Ultrasonic image federal learning method and system fused with anatomical prior knowledge

Abstract

The invention relates to an ultrasonic image federated learning method and system fusing anatomical prior knowledge, computer equipment, a storage medium and a computer program product. Global model parameters and a global anatomical structure diagram are initialized through a server; the server issues global model parameters and a global anatomical structure chart to each client; the client calculates detection loss based on the local data and performs feature enhancement on the local data; the technical problems that in the prior art, a traditional centralized or classical FL method is difficult to consider global performance and local adaptability at the same time and neglect topological priori between organ structures, so that the generalization ability of a model in complex structure recognition is limited, and deep fusion with medical knowledge is lacked are solved. The generalization ability of the model in a complex structure is improved, and deep fusion of the model and medical knowledge is enhanced.

Description

Technical Field

[0001] This application relates to the field of computer software technology, and in particular to a method, system, storage medium, and computer program product for federated learning of ultrasound images that integrates prior anatomical knowledge. Background Technology

[0002] With ultrasound screening becoming an important and widely used tool for managing maternal and fetal health globally, its ability to provide real-time, non-invasive imaging makes it a cornerstone of prenatal care. In particular, fetal structural screening using ultrasound technology enables clinicians to comprehensively assess fetal growth and development, thereby enabling early detection of potential structural abnormalities and providing crucial support for clinical decision-making.

[0003] In early pregnancy, the absence of nasal bone structures in the nuchal translucency (NT) plane is closely associated with an increased risk of Down syndrome, while a low-lying conus medullaris (CM) indicates spina bifida. In mid-to-late pregnancy, assessing fetal cardiac health involves identifying structural abnormalities such as atrial septal defect (ASD) or ventricular septal defect (VSD) in the four-chamber view, key indicators of congenital heart disease. Similarly, the absence of the cavum septum pellucidum (CSP) or posterior fossa cistern (PFC) in fetal brain assessment may indicate abnormal brain development. Furthermore, various clinical tasks, including standard plane localization, fetal growth parameter measurement, and quality control, fundamentally rely on the accurate identification of anatomical structures. Therefore, accurate identification of anatomical structures is fundamental for clinicians to assess and manage maternal and infant health, and has become an indispensable part of modern obstetric care, making a significant contribution to reducing neonatal morbidity and mortality.

[0004] While anatomical structure recognition is crucial, training robust and scalable models is often hampered by data privacy concerns and information barriers within medical centers. In large-scale cross-institutional model training, data privacy regulations (such as HIPAA) prevent the sharing of raw images, making it difficult for traditional centralized training to acquire representative "global knowledge." Federated learning (FL) offers a solution, enabling collaborative training of shared models while retaining local data across medical centers, thus achieving collaborative intelligent system construction under privacy protection. Nevertheless, three core challenges remain in real-world medical environments: (1) Differences in hospitals, equipment, and procedures lead to highly heterogeneous distribution of ultrasound images, making it difficult for traditional centralized or classic FL methods to simultaneously achieve global performance and local adaptability; (2) Existing methods often ignore topological priors between organ structures, such as cardiac chamber connections, hepatobiliary branching structures, or glandular hierarchical structures, resulting in limited generalization ability of models in complex structure recognition; (3) Mainstream FL methods focus on parameter aggregation and personalized strategies, lacking deep integration with medical knowledge (anatomical structures, clinical measurement rules, topological constraints). Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, computer device, storage medium, and computer program product for federated learning of ultrasound images that integrates prior anatomical knowledge to address the aforementioned technical problems.

[0006] In a first aspect, this application provides a federated learning method for ultrasound images that integrates prior anatomical knowledge, the method comprising: The server initializes global model parameters and global anatomical structure diagram; The server sends global model parameters and global anatomical structure diagrams to each client. The client calculates the detection loss based on local data and performs feature enhancement on the local data; The client creates a local anatomical diagram; The client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on Wasserstein distance and using Sinkhorn approximation to guide local training; The client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; The server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The server updates the global anatomical structure diagram and global model parameters.

[0007] In one implementation, the client calculates the detection loss based on local data and performs feature enhancement on the positive sample ROI, which consists of an online network and a contrastive target network. The feature enhancement updates the contrastive target network through an exponential moving average (EMA) and minimizes the Kullback-Leibler divergence of the softmax normalized ROI feature distribution.

[0008] In one implementation, the client creates a local anatomical diagram, including: The client extracts multi-scale features based on the feature pyramid network, constructs anatomical nodes by uniformly sampling the spatial locations of organs, and constructs an adjacency matrix based on a hybrid criterion of spatial proximity and feature affinity to obtain a local anatomical map.

