Domain generalization medical image segmentation method and system based on prior knowledge

Abstract

The present invention provides a domain-generalized medical image segmentation method and system based on prior knowledge, which belongs to the technical field of clinical intelligent medical equipment. During each round of training, each client receives the same global model weight from a central server, and updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning. The central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model. The present invention is based on a domain-generalized method to train a universal model, learn information from multiple source domains, and can show good segmentation performance in unknown domains without the need for additional learning; in the model learning process, shape-based prior knowledge is integrated into the segmentation vector to capture the rich domain knowledge between each source domain and narrow the domain differences; based on prior knowledge, the output of the model in the unknown domain is optimized to enhance the visualization effect of the model and calculate the morphological characteristics of the segmentation results.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical intelligent medical equipment, and in particular to a domain-generalized medical image segmentation method and system based on prior knowledge. Background Art

[0002] The goal of domain generalization-based models is to improve the generalization ability of convolutional neural networks to unknown target domains by learning knowledge from multiple source domains. The typical design framework for domain generalization-based models is to extract features from multiple datasets with domain differences and characterize the domain shifts present in specific medical imaging using computer-related technologies (such as machine learning and deep learning methods), thereby improving the model's generalization ability. The main methods for constructing existing domain generalization-based medical image segmentation models are as follows:

[0003] Traditional domain generalization methods aim to minimize domain differences between multiple source domains to learn domain-invariant representations. For example, Motiian et al. used contrastive loss to minimize the distance between samples from the same category but different domains. YC Hsu et al. formulated this as a clustering problem. In addition to features, similarity information is transmitted to the model, allowing the model to learn similarity functions and clustering networks to perform domain generalization tasks. H Li et al. applied the maximum mean difference metric to align the distributions between different domains. These strategies provide ideas for addressing domain differences in medical imaging. However, such strategies typically require the use of multi-domain datasets for learning, lacking privacy protection for dispersed datasets.

[0004] Unlike domain generalization on natural images, image enhancement methods can better address domain differences between different medical image datasets. Zhao et al. used a model that can learn independent spatial and appearance transformations to capture changes such as nonlinear deformation and imaging intensity in different medical imaging datasets, and then used these new examples to synthesize new labeled examples, thereby expanding the labeled dataset and promoting the progress of existing work on brain MRI image segmentation. Chen et al. carefully designed a data normalization and enhancement strategy to improve the generalization ability of convolutional neural networks. BigAug enables the model to better generalize to unknown domains by stacking more transformations. These methods do not require data centralization and can be used as regularization for local training with a single source domain data. However, there is still a lack of better utilization and learning of rich data distribution knowledge across domains.

[0005] Liu et al. performed contextual learning in continuous frequency space, connecting and exploiting the different distributions of multiple source domains. Wang et al. proposed a DoFE framework to embed domain prior knowledge into the network of interest. DoFE uses an attention mechanism to combine memory domain features with image features for generalizable fundus image segmentation.

[0006] Traditional domain generalization models typically improve generalization by minimizing domain differences between multiple source domains to learn domain-invariant representations. This approach requires the aggregation of multi-domain datasets for learning, lacking privacy protection for dispersed datasets. Furthermore, model performance depends on the quality and quantity of the training dataset. While traditional data augmentation methods can address the data centralization issue, they fail to effectively leverage the rich domain knowledge in diverse datasets. These strategies typically rely solely on the network to learn feature representations in the data, without integrating prior knowledge to optimize feature representations. Summary of the Invention

[0007] The object of the present invention is to provide a domain-generalized medical image segmentation method and system based on prior knowledge to solve at least one technical problem existing in the above-mentioned background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In one aspect, the present invention provides a domain-generalized medical image segmentation method based on prior knowledge, comprising:

[0010] Obtain the image to be segmented;

[0011] The image to be segmented is processed using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0012] Optionally, the aggregation method is the federated average algorithm, and the aggregated parameter θ is expressed as:

[0013] N is the number of local clients participating in the training, and the aggregation weight depends on the number of samples of each client.

