A Cardiovascular Segmentation Method and System Based on Data-Driven Iterative Learning

Abstract

The present disclosure provides a cardiovascular segmentation method and system based on data-driven iterative learning, which relates to the technical field of image segmentation and includes: inputting the cardiovascular image to be segmented into the learned segmentation model to obtain a segmentation result; wherein, the segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set trained in this round, according to the blur degree and edge pixel ratio of the image patches, a non-edge set and an edge set are constructed; then, the non-edge set is first used to train the segmentation model under the constraint of the segmentation loss function. Based on the model parameters after training with the non-edge set, the edge set is then used to continue training under the coordination of the distribution alignment loss function. Finally, the model parameters trained in this round are used for the parameter initialization of the next round; the present invention realizes high-precision cardiovascular image segmentation with high training efficiency to promote the early diagnosis, treatment planning and efficacy evaluation of cardiovascular diseases, and at the same time adapts to the clinical environment with limited computing resources.

Description

Technical Field

[0001] The present disclosure relates to the field of image segmentation technology, and particularly to a cardiovascular segmentation method and system based on data-driven iterative learning. Background Art

[0002] With the rapid development of medical imaging technology, various imaging modalities provide rich data support for the diagnosis and research of cardiovascular diseases and have become an indispensable tool in clinical diagnosis and disease research; for example, in the diagnosis of coronary heart disease, CT angiography (CTA) can accurately detect the degree of stenosis of the coronary arteries and provide an important basis for the selection of treatment plans. However, due to the complexity of the heart structure, such as different chambers of the heart having different shapes and functions, and tissues such as myocardium and blood vessels being intertwined with each other, the segmentation of the cardiovascular system in CT images becomes extremely challenging; in addition, there are differences in the contrast between different tissues in CT images, for example, the contrast between myocardium and blood vessel walls and the blood in the heart chambers is different, which may lead to difficulties in accurately defining the boundaries between the cardiovascular structure and surrounding tissues during the segmentation process.

[0003] In recent years, certain achievements have been made in the research on cardiovascular segmentation based on traditional methods; for example, the region-growing-based method (such as application publication number: CN117115186A) locates the cardiovascular system by gradually growing regions in the image according to criteria such as similarity, but this method is sensitive to the selection of seed points, vulnerable to complex image features and noise, and the stability of the segmentation results is poor; another example is a cardiovascular detection system and method based on image analysis combined with electrocardiogram information (application publication number: CN118212237A), which locates the position of the cardiovascular system with the help of electrocardiogram information, but the matching between electrocardiogram and image information may not be precise enough, affecting the segmentation accuracy; there is also a method for intelligent segmentation of cardiovascular CT images (application publication number: CN117952994A), which performs segmentation based on relevant feature parameters of region blocks, but the computational complexity is relatively high, facing challenges in actual clinical applications.

[0004] With the rapid development of deep learning technology, image segmentation methods based on deep learning have made remarkable progress in the field of medical imaging, providing new solutions for cardiovascular segmentation. Compared with traditional segmentation methods, deep learning methods can automatically learn complex features in images without relying on manually designed features, thus significantly improving the segmentation accuracy. For example, segmentation models based on convolutional neural networks (such as U-Net, DeepLab, etc.) can effectively capture the detailed information of the cardiovascular system through multi-level convolutional and pooling operations. In addition, models based on Transformer (such as ViT, Swin Transformer, etc.) further enhance the model's ability to model long-range dependencies through the global attention mechanism and show stronger robustness when dealing with cardiovascular structures. However, deep learning models also face challenges such as long training time, high computational resource requirements, and difficult clinical deployment. In this context, developing a CTA image cardiovascular segmentation system with both high segmentation accuracy and high training efficiency not only has important academic value but also has profound clinical application significance.