[0009] In one implementation, the server updates the global anatomical structure diagram and global model parameters, including: The server generates aggregate weights based on the matching scores, performs a weighted average on the domain-level anatomy map to update the global anatomy map, and uses the same aggregate weights to perform weighted aggregation on the client model parameters to update the global model parameters.

[0010] In one embodiment, the method further includes: When the preset conditions are not met, the server will send the updated global model and global graph to each client to start the next round of training; When the preset conditions are met, the server outputs the final model and the global anatomical structure prior. The preset condition is reaching a preset number of rounds or reaching a convergence condition.

[0011] In one implementation, the client calculates the detection loss based on local data and performs feature enhancement on the positive ROIs, consisting of an online network and a contrastive target network. The feature enhancement updates the contrastive target network using an exponential moving average (EMA) and minimizes the Kullback-Leibler divergence of the softmax-normalized ROI feature distribution, including: The input image is forward-predicted based on a shared encoder and a detection head, and a standard detection loss is calculated, which includes classification loss and localization loss. The online network and the target comparison network form a two-branch structure; The target network parameters are updated using the exponential moving average (EMA) of the online network parameters. Apply feature consistency constraints to positive ROIs; Minimize the Kullback-Leibler (KL) divergence between the feature distributions of the softmax normalized ROI to encourage enhanced invariance.

[0012] Secondly, this application also provides a federated learning system for ultrasound images that integrates prior anatomical knowledge, the system comprising: The initialization module is used to initialize the server's global model parameters and global anatomical structure diagram. The sending module is used by the server to send global model parameters and global anatomical structure diagrams to each client. The first processing module is used by the client to calculate the detection loss based on local data and to perform feature enhancement on the local data; A module is created for the client to generate local anatomical diagrams; The second processing module is used by the client to supervise the local anatomical map using the global anatomical map, and to guide local training by calculating the matching score between the global anatomical map and the local anatomical map based on the Wasserstein distance and using the Sinkhorn approximation. The upload module is used by the client to aggregate the trained local anatomical maps into a domain-level anatomical map and upload it to the server. The third processing module is used by the server to calculate the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The update module is used by the server to update the global anatomical structure diagram and global model parameters.

[0013] Thirdly, this application also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps: The server initializes global model parameters and global anatomical structure diagram; The server sends global model parameters and global anatomical structure diagrams to each client. The client calculates the detection loss based on local data and performs feature enhancement on the local data; The client creates a local anatomical diagram; The client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on Wasserstein distance and using Sinkhorn approximation to guide local training; The client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; The server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The server updates the global anatomical structure diagram and global model parameters.

[0014] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps: The server initializes global model parameters and global anatomical structure diagram; The server sends global model parameters and global anatomical structure diagrams to each client. The client calculates the detection loss based on local data and performs feature enhancement on the local data; The client creates a local anatomical diagram; The client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on Wasserstein distance and using Sinkhorn approximation to guide local training; The client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; The server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The server updates the global anatomical structure diagram and global model parameters.

[0015] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps: The server initializes global model parameters and global anatomical structure diagram; The server sends global model parameters and global anatomical structure diagrams to each client. The client calculates the detection loss based on local data and performs feature enhancement on the local data; The client creates a local anatomical diagram; The client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on Wasserstein distance and using Sinkhorn approximation to guide local training; The client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; The server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The server updates the global anatomical structure diagram and global model parameters.

[0016] The aforementioned federated learning method, system, computer equipment, storage medium, and computer program product for ultrasound images, which integrates prior anatomical knowledge, involves the following steps: The server initializes global model parameters and a global anatomical map; the server distributes the global model parameters and global anatomical map to each client; the client calculates the detection loss based on local data and performs feature enhancement on the local data; the client builds a local anatomical map; the client supervises the local anatomical map using the global anatomical map, calculating the matching score between the global anatomical map and the local anatomical map locally based on Wasserstein distance and using the Sinkhorn approximation to guide local training; the client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; the server calculates the matching score between the domain-level anatomical map and the global anatomical map based on Wasserstein distance and using the Sinkhorn approximation; and the server updates the global anatomical map and global model parameters. This invention addresses the technical problem that existing traditional centralized or classical FL methods struggle to simultaneously balance global performance and local adaptability, neglect topological priors between organ structures, resulting in limited generalization ability of the model in complex structure recognition and a lack of deep integration with medical knowledge. It improves the model's generalization ability in complex structures and enhances the deep integration of the model with medical knowledge. Attached Figure Description

