Flow field data correction method and device, data processing equipment and storage medium

Abstract

The application provides a flow field data correction method, which comprises the following steps: obtaining an original measurement velocity field of a measurement region of a blood vessel image; performing image segmentation on the original measurement velocity field to obtain each sub-block; processing each sub-block by using a trained neural network model, and correcting the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following data: a corrected velocity field; a downstream calculation field associated with the corrected velocity field; and index data of a flow region surface. The application also provides a flow field data correction device, a data processing equipment and a storage medium.

Description

TECHNICAL FIELD

[0001] The present application relates to, but is not limited to, the field of new generation information technology, and in particular, relates to a flow field data correction method, a flow field data correction device, a data processing device, and a storage medium. BACKGROUND

[0002] In recent years, flow field visualization technology has been continuously developed and plays an increasingly important role in the industry and medical industry. For example, the particle image velocimetry (PIV) technology obtains the velocity field in the measurement region by analyzing the imaging of tracer particles on the camera. With the development of measurement technologies such as tomographic particle image velocimetry (TPIV), it is more and more convenient to obtain all three velocity components (3D3C) in a three-dimensional space. In clinical medical diagnosis, color Doppler ultrasound imaging, magnetic resonance imaging (MRI), and digital subtraction angiography (DSA) have been widely used for blood flow velocity measurement. These technologies use different principles to perform perspective imaging of the human body to obtain gray-scale image information representing the human body's internal organs, and can capture information about the blood flow in the human body to measure the blood flow velocity. Doctors analyze the patient's case based on the organ tissue morphology and blood flow velocity obtained by medical imaging and give a diagnosis. Therefore, accurate flow field measurement technology is of great significance to industrial production and medical diagnosis.

[0003] Although the above-mentioned flow field visualization technology has been continuously developed and improved in recent years, it is restricted by many factors such as imaging principles, contrast technology, and hardware device performance, and the quality of the fluid flow velocity measurement is not high. Currently, the optimization and correction of the measured flow field can be achieved by the following way: based on the fluid mechanics continuity equation, the velocity field is corrected to obtain the corrected velocity field.

[0004] However, in the real-time calculation process based on the fluid mechanics continuity equation, there is a problem of low calculation efficiency. SUMMARY

[0005] The present application provides a flow field data correction method, a flow field data correction device, a data processing device, and a storage medium.

[0006] The technical solution of the present application embodiment is implemented as follows:

[0007] A method for correcting flow field data, the method comprising:

[0008] obtaining an original measured velocity field of a measurement region of a blood vessel image;

[0009] performing image segmentation on the original measured velocity field to obtain each sub-block;

[0010] processing each sub-block by using a trained neural network model, and correcting the original measured velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following data:

[0011] a corrected velocity field;

[0012] a downstream calculation field associated with the corrected velocity field;

[0013] index data of a flow region surface.

[0014] A device for correcting flow field data, the device comprising:

[0015] an obtaining module configured to obtain an original measured velocity field of a measurement region of a blood vessel image;

[0016] a segmentation module configured to perform image segmentation on the original measured velocity field to obtain each sub-block;

[0017] a processing module configured to process each sub-block by using a trained neural network model, and correct the original measured velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following data:

[0018] a corrected velocity field;

[0019] a downstream calculation field associated with the corrected velocity field;

[0020] index data of a flow region surface.

[0021] Embodiments of the present application provide a data processing device, comprising a processor, a memory and a communication bus;

[0022] the communication bus is configured to realize communication connection between the processor and the memory;

[0023] the processor is configured to execute a correction program of flow field data stored in the memory, so as to realize steps of the correction method of flow field data.

[0024] Embodiments of the present application provide a storage medium, the storage medium stores one or more programs, the one or more programs can be executed by one or more processors to realize steps of the correction method of flow field data.