[0005] Currently, developing such a system faces the following key problems: The density difference between the blood vessel wall and the surrounding tissues is small, which appears as an edge region with unclear and irregular edges in medical images, and the rest is the non-edge region. The boundary of the edge region is unclear, making it difficult to extract features and segment, while the features of the non-edge region are clear. Existing methods usually train the two together, resulting in the model needing to process two completely different feature distributions simultaneously, increasing the training difficulty and reducing the efficiency. Therefore, the lack of a differentiated training strategy for the edge region and the non-edge region has become one of the key factors affecting the system performance. Summary of the Invention

[0006] To solve the above problems, the present disclosure proposes a cardiovascular segmentation method and system based on data-driven iterative learning to achieve high-precision cardiovascular image segmentation with high training efficiency, promote the early diagnosis, treatment planning, and efficacy evaluation of cardiovascular diseases, and at the same time adapt to the clinical environment with limited computing resources.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] A cardiovascular segmentation method based on data-driven iterative learning, comprising:

[0009] Obtain a cardiovascular image to be segmented;

[0010] Input the cardiovascular image into the learned segmentation model to obtain a segmentation result;

[0011] Among them, the segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set trained in this round, a non-edge set and an edge set are constructed according to the blur degree of the image patches and the proportion of edge pixels; then, the non-edge set is first used to train the segmentation model under the constraint of the segmentation loss function. Based on the model parameters after training with the non-edge set, the edge set is then used to continue training under the coordination of the distribution alignment loss function. Finally, the model parameters trained in this round are used for the parameter initialization of the next round.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] A cardiovascular segmentation system based on data-driven iterative learning, comprising:

[0014] An image acquisition module, configured to: acquire a cardiovascular image to be segmented;

[0015] An image segmentation module, configured to: input the cardiovascular image into the learned segmentation model to obtain a segmentation result;

[0016] Among them, the segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set trained in this round, a non-edge set and an edge set are constructed according to the blur degree of the image patches and the proportion of edge pixels; then, the non-edge set is first used to train the segmentation model under the constraint of the segmentation loss function. Based on the model parameters after training with the non-edge set, the edge set is then used to continue training under the coordination of the distribution alignment loss function. Finally, the model parameters trained in this round are used for the parameter initialization of the next round.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] A computer program product, including a computer program, where when the computer program is executed by a processor, it implements the described cardiovascular segmentation method based on data-driven iterative learning.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] A non-transitory computer-readable storage medium, where the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, it implements the described cardiovascular segmentation method based on data-driven iterative learning.

[0021] According to some embodiments, the present disclosure adopts the following technical solutions:

[0022] 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 runs, the processor executes the computer program stored in the memory to enable the electronic device to execute a cardiovascular segmentation method implementing the described data-driven iterative learning.

[0023] Compared with the prior art, the beneficial effects of the present disclosure are:

[0024] The present invention proposes a cardiovascular segmentation method and system based on data-driven iterative learning, which is specifically designed for the CTA cardiovascular image segmentation task. Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0025] (1) Strong pertinence, optimizing the segmentation accuracy of the edge region:

[0026] The present invention divides the CTA image into a non-edge image and an edge set image, and adopts different training strategies respectively, significantly improving the segmentation accuracy of the edge region (where the density difference between the blood vessel wall and the surrounding tissue is small and the boundary is not clear); the edge set loss function designed for the edge region (including segmentation loss and distribution alignment loss) further enhances the ability to capture fuzzy boundaries, solving the problem of poor segmentation effect of traditional methods in the edge region.

[0027] (2) High training efficiency and fast model convergence speed:

[0028] By preferentially training the non-edge region and further training the edge region based on its model parameters, the present invention significantly reduces the learning difficulty of the model and improves the training efficiency.

[0029] (3) Feature distribution alignment, enhancing the generalization ability of the model:

[0030] The present invention introduces a distribution alignment module, which realizes the alignment of the feature distributions of the two types of regions by calculating and minimizing the Wasserstein distance between the feature distributions of the non-edge set and the edge set; this design effectively integrates the feature information of the non-edge region and the edge region, enhances the model's ability to model the overall cardiovascular structure, and at the same time improves the generalization ability of the model, enabling it to adapt to CTA images under different patients and imaging conditions.

[0031] (4) Low computational resource requirements, suitable for clinical deployment:

[0032] By preferentially training the non-edge region and optimizing the edge region training strategy, the present invention significantly reduces the computational complexity of the model; at the same time, the lightweight design of the segmentation model and the efficient training process enable the system to operate efficiently in a clinical environment with limited computational resources, meeting the actual application requirements.

[0033] (5) Support iterative learning and gradually optimize the segmentation performance:

[0034] The present invention adopts an iterative learning mechanism to gradually optimize the model performance through multiple rounds of training; this iterative learning method can continuously correct the segmentation error and further improve the segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of this disclosure. The illustrative embodiments and descriptions thereof of this disclosure are used to explain this disclosure and do not constitute an improper limitation of this disclosure.

