Cerebral vessel recognition model construction method and network construction method

Abstract

The application discloses a cerebral vascular recognition model construction method and a network construction method, and is suitable for magnetic resonance image processing and analysis in the field of neuroscience and clinical treatment. The method comprises the following steps: acquiring magnetic resonance image training data of cerebral vessels, inputting the magnetic resonance image training data into an initial recognition model constructed based on a deep learning segmentation framework to perform training, and obtaining a target recognition model capable of outputting prediction results of each anatomical region. The application realizes accurate extraction, integration and recognition of cerebral vascular information in different anatomical regions in the magnetic resonance image, can effectively solve the inclusion and exclusion relationship problems between corresponding structures of each anatomical region, and accurately segments various structural regions of the cerebral vessels in the image. After processing the prediction results output by the segmentation model, the nodes and edges of the cerebral vascular network are further analyzed, and information such as vascular morphological characteristics, anatomical regions, network hierarchical structure and blood flow direction is embedded as graph attributes to form a cerebral vascular information graph.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a cerebral vessel recognition model construction method and a network construction method. BACKGROUND

[0002] The efficiency of the brain's operation relies heavily on its complex and delicate vascular network. This network bears the responsibility of transporting oxygen and nutrients to meet the brain's continuous high energy consumption. Notably, changes in the cerebral vascular network are often closely related to the development of various neurological diseases, including primary vascular diseases, traumatic brain injury, vascular occlusion, and stroke. These diseases can have a profound impact on the function of the cerebral vascular network.

[0003] The complexity of the human brain surpasses that of other biological systems. The spatial stratification and structuring of brain regions make it a supergraph constructed by neuronal networks. Meanwhile, the close connection between neuronal networks and vascular networks determines the maximum energy load of the brain and the growth pattern of neurons. Changes in the vascular network, such as changes in vessel size and connection number, can have a significant impact on different brain regions, and these changes are often important indicators of early signs of specific diseases.

[0004] In the process of identifying, preventing, and predicting cerebral vascular diseases, accurately recognizing and segmenting cerebral vascular images, and deeply analyzing the features of the cerebral vascular network are key steps. Magnetic resonance angiography (MRA) has been widely adopted to depict the anatomical structure and pathological features of cerebral vessels. Precise segmentation of these MRA images and construction of individualized cerebral vascular networks can provide rich information to identify possible vascular dementia or stroke earlier.

[0005] Although deep learning methods such as convolutional neural networks (CNN) and Transformer variants have made significant progress in MRI image processing, especially in the rapid segmentation and labeling of brain regions, these neural network algorithms still show deficiencies in segmentation accuracy when faced with complex and irregular cerebral vascular structures, making it difficult to accurately identify multiple anatomical regions of cerebral vessels, which undoubtedly poses a major challenge in the construction of individualized cerebral vascular networks. SUMMARY

[0006] The present application aims to solve the problem of accurately identifying the anatomical regions of cerebral vessels in magnetic resonance images. To solve the above problem, in a first aspect, the present application provides a cerebral vessel recognition model construction method, comprising:

[0007] obtaining training data, the training data including magnetic resonance images, the magnetic resonance images including the anatomical regions of cerebral vessels;

[0008] inputting the training data into an initial recognition model based on an nnU-Net framework to obtain a target recognition model, the target recognition model being used to output a prediction result of each of the anatomical regions in the magnetic resonance image, the prediction result being used to represent position information and a category of each of the anatomical regions, and the nnU-Net framework being embedded with a graph neural network, the graph neural network being used to identify a structural relationship of each of the anatomical regions in the brain blood vessel;

[0009] creating a network structure of the brain blood vessel according to the prediction result, the network structure being used to represent structural information of the brain blood vessel in the magnetic resonance image.

[0010] In a second aspect, the present application provides a method for identifying a network of brain blood vessels, comprising:

[0011] obtaining a magnetic resonance image to be processed, the magnetic resonance image comprising each anatomical region of a brain blood vessel;

[0012] inputting the magnetic resonance image to be processed into a target recognition model constructed by the method for constructing a brain blood vessel recognition model according to the first aspect, and outputting a prediction result corresponding to the magnetic resonance image, the prediction result being used to represent a position and a category of each of the anatomical regions;

[0013] constructing the network structure of the brain blood vessel according to the prediction result, the network structure being used to represent structural information of the brain blood vessel.

[0014] It can be understood that the method for constructing a brain blood vessel recognition model and the method for network identification provided by the present application can use an nnU-Net framework as a backbone structure, wherein a novel neural network structure embedded with a graph neural network is used to train the obtained magnetic resonance image of the brain. This enables the target recognition model to extract and integrate the features of each anatomical region of the brain blood vessel by using the embedded graph neural network when identifying each anatomical region of the brain blood vessel in the magnetic resonance image to be processed, so as to identify the structural relationship of each anatomical region of the brain blood vessel in the magnetic resonance image, thereby solving the inclusion and exclusion relationship between the corresponding structures of each anatomical region. This process ultimately realizes high-precision identification and segmentation of various structural regions of the brain blood vessel in the magnetic resonance image, constructs a brain blood vessel network with rich features, and effectively reduces the variability between individuals, thereby providing accurate and comprehensive basis for large-scale clinical data analysis and medical research. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 a magnetic resonance image of a brain blood vessel;

[0016] Figure 2 a schematic diagram of a brain blood vessel network structure specimen;

[0017] Figure 3A segmentation result diagram of a cerebral blood vessel in the related art is shown;

[0018] Figure 4 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0019] Figure 5 A flowchart framework of the cerebral blood vessel recognition model construction method in some embodiments of the present application is shown;

[0020] Figure 6 A cerebral blood vessel network construction diagram in some embodiments of the present application is shown;

[0021] Figure 7 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0022] Figure 8 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0023] Figure 9 A structure diagram of a graph neural network unit in some embodiments of the present application is shown;

[0024] Figure 10 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0025] Figure 11 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0026] Figure 12 A topological interaction and binary topological interaction diagram of some embodiments of the present application is shown;

[0027] Figure 13 A cerebral blood vessel recognition model construction method flowchart of some embodiments of the present application is shown;

[0028] Figure 14 A cerebral blood vessel network construction method flowchart of some embodiments of the present application is shown;

[0029] Figure 15 A structure diagram of a cerebral blood vessel network recognition model construction device of some embodiments of the present application is shown;

[0030] Figure 16 A structure diagram of a magnetic cerebral blood vessel network recognition device of some embodiments of the present application is shown;

[0031] Figure 17 A structure diagram of a computer system provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0033] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0034] MRI technology can generate multimodal images that reflect the morphology of human organs and tissues, such as medical images of brain morphology, lung morphology, or bone morphology. Specifically, it can generate multimodal image data including T1-weighted imaging, T2-weighted imaging, T1ce imaging, and free water suppression sequence (FLAIR).

[0035] For example, for Figure 1 The magnetic resonance images of cerebral vasculature shown in the figure are important preliminary operations for standardized and non-invasive analysis of cerebral vasculature in magnetic resonance angiography (MRA).

[0036] Among related technologies, convolutional neural networks (CNNs) and Transformer variants have become the main methods in the field of medical image segmentation.

