Personalized soft-mask aggregation federated segmentation system and method guided by primary basis alignment

Abstract

The application discloses a kind of personalized soft mask aggregation federal segmentation systems and methods guided by main base alignment, it is related to medical image intelligent analysis technical field, including data collection, data preprocessing, server side global parameter issue, client double main base extraction, main base alignment loss and soft mask generation, local model training and parameter upload and global model aggregation and parameter update module.The personalized soft mask aggregation federal segmentation systems and methods guided by main base alignment of the above application guide local model training and the strategy that global model is weighted bias to specified layer with soft mask, effectively suppresses cross-center model negative transfer, improves global model generalization ability and local personalized model segmentation performance, friendly to small sample center;The application is superior to federal average baseline and existing mainstream personalized federal learning method, with stronger cross-center robustness and scalability.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, and in particular to a personalized soft mask aggregation federated segmentation system and method guided by principal basis alignment. Background Technology

[0002] Medical image segmentation, a core technology of intelligent diagnostic and treatment systems, provides quantitative evidence for clinical diagnosis and treatment planning by accurately locating anatomical structures and pathological regions. Fully automated segmentation methods based on deep learning, such as U-Net and Transformer, have achieved expert-level accuracy in multiple tasks. However, due to the privacy concerns surrounding medical data, including patient identity information and pathological characteristics, regulations such as GDPR (EU General Data Protection Regulation) impose strict constraints and controls on data collection and processing, posing a risk of data privacy breaches to traditional centralized training. For a single center, limited sample size and uneven category distribution make it difficult to independently train high-quality models. To address the need for medical data privacy protection and the bottleneck of data silos, federated learning (FL) has emerged. As an emerging distributed learning method, it allows different institutions to share learning models rather than the data itself through collaborative training of models using distributed devices, providing a feasible path for cross-institutional joint training of high-performance segmentation models.

[0003] The classic Federated Averaging (FedAvg) algorithm represents the baseline algorithm for federated learning. Although it achieves the basic task of cross-institutional privacy collaboration through parameter sharing, its reliance on a "single global model" still presents challenges in the field of medical image segmentation due to data heterogeneity caused by differences in different medical imaging equipment, the complexity of the structure of different modal data, and different doctors' annotation habits, as well as the need for model personalization. To alleviate these problems, Personalized Federated Learning (PFL) allows each client to maintain its unique personalized model and customize parameters according to its local data distribution, while sharing knowledge among different clients. The FEDGKD framework proposed in 2021 uses the knowledge of multiple rounds of historical global models to form a "global teacher" to guide local model training, thus mitigating the client drift problem under heterogeneous data. The Ditto method, also proposed in 2021, personalizes the model through globally regularized federated multi-task learning, improving the model's accuracy, fairness, and adversarial robustness without changing the privacy and efficiency of standard FL. In 2022, the FedBABU framework achieved decoupled personalization of the subject head. In the federated stage, only the feature subjects are aggregated, while the classification head is randomly initialized and fine-tuned locally, balancing personalization adaptation with global representation quality. The Fed-LC strategy proposed in 2024, through local channel calibration and prediction inconsistency-guided supervision, outperforms existing PFL baselines in multi-center medical segmentation.

[0004] However, existing personalized federated learning strategies rely on static hierarchical partitioning based on manual priors, lacking mechanisms for adaptive adjustment with rounds and data drift. Even when introducing signals such as prediction inconsistency for personalized calibration, additional forward feeds, historical models, or multi-model comparisons are often required, increasing training and communication overhead and primarily affecting the prediction end, making it difficult to accurately reach the mid-to-high-level semantic features where domain drift truly occurs. Furthermore, most existing aggregations are weighted by sample size, lacking robust deweighting strategies for strong domain drift, noisy annotations, or unstable clients, thus easily leading to negative transfer. Therefore, a personalized federated segmentation method is needed that can adaptively combat multi-center domain drift, improve generalization of highly heterogeneous centers and centers with few samples, and dynamically adjust to maintain the advantage of well-performing local models, while effectively strengthening less performing local models through global knowledge. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized soft mask aggregation federated segmentation system and method guided by master base alignment, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a personalized soft mask aggregation federated segmentation system guided by primary base alignment, comprising:

[0007] The data collection module is used to collect and store data, achieving multi-center isolation;

[0008] The data preprocessing module is used to preprocess the data and reduce data skew.