[0036] Figure 1 It is the internal structure diagram of the segmentation model for Embodiment 1.

[0037] Figure 2 It is the internal structure diagram of the encoder for Embodiment 1.

[0038] Figure 3 It is the internal structure diagram of the decoder for Embodiment 1.

[0039] Figure 4 It is the learning flow chart for Embodiment 1.

[0040] Figure 5 It is the data partitioning flow chart for Embodiment 1.

[0041] Figure 6 It is the non-edge set segmentation flow chart for Embodiment 1.

[0042] Figure 7 It is the edge set segmentation flow chart for Embodiment 1.

[0043] Figure 8 It is the distribution alignment flow chart for Embodiment 1.

[0044] Figure 9 It is an example diagram of the cardiovascular image and the label image for Embodiment 1, where Figure 9 (a) and (b) in it are the cardiovascular image and the corresponding label image respectively. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following further describes the present disclosure in conjunction with the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are all exemplary and are intended to provide a further description of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should also be understood that when the terms "comprise" and / or "comprising" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] Embodiment 1

[0049] In an embodiment of the present disclosure, a cardiovascular segmentation method based on data-driven iterative learning is provided to achieve high-precision cardiovascular image segmentation with efficient training, so as to facilitate the early diagnosis, treatment planning, and efficacy evaluation of cardiovascular diseases, and at the same time adapt to the clinical environment with limited computing resources, including:

[0050] Obtain the cardiovascular image to be segmented;

[0051] Input the cardiovascular image into the learned segmentation model to obtain the segmentation result;

[0052] Among them, the segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set of this training, according to the blur degree and edge pixel ratio of the image patches, a non-edge set and an edge set are constructed; then, the non-edge set is used to train the segmentation model under the constraint of the segmentation loss function. Based on the model parameters after training with the non-edge set, the edge set is used to continue training under the coordination of the distribution alignment loss function. Finally, the model parameters trained in this round are used for the parameter initialization of the next round.

[0053] As an embodiment, the segmentation model here adopts a structure as Figure 1 shown, an N-layer convolutional encoder, an N-layer convolutional decoder, and a segmentation prediction head connected in sequence. There is a residual connection between each layer of the encoder and the decoder. The encoder and the decoder are used for feature extraction. The features output by the last layer of the decoder are input into the segmentation prediction head to obtain the segmented cardiovascular image.

[0054] The encoder and the decoder adopt conventional structures, such as Figure 2 , Figure 3 shown. The encoder is composed of a convolutional layer, a normalization layer, an activation layer, an activation layer, a Dropout layer, and a downsampling layer in sequence, while the decoder is composed of a 1*1 convolutional layer, an upsampling layer, a feature splicing layer, a convolutional layer, a normalization layer, an activation layer, and a Dropout layer in sequence. The segmentation prediction head is composed of a fully connected layer.

[0055] Based on the above model structure, a data-driven iterative learning method is used to batch-train the segmentation model. The following is from Figure 4Describe the specific learning process of each iteration in terms of the four aspects of data partitioning, non-edge segmentation, edge segmentation, and distribution alignment shown below.

[0056] I. Data Partitioning

[0057] As Figure 5 shown, obtain the cardiovascular image set for this iteration , where N is the number of samples, is the th sample, is the original cardiovascular image, is the label image corresponding to the cardiovascular image. The samples are as Figure 9 shown, Figure 9 where (a) and (b) in

[0058] are the cardiovascular image and the corresponding label image respectively, and the cardiovascular (coronary artery) to be recognized is marked in red; each image is divided into a non-edge image and an edge image according to the blur degree and the proportion of edge pixels of the image patch. The non-edge image corresponds to the area with clear features and boundaries, while the edge set image corresponds to the area with small density difference between the blood vessel wall and the surrounding tissue and unclear boundaries. The specific steps are as follows:

[0059] (1) Divide each cardiovascular image and its corresponding label image in the image set into several image patches; Specifically, divide each sample into image patches, and divide the cardiovascular image and the label image into image patches of size

[0060] (2) Calculate the blur degree of each image patch in the cardiovascular image ;

[0061] Use the average entropy to characterize the blur degree, and calculate the average entropy of each image patch in the cardiovascular image using the following formula:

[0062]

[0063] where represents the probability of the th pixel in the th image patch of the image . The higher the average entropy value , the more chaotic the pixel distribution of the image patch, that is, the higher the blur degree.