[0017] Figure 1 and Figure 2 This is a diagram of the overall training framework of the present invention; Figure 3 A flowchart of a federated learning method for ultrasound images that integrates prior anatomical knowledge, provided as an embodiment of the present invention; Figure 4 An architecture diagram of an ultrasound imaging federated learning system that integrates prior anatomical knowledge, provided as an embodiment of the present invention; Figure 5 Internal evaluation results for the EP and M2LP datasets; Figure 6 This is an example diagram illustrating the clinical application of the model of the present invention; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] This invention proposes a federated learning method for ultrasound images that integrates prior anatomical knowledge. This method can support various downstream applications in cross-platform, multi-center clinical environments, including standardized image recognition, quality control, key parameter measurement, and intelligent diagnosis of multiple diseases in general ultrasound tasks. The method has also been validated in the field of prenatal ultrasound, enabling key functions such as fetal standard plane recognition, image quality management, structural parameter measurement, and developmental abnormality prediction. As a non-invasive, real-time, radiation-free, and highly accessible mainstream medical imaging technology, ultrasound is widely used in multi-organ scenarios such as cardiovascular, abdominal, thyroid, breast, urinary system, and prenatal screening. Its diagnostic reliability highly depends on the accurate identification of anatomical structures and topological relationships.

[0020] This invention aims to learn a domain-invariant detection model under the condition of strict localization of the original data. The method represents anatomical structures as graphs, achieves cross-domain structure alignment through optimal transport, and introduces feature-level consistency enhancement before graph construction to stabilize the representation space; it also achieves the co-evolution of global model parameters and the global structure graph through an anatomically guided weighting scheme. The overall training framework of the algorithm is shown in Figure 1, where Figure 1 represents the training content of our model. Our model contains three key components: graph matching (GM), adaptive aggregation mechanism (Agg), and feature enhancement (FE), and was trained on datasets from five hospitals. Figure 2 This section describes our model details. Each client extracts anatomical features from a local image, constructs a domain-level graph using a GNN, and applies feature consistency enhancement to stabilize the local representation. The server receives the client's graph, computes the best transport graph match to obtain a similarity score, and updates the global model parameters and global graph through weighted aggregation. The global structural prior is then sent back to the client to guide local training using graph matching loss, ensuring cross-center anatomical consistency.

[0021] The following describes in detail the ultrasound image federated learning method that integrates prior anatomical knowledge according to the present invention.

[0022] Please refer to Figure 3 This application provides a federated learning method for ultrasound images that integrates prior anatomical knowledge, the method comprising: S100, the server initializes global model parameters and global anatomical structure diagram.

[0023] Consider a federation based on N clients, where each client Holding a local data Distribution from its specific domain Mid-sampling. Each input It is an image. Represents the corresponding real labels, where the label set Shared between clients. Due to the heterogeneity of acquisition equipment, imaging conditions, and anatomical features, we assume... for This reflects the domain shift in the real world.

[0024] make This represents a shared feature encoder parameterized by θ. This represents the object detection head. For any image x, the local prediction is:

[0025] The server-side establishes a cross-domain federated learning (CDFL) training framework to coordinate iterative communication among multiple clients; it also initializes global model parameters. With global anatomical structure diagram (Can be empty or based on prior initialization). In the CDFL setup, a central server coordinates model training, while all raw data remains strictly localized on the client, respecting data privacy, regulatory constraints, and limited communication / computing resources. Each client i can only exclusively access its own target domain dataset. This reflects the scarcity of data and isolated deployment scenarios in the real world. CDFL aims to learn global parameters that generalize across domains. Instead of sharing the original data, the formula is:

[0026] Where x is the image input. This represents a shared feature encoder parameterized by θ. Represents the object detection header. The aggregate weight of client i is typically set to the size of the local dataset. proportional (i.e.) ), This is the standard detection loss (classification and localization). Expected. Data distribution taken from client i The goal is collaboratively optimized through iterative communication rounds between the server and client. In each round, the client updates its local model based on its own data. The server aggregates these local updates to produce a new global model. The aim is to minimize the global objective under the constraint of data locality.

[0027] S200: The server sends global model parameters and global anatomical structure diagrams to each client.

[0028] The server sends the global model parameters for this round to each client. With global anatomical structure diagram If this is the first training iteration, initialized global model parameters and a global anatomical structure graph are issued for use in this round of local joint training and graph matching constraints.

[0029] In the S300, the client calculates the detection loss based on local data and performs feature enhancement on the local data.