[0037] It is understandable that the structure of cerebral blood vessels in living organisms is complex and varied, showing irregular characteristics, such as Figure 2 The schematic diagram of the cerebral vascular network structure specimen shown includes anatomical regions corresponding to various different vascular types, namely anatomical types.

[0038] The anatomical region represents the specific outlines and structural positions of various types of cerebral blood vessels in the magnetic resonance image.

[0039] For example, Figure 3 As shown on the left, the cerebral blood vessels mainly include two groups of blood vessels supplying the brain (ICA and VA-BA); three groups of first-order branch blood vessels supplying the telencephalon and diencephalon above the tentorium (ACA, MCA, PCA); and three groups of first-order branch blood vessels supplying the brainstem and cerebellum below the tentorium (SCA, AICA, PICA), as well as anatomical regions of different vascular types such as veins.

[0040] However, for complex and irregular cerebral vascular structures, the combination of convolutional neural networks (CNN) and Transformer cannot effectively capture global contextual information and cannot effectively handle the inclusion and exclusion relationships between various anatomical structures, ultimately resulting in low accuracy of the segmentation results.

[0041] For example, Figure 3As shown, the left is the real situation of the cerebral blood vessels, and the right is the recognition result of the CNN and Transformer hybrid architecture, which are compared on the cerebral blood vessel multi-class segmentation task. Among them, circle 1 represents the interruption of connection between small blood vessels, and circle 2 represents the label confusion leading to ICA anatomical type error.

[0042] Therefore, in the present application, in order to improve the accuracy of the segmentation of various anatomical regions of cerebral blood vessels in the magnetic resonance image, a hybrid neural network architecture based on the nnU-Net framework is constructed, that is, a graph neural network module is embedded into the traditional nnU-Net framework to construct a new recognition model of cerebral blood vessel network, so that when the recognition model is used to recognize the various anatomical regions in the cerebral blood vessels of the magnetic resonance image, the embedded graph neural network module can integrate information from different anatomical regions and extract tubular structures to solve the structural relationship between various anatomical regions, such as inclusion and exclusion relationship, to ensure accurate prediction results, and realize the recognition of the characteristics of each anatomical region of the cerebral blood vessels from the whole, which can provide valuable cerebral blood vessel structure information for large-scale clinical data analysis.

[0043] In order to better understand and illustrate the magnetic resonance image cerebral blood vessel recognition model construction method provided by the present application, the following will be described in detail with the help of the accompanying drawings.

[0044] Figure 4 As shown in the flowchart of the cerebral blood vessel recognition model construction method provided by the embodiment of the present application, Figure 4 As shown, the method specifically includes:

[0045] S110, acquiring training data, the training data including a magnetic resonance image, the magnetic resonance image including various anatomical regions of cerebral blood vessels.

[0046] S120, inputting the training data into an initial recognition model constructed based on the nnU-Net framework, training the initial recognition model to obtain a target recognition model, the target recognition model being used to output a prediction result of each anatomical region in the magnetic resonance image, the prediction result being used to represent position information and category of each anatomical region, and a graph neural network being embedded in the nnU-Net framework, the graph neural network being used to recognize the structural relationship of each anatomical region in the cerebral blood vessels.

[0047] Specifically, in combination with Figure 5 As shown, in the embodiment of the present application, when constructing the cerebral blood vessel network structure recognition model of the magnetic resonance image, first, the training data can be acquired to train the constructed initial recognition model.

[0048] The training data can include a plurality of magnetic resonance images of the brain, i.e., brain image data collected by nuclear magnetic resonance technology, each magnetic resonance image including each anatomical region of the cerebral vessels of the collection object, i.e., the region where each type of blood vessel is located, and the corresponding contour.

[0049] In the embodiment of the present application, the initial identification model is a new neural network based on the neural network model of the nnU-Net framework, such as the efficient topology-aware U-Net (NexToU) network structure.

[0050] That is, in the new neural network structure, in order to realize the integration of each anatomical region in the cerebral vessels and identify the structural relationship between each anatomical region, so as to improve the accuracy and efficiency of the recognition of each type of cerebral vessel, a graph neural network, i.e., a visual GNN variant, is embedded in the backbone neural network structure of the nnU-Net framework, so that the visual GNN variant can integrate information from different anatomical regions of the cerebral vessels and extract tubular structures, solve the inclusion and exclusion of structural relationships between each anatomical structure, and realize accurate segmentation of the cerebral vessel structure.

[0051] It can be understood that the target identification model constructed in the embodiment can output a prediction result of each anatomical region in the magnetic resonance image, and the prediction result can be a mask image of the anatomical structure corresponding to each anatomical region of the cerebral vessels, to represent the position information and the type of each anatomical region.

[0052] That is, the prediction result can represent the position information and the anatomical type of each anatomical region of the cerebral vessels in the magnetic resonance image. The position information can specifically include the contour information and the position information of each cerebral vessel. The anatomical type can include specific type attributes, such as ICA, VA-BA, etc. mentioned above.

[0053] Further, in some embodiments of the present application, for complex cerebral vessel structures, in order to analyze and comprehensively understand the structural information of the cerebral vessels, to extract the anatomical type, blood flow direction and network level features of the cerebral vessels, the identification model constructed can further include a generation module, i.e., a generation module can be created during the training process or after the training is completed, i.e., a program for constructing a cerebral vessel structure network based on the segmentation result output by the target segmentation model is created, to realize the visual display of the network structure of the cerebral vessels.

[0054] That is, the cerebral vessel identification model constructed in the embodiment of the present application can include a segmentation model for identifying and segmenting each type of anatomical structure of the cerebral vessels, and a creation model for creating a cerebral vessel network based on the segmentation result, i.e., the prediction result. That is, first, a segmentation model for identifying and segmenting each anatomical region of the cerebral vessels is constructed by using the training data, and then a program package corresponding to the creation model is encapsulated in the identification model.

[0055] The method further comprises:

[0056] S130, generating a creation module for creating the network structure of the cerebral blood vessels according to the prediction result.

[0057] Specifically, after the training of the segmentation model in the recognition model is completed through the above embodiment, a piece of program code, i.e., a creation module, is further embedded in the segmentation model. For example, a code input by a programmer or a software package is input as the creation module embedded in the segmentation model to communicate the recognition model constituted by the magnetic resonance image.

[0058] The creation module is used to construct the cerebral blood vessel network structure of the magnetic resonance image according to the output result of the segmentation model, i.e., the prediction result, as shown in Figure 6 .

[0059] It can be understood that, in the embodiment of the present application, the new neural network structure in which the graph neural network is embedded is trained by taking the nnU-Net framework as the backbone structure based on the acquired brain magnetic resonance image, so as to construct a target recognition model for recognizing the cerebral blood vessel network structure. Therefore, when the cerebral blood vessel network in the brain magnetic resonance image to be processed is recognized and constructed by using the constructed target recognition model, the graph neural network module embedded therein is used to extract and integrate the cerebral blood vessel information of each different anatomical region of the cerebral blood vessels in the magnetic resonance image, so as to solve the inclusion and exclusion relationship between the anatomical structures, and finally recognize and segment the accurate segmentation result of the cerebral blood vessels, thereby effectively reducing the individual variability and providing accurate and comprehensive scientific basis for large-scale clinical data analysis.