[0025] The method for correcting flow field data, the device for correcting flow field data, the data processing equipment and the storage medium provided by the application obtain an original measurement velocity field of a measurement region of a blood vessel image; perform image segmentation on the original measurement velocity field to obtain each sub-block; process each sub-block by using a trained neural network model, and correct the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data includes at least one of the following data: a corrected velocity field; a downstream calculation field associated with the corrected velocity field; and index data of a flow region surface. That is, the method for correcting flow field data provided by the application segments the original measurement velocity field to obtain a series of to-be-calculated sub-blocks, then inputs each sub-block into a trained neural network model for calculation, corrects the original measurement velocity field, and thus obtains corrected flow field data, thereby achieving the purpose of correcting the original measurement velocity field by using a neural network model for flow field optimization calculation, and improving the calculation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 Flowchart of the method for correcting flow field data provided by the embodiment of the application Figure 1 ;

[0027] Figure 2 Flowchart of the method for correcting flow field data provided by the embodiment of the application Figure 2 ;

[0028] Figure 3 Flowchart of the method for correcting flow field data provided by the embodiment of the application Figure 3 ;

[0029] Figure 4 Structural diagram of a first neural network provided by the embodiment of the application;

[0030] Figure 5 Structural diagram of a second neural network provided by the embodiment of the application;

[0031] Figure 6 Structural diagram of a third neural network provided by the embodiment of the application;

[0032] Figure 7 Structural diagram of a fourth neural network provided by the embodiment of the application;

[0033] Figure 8 Flowchart of the method for correcting flow field data provided by the embodiment of the application Figure 4 ;

[0034] Figure 5 Flowchart of the method for correcting flow field data provided by the embodiment of the application Figure 6 ;

[0035] Figure 7 A segmentation schematic diagram provided for an embodiment of the present application;

[0036] Figure 8 A structure schematic diagram of a flow field data correction device provided for an embodiment of the present application;

[0037] Figure 9 A structure schematic diagram of a data processing device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the personnel in the technical field better understand the present application scheme, the technical scheme in the present application embodiment will be described clearly and completely in the following with reference to the drawings in the present application embodiment. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor are within the scope of protection of the present application.

[0039] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, not to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0040] In this paper, the phrase "embodiment" means that the specific features, structures or characteristics described in conjunction with the embodiment can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment to other embodiments. The skilled person in the art explicitly and implicitly understands that the embodiments described herein can be combined with other embodiments.

[0041] The data processing device provided by the embodiments of the present application can be implemented as any data processing device such as a notebook computer, a tablet computer, a desktop computer, a mobile device (for example, a personal digital assistant, a dedicated message device), a smart robot, etc., or as a server. In the following, an exemplary application of the data processing device when implemented as a data processing device will be described.

[0042] The embodiments of the present application provide a flow field data correction method, which is applied to a data processing device, as shown in Figure 10 The method comprises the following steps.

[0043] Step 101, obtaining an original measurement velocity field of a measurement region of a blood vessel image.

[0044] In the embodiments of the present application, the data processing device can be directly connected to a flow field measurement device, such as a PIV measurement device, a magnetic resonance device, etc., to obtain the original measurement velocity field of the measurement region of the blood vessel image.

[0045] Step 102, image segmentation is performed on the original measurement velocity field to obtain each sub-block.

[0046] Step 103, the neural network model trained is used to process each sub-block, and the original measurement velocity field is corrected to obtain the corrected flow field data.

[0047] In the embodiments of the present application, each sub-block segmented is input to the neural network model trained, and the trained neural network model is used to calculate a series of sub-blocks to obtain the calculation results of each sub-block. Finally, the results of each sub-block are combined to form a complete calculation result, so as to correct the original measurement velocity field to obtain the corrected flow field data. The overlapping parts between adjacent sub-blocks adopt the average values of the overlapping parts.

[0048] The corrected flow field data includes at least one of the following data:

[0049] The corrected velocity field;

[0050] The downstream calculation field associated with the corrected velocity field;

[0051] The index data of the flow region surface.

[0052] In the embodiments of the present application, the downstream calculation field includes but is not limited to the pressure field, the vorticity field, and the wall shear stress. The index data of the flow region surface includes but is not limited to the average pressure, the residual fraction of a certain fluid of multiphase flow, and the flow field velocity distribution uniformity index. That is, the neural network model trained can not only obtain the corrected velocity field, but also can directly calculate the pressure field, the vorticity field, etc. downstream calculation field based on the corrected velocity field, and / or obtain the flow field macroscopic statistical index, i.e. the index data of the flow region surface, by using the trained neural network model. It can be seen that the neural network model for flow field optimization calculation is used to process the original measurement velocity field, which is not only flexible in calculation, but also can meet the various calculation requirements in actual scenes.