[0030] To reduce the interference of domain offset on anatomical structure modeling, this invention stabilizes cross-domain ROI representation by introducing a feature enhancement mechanism that relies solely on local positive samples before graph construction.

[0031] Specifically, the client of this invention calculates the detection loss based on local data and performs feature enhancement on positive sample ROIs consisting of an online network and a contrast target network. The feature enhancement updates the contrast target network through exponential moving average (EMA) and minimizes the Kullback-Leibler divergence of the softmax normalized ROI feature distribution.

[0032] First, based on the shared encoder and detection head, forward prediction is performed on the input image, and the standard detection loss is calculated to obtain the local detection loss. The ROI features of candidate anatomical structures / organs are considered. The standard detection loss includes classification loss and localization loss, and the detection model can be, for example, the YOLO series (YOLOv8, YOLO11, etc.).

[0033] Then, we stabilize cross-domain representations through feature enhancement. A two-branch structure is constructed with an online network and a contrastive target network. The parameters of the contrastive target network are updated using the exponential moving average (EMA) of the online network parameters. Feature consistency constraints are applied to positive ROIs, minimizing the Kullback-Leibler (KL) divergence between the softmax-normalized ROI feature distributions to encourage enhancement invariance.

[0034] Finally, we obtain the feature enhancement loss. The stable ROI features are multi-scale. The feature enhancement loss formula is:

[0035] Where n is the number of targets detected in the batch. represents the softmax normalized probability distribution of ROI features from the online branch and the contrastive branch, respectively. After the initial training phase, the contrastive network C is discarded to reduce memory overhead for resource-constrained clients, while the representations learned from the online network O are retained for subsequent joint rounds. S400, the client creates a local anatomical diagram.

[0036] The client extracts multi-scale features based on the feature pyramid network, constructs anatomical nodes by uniformly sampling the spatial locations of organs, and constructs an adjacency matrix based on a hybrid criterion of spatial proximity and feature affinity to obtain a local anatomical map.

[0037] Specifically, the client uniformly samples from the detected organ-related spatial locations to construct M anatomical nodes, forming a node matrix. To maintain a consistent embedding space, all clients use the same globally shared encoder to generate node representations.

[0038] An adjacency matrix is ​​constructed based on a hybrid criterion of "spatial proximity + feature affinity". When the normalized Euclidean distance between two nodes is below the spatial threshold Furthermore, the cosine similarity exceeds the feature threshold. Connections are established in real time. Structural reasoning is then performed using a GNN to obtain a local structural graph representation with anatomical semantics.

[0039] Finally, batch-level / sample-level local anatomical structure diagrams were obtained for the client. .

[0040] By explicitly modeling the spatial semantic relationships between organs in the form of anatomical diagrams and constructing edge connections using a hybrid criterion of spatial proximity and feature affinity, the structural expressive power is improved.

[0041] In the S500, the client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on the Wasserstein distance and using the Sinkhorn approximation to guide local training.

[0042] First, the client measures the difference in node feature distribution between the local graph and the global graph based on the Wasserstein distance metric, and uses the Sinkhorn algorithm for effective approximation.

[0043] Then, the client calculates the graph matching loss, which is used as a consistency loss to guide local training. The formula for the graph matching loss is:

[0044] in It is a set of double random matrices. It is the optimal transmission plan calculated through Sinkhorn iteration. This represents the j-th node (a d-dimensional vector).

[0045] Finally, the client minimizes the overall objective: .in It forces the alignment between the local map and the global anatomical prior, while This forces alignment between the two branches in the local client. The updated local model parameters are then obtained. With graph matching measure (i.e. graph matching loss score).

[0046] This invention achieves cross-domain anatomical alignment based on Wasserstein distance and Sinkhorn approximation, serving as both a local consistency constraint and a basis for server-side aggregation weights, thereby enhancing the anatomical fidelity of global learning.

[0047] In the S600, the client aggregates the trained local anatomical images into a domain-level anatomical image and uploads it to the server.

[0048] The client will use local graphs from all training. A single representative domain-level graph of domain i is aggregated using averaging or pooling methods. This allows for the summarization of all local anatomical structures and a reduction in communication overhead.

[0049] In this invention, the client does not upload the original features or the complete map; it only transmits the domain-level local map to the server to reduce the risk of privacy leaks and manipulation. The server obtains... .

[0050] S700, the server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation.

[0051] After obtaining the domain-level anatomical map, the server also needs to calculate the matching score. First, the difference in node feature distribution between the local and global graphs is measured based on the Wasserstein distance, and the Sinkhorn algorithm is used for effective approximation.