[0009] The server-side global parameter distribution module is used to distribute global model parameters and provide training control.

[0010] The client-side dual-primary-base extraction module is used to obtain the local primary base and the global primary base. The client-side dual-primary-base extraction module includes a local model primary base extraction module and a distributed global model primary base extraction module.

[0011] The principal basis alignment loss and soft mask generation module is used for alignment constraints and soft mask weighted aggregation;

[0012] The local model training and parameter upload module is used to train and optimize local models.

[0013] The global model aggregation and parameter update module is used to update the global model.

[0014] A method for a principal-base alignment-guided personalized soft-mask aggregation federated segmentation system includes the following steps:

[0015] S1. Acquire similar medical images and store them in different centers;

[0016] S2. Normalize and crop the medical images, and divide them into training and test datasets. Perform data augmentation on the training dataset.

[0017] S3. The server sends the current global model parameters to the client and provides training control for each round of global training.

[0018] S4. Extract features of the same scale in the specified feature decomposition layer, perform adaptive average pooling compression and singular value decomposition on the features, and adaptively determine the local principal basis and the global principal basis.

[0019] S5. Transform the subspace consistency of the local master base and the global master base into an alignment loss, and generate soft mask weights through monotonic mapping;

[0020] S6. The client trains and optimizes the local model and uploads the updated parameters and corresponding soft mask weights to the server.

[0021] S7. The server updates the global model parameters based on the parameters returned by each client and their sample size.

[0022] Preferably, S1 specifically comprises:

[0023] Collect the same type of medical image data from different medical institutions and store them locally respectively;

[0024] Among them, the images containing fundus optic cup-optic disc segmentation tasks came from four different centers, the endoscopic images containing polyp lesions came from six different centers, and the colposcopy images containing cervical lesions came from two different centers.

[0025] Preferably, the specific steps of S2 are as follows:

[0026] S21. Using the data preprocessing module, the collected local data from each center are normalized, cropped, and size standardized in sequence.

[0027] S22. Divide the training set and test set at each center according to an 8:2 ratio;

[0028] S23. Perform data augmentation on the training set data, including color enhancement, random flipping, scaling, affine transformation, and lighting transformation.

[0029] Preferably, S3 specifically includes:

[0030] Using the server-side global parameter distribution module, several clients are randomly selected at the beginning of each global training round on the server side;

[0031] In the first round, the server sends the initial global model parameters to the selected clients. In subsequent rounds, the server sends the current global model parameters, which were aggregated in the previous round, to the selected clients. Clients that were not selected keep their global model parameters unchanged in this round.

[0032] Preferably, the specific steps of S4 are as follows:

[0033] S41. Using the client-side dual principal basis extraction module, on the same small batch of local samples of the selected client, extract features of the same scale at the specified feature decomposition layer using both the local model and the distributed global model, and calculate the principal basis for the features.

[0034] S42. The client-side dual principal basis extraction module performs adaptive average pooling on the extracted feature tensors, compressing the four-dimensional spatial features into a two-dimensional feature matrix.

[0035] S43. Perform singular value decomposition on the obtained two-dimensional feature matrix to obtain the left singular vector matrix of the local model and the global model;

[0036] S44. Adaptively select the first few columns as the local principal basis and the global principal basis respectively, wherein the cumulative singular value energy of the local principal basis and the global principal basis is not less than 95% of the threshold.

[0037] Preferably, the specified feature decomposition layer is a dual-scale feature fusion layer, which is obtained by adding the bottleneck layer of the backbone U-Net encoder and its previous layer through lateral connection and upsampling, and then smoothing it through convolution;

[0038] The parameters of the specified feature decomposition layer of the global model are as follows: in the first round, they are the initialization parameters of the corresponding layer; from the second round onwards, they are the current global parameters obtained by the server based on the corresponding layer parameters uploaded by the selected client in the previous round, and weighted and aggregated according to predetermined rules.