[0060] It can be understood that, in the embodiment of the present application, the finally constructed cerebral blood vessel recognition model can not only accurately recognize each anatomical region of the cerebral blood vessels in the magnetic resonance image to be processed, but also can further produce the network structure of the cerebral blood vessels by using the recognized anatomical regions with the help of the creation module, so that the network structure can represent the structural information of the cerebral blood vessels as a whole, i.e., the structural information of the cerebral blood vessels can be extracted from the network structure, such as the anatomical type, blood flow direction and network hierarchy features of the cerebral blood vessels.

[0061] In detail, when the creation module is generated according to the position information and category attributes of each anatomical region in the prediction result to create the cerebral blood vessel network structure, the prediction result can be subjected to cerebral blood vessel structure information graph extraction to obtain the nodes of the information graph and the edges of the information graph, respectively. Furthermore, the vascular morphological features, vascular anatomical regions, vascular network hierarchy structure and blood flow direction and other information are embedded as graph attributes on the edges, as shown in Figure 6 .

[0062] The node features of the obtained cerebral vascular network can include node serial number , node coordinates , radius , node degree , anatomical type , bifurcation point anatomical type , blood flow map level , and blood flow path distance , etc. The obtained edge features can include edge serial number , starting node serial number , ending node serial number , length , distance , curvature , volume , average cross section , average radius , roundness average value , starting node degree , ending node degree , voxel number , surface area , vascular anatomical category , segmentation index , blood flow direction , etc.

[0063] Table 1. Features of cerebral vascular network and measurement thereof

[0064]

[0065] It can be understood that the new recognition model constructed by the method can automatically output the network structure of the cerebral vessels of the magnetic resonance image in actual use, and can be applied to image analysis of various cerebral vascular lesion types such as stroke, vascular dementia, aneurysm and cerebral small vessel disease, which helps to better understand the influence of aging and other pathophysiological conditions on cerebral vessels and hemodynamics.

[0066] Optionally, in some embodiments of the present application, after obtaining the training data, the training data can be preprocessed, that is, for the obtained magnetic resonance image data, before training the initial recognition model built therefrom, the magnetic resonance image in the training data can be preprocessed such as cropping and data enhancement, to discard the surrounding invalid area and highlight the anatomical area where the cerebral vessels are located.

[0067] Optionally, in the embodiment of the present application, the segmentation model in the newly built recognition model is mainly based on the nnU-Net framework, and the graph neural network module is embedded in the corresponding encoder and decoder module.

[0068] Specifically, asFigure 4 As shown in the figure, the segmentation model taking the nnU-Net framework as the backbone structure includes an encoder module and a decoder module, and the encoder module includes a plurality of convolution layers, and the decoder module includes a plurality of deconvolution layers. The plurality of convolution layers in the encoder module realizes down-sampling on the input training data, and the plurality of deconvolution layers in the decoder module realizes up-sampling on the input training data to output the final prediction result.

[0069] Correspondingly, the convolution layers in the encoder module and the decoder module are embedded with a graph neural network (GNN) module in the recognition model designed and defined in the embodiment of the application, so as to integrate and extract the vascular region information of different anatomical structures of the cerebral vessels, assist the main structure of the U-Net neural network model, and realize the final efficient and accurate segmentation of various cerebral vessels.

[0070] As shown in the figure, Figure 7 In S120, the training of the initial recognition model, that is, the analysis and processing process of the input training data, specifically includes:

[0071] S121, input the training data into the encoder module to extract the anatomical region features of the cerebral vessels in the training data and output the down-sampling result.

[0072] S122, input the down-sampling result into the decoder module to integrate the features of the down-sampling result and output a plurality of prediction results.

[0073] S123, according to the prediction result, determine the loss function corresponding to the prediction result, and the loss function is used for parameter updating of the initial recognition model to obtain the target recognition model.

[0074] Specifically, when training the initial segmentation model with the training data, that is, when recognizing and segmenting various anatomical regions in the cerebral vessels from the training data, first, the training data is subjected to convolution processing to realize the down-sampling stage, that is, the training data is input into each convolution layer of the encoder module, and the embedded graph neural network is used to assist each convolution layer to extract and integrate the features of each anatomical region of the cerebral vessels, and then to enhance the features of various cerebral vessels in the magnetic resonance image to obtain the down-sampling result.

[0075] Further, after completing the down-sampling of the magnetic resonance image, each deconvolution layer can be used to up-sample the down-sampling result, that is, the down-sampling result is first input into the lowest layer of the deconvolution layer, and then iteratively processed to use the graph neural network embedded in the deconvolution layer to integrate the features of various cerebral vessels in the magnetic resonance image, and then use the backbone structure in the nnU-Net to enlarge the down-sampling result, and finally output the final prediction result in the uppermost layer of the deconvolution layer.

[0076] It can be understood that in the up-sampling process, the output corresponding prediction result is obtained when up-sampling is performed in each deconvolution layer, that is, the segmentation result of each type of cerebral blood vessel in the magnetic resonance.

[0077] In the training process of the neural network, in order to realize supervised machine learning, the loss function is used for constraint, that is, for each output prediction result, the corresponding loss function can be determined. Then the loss function can be used to perform back propagation to update the parameters of the identification model.

[0078] It can be understood that in the training process, the training epoch can be set in advance, so that when the set epoch is gradually approached, the loss function of the corresponding iteration will converge, and finally decrease to stable, and then the training is stopped. At this time, the corresponding neural network model is the target identification model, that is, the training and construction of the model are completed.

[0079] For example, in some embodiments, 1000 epochs can be set for training, and when the propagation and back propagation of 1000 times of down-sampling and up-sampling are completed, the corresponding identification model is the target identification model.

[0080] It can be understood that the construction of the segmentation model in the identification model of the embodiment of the present application is to use the training data to perform supervised machine learning on the initial identification model, and finally construct an identification model that can accurately segment each type of cerebral blood vessel in the magnetic resonance image.

[0081] Further, in order to better understand the process of training the training data in the encoder module and the decoder module combined with the embedded graph neural network in the embodiment of the present application, the following embodiments are described in detail.

[0082] Optionally, in some embodiments of the present application, the encoder module can include a first down-sampling module, a position encoding module and a second down-sampling module, and the graph neural network is embedded in the second down-sampling module.

[0083] Correspondingly, combined with Figure 5 As shown in the figure, the processing of the training data in the down-sampling process can include two stages, that is, the convolution processing stage (N1) corresponding to the first down-sampling module, and the feature enhancement processing stage (N2) corresponding to the second down-sampling module.

[0084] That is, for the large size magnetic resonance image block in the embodiment of the present application, the calculation amount will be too large when constructing the graph based on the k-nearest neighbors of each pixel point. In order to extract the shallow feature, first, the convolution module can be used to perform the first stage of convolution processing operation, and then the GNN unit can be used to perform the second stage of processing.

[0085] That is Figure 8 As shown, S121 specifically includes the following steps:

[0086] S01, input the training data into the first downsampling module, perform convolution processing on the training data, and obtain a first downsampling result corresponding to the training data.

[0087] Specifically, combined Figure 5 As shown, the first downsampling module may include a first convolution unit, a second convolution unit, a downsampling unit and a third convolution unit in the order of data processing.