[0064] (3) Calculate the proportion of edge pixels of each image patch in the label image :

[0065]

[0066] Among them, is an indicator function, indicating that the th pixel in the th image patch of the label image

[0067] is an edge pixel. th sample's non-edge set mask and edge set mask .

[0068] Specifically, is the non-edge set mask of the th sample, which consists of L ; is the non-edge set mask of the th image patch in the th sample; is the edge set mask of the th sample, which consists of L ; is the edge set mask of the th image patch in the th sample, and are calculated as follows and expressed by the formula:

[0069]

[0070]

[0071] Among them, is a hyperparameter, which is set to 0.8 in this embodiment; is defined as the maximum value of the average entropy of the image and the edge pixel ratio of the label .

[0072] Based on the masks, generate the non-edge set and edge set.

[0073] Using the masks and of the th sample, generate non-edge images and edge images for the cardiovascular image and label image respectively, and express it by the formula:

[0074] ,

[0075] ,

[0076] ,

[0077]

[0078] Among them, and are the non-edge image and the edge image of the cardiovascular image respectively, and and are the non-edge image and the edge image of the label image respectively.

[0079] Based on the non-edge images and edge images generated for each sample , an edge set and a non-edge set are constructed for the following training.

[0080] II. Non-edge set segmentation

[0081] As Figure 6 shown, the non-edge set images are input into the segmentation model as Figure 3 shown, and the segmentation loss function is used for parameter training to obtain the trained model parameters and non-edge set features. Here, the non-edge set features are the features output by the last layer of the decoder.

[0082] The segmentation loss function consists of cross-entropy loss and Dice loss to enhance the model's ability to capture fuzzy boundaries and segmentation accuracy. By preferentially training the non-edge regions, the model's learning ability for clear features can be quickly optimized, laying a foundation for the subsequent training of the edge regions.

[0083] Specifically, the segmentation loss function for the non-edge set is:

[0084] ,

[0085] ,

[0086] ,

[0087] Among them, is the cross-entropy loss, is the Dice loss, is the label image, is the probability value predicted by the model, N is the number of samples, is a small constant, set to 。

[0088] The parameter learning process of the model is as follows:

[0089] ,

[0090] Among them, is the model parameter of the th iteration training of the non-edge set, represents Figure 3 the segmentation model shown, represents the cardiovascular image in the th image patch, represents the label image in the th image patch, is the model parameter of the th iteration training of the edge set. When , , which is the model initialization parameter.

[0091] III. Edge set segmentation

[0092] As Figure 7 shown, based on the model parameters trained on the non-edge set, the edge set image is input into the segmentation model, and the parameter training is performed using the edge set loss function to obtain the trained model parameters and edge set features. Here, the edge set features are also the features output by the last layer of the decoder.

[0093] The edge set loss function mentioned above includes a segmentation loss function and a distribution alignment loss function. Specifically, the edge set loss function is:

[0094]

[0095] Among them, is the segmentation loss, is the distribution alignment loss obtained by distribution alignment, is a hyperparameter, which is set to 0.1 in this embodiment.

[0096] The parameter learning process of the model is as follows:

[0097]

[0098] Among them, is the model parameter of the th iteration training of the edge set, represents Figure 3 the segmentation model shown, is the model parameter of the th iteration training of the non-edge set.

[0099] First non-edge set segmentation and then edge set segmentation. The advantage of this arrangement is that:

[0100] Edge regions usually contain blurred and complex details, and direct segmentation is likely to introduce noise and errors; the features of non-edge regions are more uniform and the structure is clearer. Segmenting the non-edge regions first can obtain a relatively stable and accurate preliminary segmentation result, laying a good foundation for the overall segmentation task.

[0101] IV. Distribution Alignment

[0102] As Figure 8 shown, input the output features of the last layer decoder of the segmentation model in the non-edge set segmentation and the edge set segmentation, calculate the Gaussian distributions of the two, and minimize the Wasserstein distance between the two Gaussian distributions to obtain the distribution alignment loss function, thereby realizing the alignment of the non-edge set feature distribution and the edge set feature distribution. The specific steps are as follows:

[0103] (1) Use the segmentation model to extract features from the non-edge set and the edge set respectively;

[0104] Obtain the features output by the last layer decoder of the segmentation model in the non-edge set and the features output by the last layer decoder of the segmentation model in the edge set segmentation .