[0039] Preferably, the specific steps of S5 are as follows:

[0040] S51. Using the principal basis alignment loss and soft mask generation module, calculate the cosine similarity of the corresponding column vectors of the local principal basis and the global principal basis one by one to construct subspace consistency.

[0041] S52. The complementary quantity of subspace consistency is used as the alignment loss and transformed into soft mask weights through monotonic mapping. The soft mask weights are between 0 and 1.

[0042] Preferably, the specific steps of S6 are as follows:

[0043] S61. Using the local model training and parameter upload module, the client trains and optimizes the local model based on the joint loss, which consists of cross-entropy loss, Dice loss and alignment loss.

[0044] S62. The client uploads the trained model parameters and corresponding soft mask weights to the server, without uploading any original images or reversibly reconstructable intermediate features.

[0045] Preferably, the specific steps of S7 are as follows:

[0046] Using the global model aggregation and parameter update module, the server uses soft mask weights and sample size weighted aggregation for the feature decomposition layer parameters, and sample size weighted averaging for the parameters of other layers.

[0047] Therefore, the present invention employs the above-mentioned personalized soft mask aggregation federated segmentation system and method guided by primary base alignment, which has the following beneficial effects:

[0048] (1) This invention innovatively proposes an interpretable principal basis alignment personalization mechanism. The principal basis of the local model and the global model are extracted on the client through SVD decomposition. The consistency of the dual principal basis subspace is used as a unified measurable basis to generate alignment loss constraints to guide the training of the local model, thereby achieving domain drift suppression at the semantic level.

[0049] (2) This invention constructs a two-stage collaborative and hierarchical adaptive integrated framework for federated segmentation. During the local training phase, the client maintains personalized representation through principal basis alignment loss, suppressing drift while maintaining site specificity. During the aggregation phase on the server side, a hierarchical soft mask weighting strategy is implemented to form a collaborative process of “training alignment - aggregation bias”, which takes into account both global generalization and local personalization, improves cross-center fairness, and is more friendly to sites with few samples.

[0050] (3) The invention has been validated on multiple disease segmentation tasks in multiple centers. The results show that the overall performance of the invention is better than that of federated average and mainstream personalized methods, improves global generalization and local segmentation accuracy, and has stronger cross-center robustness and application feasibility.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] Figure 1 This is an overall flowchart of an embodiment of a personalized soft mask aggregation federated segmentation system and method guided by primary base alignment according to the present invention;

[0053] Figure 2 This is a flowchart illustrating the interaction between the federated learning server and the client in an embodiment of a personalized soft mask aggregation federated segmentation system and method guided by primary base alignment according to the present invention.

[0054] Figure 3This is a schematic diagram of client-side dual-primary base extraction in an embodiment of a personalized soft mask aggregation federated segmentation system and method guided by primary base alignment according to the present invention;

[0055] Figure 4 This is a schematic diagram of guest-master basis alignment loss and soft mask weight generation in an embodiment of a personalized soft mask aggregation federated segmentation system and method guided by master basis alignment according to the present invention.

[0056] Figure 5 This is a schematic diagram of server-side soft mask weighted aggregation in an embodiment of a personalized soft mask aggregation federated segmentation system and method guided by primary base alignment according to the present invention.

[0057] Figure 6 This is a schematic diagram illustrating the experimental results of a fundus optic cup-optic disc segmentation task based on an embodiment of the principal base alignment-guided personalized soft mask aggregation federated segmentation system and method of the present invention. Detailed Implementation

[0058] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0060] Example

[0061] Please see Figures 1-6 This invention provides a personalized soft mask aggregation federated segmentation system guided by master base alignment, comprising:

[0062] The data collection module is used to collect similar medical images from multiple medical institutions and store the images in different centers according to different medical institutions or different imaging devices to achieve multi-center isolation.

[0063] The data preprocessing module is used to normalize and crop medical images from different centers, split the training dataset and test dataset at a fixed ratio in each center, and perform data augmentation on the images in the training dataset to increase the richness of the samples and reduce the bias caused by differences in imaging equipment from different institutions and annotation habits of different doctors.