[0053] The neural network model for flow field optimization calculation is used to correct the original measurement velocity field, the fast optimization calculation of the original measurement velocity field is realized, the required computing resources and computing time are reduced, and real-time calculation can be realized in actual application. The calculation method is simple and efficient, the steps are simple, and the calculation process after the training of the neural network model does not require human intervention and optimization.

[0054] The method for correcting flow field data provided in the embodiments of the present application obtains an original measurement velocity field of a measurement region of a blood vessel image, performs image segmentation on the original measurement velocity field to obtain each sub-block, processes each sub-block by using the trained neural network model, corrects the original measurement velocity field, and obtains corrected flow field data, wherein the corrected flow field data includes at least one of the following data: a corrected velocity field, a downstream calculation field associated with the corrected velocity field, and index data of a flow region surface. That is, the method for correcting flow field data provided in the present application segments the original measurement velocity field to obtain a series of to-be-calculated sub-blocks, then inputs each sub-block into the trained neural network model for calculation, corrects the original measurement velocity field, and obtains corrected flow field data, thereby achieving the purpose of correcting the original measurement velocity field by using the neural network model for flow field optimization calculation, and improving the calculation efficiency.

[0055] In some embodiments of the present application, the corrected velocity field satisfies the following constraint condition: the divergence of the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity at the geometric boundary of the flow field. The neural network model trained in the present application is used to correct the original measurement velocity field, and the corrected velocity field satisfies the above constraint condition. It can be understood that, for the corrected velocity field obtained in the present application and satisfying the above constraint condition, if the geometric boundary of the flow field is stationary, the velocity at the boundary of the flow field is zero, and if the geometric boundary of the flow field is moving, the velocity at the boundary of the flow field is equal to the moving velocity at the geometric boundary of the flow field. In the process of training the neural network model, different boundary condition datasets can be used to train the neural network model, so that the trained neural network model can process flow field optimization problems under different boundary conditions.

[0056] In some embodiments of the present application, the model loss function of the trained neural network model comprises a first loss function and a second loss function, wherein the first loss function is a loss function constructed by the mean square error between the predicted value calculated by inputting the segmented sample measured velocity field into the neural network model and the segmented sample flow field data, and the second loss function is a loss function constructed by the mean square error between the gradient of the predicted value calculated by inputting the segmented sample measured velocity field into the neural network model and the gradient of the segmented sample flow field data; wherein the segmented sample flow field data comprises data obtained by image segmentation on the sample index data of the corrected sample velocity field, the downstream calculation field calculated based on the corrected sample velocity field, and the surface of the flow field. It can be seen that the loss function of the neural network model of the present application is composed of two parts of loss, for example, the sum of the first loss function and the second loss function constitutes the total loss function, and then the sum is optimized by training, so as to obtain the required neural network model. It should be noted that the segmented sample flow field data can also be referred to as sample label value, which can be calculated by a traditional non-dispersion optimization method.

[0057] For example, in the embodiments of the present application, the total loss function is represented by L total , the first loss function is represented by L mse , and the second loss function is represented by L gradient , wherein L total =L mse +L gradient .

[0058] In the three-dimensional case,

[0059]

[0060]

[0061] In the two-dimensional case,

[0062]

[0063] wherein are the predicted values in x, y, and z directions respectively, are the label values in x, y, and z directions respectively, are the gradients of the predicted values in x, y, and z directions respectively, are the gradients of the label values in x, y, and z directions respectively, and the gradient in the present application can be calculated by using the central difference method. W, H, and D are the sizes of the x, y, and z directions of the cube respectively, and according to the above calculation formula, in the two-dimensional case, there is no z direction and only H and W are included. In the discrete case, the gradient in the loss function can be calculated by using the re-difference format.

[0064] In one implementable scenario of flow field data correction, a neural network model can be constructed, and then the constructed neural network model is trained, and finally, the trained neural network model is used to process the blood vessel image, so as to correct the original measurement velocity field of the measurement region of the blood vessel image. Here, the construction of the neural network model is described in combination with the following flow field correction process. Figure 11 The processing process in this scenario is described as follows:

[0065] Step 201, constructing a neural network model for flow field optimization calculation.