[0105] (2) Based on the extracted features, estimate the Gaussian distribution to obtain the multivariate Gaussian distribution of the non-edge set features and the multivariate Gaussian distribution of the edge set features;

[0106] Specifically, estimate the multivariate Gaussian distribution of each category of the non-edge set features , and the multivariate Gaussian distribution of each category of the edge set features . Among them, , in this embodiment, the category set includes two categories: cardiovascular and background. , in this embodiment, the category set includes two categories: cardiovascular and background.

[0107] (3) Add perturbations to the covariance matrix of the multivariate Gaussian distribution;

[0108] Considering that the multivariate Gaussian distribution of each category obtained from the non-edge set may not fully cover all the data of this category, so the obtained distribution has a deviation, that is, the phenomenon of covariate shift. For this reason, in this embodiment, appropriate perturbations are added to the estimated empirical covariance matrix to reduce the impact of covariance shift. For category , its covariance is updated as:

[0109] ​​​​

[0110] Among them, represents a matrix of all 1s, is a random number that controls the degree of perturbation to obtain the updated distribution .

[0111] (4) After adding the perturbation, calculate the Wasserstein distance between the two multivariate Gaussian distributions;

[0112] Define the Wasserstein distance between the two multivariate Gaussian distributions, which is used to measure the difference in the representative distributions of the fuzzy and non-fuzzy region sets; for each category , the Wasserstein distance is:

[0113]

[0114] Among them, tr is the trace of the matrix.

[0115] (5) Based on the Wasserstein distance, construct a distribution alignment loss function.

[0116] Finally, the overall distribution alignment loss is defined as the sum of the Wasserstein distances of each category:

[0117] .

[0118] By minimizing the Wasserstein distance between the feature distributions of the non-marginal set and the marginal set, the feature distributions of the two parts can be effectively aligned, reducing the model bias caused by the data distribution difference, thereby improving the recognition consistency of the model in different regions; by achieving the consistency of the feature distributions of the fuzzy region and the non-fuzzy region, the model can maintain stable and excellent segmentation performance when processing the cardiovascular marginal fuzzy region, enhancing the robustness and generalization ability of the model.

[0119] Embodiment 2

[0120] In an embodiment of the present disclosure, a cardiovascular segmentation system based on data-driven iterative learning is provided, including:

[0121] An image acquisition module, configured to: acquire a cardiovascular image to be segmented;

[0122] An image segmentation module, configured to: input the cardiovascular image into the learned segmentation model to obtain a segmentation result;

[0123] Among them, the segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set trained in this round, a non-edge set and an edge set are constructed according to the blur degree and edge pixel ratio of the image patches. Then, the non-edge set is first used to train the segmentation model under the constraint of the segmentation loss function. Based on the model parameters after training with the non-edge set, the edge set is then used to continue training under the coordination of the distribution alignment loss function. Finally, the model parameters trained in this round are used for the parameter initialization of the next round.

[0124] Embodiment 3

[0125] In an embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, it implements the data-driven iterative learning-based cardiovascular segmentation method described above.

[0126] Embodiment 4

[0127] In an embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided. The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, it implements the data-driven iterative learning-based cardiovascular segmentation method described above.

[0128] Embodiment 5

[0129] In an embodiment of the present disclosure, an electronic device is provided, including: 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 runs, the processor executes the computer program stored in the memory so that the electronic device executes and implements the data-driven iterative learning-based cardiovascular segmentation method described above.

[0130] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination 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 the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or a device for implementing the specified functions in one block or multiple blocks.

[0131] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions for implementing the process Figure 1 in one process or a plurality of processes and / or boxes Figure 1 steps for the functions specified in one box or a plurality of boxes.

[0132] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they are not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that various modifications or variations that can be made without creative efforts on the basis of the technical solutions of the present disclosure are still within the scope of protection of the present disclosure.