[0064] The server-side global parameter distribution module is used to distribute the current global model parameters to the selected clients participating in the current training round at the beginning of each global training round, providing overall training control for each round.

[0065] The client-side dual principal basis extraction module is used to extract features of the same scale at a specified feature decomposition layer using both the local model and the distributed global model on the same small batch of local samples. After the features are compressed by adaptive average pooling, singular value decomposition is performed, and the principal basis dimension is adaptively determined according to the energy proportion threshold to obtain the local principal basis and the global principal basis, which are used as measurable basis for subsequent alignment constraints and soft mask weighted aggregation.

[0066] The principal basis alignment loss and soft mask generation module is used to transform the subspace consistency of the local principal basis and the global principal basis into an alignment loss, and to generate the soft mask weights of the local end specified feature decomposition layer through monotonic mapping of the consistency.

[0067] The local model training and parameter upload module is used by the client to train and optimize the local model based on the joint loss composed of cross-entropy loss, Dice loss and alignment loss, and upload the updated parameters and corresponding soft mask weights to the server.

[0068] The global model aggregation and parameter update module is used by the server to update the global model parameters based on the parameters returned by each client and their sample size. For the parameters of the specified feature decomposition layer, soft mask weighted aggregation is used, and for the parameters of other layers, sample size weighted averaging is used to complete the global model update.

[0069] No local original image data is included in the entire model aggregation and transmission process.

[0070] This invention effectively suppresses negative transfer of cross-center models and improves the generalization ability of global models and the segmentation performance of local personalized models by guiding local model training through the alignment loss constraint of local and global dual principal basis subspaces and the soft mask weighting bias of global models on specified layers, without transmitting the original images throughout the process. It is also friendly to small sample centers.

[0071] This invention also provides a method for a personalized soft mask aggregation federated segmentation system guided by master base alignment, comprising the following steps:

[0072] S1, Data Collection.

[0073] The data collection module uses data from multiple centers for similar tasks, with different tasks being trained and tested separately based on their respective centers.

[0074] Three datasets were constructed, each containing multiple centers. The fundus optic cup and optic disc segmentation dataset came from three different hospitals (A, B, and C). The dataset in center C came from two different devices. Based on this, a total of four datasets were constructed, with the numbers {800, 400, 400, 791}. The endoscopic polyp dataset containing lesions came from six different centers, with the numbers {1000, 251, 456, 146, 206, 83}. The colposcopy cervical dataset containing lesions came from two different centers, with the numbers {971, 353}.

[0075] S2, Data Preprocessing.

[0076] The data preprocessing module was used to normalize the data of each center to the range [0, 1]. The images and masks of the polyp dataset and the cervical dataset were uniformly scaled to 384*384. The fundus dataset was first cropped to 800*800 and then scaled to 384*384. The image scaling used bilinear interpolation and the mask scaling used NEAREST interpolation. The dataset of each center dataset was divided into two parts in an 8:2 ratio. Data augmentation was performed on the training set, including random flipping, rotation ±30°, small affine transformation (translation ≤5%, scaling 0.95–1.05, cropping ±5°), brightness or contrast ±10%, and Gaussian noise (σ=0.01) to increase the richness of the samples.

[0077] To ensure fairness, data augmentation was disabled on the test set; only normalization and scaling were performed, without any data augmentation.

[0078] S3, Server-side global parameter distribution.

[0079] like Figure 2 As shown, at the beginning of each global training round, the server selects a predetermined number of clients to participate in training in each round using random sampling without replacement, and sends the global model parameters used in this round to the selected clients: the initial global model parameters are sent in the first round, and the current global model parameters obtained from the aggregation in the previous round are sent from the second round onwards, along with the local training configuration: local training round, batch size, learning rate (0.0001), etc.; and the configuration required for alignment and aggregation: alignment loss weight (0.2), principal basis energy threshold (0.95), and the specified feature decomposition layer used for principal basis extraction and soft mask calculation; the local parameters of clients not selected remain unchanged in this round.

[0080] S4, Client-side dual-primary-base extraction.