[0014] Optionally, the local client trains the local model, and the shared information is integrated on the central server. Considering that sharing the original image is prohibited, there are N clients in total, and the image stored on each client is recorded as Each sample first obtains low-level distribution information and high-level semantic information in the image through image transformation, and uploads the low-level distribution information to the central server for sharing; obtains the general knowledge issued by the central server, and uses the client feature pool to simulate and generate images with different domain differences, which are recorded as Use general knowledge to constrain the learning process and train the model.

[0015] Optionally, the ultimate goal of each local client training is to minimize the segmentation loss and prior knowledge loss The model parameters θ of the sum k , expressed as:

[0016]

[0017] Refers to the classic dice loss in medical images. The initial parameters of the local segmentation model are updated by dice loss to obtain The calculation is expressed as:

[0018]

[0019] First, we segment the domain difference image S k , calculate the dice loss, based on prior knowledge, calculate the compactness loss and shape loss

[0020] Optional, compactness loss Is to ensure that the segmentation result is a closed graph, m i Segmentation for the model The output of P is the segmentation shape m i The perimeter of the segmented shape m i The area of ​​the ​​object, λ is a hyperparameter, and its value is determined by the experimental settings. Expressed as:

[0021]

[0022] Optional, shape loss It is to maximize the difference between the edge and background of the segmentation result, with the help of model segmentation and As a result, the contour feature h is extracted bd and background features h bg , there are 2N shape features in total, from which two shape features h are randomly selected j and h t , if h j and h tare all contour features or background features, then sgn(h j ,h t ) is 1, otherwise it is 0, and h is calculated based on whether they belong to the same category. j ,h t The contrast loss InfoNCE between them is recorded as Expressed as:

[0023]

[0024] After each round of local client learning is completed, the local parameters θ from all clients are k The global model parameters θ will be aggregated on the central server.

[0025] In a second aspect, the present invention provides a domain-generalized medical image segmentation system based on prior knowledge, comprising:

[0026] An acquisition module, used for acquiring an image to be segmented;

[0027] The segmentation module is used to process the image to be segmented using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weights from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0028] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the domain-generalized medical image segmentation method based on prior knowledge as described above is implemented.

[0029] In a fourth aspect, the present invention provides a computer device comprising a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the domain-generalized medical image segmentation method based on prior knowledge as described above.

[0030] In a fifth aspect, the present invention provides an electronic device comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the domain generalized medical image segmentation method based on prior knowledge as described above.

[0031] The beneficial effects of the present invention are as follows: Based on the domain generalization method, a universal model is trained to learn information from multiple source domains (organizations), and can show good segmentation performance in unknown domains without the need for additional learning. In the process of model learning, shape-based prior knowledge is integrated into the segmentation vector, allowing the model to capture rich domain knowledge between various source domains and narrow domain differences. Based on prior knowledge, the output of the model in the unknown domain is optimized to enhance the visualization effect of the model, and the morphological characteristics of the segmentation results are calculated, including indicators such as area, perimeter, and circularity.

[0032] Additional advantages of the present invention will be more clearly given in the following description or learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 This is a functional framework diagram of the domain-generalized medical image segmentation system based on prior knowledge described in an embodiment of the present invention.

[0035] Figure 2 Schematic diagram of four public fundus retinal image datasets from different scanning instruments or clinical centers used as standard datasets according to an embodiment of the present invention.

[0036] Figure 3 Schematic diagram of qualitative comparison of generalization results of incorporating prior knowledge in optic disc and optic cup segmentation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0038] Those skilled in the art will understand that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0039] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0040] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0041] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.

[0042] Domain generalization is the process of learning a model from one or more related but distinct domains and applying the model generalization to unknown test domains. This property has attracted widespread attention in research areas such as healthcare and smart medicine. Typically, in actual clinical work, medical images are often affected by different institutions, different types of scanning equipment, and different individual patients, exhibiting differences in image quality, field of view, and appearance. To better adapt to images from different distributions, the ideal solution is to collect data from all possible institutions to train the model. However, the process of collecting and annotating medical imaging data from multiple institutions is very expensive and time-consuming, and the method of combining data from multiple institutions for training ignores the privacy of patient data.

[0043] Prior knowledge is a preconceived assumption or model based on past experience and known information. By incorporating prior knowledge into image processing algorithms, it can provide additional constraints and guidance for the model, improving performance in tasks such as image segmentation, registration, and feature extraction. The application of prior knowledge can help deep learning methods better adapt to different types of image data and patient variations, playing a crucial role in computer-assisted medical image analysis.