[0088] Correspondingly, in the first stage of processing the training data, it can first be processed through two three-dimensional convolution units, such as two three-dimensional convolution units with a convolution kernel size of 3x3x3, that is, through the first convolution unit and the second convolution unit to obtain the intermediate feature representation f1 of the training data, which is used to retain the local feature information of the input image.

[0089] Furthermore, the intermediate feature representation f1 is input into a downsampling unit, such as a downsampling unit with a convolution kernel size of 3x3x3, to expand the dimension of the features while reducing the size of the image to enhance the semantic information of the features. After being processed by the third convolution layer, the feature representation f2 is finally obtained, which is the first downsampling result.

[0090] Specifically, when processing f1, the intermediate feature representation f1 can first be processed through a downsampling unit with a convolution kernel size of 3x3x3, thereby expanding the dimension while reducing the image size, thereby obtaining an intermediate feature representation f21 with enhanced feature semantic information; the intermediate feature representation f21 is then input into a three-dimensional convolution unit with a convolution kernel size of 3x3x3, i.e., the third convolution unit, for processing, thereby obtaining a feature representation f22, i.e., the first downsampling result;

[0091] In practice, the steps of the third convolution unit and the downsampling unit operation may be repeated, for example, twice, to obtain a better sampling result.

[0092] S02 , using a position encoding module, adding position reference information of each anatomical region of the cerebral blood vessels to the first downsampling result to obtain a second downsampling result.

[0093] S03: Input the second down-sampling result into the second down-sampling module, perform convolution and integration processing on the second down-sampling result to obtain a down-sampling result.

[0094] Specifically, if Figure 5As shown, in the down-sampling process in the second stage, the second down-sampling module can include a position encoding module and a plurality of first feature processing modules. The first feature processing module includes a down-sampling unit and a graph neural network unit.

[0095] Correspondingly, in this step, for the input intermediate feature, i.e. the first down-sampling result, first, the position encoding module can be used to add the learnable position encoding to obtain the feature representation f3 with position information, i.e. the second down-sampling result, to improve the model's processing ability for spatial information.

[0096] Further, for the feature representation with position information, i.e. the second down-sampling result, it can be input into the first feature processing module in the second down-sampling module to obtain a new feature representation, and the new feature representation is input into the subsequent first feature processing module to gradually process the feature representation second down-sampling result, and finally output the down-sampling result.

[0097] That is, the second down-sampling module includes a plurality of first feature processing modules with iterative relationship to iteratively process the input second down-sampling result.

[0098] As shown in Figure 5 Each of the first feature processing modules includes a down-sampling unit and a graph neural network unit, i.e. a first graph neural network unit.

[0099] In S03, the processing of the first feature processing module on the input second down-sampling result specifically includes:

[0100] S031, input the second down-sampling result into the down-sampling unit to perform feature enhancement on the second down-sampling result to obtain a processing result after feature enhancement.

[0101] S032, input the enhanced processing result into the first image neural network unit to perform integrated recognition on the brain blood vessel region in the processing result to obtain the down-sampling result.

[0102] Specifically, for the down-sampling unit and the first graph neural network unit included in each first feature processing module, the processing of the input, such as the processing of the second down-sampling result, first inputs the intermediate feature representation f3, i.e. the second down-sampling result, into the down-sampling unit with a convolution kernel size of 3x3x3 to expand the dimension of the feature while reducing the size of the image to enhance the semantic information of the feature, and obtain the processing result after enhancement. Then, through the efficient graph neural network unit, the brain blood vessel region in the processing result is integrated and recognized to obtain the feature representation f4.

[0103] For example, in some embodiments, three first feature processing modules can be included in the model, i.e., when the feature representation f3 is processed in the second stage N2, it can be repeated three times to obtain the final down-sampling result.

[0104] Optionally, as shown in some embodiments, the graph neural network unit, i.e., the first graph neural network unit, can include a pooling graph neural network unit and a shift window graph neural network unit. Among them, the pooling graph neural network unit includes a pooling graph unit and a feed forward network unit (FFN); the shift window graph neural network unit includes a shift window graph unit and a shift window feed forward network unit (FFN). Figure 9

[0105] Correspondingly, as shown in the above input to the first feature processing module to obtain a new feature representation, the specific steps can include the following steps: Figure 10

[0106] S001, based on the pooling graph unit, the enhanced processing result is extracted based on the global topology structure, and the extracted result is added to the processing result to obtain the pooling result.

[0107] S002, based on the pooling feed forward network unit, the pooling result is processed by multiple layers, and the multiple layer processing results are added to the pooling result to obtain the first graph neural network processing result.

[0108] S003, based on the shift window graph unit, the first graph neural network processing result is shifted, and the shift processing result is added to the first graph neural network processing result to obtain the shift window processing result.

[0109] S004, based on the shift window feed forward neural unit, the shift window processing result is processed by multiple layers, and the multiple layer processing results are added to the shift window processing result to obtain the second graph neural network processing result.

[0110] Specifically, in combination with Figure 9 As shown in the data processing process in some layers of the pooling graph neural network unit and the shift window graph neural network unit: in the pooling graph neural network unit, the second down-sampling result, i.e., f31, is processed by the pooling graph unit in the pooling GNN, then the processed result is added to the second down-sampling result, and then the feed forward network unit in the pooling GNN is added to the original feature to obtain the feature representation f9, i.e., the first graph neural network processing result.

[0111] ​​Further, in the shift window graph neural network unit, the feature f9, i.e. the first graph neural network processing result is processed by the shift window graph unit in the shift window GNN and added to the feature f9 to obtain a shift window processing result. Then the shift window processing result is processed by the FFN in the shift window GNN and added to the feature shift window processing result to obtain a second graph neural network processing result.

[0112] It can be understood that each unit in the above-mentioned graph neural network includes multi-layer iterative processing, and the processing in the above-mentioned embodiment is only a specific description of a certain layer, for example, the processing of the zl-1 layer and the zl layer in the pooling graph neural network unit. Correspondingly, the output result of the pooling graph neural network unit is input into the shift window graph neural network, and the processing process thereof at the zl layer and the zl+1 layer is described.

[0113] It can also be understood that the second down-sampling result is finally output as a corresponding down-sampling result after being processed by the pooling graph neural network unit and the shift window graph neural network unit for processing by the decoder.

[0114] It can be understood that by the above-mentioned operation, in the encoder structure under the nnU-Net neural framework, the GNN module is used to extract the cerebral vascular topological structure in the magnetic resonance image, which can solve the exclusion and inclusion between different anatomical regions, and finally realize the complete recognition and segmentation of the cerebral vascular topological structure image.

[0115] Optionally, in some embodiments of the present application, in the process of constructing the segmentation model, i.e. after the encoder module processes the training data and outputs the down-sampling result, the decoder is used to perform decoding operation on the down-sampling result in S122. Similarly, on the basis of the U-Net structure, the up-sampling unit and the visual GNN unit are used to realize.