[0081] like Figure 3 As shown, after receiving the global parameters, the selected client sends small batches of local samples with a fixed batch size of 2 to both the local model and the distributed global model, extracting features of the same scale at a specified feature decomposition layer. The specified feature decomposition layer is a dual-scale fusion layer p4 obtained by sequentially passing the bottleneck layer b of the U-Net encoder through 1*1 convolutions and bilinear upsampling to the same resolution as the previous layer e4 of the encoder, adding it to the 1*1 lateral mapping of e4, and finally passing it through a 3*3 smooth convolution. Compared to the original single network layer, it acquires richer feature information, primarily based on deeper information. In the first round of training, the parameters of the global model at this layer are the network initialization parameters. From the second round onwards, its parameters are obtained from the current global parameters aggregated by the server based on predetermined rules of the parameters returned by each client.

[0082] To eliminate the influence of spatial scale, the obtained dual-scale four-dimensional feature F∈R B*C*H*W Performing adaptive average pooling yields a two-dimensional matrix Z∈R B*C Then transpose it to X=Z T ∈R C*B Singular value decomposition is performed on X according to formula (1), which is as follows:

[0083] X=U∑V T ,U∈R C*C ,∑=diag(σ1≥...≥σ n ),V∈R B*B (1)

[0084] Where X is the transposed two-dimensional matrix, and the column vectors of the left singular vector matrix U form the orthogonal principal directions of the channel subspace, reflecting the common information of the local models of each client. This information is more beneficial for updating the global model; ∑ is the singular value diagonal matrix; σ i Let U be the i-th singular value, sorted in descending order; V is the right singular vector matrix, corresponding to the sample subspace direction, reflecting the personalized information differences caused by imaging devices and other factors among different clients. T is the matrix transpose symbol, R represents the real number field, C is the number of channels, B is the batch size, and H and W are the height and width of the feature map, respectively. Here, the obtained left singular vector matrix U is adaptively selected based on the cumulative singular value energy not being less than a threshold of 95%, with the first k columns of the principal basis dimensions being used as the local principal basis and the global principal basis, respectively denoted as Ui. lacal and U global This is used for subsequent principal base alignment loss and soft mask generation.

[0085] S5, Principal Basis Alignment Loss and Soft Mask Generation.

[0086] like Figure 4 As shown, the cosine similarity of the obtained local and global principal bases is calculated column-by-column using the principal base alignment loss and soft mask generation module, and the average is taken to obtain the corresponding subspace consistency. The formula is as follows:

[0087]

[0088] Where m is the number of principal basis columns participating in the averaging (i.e., k mentioned above), i represents the column index of the principal basis (the i-th principal direction), and <·,·> represent the inner basis. The complementarity of this consistency is used as the alignment loss; the higher the consistency, the smaller the loss.

[0089] The obtained subspace consistency is mapped to the alignment loss constraint of the local model training according to formula (3), and then transformed into soft mask weights using monotonic mapping according to formula (4), as follows:

[0090]

[0091] Among them, L align is the alignment loss, and w is the soft mask weight. It can be seen that the higher the consistency, the smaller the loss, and the larger the soft mask weight.

[0092] S6. Local model training and parameter upload.

[0093] In the local model training phase, a joint loss consisting of cross-entropy loss, Dice loss, and alignment loss guides the training of the local model. The Dice loss is calculated based on the segmentation probability to maintain differentiability. The alignment loss weight is set to 0.2 to guide the local model closer to the global representation while ensuring its individuality. To avoid a lack of stable global representation in the first round, the alignment loss is set to 0 in the first round and enabled from the second round onwards. During training, the Adam optimizer is used with a learning rate of 0.0001. After each round of local training, the client uploads the updated parameters and corresponding soft mask weights to the server, without uploading any original local images in between.

[0094] S7, Global Model Aggregation and Parameter Update.

[0095] As attached Figure 5 As shown, the server collects the parameters, soft mask, and sample size returned by the selected client. It then uses the soft mask weight and sample size to weight the parameters of the specified feature decomposition layer to highlight the client contribution that is consistent with the global representation. The remaining layers are weighted by the sample size and then aggregated to update the global model. The aggregated and updated global parameters can be used for the next round of distribution. Only the model parameters and soft mask are exchanged throughout the process, and the original image is not transmitted.