[0052] Then, the graph matching loss is calculated, which provides the basis for calculating the aggregation weights on the server side. The formula for the graph matching loss is:

[0053] in It is a set of double random matrices. It is the optimal transmission plan calculated through Sinkhorn iteration. This represents the j-th node (a d-dimensional vector).

[0054] To enhance security and robustness, each [item] is calculated by the server. and The graph matching score is used to prevent malicious clients from forging matching scores and interfering with the model results.

[0055] S800, the server updates the global anatomical structure diagram and global model parameters.

[0056] Specifically, the server generates aggregation weights based on the matching scores, performs a weighted average on the domain-level anatomy map to update the global anatomy map, and uses the same aggregation weights to perform weighted aggregation on the client model parameters to update the global model parameters.

[0057] Update global anatomical structure diagram: The server generates aggregate weights based on graph matching scores. A weighted average of the domain-level graphs is used to form a robust local graph. This serves as a structural prior for the next round of local training and facilitates cross-client knowledge transfer. The weights of client i... Defined as:

[0058] in This represents the local image uploaded by client i. Represents a global graph. This represents the graph matching similarity score. In the first round of communication... It is initialized to the uniform average value of all local graphs. The updated global structure graph is obtained. .

[0059] Update global model parameters: By applying a weighted federated average to the local model parameters using the same weights as graph aggregation, clients with higher structural consistency contribute more to global model updates, thereby achieving a dual consistency evolution between model parameters and the structural graph. Global model parameters Updated to:

[0060] in These are the local model weights from client i, where ∂i is the same weight used in graph aggregation. This dual consistency between model parameters and the structure graph ensures that the evolution of the global model remains consistent with the shared anatomical priors. Updated global model parameters. .

[0061] This invention improves generalization performance across centers and devices by using anatomically guided graph-model dual-weighted aggregation to ensure that the global model evolution is consistent with shared anatomical priors.

[0062] Furthermore, after updating the global anatomical structure diagram and global model parameters, it will determine whether further training is needed. Specifically, if the preset conditions are not met, the server will send the updated global model and global diagram to each client to start the next round of training, i.e., cyclically execute steps S200-S800; if the preset conditions are met, the server will output the final model and global anatomical structure prior; the preset conditions are reaching a preset number of rounds or reaching a convergence condition.

[0063] Below, we will demonstrate the effectiveness of the training method of this invention through experiments.

[0064] Dataset description and experimental setup: We used a large-scale prenatal ultrasound dataset of early and mid-to-late pregnancy from seven hospitals (AGs). This ultrasound data was collected using equipment from multiple manufacturers, including Mindray, Sonoscape, Samsung, and Philips. The gestational age range of the collected data was 11–14 weeks in the early pregnancy and 14–40 weeks in the mid-to-late pregnancy. Data collection complied with clinical ultrasound guidelines. Early pregnancy fetal screening primarily included two planes: the crown-rump length (CRL) plane and the nuchal translucency (NT) plane. Late pregnancy screening included 22 standardized planes, each containing multiple anatomical structures. For example, the 4C view included the left ventricle, left atrium, right ventricle, right atrium, ribs, descending aorta, ventricular septum, and atrial septum. All experiments involving the prenatal ultrasound datasets (including early and late pregnancy) were approved by the local hospital ethics committee. The datasets were divided into two pregnancy-related datasets: early pregnancy (EP) and mid-to-late pregnancy (M2LP). For example, the M2LP dataset contains 60 anatomical structures, 88,961 ultrasound images, and a total of 786,220 anatomical structures; the EP dataset contains 14 anatomical structures, 9,879 ultrasound images, and a total of 195,332 anatomical structures. The data annotation process was performed by experienced senior ultrasound physicians.

[0065] In our invention, our evaluation metric uses mAP (mean average precision) with IoU thresholds of 0.5 (mAP50), 0.75 (mAP75), and 0.95 (mAP95) to assess model performance at different precision levels. This is a widely accepted and robust metric for evaluating object detection tasks. For the diagnosis of single ventricle and single atrium, we leverage accuracy to diagnose the disease. We compare our method with state-of-the-art methods such as FedAvg, FedBN, FedProx, FedDyn, Fedbabu, GradMA, GPFL, FedFA, FedPAC, and CDFOD to demonstrate the effectiveness of our approach. For internal evaluation, five hospitals (A, B, C, D, and E) were selected to represent individual clients within the FL framework. Local datasets from each hospital were used for model training and internal testing to evaluate performance in the FL paradigm. This approach ensures robust evaluation of the model's generality and adaptability across different medical center datasets.