[0044] How to incorporate shape-based prior knowledge into segmentation masks during the learning process is a research hotspot in the field of medical image segmentation. This paper considers the characteristics of domain generalization, extracts valuable features between multiple source domains (organisms), integrates existing shape prior knowledge, supplements and constrains it during model training, and alleviates the impact of data imbalance between different domains, thereby obtaining a more robust model that generalizes to unknown domains.

[0045] This paper proposes a domain-generalized medical segmentation model based on prior knowledge, which takes advantage of domain generalization to reduce the dependence of traditional convolutional neural networks on datasets, and further integrates prior knowledge of shape-related features to improve the generalization performance and interpretability of the model in unknown domains.

[0046] Example 1

[0047] In this embodiment 1, a domain-generalized medical image segmentation system based on prior knowledge is first provided, including: an acquisition module for acquiring an image to be segmented; a segmentation module for processing the image to be segmented using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from a central server, updates local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0048] In this embodiment 1, the above-mentioned system is used to implement a domain-generalized medical image segmentation method based on prior knowledge, including: obtaining an image to be segmented; processing the image to be segmented using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, the central server obtains the local model parameters of all clients, and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0049] Among them, the aggregation method for local model parameters is the federated averaging algorithm, and the aggregated parameter θ is expressed as: N is the number of local clients participating in the training, and the aggregation weight depends on the number of samples of each client.

[0050] The local client trains the local model, and the shared information is integrated on the central server. Considering that sharing the original image is prohibited, there are N clients in total, and the image stored on each client is recorded as Each sample first obtains low-level distribution information and high-level semantic information in the image through image transformation, and uploads the low-level distribution information to the central server for sharing; obtains the general knowledge issued by the central server, and uses the client feature pool to simulate and generate images with different domain differences, which are recorded as Use general knowledge to constrain the learning process and train the model.

[0051] The ultimate goal of each local client training is to minimize the segmentation loss and prior knowledge loss The model parameters θ of the sum k , expressed as:

[0052]

[0053] Refers to the classic dice loss in medical images. The initial parameters of the local segmentation model are updated by dice loss to obtain The calculation is expressed as:

[0054]

[0055] First, we segment the domain difference image S k , calculate the dice loss, based on prior knowledge, calculate the compactness loss and shape loss

[0056] Compactness loss Is to ensure that the segmentation result is a closed graph, m i Segmentation for the model The output of P is the segmentation shape m i The perimeter of the segmented shape m i The area of ​​the ​​object, λ is a hyperparameter, and its value is determined by the experimental settings. Expressed as:

[0057]

[0058] Shape loss It is to maximize the difference between the edge and background of the segmentation result, with the help of model segmentation and As a result, the contour feature h is extracted bd and background features h bg , there are 2N shape features in total, from which two shape features h are randomly selected j and h t , if h j and h t are all contour features or background features, then sgn(hj ,h t ) is 1, otherwise it is 0, and h is calculated based on whether they belong to the same category. j ,h t The contrast loss InfoNCE between them is recorded as Expressed as:

[0059]

[0060] After each round of local client learning is completed, the local parameters θ from all clients are k The global model parameters θ will be aggregated on the central server.

[0061] Example 2

[0062] In this embodiment 2, a domain generalization medical segmentation model PKDG (DomainGeneralization Medical Segmentation Model Based on Prior Knowledge) is proposed. Figure 1 As shown. In the framework of this method, PKDG mainly simulates the domain offset existing in multiple domains in the network by combining multi-source distribution characteristics in each source domain (local clients 1~N), enriches the diversity of data, and uses prior knowledge to constrain the training of models in each source domain, and timely integrates the multi-source trained models, and finally applies them to the unknown domain (client N+1) to test the generalization performance of the model. In this embodiment, prior knowledge is integrated into the medical imaging model based on domain generalization. Under the premise of protecting the privacy of each source domain data, the information in the multi-source domain is fully utilized to improve the generalization ability of the model, and excellent segmentation performance is achieved on the existing standard dataset.