[0116] As shown in the above-mentioned embodiment, the decoder structure of the up-sampling includes a plurality of deconvolution layers, and a visual GNN module is embedded in part or each convolution layer. Figure 5

[0117] That is, the decoder module can include a plurality of second feature processing modules, and the plurality of second feature processing modules iteratively process the output result of the encoder in the above-mentioned embodiment, i.e. the down-sampling result. Each second feature processing module includes an up-sampling unit, a fusion unit, a first deconvolution unit, a second graph neural network unit and a second deconvolution unit.

[0118] Similarly, as shown in the above-mentioned embodiment, similar to the feature processing module in the encoder structure, the second graph neural network unit includes a pooling GNN unit and a shift window GNN unit. Figure 5 Correspondingly, as shown in the above-mentioned embodiment, the output result of the pooling GNN unit in the second graph neural network unit is input into the shift window GNN unit, and the processing process thereof at the zl layer and the zl+1 layer is described.

[0119] Figure 11 ​​As shown, when upsampling the down-sampled sparse output result using the decoder structure described above, for the deconvolution processing of each layer, the following steps are specifically included:

[0120] S04, input the down-sampled result into the up-sampling unit for feature space amplification processing, and output the amplification result.

[0121] S05, based on the fusion unit, fuse the amplification result and the output down-sampled result of the corresponding first feature processing module, and output the fusion result;

[0122] S06, input the fusion result into the first deconvolution unit for deconvolution processing, and obtain the processing result after deconvolution processing;

[0123] S07, input the processing result after deconvolution processing into the second graph neural network for feature integration processing, and obtain the result after integration processing;

[0124] S08, input the result after integration processing into the second deconvolution unit for convolution processing, and output the prediction result.

[0125] Specifically, for the output result of the encoder in the above embodiment, i.e. the down-sampled result (feature representation f4), first, it can pass through the up-sampling unit for feature space amplification processing, such as being input into the up-sampling module with a convolution kernel size of 2x2x2 for processing, amplifying the feature space to obtain the amplification result, i.e. the feature representation f5.

[0126] Further, as shown, Figure 5 the amplification result is fused with the down-sampled result of the corresponding level in the encoder, i.e. the feature output from the graph neural network unit and the feature representation f5, and the channel number of the feature tensor before and after fusion is doubled, and the image size remains unchanged, to obtain the fusion result, i.e. the feature representation f6.

[0127] As can be understood, Figure 5 the output of the encoder is fused in the decoder, i.e. the output of a certain down-sampling in the encoder is input into the corresponding up-sampling structure for up-sampling fusion processing.

[0128] Further, the feature representation f6 is input into the first deconvolution unit, such as passing through the convolution unit with a convolution kernel size of 3x3x3, to output the deconvolution result. For example, the feature representation f6 is processed by the convolution module with a convolution kernel size of 3x3x3, the channel number is halved, the image size remains unchanged, and the feature representation f61 is obtained, i.e. the processing result after deconvolution processing.

[0129] Further, the intermediate feature representation f61 is input to an efficient visual GNN module including a pooling GNN and a shift window GNN for processing to obtain a result after integration processing, i.e., a feature representation f7.

[0130] Finally, the feature representation f7 is input to a second deconvolution unit, e.g., a convolution module with a convolution kernel size of 1x1x1 for processing to obtain a corresponding prediction result, i.e., a segmentation result.

[0131] It can be understood that the output segmentation result of each second feature processing module, i.e., each time in the segmentation model, has a channel number consistent with the number of anatomical types of the last cerebral blood vessels, a total of 18 categories. In practice, the output result of the uppermost layer can be taken as the segmentation result of the target.

[0132] It can also be understood that the specific structure of the second graph neural network unit in the embodiment is similar to the first graph neural network structure in the encoder in the above embodiment, i.e., the processing of the input data is also through the processing of the pooling graph unit, the pooling feedforward network unit, the shift window graph unit and the shift window unit, which will not be described here.

[0133] It can be understood that in the segmentation model construction in the embodiment of the application, as supervised machine learning, the output result of each layer processing in the decoder, i.e., the prediction result of each layer, can calculate the corresponding loss function to perform back propagation and update the model parameters of the layer.

[0134] For example, in the embodiment of the application, the loss function defined by the neural network with the nnU-Net framework as the backbone structure can specifically include a cross-entropy loss function L ce and a similarity (Dice) loss function L dice .

[0135] i.e., L total = L ce + L dice .

[0136] Further, in some embodiments of the application, in order to efficiently and accurately distinguish different blood vessel categories, reduce the misclassification of adjacent blood vessels, and ensure that the topology of the segmentation is correct, the definition of the loss function in the segmentation model introduces a binary topology interaction loss , i.e., L_pixel is used as a pixel-level loss function to improve the performance of the model.

[0137] That is, the loss function in the embodiment of the application can include three parts: a cross-entropy loss L ce , a binary topology interaction loss , which is expressed as follows:

[0138]

[0139] From the above formula, the loss function in the embodiment of the application combines the cross-entropy loss L ce and the Dice loss with adjustable weights and For the 3D brain blood vessel multi-class segmentation task, The value can be 1e-6.

[0140] The binary topological interaction loss can be calculated by the following formula:

[0141]

[0142] Wherein, f is the prediction result of each layer up-sampling output in the decoder module, V is the key pixel map processed by f through the binary topological interaction module, and g is the configured segmentation standard map, that is, the segmentation ground truth labeled by an expert.

[0143] That is, as shown in Figure 13 In some embodiments, the calculation of the loss function of the prediction result output by each layer up-sampling module in the decoder during each training process specifically includes:

[0144] S005, based on the binary topological interaction module, the structure relationship of each anatomical region in the prediction result is encoded to obtain the processed key pixel map.

[0145] S006, the key pixel map is point multiplied with the configured segmentation standard map to determine the corresponding point multiplication result as the first loss value;

[0146] S007, the prediction result is point multiplied with the key pixel map to determine the corresponding point multiplication result as the second loss value.

[0147] S008, the first loss value and the second loss are fused to determine the fusion result as the binary topological interaction loss.

[0148] S009, the binary topological interaction loss, the cross-entropy loss and the similarity loss are fused, and the corresponding fusion result is the loss function corresponding to the prediction result.

[0149] Specifically, as Figure 12As shown, in the model training process, for each output prediction result, when calculating the corresponding loss function, first, the prediction result is passed through the binary topological interaction module, the inclusive and exclusive relationship between different anatomical structures is quickly coded, and the key pixel graph V is obtained. Further, the key pixel graph V obtained in the above step is multiplied by the segmentation gold standard true value g obtained by expert annotation to obtain a first loss value. Further, the prediction result, i.e., the probability graph f, is multiplied by the key pixel graph V obtained in the above step as a second loss value. Finally, the two loss values obtained are fused to obtain the binary topological interaction loss function in the model training process.

[0150] It can be understood that the binary topological interaction module introduced in the embodiment of the present application, as shown in Figure 12 For a task with a total of four categories c=4 (α, β, γ, δ), compared with the traditional calculation, 4•(4-1) convolution operations are required; and after introducing the binary topological interaction module, i.e., the improved calculation method, only 2•(4-1) convolution operations are required.

[0151] It can be understood that after obtaining the binary topological interaction loss function value, it can be brought into the above total loss function calculation formula to determine the loss function value corresponding to the upsampling layer, and then the parameters in the segmentation model are updated.