[0066] Experimental verification: Our experiments were conducted on both the EP and M2LP datasets. For internal validation, the joint learning framework was deployed across five centers (AEs), and performance was evaluated within the same institution. Experimental results demonstrate state-of-the-art detection performance on both the EP and M2LP datasets in our internal evaluation. Figure 5 To validate broader clinical applicability, we extended the framework to cardiovascular disease classification tasks. Our method achieved diagnostic accuracy of 93.07% and 85.27% for single ventricle and single atrium in two hospitals, respectively (Table 2), highlighting its practicality in various clinical identification challenges.

[0067] Table 1. Comparative analysis of cross-domain detection performance between traditional non-FL methods and our Fed-Topo method. ; For the Fed-Topo advanced detection performance experiment (Table 1), the experimental design involved three configurations to evaluate model performance and generalization ability across medical centers. (1) All-in-one training: The datasets from center A to center E were pooled together for joint model training, providing upper limit performance for combined multi-center data. (2) Single-center training and cross-center validation: The model was trained on data from one institution (e.g., center A) and evaluated on another institution (e.g., A→B), highlighting the performance degradation caused by data variation between institutions. (3) Cross-center validation with data privacy protection (ours): To address data privacy constraints, the model was collaboratively trained across institutions without sharing the original data, achieving fair cross-center validation while protecting privacy. The experimental results revealed a significant performance degradation in cross-center testing, indicating significant domain shifts between datasets from different medical centers. For example, while the model achieved 86.03% mAP50 when trained and tested on center A (A→A) of the EP dataset, its performance degraded when tested on center BG. Similarly, model performance degrades when each center in the BG is used as a training setting for predictions of other centers. Notably, the proposed method demonstrates substantial improvements in cross-domain detection performance without access to actual data from the target institution. Overall, our method outperforms models trained on only a single-center dataset (AG) in terms of mAP50 across individual test centers on the EP dataset, with average advantages of 12.50%, 13.96%, 14.01%, 14.83%, 8.17%, 15.80%, and 21.96%. Likewise, our method outperforms models trained on only a single-center dataset (AG) in terms of mAP50 on the M2LP dataset, with average advantages of 16.65%, 16.08%, 17.63%, 15.98%, 18.16%, 17.60%, and 17.66%, thus validating its effectiveness in addressing the challenges of domain adaptation. While our method falls short of the upper limit established by unified centralized training, such a paradigm remains frequently unattainable in clinical practice due to stringent constraints imposed by medical ethics and privacy regulations.

[0068] Table 2. Comparison of different diagnostic methods for single ventricle / single atrium and the anatomical accuracy of four-chamber cross-section.

[0069] ; To comprehensively assess the model's robustness across different clinical settings, we conducted internal evaluations in three configurations: central AE, CG, and AG, representing an increasingly broader range of data integration scenarios. Figure 5As shown, the proposed framework consistently outperforms the top 10 state-of-the-art federated learning methods across all configurations and evaluation thresholds. In EP, our method achieves significant progress over the FedAvg baseline, reaching up to 26.8% at mAP@75, and improves mAP@50 by 2.1–4.6% compared to the second-best method (CDFOD). A similar trend is observed for M2LP, with absolute gains of 3.0–8.5% in mAP@50 under the AG configuration, and continuous improvement at more stringent localization thresholds. These results demonstrate that the framework can capture fine-grained anatomical variations and maintain high detection fidelity in heterogeneous structure data.

[0070] in conclusion: Our Fed-Topo method can be deployed in the cloud to detect anatomical structures across multiple hospitals, achieving high-precision detection performance while protecting the privacy of datasets from various medical institutions, thus effectively serving clinical tasks, such as... Figure 6 As shown in the figure. In this invention, we propose Fed-Topo, a novel federated learning framework that incorporates anatomical topological priors for large-scale, multi-center prenatal anatomical structure detection. Our topology-aware federated paradigm demonstrates superior performance in bridging cross-institutional domain transfers while maintaining strict data privacy. Comprehensive evaluations on intra-center and inter-center test sets show that Fed-Topo improves average detection accuracy by 3.7–25.4% compared to isolated-center training, and outperforms existing federated learning baselines by 1.5–8.2% in cross-domain generalization mAP50. Notably, the framework shows particular promise for the early detection of congenital heart defects, achieving 93.1% accuracy in identifying ventricular septal defects from a standard four-chamber view. This breakthrough demonstrates the potential of Fed-Topo as a collaborative medical imaging analysis infrastructure, particularly in resource-constrained prenatal care scenarios.