[0063] (1) Technical framework

[0064] This method framework consists of three modules. The specific process is described as follows:

[0065] 1) Central server integration module

[0066] The central server-side integration module mainly realizes the exchange of distribution information across clients, so that each local client can access multi-source data distribution, and train the global segmentation network to learn generalizable parameters. The prior knowledge integrated by the central server includes the client feature pool and general knowledge. The client feature pool consists of the low-level distribution information of each local client. General knowledge includes information such as shape, compactness, and quantity. During each round of training, each client will receive the same global model weights from the central server, and with the help of the prior knowledge integrated by the central server, combine local data learning to update the model parameters. The central server will then collect local parameters from all clients and aggregate them to update the global model. The aggregation method in this technology selects the federated average algorithm, which can be expressed as:

[0067]

[0068] N is the number of local clients participating in the training, and the aggregation weight depends on the number of samples from each client. To improve the generalization ability of the global model, we integrate prior knowledge under the premise of the same task and different data, optimize the visualization and interpretability of the model output, and obtain the final prediction results.

[0069] 2) Local client training module

[0070] In order to protect the privacy of each dataset, the local client module needs to be trained locally, and the shared information is integrated at the center. Considering that sharing original images is prohibited, suppose there are N clients in total, and the images stored on each client are recorded as Each sample first obtains low-level distribution information (style) and high-level semantic information in the image through image transformation, and uploads the low-level distribution information to the central server for sharing. Then, by obtaining the knowledge issued by the central server, the client feature pool is used to simulate and generate images with different domain differences, which are recorded as The general knowledge is used to constrain the learning process and train the model. The parameters of each round of training are uploaded to the central end for aggregation. The ultimate goal of each local client training is to minimize the segmentation loss. and prior knowledge loss The model parameters θ of the sum k , which can be expressed as:

[0071]

[0072] Refers to the classic dice loss in medical images. The initial parameters of the local segmentation model are updated by dice loss to obtain The calculation can be expressed as:

[0073]

[0074] First, we segment the domain difference image S k , calculate the dice loss, and then calculate the compactness loss based on prior knowledge and shape loss Compactness loss The main function is to ensure that the segmentation result is a closed graph, where m i Segmentation for the model The output of P is the segmentation shape m i The perimeter of the segmented shape m i The area of ​​the ​​object, λ is a hyperparameter, and its value depends on the experimental settings and can be expressed as:

[0075]

[0076] Shape loss In order to better maximize the difference between the edge and background of the segmentation result, the model segmentation and As a result, the contour feature h is extracted bd and background features h bg , there are 2N shape features (contour features, background features), from which two shape features h are randomly extracted j and h t , if h j and h t are all contour features or background features, then sgn(h j ,h t ) is 1, otherwise it is 0, and h is calculated based on whether it belongs to the same category (contour, background) j ,h t The contrast loss InfoNCE between them is recorded as It can be expressed as:

[0077]

[0078] After each round of local client learning is completed, the local parameters θ from all clients are k will be aggregated on a central server to update the global model parameters θ.

[0079] 3) Local client migration module

[0080] The generalization capability of the global model is verified by migrating data from clients that did not participate in training (unknown domains). By feeding the untrained data into the model delivered by the server, segmentation results are obtained. Prior knowledge is used to enhance the visualization of the segmentation results, and their morphological features are calculated to provide an interpretable scale to facilitate downstream tasks.

[0081] (2) Experimental verification

[0082] 1) Benchmark Dataset

[0083] This technology uses four public fundus retinal image datasets, 'Drishti-GS', 'RIM-ONE-r3', 'REFUGE (train)', and 'REFUGE (val)', from different scanning instruments or clinical centers as standard datasets. Figure 2 As shown in Figure 2. These data are pre-processed into a size of 384×384 in order to serve as network input.

[0084] 2) Performance comparison on the dataset

[0085] The goal of this technology is to improve the generalization performance of the model in medical image segmentation. In order to evaluate the performance of the method, we use the dice similarity coefficient (dice) as the evaluation indicator of the experiment, which can be expressed as:

[0086]

[0087] Where A and B are the segmentation result and segmentation mask respectively. The overall performance comparison of the proposed model PKDG and other baseline methods is shown in Table 1.

