift sensitivity greater than this retention threshold are selected to form the retention set.

[0176] Calculate the concentration factor: Concentration factor = (sum of the sensitivity of all components in the retention set) / (sum of the sensitivity of all components). The concentration factor is between 0 and 1, reflecting the degree of aggregation of high-value components.

[0177] The original noise variance is obtained by squaring the determined noise amplitude. The original noise variance is then multiplied by the concentration factor to obtain the total noise variance of the retained set. Each component in the retained set is then traversed to calculate the proportion of the sensitivity of that component to the total sensitivity of the retained set (sensitivity percentage).

[0178] The allocation noise variance of this component is equal to the total noise variance of the retained set × (1 / sensitivity ratio). That is, the higher the sensitivity of the component, the smaller the noise variance. For components that are not in the retained set, their allocation noise variance is set to zero.

[0179] Arrange the noise variance of all components in their original positions to form a list with the same dimension as the gradient vector, and denote it as the component-level noise variance allocation set;

[0180] It should be noted that the logic of domain-sensitive gradient sparsity and inverse noise budget allocation is as follows: the limited privacy noise budget is concentrated on non-critical gradient components, while minimal noise is applied to critical components that carry rare features, thereby protecting rare signals without increasing the overall privacy consumption.

[0181] Thirdly, it should be noted that local noise addition and gradient uploading are specifically as follows:

[0182] Each center generates Gaussian noise (mean 0, variance of the assigned value) independently for each component of the current gradient vector based on the generated component-level noise variance allocation set, and appends it to the gradient component to obtain a noisy gradient vector.

[0183] The noisy gradient vector, the instantaneous signal-to-noise ratio value after this round of adjustment, and the central domain feature difference coefficient are encrypted and uploaded to the aggregation server together.

[0184] Step 5: Calculate the cosine similarity of the gradients at each center, and perform clustering to divide the drift direction clusters. Obtain the calibration gradient by random flipping of the median residual to reduce noise. Combine the static clusters to calculate the aggregation weights, and obtain the global gradient by weighted aggregation within and between clusters. Update the model parameters.

[0185] In step five, the calculation process of the calibration gradient includes:

[0186] For the noisy gradient vectors of all centers collected by the server in this round, the gradient vector of each center is expanded into a one-dimensional numerical sequence to obtain the gradient sequence;

[0187] For any two centers, calculate the cosine similarity of their gradient sequences = the sum of the products of the elements at the same position in the two sequences / (the product of their respective moduli);

[0188] Arrange all cosine similarity values ​​in ascending order and take the median as the clustering threshold.

[0189] Center pairs with cosine similarity greater than the threshold are grouped into the same drift direction cluster. The connected component algorithm is used: if the similarity between center A and B is greater than the threshold, and the similarity between B and C is also greater than the threshold, then A, B, and C belong to the same cluster.

[0190] Within the drift direction cluster, all centers are sorted from high to low according to their signal-to-noise ratio to form a sequence. Each center is defined with its neighboring centers as the first and last centers in the sequence (the first and last centers have only one neighboring center).

[0191] For each drift direction cluster, extract the elements at the same position of all central gradient sequences within the drift direction cluster, calculate the median at that position, and obtain the median vector of the drift direction cluster. Subtract the median vector from the gradient sequence of each center within the drift direction cluster to obtain the residual vector.

[0192] Calculate the magnitude of the residual vector and the median of the absolute values ​​of all elements. If the magnitude of the residual is greater than the median of the absolute values ​​of the residual, then the residual is determined to be mainly caused by differential privacy noise, i.e., it is a noise-dominated residual vector.

[0193] For a residual vector r determined to be noise-dominated, a random symbol vector s with the same dimension as r is generated, where each element independently takes +1 with a probability of 0.5 and -1 with a probability of 0.5. The noise-dominated residual vector r is multiplied element by element by the random symbol vector to obtain the flipped residual vector.

[0194] For the current center, find the neighboring centers in the sorted sequence within the cluster of the drift direction. If there are two neighboring centers, take the one with higher signal-to-noise ratio and obtain the flipped residual vector at the same position of the neighboring centers.

[0195] Calculate the mean of the flipped residual vectors between the current center and the adjacent centers, and then multiply it element-wise with the random symbol vector to recover the sign, thus obtaining the noise estimate;

[0196] Subtract the noise estimate from the original noisy gradient vector to obtain the calibrated gradient vector;

[0197] If there is only one center in the cluster, the original noisy gradient is used directly as the calibration gradient; if there are only two centers in the cluster, each center uses the flipped residual of the other for averaging, and does not use its own.

[0198] In step five, the calculation process of the global gradient includes:

[0199] Using the domain feature difference coefficient (same-cluster / different-cluster label), the static cluster affiliation of each center is reconfirmed. It should be noted that this static cluster is different from the drift direction cluster. The static cluster is based on the image grayscale distribution, while the drift direction cluster is based on the current gradient direction.