[0071] Our method also achieved state-of-the-art performance in diagnosing structural absence abnormalities (single ventricle and single atrium) at two independent medical centers. Diagnostic accuracies reached 93.07% (center F) and 85.27% (center E), representing significant improvements of 23.91% and 19.89% respectively compared to the FedAvg method. These promising results demonstrate that integrating topological knowledge into federated learning can improve the accuracy of disease diagnosis. This improvement is likely due to the difference in topological features between pathological and normal images, allowing for effective differentiation. As shown in Table 2, the high accuracy in anatomical structure and disease diagnosis underscores the effectiveness and potential of our method. Furthermore, this method could be extended in the future to diagnose other diseases caused by anatomical absences, such as renal dysplasia (absence of the kidney).

[0072] In summary, this application initializes global model parameters and a global anatomical structure map on the server; the server distributes the global model parameters and the global anatomical structure map to each client; the client calculates the detection loss based on local data and performs feature enhancement on the local data; the client builds a local anatomical map; the client uses the global anatomical map to supervise the local anatomical map, calculating the matching score between the global anatomical structure map and the local anatomical map locally based on Wasserstein distance and using Sinkhorn approximation to guide local training; the client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; the server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on Wasserstein distance and using Sinkhorn approximation; the server updates the global anatomical structure map and global model parameters. This solves the technical problems of existing traditional centralized or classic FL methods, which struggle to simultaneously consider global performance and local adaptability, neglect topological priors between organ structures, resulting in limited generalization ability in complex structure recognition and a lack of deep integration with medical knowledge; it improves the model's generalization ability in complex structures and enhances the deep integration of the model with medical knowledge.

[0073] At least some steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0074] Based on the same inventive concept, this application also provides an ultrasound image federated learning system for implementing the aforementioned fusion of anatomical prior knowledge. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more ultrasound image federated learning system embodiments for fusing anatomical prior knowledge provided below can be found in the limitations of the ultrasound image federated learning method for fusing anatomical prior knowledge described above, and will not be repeated here.

[0075] In one embodiment, such as Figure 4 As shown, a federated learning system for ultrasound images that integrates prior anatomical knowledge is provided, including: Initialization module 100 is used to initialize the global model parameters and global anatomical structure diagram of the server.

[0076] The sending module 200 is used by the server to send global model parameters and global anatomical structure diagrams to each client.

[0077] The first processing module 300 is used by the client to calculate the detection loss based on local data and to perform feature enhancement on the local data.

[0078] Module 400 is created for clients to create local anatomical diagrams.

[0079] The second processing module 500 is used by the client to supervise the local anatomical map using the global anatomical map, and to guide local training by calculating the matching score between the global anatomical map and the local anatomical map based on the Wasserstein distance and using the Sinkhorn approximation.

[0080] The upload module 600 is used by the client to aggregate the trained local anatomical images into a domain-level anatomical image and upload it to the server.

[0081] The third processing module 700 is used by the server to calculate the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation.

[0082] Update module 800 is used to update the global anatomical structure diagram and global model parameters on the server.

[0083] In one implementation, the first processing module 300 is configured to: calculate the detection loss based on local data and perform feature enhancement on the positive sample ROI consisting of an online network and a contrast target network, wherein the feature enhancement updates the contrast target network through an exponential moving average EMA and minimizes the Kullback-Leibler divergence of the softmax normalized ROI feature distribution.

[0084] In one implementation, the establishment module 400 is further configured to: extract multi-scale features based on the feature pyramid network, construct anatomical nodes by uniformly sampling the spatial locations of organs, and construct an adjacency matrix based on a hybrid criterion of spatial proximity and feature affinity to obtain a local anatomical map.

[0085] In one implementation, the update module 800 is further configured to: generate aggregate weights based on the matching scores, perform weighted averaging on the domain-level anatomy map to update the global anatomy map, and use the same aggregate weights to perform weighted aggregation on the client model parameters to update the global model parameters.

[0086] In one embodiment, the sending module 200 is further configured to: when a preset condition is not met, the server sends the updated global model and global graph to each client to start the next round of training; when the preset condition is met, the server outputs the final model and global anatomical structure prior; the preset condition is reaching a preset number of rounds or reaching a convergence condition.

[0087] In one implementation, the first processing module 300 is further configured to: perform forward prediction on the input image based on the shared encoder and the detection head and calculate a standard detection loss, wherein the standard detection loss includes classification loss and localization loss; The online network and the target comparison network form a two-branch structure; The target network parameters are updated using the exponential moving average (EMA) of the online network parameters. Apply feature consistency constraints to positive ROIs; Minimize the Kullback-Leibler (KL) divergence between the feature distributions of the softmax normalized ROI to encourage enhanced invariance.