[0152] It can be understood that in the embodiment of the present application, on the basis of realizing high-precision segmentation of multiple types of brain blood vessels, the binary topological interaction module greatly saves the calculation amount in the classification process of each type of brain blood vessel.

[0153] It can also be understood that the brain blood vessel recognition model constructed by the above method can be used for various recognition and segmentation of brain blood vessels in actual magnetic resonance images in some embodiments, and further the network structure of the brain blood vessels is constructed using the segmentation result to comprehensively analyze the structural information of the brain blood vessels in the magnetic resonance image.

[0154] As shown in Figure 13 The method specifically comprises:

[0155] S210, acquiring a magnetic resonance image to be processed, the magnetic resonance image comprising anatomical regions of brain blood vessels;

[0156] S220, inputting the magnetic resonance image into a pre-constructed segmentation model to output a brain blood vessel prediction result of the magnetic resonance image, the prediction result being used to represent position information and categories of each anatomical region, the segmentation model being constructed based on an nnU-Net framework, a graph neural network being embedded in the nnU-Net framework, and the graph neural network being used to fuse the anatomical regions.

[0157] S230, according to the prediction result, the network structure of the brain blood vessel in the magnetic resonance image is constructed, and the network structure is used to represent the structural information of the brain blood vessel.

[0158] Specifically, in the embodiment of the present application, in order to realize accurate recognition and segmentation of each anatomical region of the brain blood vessel in the magnetic resonance image, and to construct a comprehensive brain blood vessel network structure as a whole, a new type of neural network is used, that is, a neural network model based on the nnU-Net framework, such as the efficient topological perception U-Net (NexToU) network structure.

[0159] In the new type of neural network structure, in order to realize the integration of each anatomical region in the brain blood vessel and improve the accuracy and efficiency of brain blood vessel recognition, a graph neural network, that is, a visual GNN variant, is embedded in the backbone neural network structure of the nnU-Net framework, so as to integrate information from different anatomical regions of the brain blood vessel and extract tubular structures by using the visual GNN variant, solve the inclusion and exclusion relationship of each anatomical structure, and realize accurate segmentation of the brain blood vessel structure.

[0160] Further, after accurately recognizing each anatomical region of the brain blood vessel by using the constructed segmentation model and obtaining the corresponding prediction result, the prediction result of the recognition can be used to further construct the network structure of the brain blood vessel, so that the network structure can represent the structural information of the brain blood vessel as a whole, that is, the structural information of the brain blood vessel can be extracted from the network structure. For example, the anatomical type, blood flow direction and network level features of the brain blood vessel can be extracted from the constructed network structure.

[0161] Specifically, in combination with Figure 6 As shown in the figure, when the brain blood vessel network is constructed by using the prediction result output by the segmentation model, the following four stages can be realized:

[0162] In the first stage, a topology sparsification algorithm is used to convert the binary segmentation volume into a skeleton-based representation, which is simplified into a simpler structure for subsequent processing and analysis. In the second stage, a memory-efficient algorithm is used to extract the blood vessel centerline, and the sparse blood vessel structure is analyzed and processed to extract the blood vessel foreground centerline. In the third stage, the mapping relationship between the prototype graph and the pixels of the binary segmentation is calculated, and the extracted blood vessel centerline is one-to-one corresponding to the pixels in the binary segmented blood vessel image. In the fourth stage, based on the prototype graph and the pixel branch assignment result, a plurality of features are calculated from the foreground segmentation, including the diameter, length, curvature, branch point and intersection point of the blood vessel.

[0163] Further, when the prediction result output by the segmentation model is subjected to brain blood vessel structure information graph extraction, the nodes of the information graph and the edges of the information graph are obtained respectively. Then, the blood vessel morphological features, the blood vessel anatomical regions, the blood vessel network hierarchical structure and the blood flow direction are embedded as graph attributes on the edges to obtain the network structure of the brain blood vessels. That is, the brain blood vessels of different anatomical regions recognized can be combined with the network structure constructed, and the node features and edge features of the brain blood vessel network are added.

[0164] The node features can include node serial number, node coordinates, radius, node degree, anatomical type, bifurcation point anatomical type, blood flow graph level and blood flow path distance. The edge features can include edge serial number, starting node serial number, ending node serial number, length, distance, curvature, volume, average cross section, average radius, roundness average value, starting node degree, ending node degree, voxel number, surface area, blood vessel anatomical category, segmentation index and blood flow direction.

[0165] The node features of the brain blood vessel network obtained can include node serial number , node coordinates , radius , node degree , anatomical type , bifurcation point anatomical type , blood flow graph level and blood flow path distance , etc. The edge features obtained can include edge serial number , starting node serial number , ending node serial number , length , distance , curvature , volume , average cross section , average radius , roundness average value , starting node degree , ending node degree , voxel number , surface area , blood vessel anatomical category , segmentation index , blood flow direction , etc.

[0166] It can be understood that, in the embodiments of the present application, for each type of anatomical region of the brain blood vessels of the acquired brain magnetic resonance image to be processed, when the new neural network structure with the nnU-Net framework as the backbone structure and embedded with a graph neural network is used for segmentation and identification, the graph neural network module embedded therein can be used to extract and integrate the brain blood vessel structures of each different anatomical region of the brain blood vessels in the magnetic resonance image, so as to solve the inclusion and exclusion relationship between each anatomical structure, and finally identify and segment the accurate prediction result of the brain blood vessels, effectively reducing the inter-individual variability.

[0167] Further, in the embodiments of the present application, by using the creation module in the identification model, the prediction result output by the segmentation model is used to construct a corresponding network structure that can reflect the brain blood vessel structure information, so as to comprehensively describe the brain blood vessel network structure and provide accurate and comprehensive scientific basis for large-scale clinical data analysis.

[0168] That is, the new neural network constructed in the present application automatically outputs the network structure of the brain blood vessels, which is suitable for image analysis of various brain blood vessel diseases such as stroke, vascular dementia, aneurysm and cerebral small vessel disease, and helps to better understand the influence of aging and other pathophysiological conditions on the brain blood vessels and their hemodynamics.

[0169] On the other hand, as shown in Figure 15 The present application provides a brain blood vessel identification model construction device, which comprises:

[0170] The first acquisition module 110 is configured to acquire training data, wherein the training data comprises a magnetic resonance image, and the magnetic resonance image comprises each anatomical region of brain blood vessels.

[0171] The construction module 120 is configured to input the training data into an initial identification model constructed based on an nnU-Net framework, train the initial identification model, and obtain a target identification model, wherein the target identification model is used to output a prediction result of each anatomical region in the magnetic resonance image, the prediction result is used to represent the position and category of each anatomical region, and the nnU-Net framework is embedded with a graph neural network, and the graph neural network is used to identify the structural relationship of each anatomical region in the brain blood vessels.

[0172] The generation module 130 is configured to generate a creation module, wherein the creation module is used to create a network structure of the brain blood vessels according to the prediction result, the network structure is used to represent the structural information of the brain blood vessels in the magnetic resonance image, and the structural information comprises anatomical types, blood flow directions and hierarchical features.