[0088] Based on these results, PKDG achieved the highest average dice in the optic disc and cup segmentation tasks, reaching 92.85% and 87.22%, respectively, significantly improving prediction performance compared to other models. Table 1 compares these methods from two perspectives: domain generalization methods commonly used in natural images and domain generalization methods commonly used in medical imaging.

[0089] Table 1 Performance comparison of models on the optic cup / disc segmentation task (%) (A = Drishti-GS, B = RIM-ONE-r3, C = REFUGE (train), D = REFUGE (val))

[0090]

[0091] This model is compared with two classic domain generalization methods for natural images: M-Mixup and CutMix. Observing the tabular data, CutMix achieves the highest scores on some tasks. This is because CutMix centrally learns information from multiple source domains during training without considering data privacy. As a result, it learns more features, resulting in higher performance.

[0092] We then compared PKDG with two DG methods for medical images, BigAug and ELCFS. To simulate data independence, BigAug, ELCFS, and PKDG all used the FedAvg aggregation method to update global model parameters. Compared to FedAvg, we found that different domain generalization methods can more or less improve overall generalization performance. BigAug performs data augmentation on images by stacking a stack of carefully designed image transformations. Compared to domain generalization methods for natural images, it shows some improvement in performance. However, the new domains generated by this image augmentation method do not fully cover all used datasets, resulting in poor generalization performance on some domains. ELCFS can generate images with features from other domains through a frequency-space interpolation mechanism, significantly improving its performance compared to other models. Our PKDG further improves ELCFS by integrating prior knowledge during training, enhancing the model's learning of domain-invariant features, and optimizing segmentation results during generalization. Ultimately, PKDG's performance is slightly better than ELCFS. Although the improvement is small, the visualization and solvability of the results are significantly improved due to the integration of prior knowledge during generalization.

[0093] In order to further observe the impact of prior knowledge on the generalization effect of the optic disc and cup segmentation task, and compare the different visualization effects after incorporating knowledge in the training and generalization processes, we designed PKDG-V1, which only incorporates knowledge in the training model, to assist in demonstrating the importance of incorporating knowledge in the generalization process. Figure 3 As shown in the figure, we can see that when the segmentation is basically correct, prior knowledge can smooth the edges of the segmentation results, but the performance improvement is subtle. When the segmentation performance is average or poor, optimizing the shape of the segmented region based on prior knowledge (shape, number, compactness, etc.) significantly improves generalization performance and visualization effects compared to PKDG-V1. Therefore, incorporating prior knowledge into the generalization process plays a role in improving performance and enhancing visualization effects.

[0094] Example 3

[0095] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, a domain-generalized medical image segmentation method based on prior knowledge is implemented. The method includes:

[0096] Obtain the image to be segmented;

[0097] The image to be segmented is processed using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0098] Example 4

[0099] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute a domain-generalized medical image segmentation method based on prior knowledge, the method comprising:

[0100] Obtain the image to be segmented;

[0101] The image to be segmented is processed using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0102] Example 5

[0103] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing a domain-generalized medical image segmentation method based on prior knowledge. The method includes:

[0104] Obtain the image to be segmented;

[0105] The image to be segmented is processed using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, updates the local model parameters based on the prior knowledge integrated by the central server and combined with local data learning, and the central server obtains the local model parameters of all clients and aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information.

[0106] In summary, the embodiment of the present invention proposes a method based on domain generalization, which learns information from multiple source domains and integrates prior knowledge in the model training process to obtain a robust model that can perform good segmentation performance in unknown domains; by optimizing the shape of the output of the model generalized to the unknown domain based on prior knowledge, the visualization effect of the model is improved, the morphological features of the segmentation results are provided, and the interpretability of the results is enhanced. The present invention proposes for the first time a domain generalized medical image segmentation model PKDG based on prior knowledge. Experimental results show that the performance of this method is better than the existing baseline method. For the first time, prior knowledge is integrated into the model generalization process, which enhances the visualization and interpretability of the model output and effectively improves the generalization ability of the model on different data sets for the same task.

[0107] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide the functions for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0110] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solutions disclosed in the present invention without the need for creative work should be included in the scope of protection of the present invention.