Claims

1. A data-driven iterative learning cardiovascular segmentation method, characterized in that: include: Acquire a cardiovascular image to be segmented; Inputting the cardiovascular image into the learned segmentation model to obtain the segmentation result; The segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set of this training, according to the blur degree of the image block and the edge pixel ratio, a non-edge set and an edge set are constructed; then, the segmentation model is first trained with the non-edge set under the constraint of the segmentation loss function, and based on the model parameters after the non-edge set training, the edge set is used to continue training under the coordination of the distribution alignment loss function; finally, the model parameters trained in this round are used for parameter initialization of the next round; The loss function of the non-edge set is a segmentation loss function composed of a cross entropy loss and a Dice loss; The loss function of the edge set is an edge set loss function composed of a segmentation loss function and a distribution alignment loss function; The distribution alignment loss function is used to minimize the Wasserstein distance between the feature distributions of the non-marginal set and the marginal set; The construction process of the distribution alignment loss function is: Using the segmentation model, features are extracted from the non-edge set and the edge set respectively; Based on the extracted features, the Gaussian distribution is estimated to obtain the multivariate Gaussian distribution of the non-marginal set features and the multivariate Gaussian distribution of the marginal set features; Add perturbations to the covariance matrix of a multivariate Gaussian distribution; After adding the perturbation, the Wasserstein distance between two multivariate Gaussian distributions is calculated; Based on the Wasserstein distance, a distribution alignment loss function is constructed.

2. A data-driven iterative learning cardiovascular segmentation method according to claim 1, characterized in that: The image set consists of a plurality of cardiovascular images and corresponding label images; The specific steps of constructing the non-edge set and the edge set are as follows: Each cardiovascular image and the corresponding label image in the image set are divided into blocks to obtain a number of image blocks; Calculate the blur level of each image block in the cardiovascular image; Calculate the edge pixel ratio of each image block in the label image; Using soft threshold masking, masks of non-edge set and edge set are constructed based on blur degree and edge pixel ratio; Based on the mask, non-edge sets and edge sets are generated.

3. A data-driven iterative learning cardiovascular segmentation method as claimed in claim 2, characterized in that: The masks of the non-edge set and the edge set are respectively composed of the non-edge set masks and the edge set masks of a plurality of image blocks. The non-edge set masks and the edge set masks of the image blocks are expressed by the formula: in, For the In the sample The non-edge set mask of the image patch, For the In the sample The edge set mask of the image patch, is a hyperparameter, Defined as the blurriness of the image patch and the edge pixel ratio of the label image patch The maximum value of .

4. A data-driven iterative learning cardiovascular segmentation system, characterized in that: include: The image acquisition module is configured to: acquire a cardiovascular image to be segmented; The image segmentation module is configured to: input the cardiovascular image into the learned segmentation model to obtain a segmentation result; The segmentation model adopts a data-driven iterative learning method. In each iteration, first, based on the image set of this training, according to the blur degree of the image block and the edge pixel ratio, a non-edge set and an edge set are constructed; then, the segmentation model is first trained with the non-edge set under the constraint of the segmentation loss function, and based on the model parameters after the non-edge set training, the edge set is used to continue training under the coordination of the distribution alignment loss function; finally, the model parameters trained in this round are used for parameter initialization of the next round; The loss function of the non-edge set is a segmentation loss function composed of a cross entropy loss and a Dice loss; The loss function of the edge set is an edge set loss function composed of a segmentation loss function and a distribution alignment loss function; The distribution alignment loss function is used to minimize the Wasserstein distance between the feature distributions of the non-marginal set and the marginal set; The construction process of the distribution alignment loss function is: Using the segmentation model, features are extracted from the non-edge set and the edge set respectively; Based on the extracted features, the Gaussian distribution is estimated to obtain the multivariate Gaussian distribution of the non-marginal set features and the multivariate Gaussian distribution of the marginal set features; Add perturbations to the covariance matrix of a multivariate Gaussian distribution; After adding the perturbation, the Wasserstein distance between two multivariate Gaussian distributions is calculated; Based on the Wasserstein distance, a distribution alignment loss function is constructed.

5. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the data-driven iterative learning cardiovascular segmentation method according to any one of claims 1 to 3 is implemented.

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 cardiovascular segmentation method of data-driven iterative learning is implemented as described in any one of claims 1-3.

7. 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 so that the electronic device executes a cardiovascular segmentation method of data-driven iterative learning as described in any one of claims 1-3.

Images