[0096] S8. Model Evaluation.

[0097] The model performance was evaluated on the test set of each local endpoint, with evaluation metrics including the average segmentation Dice coefficient and the average surface distance ASSD. Comparisons were made with local training methods, traditional federated average baseline methods, and existing mainstream personalized federated learning methods. Furthermore, for the fundus optic cup / optic disc task, which uses binary segmentation, visualization results were incorporated into the evaluation and analysis for a more comprehensive assessment of the segmentation effectiveness.

[0098] Table 1. Quantitative comparison results on the polyp dataset [Dice (%) (↑) and ASSD (pixels) (↓)]

[0099]

[0100] Experimental results show that the principal basis alignment-guided personalized soft mask aggregation federated learning method of the present invention significantly improves the segmentation performance of the polyp dataset. It has the best Dice coefficient at all sites, and the average Dice coefficient and ASSD at each site are also the best. This confirms that the principal basis alignment-guided strategy can suppress model drift and enhance the generalization performance of the model while ensuring local personalized segmentation performance.

[0101] Table 2 shows the quantitative comparison results on the ocular dataset [Dice (%) (↑) and ASSD (pixels) (↓)].

[0102]

[0103] Experimental results show that the principal basis alignment-guided personalized soft mask aggregation federated learning method of this invention performs well on the optic cup-optic disc segmentation task, achieving optimal performance in both all-site and average-site performance. Figure 6 The visualization results of the visual cup-disc segmentation are presented. The qualitative and quantitative evaluation methods further confirm that the principal basis alignment guidance strategy can enable the already good model to continue to optimize in a better direction under the guidance of the global model, thus having stronger robustness.

[0104] Table 3 shows the quantitative comparison results on the cervical dataset [Dice (%) (↑) and ASSD (pixels) (↓)].

[0105]

[0106]

[0107] Experimental results show that the principal basis alignment-guided personalized soft mask aggregation federated learning method of the present invention performs well on the cervical lesion segmentation task and achieves the best result in the average Dice coefficient. Since the lesion area of ​​the cervical dataset is more complex and irregular, the ASSD value is not particularly low overall, but the effect of the present invention also reaches the second best in this index.

[0108] This invention has been extensively tested on various disease segmentation tasks, including polyp lesion segmentation, optic cup / disc segmentation, and cervical lesion region segmentation, achieving excellent performance in all cases. It provides an effective solution for deploying intelligent segmentation systems for cross-center medical image lesion regions. This invention can solve the problem of poor segmentation performance in clinics or small hospitals with limited samples, and has broad application prospects.