[0173] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application comprises an encoder module and a decoder module, the encoder module and the decoder module are embedded with the graph neural network, and the construction module specifically comprises:

[0174] a downsampling unit 121, configured to input the training data into the encoder module, extract features of each anatomical region in the training data, and obtain a downsampling result;

[0175] an upsampling unit 122, configured to input the downsampling result into the decoder module, integrate features of the downsampling result, and output a plurality of prediction results;

[0176] an updating unit 123, configured to determine a loss function corresponding to the prediction result according to the prediction result, update parameters of the initial recognition model according to the loss function, and obtain the target recognition model.

[0177] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application comprises a first downsampling module, a position encoding module and a second downsampling module, the graph neural network is embedded in the second downsampling module, and the downsampling unit is specifically configured to:

[0178] input the training data into the first downsampling module, perform convolution processing on the training data, and obtain a first downsampling result corresponding to the training data;

[0179] add position reference information of each anatomical region of the brain blood vessel in the first downsampling result by using the position encoding module, and obtain a second downsampling result;

[0180] input the second downsampling result into the second downsampling module, perform convolution and integration processing on the second downsampling result, and obtain the downsampling result.

[0181] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application comprises a plurality of first feature processing modules, the plurality of first feature processing modules perform iterative processing on the second downsampling result, the first feature processing module comprises a downsampling unit and a first graph neural network unit, and the downsampling unit is specifically configured to:

[0182] input the second downsampling result into the downsampling unit, perform feature enhancement on the second downsampling result, and obtain a processing result after feature enhancement;

[0183] input the processing result after enhancement into the image neural network unit, and integrate and recognize the brain blood vessel region in the processing result.

[0184] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application, the graph neural network unit comprises a pooling graph unit, a pooling feedforward network unit, a shift window graph unit and a shift window feedforward network unit, and the downsampling unit is specifically used for:

[0185] Based on the pooling graph unit, the global topology structure of the processing result is extracted, and the extracted result is added to the processing result to obtain a pooling result;

[0186] Based on the pooling feedforward network unit, the pooling result is processed in multiple layers, and the multiple layer processing results are added to the pooling result to obtain a first graph neural network processing result;

[0187] Based on the shift window graph unit, the first graph neural network processing result is shifted, and the shift processing result is added to the first graph neural network processing result to obtain a shift window processing result;

[0188] Based on the shift window feedforward neural unit, the shift window processing result is processed in multiple layers, and the multiple layer processing results are added to the shift window processing result to obtain a second graph neural network processing result.

[0189] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application, the decoder module comprises a plurality of second feature processing modules, and the plurality of second feature processing modules iteratively process the downsampling result. Each second feature processing module comprises an upsampling unit, a fusion unit, a second graph neural network module and a convolution unit, and the upsampling unit is specifically used for:

[0190] The downsampling result is input into the upsampling unit for feature space amplification processing, and the amplification result is output;

[0191] The amplification result and the downsampling result output by the corresponding first feature processing module are fused to output a fusion result;

[0192] The fusion result is input into the first deconvolution unit for deconvolution processing to obtain a deconvolution processed result;

[0193] The deconvolution processed result is input into the graph neural network for feature integration processing to obtain an integrated result;

[0194] The integrated result is input into the second deconvolution unit for deconvolution processing to output the prediction result.

[0195] Optionally, the brain blood vessel recognition model construction device provided by the embodiment of the present application, the loss function comprises a binary topology interaction loss, a cross-entropy loss and a similarity loss, and the updating unit is specifically used for:

[0196] Based on the binary topology interaction module, the structure relationship of each anatomical region in the prediction result is encoded, and a processed key pixel map is obtained.

[0197] The key pixel map is point multiplied with the configured segmentation standard map, and the corresponding point multiplication result is determined as a first loss value.

[0198] The prediction result is point multiplied with the key pixel map, and the corresponding point multiplication result is determined as a second loss value.

[0199] The first loss value and the second loss are fused, and the fusion result is determined as a binary topology interaction loss.

[0200] The binary topology interaction loss, the cross-entropy loss and the similarity loss are fused, and the corresponding fusion result is the loss function corresponding to the prediction result.

[0201] On the other hand, as shown in Figure 16 , the embodiment of the application provides a cerebral vascular network construction device, which comprises:

[0202] The acquisition module 210 is configured to acquire a magnetic resonance image to be processed, wherein the magnetic resonance image comprises anatomical regions of cerebral blood vessels.

[0203] The recognition module 220 is configured to input the magnetic resonance image into a target recognition model constructed by a cerebral blood vessel recognition model construction method, and output a prediction result corresponding to the magnetic resonance image, wherein the prediction result is used to represent the position and category of each anatomical region.

[0204] The creation module 230 is configured to construct a cerebral vascular network structure according to the prediction result, wherein the network structure is used to represent structure information of the cerebral blood vessels, and the structure information comprises anatomical types, blood flow directions and hierarchical features.

[0205] On the other hand, the computer device provided by the embodiment of the application further comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the image method as described above when executing the program.

[0206] The following refers to Figure 17 , Figure 17 the structural schematic diagram of the computer device of the embodiment of the application.

[0207] As Figure 17As shown, the electronic device includes a central processing unit (CPU) 301 that can perform various appropriate actions and processes in accordance with a program stored in a read only memory (ROM) 302 or a program loaded from a storage section 308 into a random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The CPU 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304. In some embodiments, the following components are connected to the I / O interface 305: an input section 306 including a keyboard, a mouse, etc.; an output section 307 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as necessary. A removable media 311 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 310 as necessary, so that a computer program read therefrom is installed into the storage section 308 as necessary. In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program in accordance with embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a machine-readable medium, the computer program containing program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 309, and / or installed from the removable media 311. When the computer program is executed by the central processing unit (CPU) 301, the above-described functions defined in the electronic device of the present application are performed.

[0208] The units or modules described in the embodiments of the present application can be implemented by software or by hardware. The described units or modules can also be arranged in a processor, for example, can be described as: a processor comprising an acquisition module and a construction module. In some cases, the names of these units or modules do not constitute a limitation on the units or modules themselves, for example, the construction module can also be described as: "for inputting the training data into an initial identification model constructed based on an nnU-Net framework, training the initial identification model to obtain a target identification model, the target identification model being used to output a prediction result of each anatomical region in the magnetic resonance image, the prediction result being used to represent the position and category of each anatomical region, the nnU-Net framework being embedded with a graph neural network, and the graph neural network being used to identify the structural relationship of each anatomical region in the brain blood vessel".

[0209] As another aspect, the application also provides a computer readable storage medium, which can be included in the electronic device described in the above embodiments, or can exist separately without being assembled into the electronic device. The computer readable storage medium stores one or more computer programs, which are used by one or more processors to execute the method for constructing a brain blood vessel recognition model of a magnetic resonance image provided by the above embodiments of the application, or to execute the method for recognizing brain blood vessels of a magnetic resonance image provided by the above embodiments of the application.

[0210] The above description is merely preferred embodiments of the application and a description of the principles of the technology used. Those skilled in the art should understand that the disclosed range of the application is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combinations of the above technical features or equivalent features without departing from the above disclosed concepts. For example, the above features can be replaced with technical features disclosed in the application (but not limited to) having similar functions to form technical solutions.