Claims

1. A domain-generalized medical image segmentation method based on prior knowledge, characterized in that: include: Obtain the image to be segmented; The image to be segmented is processed using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, based on the prior knowledge integrated by the central server, combined with local data learning to update the local model parameters, the central server obtains the local model parameters of all clients, aggregates them to update the global segmentation model; wherein, the prior knowledge includes a client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information; wherein, the local client trains the local model, and the shared information is integrated on the central server; considering that the sharing of the original image is prohibited, it is assumed that there are N clients in total, and the image stored on each client is recorded as Each sample first obtains low-level distribution information and high-level semantic information in the image through image transformation, and uploads the low-level distribution information to the central server for sharing; obtains the general knowledge issued by the central server, and uses the client feature pool to simulate and generate images with different domain differences, which are recorded as Use general knowledge to constrain the learning process and train the model; the ultimate goal of each local client training is to minimize the segmentation loss and prior knowledge loss The model parameters θ of the sum k , expressed as: Refers to the classic dice loss in medical images. The initial parameters of the local segmentation model are updated by dice loss to obtain The calculation is expressed as: First, we segment the domain difference image S k , calculate the dice loss, based on prior knowledge, calculate the compactness loss and shape loss 2. The domain generalization medical image segmentation method based on prior knowledge according to claim 1, characterized in that: The aggregation method is the federated average algorithm, and the aggregated parameter θ is expressed as: N is the number of local clients participating in the training, and the aggregation weight depends on the number of samples of each client.

3. The domain generalization medical image segmentation method based on prior knowledge according to claim 1, characterized in that: Compactness loss Is to ensure that the segmentation result is a closed graph, m i Segmentation for the model The output of P is the segmentation shape m i The perimeter of the segmented shape m i The area of ​​the ​​object, λ is a hyperparameter, and its value is determined by the experimental settings. Expressed as:

4. The domain generalization medical image segmentation method based on prior knowledge according to claim 3, characterized in that: Shape loss It is to maximize the difference between the edge and background of the segmentation result, with the help of model segmentation and As a result, the contour feature h is extracted bd and background features h bd , there are 2N shape features in total, from which two shape features h are randomly selected j and h t , if h j and h t are all contour features or background features, then sgn(h j ,h t ) is 1, otherwise it is 0, and h is calculated based on whether they belong to the same category. j ,h t The contrast loss InfoNCE between them is recorded as Expressed as: After each round of local client learning is completed, the local parameters θ from all clients are k will be aggregated on a central server to update the global model parameters θ.

5. A domain-generalized medical image segmentation system based on prior knowledge, characterized in that: include: An acquisition module, used for acquiring an image to be segmented; The segmentation module is used to process the image to be segmented using a pre-trained global segmentation model to obtain a segmentation result; wherein, the training of the global segmentation model includes: in each round of training, each client receives the same global model weight from the central server, based on the prior knowledge integrated by the central server, combined with local data learning to update the local model parameters, the central server obtains the local model parameters of all clients, and aggregates them to update the global segmentation model; wherein, the prior knowledge includes the client feature pool and general knowledge, the client feature pool is composed of the low-level distribution information of each local client, and the general knowledge includes shape, compactness and quantity information; wherein, the local client trains the local model, and the shared information is integrated on the central server; considering that the sharing of the original image is prohibited, it is assumed that there are N clients in total, and the image stored on each client is recorded as Each sample first obtains low-level distribution information and high-level semantic information in the image through image transformation, and uploads the low-level distribution information to the central server for sharing; obtains the general knowledge issued by the central server, and uses the client feature pool to simulate and generate images with different domain differences, which are recorded as Use general knowledge to constrain the learning process and train the model; the ultimate goal of each local client training is to minimize the segmentation loss and prior knowledge loss The model parameters θ of the sum k , expressed as: Refers to the classic dice loss in medical images. The initial parameters of the local segmentation model are updated by dice loss to obtain The calculation is expressed as: First, we segment the domain difference image S k , calculate the dice loss, based on prior knowledge, calculate the compactness loss and shape loss 6. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by the processor, the domain-generalized medical image segmentation method based on prior knowledge is implemented as described in any one of claims 1 to 4.

7. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the domain generalization medical image segmentation method based on prior knowledge as described in any one of claims 1 to 4.

8. An electronic device, characterized in that: include: A processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the domain generalized medical image segmentation method based on prior knowledge as described in any one of claims 1 to 4.

Images