[0066] In the embodiment of the present application, the overall correction process of the flow field can be described by the equation as shown below:

[0067] U c =f(U exp )

[0068] Wherein, U c is the corrected flow field data, including at least one of the following: corrected velocity field; downstream calculation field associated with the corrected velocity field; index data of the flow domain surface. U exp is the original measurement velocity field, and function f represents the entire flow field correction process. The "universal approximation theorem" of the neural network shows that a feedforward neural network with a linear output layer and at least one hidden layer with any "squeezing" activation function, such as logistic sigmoid activation function, can approximate any measurable function from one finite-dimensional space to another finite-dimensional space with sufficient number of hidden units. Therefore, a neural network can be constructed to approximate the flow field correction process function f.

[0069] In the embodiment of the present application, the constructed neural network structure includes an input layer, an output layer and an intermediate layer. The intermediate layer includes a contraction network connected with the input layer, used for feature extraction of the input; an expansion network connected with the contraction network, used for feature restoration; and a feature transfer layer connecting the contraction network and the expansion network, used for transferring the shallow detail features obtained in the feature extraction process of each sub-block by the contraction network to the expansion network of the trained neural network model. In the present application, the expansion network is directly connected with the deep abstract features, and the deep abstract features are obtained by continuous sampling, and the information will be lost. It is not enough to perform flow field restoration calculation based on the deep abstract features only. The feature transfer layer set between the contraction network and the expansion network in the present application transfers a series of features of different scales in the feature extraction stage to the feature restoration structure and combines the features of the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration ability of the neural network for the flow field.

[0070] In other embodiments of the present application, the intermediate layer can further include a flatten layer connected with the contraction network, configured to perform shape transformation on the deep abstract features representing the global features of the flow field, and flatten them into 1-dimensional vectors, and then set a fully connected layer to perform nonlinear transformation on the flattened feature vectors, and finally set a fully connected layer as the second output layer, configured to output the macroscopic statistical indicators of the flow field.

[0071] Step 202, collect flow field measurement data of different geometric basins under different working conditions to construct a data set.

[0072] Step 203, train the neural network using the data set.

[0073] Step 204, collect input data in actual application and input into the trained neural network model to calculate the required results.

[0074] In some embodiments of the present application, the above step 103 uses the trained neural network model to process each sub-block and correct the original measurement velocity field to obtain the corrected flow field data, which can be realized by the steps as shown in the following figure: Figure 12

[0075] Step 1031, take each sub-block as the input data of the trained neural network model.

[0076] Step 1032, perform feature extraction on each sub-block through the contraction network of the trained neural network model to obtain deep abstract features.

[0077] The deep abstract features include the first scale global features in the input data.

[0078] In the embodiments of the present application, the contraction network is a network structure composed of convolution layers and pooling layers alternately, configured to perform feature extraction on the input data. The convolution layer is configured to extract features from the input data, and the pooling layer is configured to down-sample the input data. The down-sampling can reduce the learning parameters in the network and increase the calculation speed of the network, and can also make the receptive field of the convolution layer expand continuously so that the convolution layer can extract a series of features with different scales in the flow field from small to large. The overall effect of the contraction network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales and degrees of abstraction. The shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field.

[0079] ​Step 1033, the shallow layer detail features obtained by the contraction network in the feature extraction process of each sub-block are transferred to the expansion network of the trained neural network model through the feature transfer layer of the trained neural network model.

[0080] The shallow layer detail features include second scale local flow features in the input data, and the first scale is greater than the second scale.

[0081] In the embodiments of the present application, the feature transfer layer transfers a series of features of different scales in the feature extraction stage to the feature restoration structure and merges the features of the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration capability of the neural network for the flow field. Finally, a convolution layer is set as the first output layer, and the output dimension of the convolution layer is the same as the input of the input layer, so as to obtain the corrected flow field and the pressure field, the vorticity field and other full flow field calculated based on the velocity field.

[0082] Step 1034, at least one of the corrected velocity field and the downstream calculation field is calculated based on the deep abstract features and the shallow detail features through the expansion network; and / or.

[0083] Step 1035, the deep abstract features are shape-transformed through the flattening layer of the trained neural network model to obtain flattened feature vectors, and the flattened feature vectors are nonlinearly transformed through the fully connected layer of the trained neural network model to obtain index data of the flow field surface.