[0088] The modules in the aforementioned ultrasound imaging federated learning system that integrates prior anatomical knowledge can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0089] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores preset data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a federated learning method for ultrasound images that integrates prior anatomical knowledge.

[0090] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0091] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the above-described federated learning method for ultrasound images that integrates prior anatomical knowledge.

[0092] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described federated learning method for ultrasound images that integrates prior anatomical knowledge.

[0093] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described federated learning method for ultrasound images that integrates prior anatomical knowledge.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An ultrasound image federated learning method fusing anatomical prior knowledge, characterized in that, The method includes: The server initializes global model parameters and global anatomical structure diagram; The server sends global model parameters and global anatomical structure diagrams to each client. The client calculates the detection loss based on local data and performs feature enhancement on the local data; The client creates a local anatomical diagram; The client uses a global anatomical map to supervise the local anatomical map, and locally calculates the matching score between the global anatomical map and the local anatomical map based on Wasserstein distance and using Sinkhorn approximation to guide local training; The client aggregates the trained local anatomical maps into a domain-level anatomical map and uploads it to the server; The server calculates the matching score between the domain-level anatomical map and the global anatomical structure map based on the Wasserstein distance and using the Sinkhorn approximation. The server updates the global anatomical structure diagram and global model parameters.

2. The method of claim 1, wherein, The client calculates the detection loss based on local data and performs feature enhancement on the local data, including: The client calculates the detection loss based on local data and performs feature enhancement on positive ROIs consisting of an online network and a contrastive target network. The feature enhancement updates the contrastive target network using an exponential moving average (EMA) and minimizes the Kullback-Leibler divergence of the softmax normalized ROI feature distribution.

3. The method of claim 1, wherein, The client establishes a local anatomical diagram, including: The client extracts multi-scale features based on the feature pyramid network, constructs anatomical nodes by uniformly sampling the spatial locations of organs, and constructs an adjacency matrix based on a hybrid criterion of spatial proximity and feature affinity to obtain a local anatomical map.

4. The method of claim 1, wherein, The server updates the global anatomical structure diagram and global model parameters, including: The server generates aggregate weights based on the matching scores, performs a weighted average on the domain-level anatomy map to update the global anatomy map, and uses the same aggregate weights to perform weighted aggregation on the client model parameters to update the global model parameters.

5. The method of claim 1, wherein, The method further includes: When the preset conditions are not met, the server will send the updated global model and global graph to each client to start the next round of training; When the preset conditions are met, the server outputs the final model and the global anatomical structure prior. The preset condition is reaching a preset number of rounds or reaching a convergence condition.

6. The method of claim 2, wherein, The client calculates the detection loss based on local data and performs feature enhancement on positive ROIs using an online network and a contrastive target network. This feature enhancement updates the contrastive target network using an exponential moving average (EMA) and minimizes the Kullback-Leibler divergence of the softmax-normalized ROI feature distribution. This includes: The input image is forward-predicted based on a shared encoder and a detection head, and a standard detection loss is calculated, which includes classification loss and localization loss. The online network and the target comparison network form a two-branch structure; The target network parameters are updated using the exponential moving average (EMA) of the online network parameters. Apply feature consistency constraints to positive ROIs; Minimizing the Kullback-Leibler (KL) divergence between softmax-normalized ROI feature distributions to encourage enhanced invariance.

7. An ultrasound image federated learning system fusing anatomical prior knowledge, characterized in that, The system comprises: An initialization module configured to initialize, by a server, a global model parameter and a global anatomical graph; A sending module configured to send, by the server, the global model parameter and the global anatomical graph to each client; A first processing module configured to calculate, by a client, a detection loss based on local data and perform feature enhancement on the local data; An establishing module configured to establish, by the client, a local anatomical graph; A second processing module configured to supervise, by the client, the local anatomical graph using the global anatomical graph, and calculate, based on a Wasserstein distance and using a Sinkhorn approximation, a matching score of the global anatomical graph and the local anatomical graph to guide local training; An uploading module configured to aggregate, by the client, a trained local anatomical graph into a domain-level anatomical graph and upload the domain-level anatomical graph to the server; A third processing module configured to calculate, by the server, a matching score of the domain-level anatomical graph and the global anatomical graph based on the Wasserstein distance and using the Sinkhorn approximation; An updating module configured to update, by the server, the global anatomical graph and the global model parameter.

8. A computer device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

Images