[0109] Therefore, this invention employs a personalized soft-mask aggregation federated segmentation system and method guided by principal basis alignment. Under non-independent and identically distributed conditions, it constrains local training through alignment loss between local and global dual principal basis subspaces, and generates soft masks based on alignment degree for hierarchical adaptive aggregation on the server side. The alignment loss is used to appropriately pull the key representation directions of the local model back to the global consensus subspace, avoiding training drift and improving cross-center transferability. Validated through multi-center fundus, endoscopic polyp, and colposcopy cervical lesion segmentation tasks, this invention outperforms the federated average baseline and existing mainstream personalized federated learning methods, exhibiting stronger cross-center robustness and scalability.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A personalized soft mask aggregation federated segmentation system guided by primary base alignment, characterized in that, include: The data collection module is used to collect and store data, achieving multi-center isolation; The data preprocessing module is used to preprocess the data and reduce data skew. The server-side global parameter distribution module is used to distribute global model parameters and provide training control. The client-side dual principal basis extraction module is used to obtain the local principal basis and the global principal basis. The client-side dual principal basis extraction module includes a local model principal basis extraction module and a distributed global model principal basis extraction module; the principal basis alignment loss and soft mask generation module is used for alignment constraints and soft mask weighted aggregation; The local model training and parameter upload module is used to train and optimize the local model; the global model aggregation and parameter update module is used to update the global model. The method applied to the aforementioned personalized soft mask aggregation federated segmentation system guided by primary base alignment includes the following steps: S1. Acquire similar medical images and store them in different centers; S2. Normalize and crop the medical images, and divide them into training and test datasets. Perform data augmentation on the training dataset. S3. The server sends the current global model parameters to the client and provides training control for each round of global training. S4. Extract features of the same scale in the specified feature decomposition layer, perform adaptive average pooling compression and singular value decomposition on the features, and adaptively determine the local principal basis and the global principal basis. The specific steps of S4 are as follows: S41. Using the client-side dual principal basis extraction module, on the same small batch of local samples of the selected client, extract features of the same scale at the specified feature decomposition layer using both the local model and the distributed global model, and calculate the principal basis for the features; S42. The client-side dual principal basis extraction module performs adaptive average pooling on the extracted feature tensors, compressing the four-dimensional spatial features into a two-dimensional feature matrix; S43. Perform singular value decomposition on the obtained two-dimensional feature matrix to obtain the left singular vector matrices of the local model and the global model; S44. Adaptively select the first few columns as the local principal basis and the global principal basis, respectively, wherein the cumulative singular value energy of the local principal basis and the global principal basis is not less than 95% of the threshold. The specified feature decomposition layer is a dual-scale feature fusion layer, which is obtained by adding the bottleneck layer of the backbone U-Net encoder and its previous layer through lateral connection and upsampling, and then smoothing it through convolution. The parameters of the specified feature decomposition layer of the global model are: the initialization parameters of the corresponding layer in the first round; from the second round onwards, the current global parameters are obtained by the server based on the corresponding layer parameters uploaded by the client in the previous round, and weighted and aggregated according to the predetermined rules. S5. Transform the subspace consistency of the local principal basis and the global principal basis into an alignment loss, and generate soft mask weights through monotonic mapping; S51. Using the principal basis alignment loss and the soft mask generation module, calculate the cosine similarity of the corresponding column vectors of the local principal basis and the global principal basis one by one to construct the subspace consistency; S52. Use the complementary amount of the subspace consistency as the alignment loss, and transform it into soft mask weights through monotonic mapping. The soft mask weights are between 0 and 1. S6. The client trains and optimizes the local model and uploads the updated parameters and corresponding soft mask weights to the server. The specific steps of S6 are as follows: S61, using the local model training and parameter upload module, the client trains and optimizes the local model based on the joint loss, which consists of cross-entropy loss, Dice loss and alignment loss; S62, the client uploads the trained model parameters and corresponding soft mask weights to the server, without uploading any original images or reversibly reconstructable intermediate features. S7. The server updates the global model parameters based on the parameters returned by each client and their sample size. S7 uses the global model aggregation and parameter update module. The server uses soft mask weight and sample size weighted aggregation for the feature decomposition layer parameters, and sample size weighted average for the parameters of other layers.

2. The personalized soft mask aggregation federated segmentation system guided by primary base alignment according to claim 1, characterized in that, S1 specifically involves: collecting medical image data of the same type from different medical institutions and storing them locally; among them, images containing fundus optic cup-optic disc segmentation tasks come from four different centers, endoscopic images containing polyp lesions come from six different centers, and colposcopy images containing cervical lesions come from two different centers.

3. The personalized soft mask aggregation federated segmentation system guided by primary base alignment according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, using the data preprocessing module, normalize, crop and standardize the size of the local data collected from each center in sequence; S22, divide the training set and test set in each center according to an 8:2 ratio; S23, perform data augmentation on the training set data, including color enhancement, random flipping, scaling, affine transformation and lighting transformation.

4. The personalized soft mask aggregation federated segmentation system guided by primary base alignment according to claim 1, characterized in that, Specifically, S3 involves using the server-side global parameter distribution module to randomly select several clients at the start of each global training round on the server side. In the first round, the server distributes the initial global model parameters to the selected clients. In subsequent rounds, the server distributes the current global model parameters, which were aggregated in the previous round, to the selected clients. Clients that were not selected retain their global model parameters unchanged in this round.

Images