Claims

1. A method for constructing a cerebral vascular identification model, characterized in that: include: Acquiring training data, the training data including magnetic resonance images, the magnetic resonance images including various anatomical regions of cerebral blood vessels; Inputting the training data into an initial recognition model constructed based on the nnU-Net framework for training to obtain a target recognition model, wherein the target recognition model is used to output prediction results for each of the anatomical regions in the magnetic resonance image, wherein the prediction results are used to characterize the position information and category of each of the anatomical regions, and wherein a graph neural network is embedded in the nnU-Net framework, and wherein the graph neural network is used to identify the structural relationship between the anatomical regions in the cerebral blood vessels; creating a network structure of cerebral blood vessels according to the prediction results, wherein the network structure is used to represent structural information of cerebral blood vessels in magnetic resonance images; The initial recognition model and the target recognition model include an encoder module and a decoder module, wherein the encoder module and the decoder module are embedded with a graph neural network. The training data is input into the initial recognition model built based on the nnU-Net framework for training, and the steps of obtaining the target recognition model include: Inputting the training data into the encoder module, extracting features of each anatomical region in the training data, and outputting a downsampling result; Inputting the down-sampling results into the decoder module, performing feature integration on the down-sampling results, and outputting multiple prediction results; Determine a loss function corresponding to the prediction result according to the prediction result, and use the loss function to update the parameters of the initial recognition model to obtain the target recognition model; The loss function includes binary topology interaction loss, cross entropy loss and similarity loss, and the loss function corresponding to the prediction result is determined according to the prediction result. Based on a binary topology interaction module, the structural relationship of each anatomical region in the prediction result is encoded to obtain a key pixel map; Performing a dot product of the key pixel map and the configured segmentation standard map, and determining a corresponding dot product result as a first loss value; Performing a dot product of the prediction result and the key pixel map, and determining a corresponding dot product result as a second loss value; Fusing the first loss value and the second loss value to determine a fusion result as a binary topology interaction loss; The binary topology interaction loss, the cross entropy loss and the similarity loss are fused, and a fusion result is determined as the loss function corresponding to the prediction result.

2. The method for constructing a cerebral blood vessel identification model according to claim 1, wherein: The encoder module includes a first downsampling module, a position encoding module, and a second downsampling module. The graph neural network is embedded in the second downsampling module. The training data is input into the encoder module, the features of each anatomical region in the training data are extracted, and the output downsampling results include: Inputting the training data into the first downsampling module, performing convolution processing on the training data, and obtaining a first downsampling result corresponding to the training data; Using the position encoding module, adding the position reference information of each anatomical region to the first downsampling result to obtain a second downsampling result; The second down-sampling result is input into the second down-sampling module, and convolution and integration processing are performed on the second down-sampling result to obtain the down-sampling result.

3. The method for constructing a cerebral blood vessel identification model according to claim 2, wherein: The second downsampling module includes a plurality of first feature processing modules, the plurality of first feature processing modules iteratively process the second downsampling result, the first feature processing module includes a downsampling unit and a first graph neural network unit, Then, the first of the first feature processing modules in the second downsampling module performs convolution and integration processing on the second downsampling result, including: Inputting the second downsampling result into the downsampling unit, performing feature enhancement on the second downsampling result, and obtaining a feature-enhanced processing result; The enhanced processing results are input into the first image neural network unit, and the anatomical regions of the cerebral blood vessels in the processing results are integrated and identified.

4. The method for constructing a cerebral blood vessel identification model according to claim 3, wherein: The first image neural network unit includes a pooling image unit, a pooling feedforward network unit, a shift window image unit, and a shift window feedforward network unit. Inputting the enhanced processing result into the first image neural network unit, and integrating and identifying the anatomical regions of the cerebral blood vessels in the processing result includes: Based on the pooling graph unit, extracting a global topological structure of the enhanced processing result, and adding the extracted result to the processing result to obtain a pooling result; Based on the pooling feedforward network unit, perform multi-layer processing on the pooling result, and add the multi-layer processing result to the pooling result to obtain a first graph neural network processing result; Based on the shift window graph unit, shift processing is performed on the first graph neural network processing result, and the shift processing result is added to the first graph neural network processing result to obtain a shift window processing result; Based on the shift window feedforward neural unit, multi-layer processing is performed on the shift window processing result, and the multi-layer processing result is added to the shift window processing result to obtain a second graph neural network processing result.

5. The method for constructing a cerebral blood vessel identification model according to claim 4, characterized in that: The decoder module includes multiple second feature processing modules, which iteratively process the downsampling results. Each of the second feature processing modules includes an upsampling unit, a fusion unit, a first deconvolution unit, a second graph neural network unit, and a second deconvolution unit. The second feature processing module performs feature integration on the downsampling result, and outputs a corresponding prediction result including: Inputting the downsampling result into the upsampling unit for feature space amplification processing, and outputting the amplified result; Based on the fusion unit, the amplification result and the corresponding downsampling result output by the first feature processing module are fused and a fusion result is output; Inputting the fusion result into a first deconvolution unit for deconvolution processing to obtain a processing result after deconvolution processing; Inputting the deconvolution result into the second graph neural network for feature integration to obtain an integrated result; The result after the integration processing is input into the second deconvolution unit for deconvolution processing, and the prediction result corresponding to the second feature processing module is output.

6. A method for constructing a cerebral vascular network, characterized in that: include: Acquiring a magnetic resonance image to be processed, wherein the magnetic resonance image includes various anatomical regions of cerebral blood vessels; inputting the magnetic resonance image to be processed into the target recognition model constructed in the method for constructing a cerebral blood vessel recognition model according to any one of claims 1 to 5, and outputting a prediction result corresponding to the magnetic resonance image, wherein the prediction result is used to characterize the position and category of each of the anatomical regions; The cerebral blood vessel network structure is constructed according to the prediction result, and the network structure is used to represent the structural information of the cerebral blood vessels.

7. The method for constructing a cerebral vascular network according to claim 6, wherein: The cerebral vascular network structure is constructed according to the prediction result, and the network structure is used to characterize the structural information of the cerebral blood vessels, including: In the first stage, the binary segmented volume is converted into a skeleton-based representation using a topological sparsification algorithm; In the second stage, a memory-efficient algorithm is used to extract the center lines of each cerebral vessel, and the sparse cerebral vascular structure is analyzed and processed to extract the vascular foreground center line; In the third stage, the mapping relationship between the prototype image and the pixels of the binary segmentation is calculated, and the extracted blood vessel centerline is matched one-to-one with the pixels in the blood vessel image after the binary segmentation; In the fourth stage, based on the prototype image and the pixel branch correspondence results, the cerebral vascular parameter features are calculated. The cerebral vascular parameter features include the diameter, length, curvature, branch points and intersection points of the blood vessels.

8. The method for constructing a cerebral vascular network according to claim 7, wherein: Also includes: The identified cerebral vessels in different anatomical regions are combined with the constructed network structure, and the node features and edge features of the cerebral vascular network are added. The node features include node sequence number, node coordinates, radius, node degree, anatomical type, bifurcation point anatomical type, blood flow map level, and blood flow path distance; The edge features include edge number, start node number, end node number, length, distance, curvature, volume, average cross section, average radius, average roundness, start node degree, end node degree, number of voxels, surface area, vascular anatomy category, segmentation index and blood flow direction.

Images