[0200] For each static cluster, count the number of cluster centers and calculate the cluster center percentage = number of cluster centers / total number of centers;

[0201] Calculate the arithmetic mean of the instantaneous signal-to-noise ratios of all centers within the cluster;

[0202] The cluster aggregation weight = cluster center proportion × arithmetic mean of instantaneous signal-to-noise ratio within the cluster;

[0203] First weighted average (within cluster):

[0204] For the calibrated gradient vectors within the same static cluster, a weighted average is calculated based on the instantaneous signal-to-noise ratio (SNR) of each center: the gradient vector of each center is multiplied by the instantaneous SNR of that center, and the sum is divided by the sum of the instantaneous SNR values ​​within the cluster to obtain the cluster aggregate gradient.

[0205] Second weighted average (inter-cluster):

[0206] The cluster aggregate gradients of all clusters are weighted and averaged according to their respective cluster aggregate weights: each cluster aggregate gradient is multiplied by the aggregate weight of that cluster, and the sum is divided by the sum of the aggregate weights of all clusters to obtain the global gradient. The global gradient is then used to update the global model parameters.

[0207] The technical solution and advantages of this application are as follows: Domain drift comparison analysis is performed on each central image, and the cross-domain consistency between domain drift and rare features is evaluated. Combined with the analysis of the degree of damage to rare signals by differential privacy noise, it is determined whether there is a rare feature information annihilation problem. If so, domain feature vectors are extracted based on the grayscale histograms of each central image, and the normalized difference between centers is calculated to generate static cluster labels and domain feature difference vectors for each center. Domain alignment is performed on each central image, and after alignment, a secondary analysis of the cross-domain consistency between domain drift and rare features is conducted. For evaluation, if depth domain alignment is required, federated adversarial domain adaptation is used for depth domain alignment, and a domain adapter is designed to participate in federated training. In each round of federated training, the centers of the same cluster are determined according to the static cluster labels, and the signal-to-noise ratio is interlocked and the domain-sensitive inverse ratio noise is allocated to the centers of the same cluster to obtain noisy gradient vectors. The cosine similarity of the gradients of each center is calculated, and the drift direction clusters are divided by clustering. The calibration gradient is obtained by random flipping of the median residual to reduce noise. The aggregation weight is calculated in combination with the static clusters, and the global gradient is obtained by weighted aggregation within and between clusters, and the model parameters are updated. This invention determines whether rare feature information is annihilated due to a combination of domain drift and differential privacy by quantifying the degree of domain drift and the risk of signal annihilation. Then, it generates cluster-to-cluster and cluster-to-cluster labels and domain feature difference vectors among centers through grayscale histogram analysis. Next, it eliminates domain drift through lightweight domain alignment, secondary evaluation, and deep domain alignment. During federated training, the learning rate and noise intensity are dynamically adjusted based on the real-time gradient signal-to-noise ratio, and domain-sensitive sparsity and differential noise allocation are performed on the gradient components. Finally, cross-center gradient residual noise interference is eliminated on the server side, and weighted aggregation is performed based on intra-cluster signal-to-noise ratio and domain similarity. This effectively protects the signal integrity of rare medical features while ensuring differential privacy, thus improving the generalization ability and security of multi-center federated learning.

[0208] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A multi-center medical heterogeneous data federated governance platform integrating privacy-preserving algorithms, characterized by: Includes the following steps: Domain drift comparison analysis was performed on images from each center, and the cross-domain consistency between domain drift and rare features was evaluated. The degree of damage to rare signals by differential privacy noise was analyzed to determine whether there is a problem of rare feature information annihilation. If they exist, extract the domain feature vector based on the grayscale histogram of each center image, calculate the normalized difference between centers, and generate static cluster labels and domain feature difference vectors for each center. Domain alignment is performed on each center image, and after alignment, the cross-domain consistency of domain drift and rare features is evaluated again. If deep domain alignment is required, federated adversarial domain adaptation is used for deep domain alignment, and a domain adapter is designed to participate in federated training. In each round of federated training, the cluster centers are determined based on the static cluster labels, and the signal-to-noise ratio is interlocked and the domain-sensitive inverse ratio noise is allocated to the cluster centers to obtain the noisy gradient vector. Calculate the cosine similarity of the gradients at each center, and perform clustering to divide the drift direction clusters. Obtain the calibration gradient by random flipping of the median residual to reduce noise. Combine the static clusters to calculate the aggregation weights, and obtain the global gradient by weighted aggregation within and between clusters, and update the model parameters.

2. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 1, characterized in that: The process of performing domain drift comparison analysis on each central image is as follows: Each center uses a pre-trained image feature extraction model to extract image feature vectors. The feature vector set is sent to the aggregation server through secure multi-party computation. The server selects a Gaussian radial basis kernel function, calculates the maximum mean difference between the feature distributions of any two centers, and takes the maximum value among all pairs as the domain drift severity index.

3. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 2, characterized in that: The process for evaluating the cross-domain consistency of domain drift and rare features is as follows: Positive and negative feature vectors are extracted from the center of positive samples with rare features to form positive and negative feature matrices. Calculate the t-statistic for each feature dimension and select dimensions with p-values ​​less than 0.05 as key dimensions; The number of times each key dimension appears in different centers is counted, and the ratio of the number of key dimensions that appear only in one center to the total number of key dimensions is calculated to obtain the single center lock-in rate.

4. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 3, characterized in that: The method for determining whether domain drift and differential privacy cause the annihilation of rare feature information is as follows: Each center trains a shallow reference model, calculates the average L2 norm of the gradient components related to rare features as the gradient strength of rare features, and calculates the noise standard deviation based on the differential privacy parameter to obtain the signal-to-noise ratio. When the domain drift severity index is greater than the preset index, the single center locking rate is greater than the preset threshold, and the signal-to-noise ratio is less than 1, it is determined that there is an information annihilation problem.

5. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 1, characterized in that: The generation process of the static cluster label and the domain feature difference vector of each center is as follows: For each image at each center, a gray-level histogram is statistically analyzed, and the peak position, peak width, and gray-level skewness coefficient are extracted to form a three-dimensional domain feature vector; Calculate the Euclidean distance and modulus of any two central feature vectors to obtain the normalized dissimilarity. Calculate the mean and standard deviation of the normalized variance for all center pairs. Pairs with a normalized variance less than the mean minus one standard deviation are labeled as clustered together, and pairs with a normalized variance greater than the mean plus one standard deviation are labeled as dissimilar. Generate a domain feature difference vector for each center, with its components corresponding to other centers, taking 1 for the same cluster and 0 for different clusters.

6. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 1, characterized in that: The process of performing domain alignment on each center image includes: Select the center with the largest sample size as the reference center, calculate the cumulative distribution function of all image pixel values ​​of the center and create a lookup table; For each image at other centers, the cumulative distribution function is calculated pixel by pixel to find the cumulative probability corresponding to the original gray value. The gray value with the same cumulative probability is found in the reference lookup table as the new gray value, thus completing the histogram matching.

7. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 6, characterized in that: The process of secondary evaluation and deep domain alignment using federated adversarial domain adaptation is as follows: After alignment, the global maximum mean difference is recalculated. If it is greater than the preset difference, the single center locking rate and the average center occurrence rate are recalculated. When the single center locking rate is greater than the preset locking rate or the average center occurrence rate is greater than the preset occurrence rate, the federated adversarial domain adaptation framework is used to jointly train the generator and discriminator, and each center only exchanges parameter gradients without transmitting the original image.

8. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 1, characterized in that: The method of signal-to-noise ratio interlocking adjustment and domain-sensitive inverse ratio noise allocation is as follows: For the same cluster center, calculate the signal strength value of the gradient tensor of the current batch and the noise basis value to obtain the instantaneous signal-to-noise ratio, maintain the signal-to-noise ratio queue of the past multiple rounds and take the 25th percentile as the coupling adjustment threshold; If the current signal-to-noise ratio is less than the threshold, increase the learning rate and decrease the noise amplitude; otherwise, perform the opposite adjustment. Calculate the Jacobian matrix of the loss function on the output feature map of the domain adapter, and obtain the domain drift sensitivity of each gradient component by combining the local contrast energy. Select the components with sensitivity greater than the mean plus one standard deviation to form the retention set. Calculate the concentration factor, which is calculated as (sum of the sensitivities of all components within the retained set) / (sum of the sensitivities of all components). Assign noise variance to each component in inverse proportion to its sensitivity percentage to generate a noisy gradient vector.

9. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 1, characterized in that: The calibration gradient is calculated as follows: The server collects all noisy gradient vectors, calculates the cosine similarity of the gradient sequences between centers, and forms a drift direction cluster using the median similarity as a threshold. For each cluster, calculate the median vector, and subtract the median vector from the central gradient to obtain the residual vector; Compare the residual magnitude with the median of the residual absolute value. If the residual magnitude is large, it is determined to be noise-dominated. Multiply each element of the residual by a random sign and perform an arithmetic mean with the corresponding positions of the adjacent centers in the same cluster to obtain the noise estimate. Subtract this from the original gradient to obtain the calibration gradient.

10. The multi-center medical heterogeneous data federated governance platform integrating privacy protection algorithms according to claim 9, characterized in that: The global gradient is calculated as follows: Using static cluster labels, calculate the center proportion of each static cluster and the arithmetic mean of the instantaneous signal-to-noise ratio of all centers within the cluster, and multiply them to obtain the cluster aggregation weight; First, a weighted average is performed within each cluster based on the instantaneous signal-to-noise ratio of each center to obtain the cluster aggregate gradient. Then, a second weighted average is performed between clusters based on the cluster aggregate weights to obtain the global gradient, which is used to update the global model parameters.

Images