[0084] In the embodiments of the present application, after feature extraction, two branches can be divided, the first branch is provided with a set of expansion network structure composed of deconvolution layers, the deconvolution layers have the functions of up-sampling and feature restoration calculation, and the velocity field is gradually restored from the encoded features through a series of up-sampling and feature restoration calculation. The second branch is provided with a flattening layer, which is used to shape-transform the deep abstract features representing the global features of the flow field and flatten them into one-dimensional vectors. Then, a fully connected layer is set as the second output layer, which is used to output the macroscopic statistical indicators of the flow field. Since the macroscopic statistical indicators represent the global macroscopic statistical information of the flow field, the second branch can calculate the macroscopic statistical indicators of the flow field from the deep abstract features representing the global features of the flow field. In order to increase the nonlinear fitting capability of the neural network and facilitate training, an activation layer and a batch normalization layer are arranged after each convolution layer and fully connected layer except the output layer.

[0085] In some embodiments of the present application, the original measured velocity field includes a three-dimensional fluid velocity field, the contraction network has a network structure composed of three-dimensional convolution layers and pooling layers alternately, and the expansion network has a network structure composed of three-dimensional deconvolution layers.

[0086] Here, the first neural network is exemplarily given, based on which the optimization calculation of the three-dimensional velocity field can be realized. Referring to Figure 1 As shown, the first neural network has an input layer and two output layers, and the input layer inputs the original measured velocity field. Then a set of shrinking network structure composed of three-dimensional convolutional layers and pooling layers alternately is set to extract features from the input original measured velocity field, wherein the convolutional layer is used to extract features from the input velocity field, and the pooling layer is used to down-sample the input velocity field. The down-sampling can reduce the learning parameters in the network and increase the calculation speed of the network, and can also make the receptive field of the convolutional layer expand continuously so that the convolutional layer can extract a series of features with different scales in the flow field from small to large. The overall effect of the shrinking network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales and degrees of abstraction. The shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field. After the feature extraction module, two branches are divided, the first branch is provided with a set of expanding network structure composed of three-dimensional deconvolutional layers. The three-dimensional deconvolutional layer has the functions of up-sampling and feature restoration calculation, and gradually restores the velocity field from the encoded features through a series of up-sampling and feature restoration calculations. In the implementation process, the present application also takes into account that the deep abstract features are directly connected with the expanding network structure, and the information will be lost due to the continuous down-sampling. Therefore, it is not enough to calculate the flow field based on the deep abstract features. In order to improve the detail restoration ability of the neural network to the flow field, a feature transfer layer is set on this basis, which is used to transfer a series of features with different scales in the feature extraction stage to the feature restoration structure and merge the features with the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration ability of the neural network to the flow field. Finally, a convolutional layer is set as the first output layer, and the output dimension of the convolutional layer is the same as the input dimension of the original velocity field input by the input layer, so as to obtain the corrected flow field and the pressure field, vorticity field and other whole flow field calculated based on the velocity field. The second branch after the feature extraction module is provided with a flattening layer, which is used to transform the shape of the deep abstract features representing the global features of the flow field into a 1-dimensional vector. Then a fully connected layer is set to perform non-linear transformation on the flattened feature vector. Finally, a fully connected layer is set as the second output layer, which is used to output the macroscopic statistical indicators of the flow field. Since the macroscopic statistical indicators represent the global macroscopic statistical information of the flow field, the second branch can calculate the macroscopic statistical indicators of the flow field from the deep abstract features representing the global features of the flow field. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are set after each convolutional layer and fully connected layer except the output layer.

[0087] Here, the second neural network is exemplarily given, and based on the neural network, the optimization calculation of the three-dimensional velocity field can also be realized. Referring to Figure 2 As shown, the second neural network has an input layer and an output layer, and the input layer inputs the original measured velocity field. Then a set of shrinkage network structure composed of three-dimensional convolution layer and pooling layer alternately is set to extract features from the input original measured velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer is used to down-sample the input velocity field. On the one hand, the down-sampling can reduce the learning parameters in the network and increase the calculation speed of the network, and on the other hand, the receptive field of the convolution layer can be continuously expanded so that the convolution layer can extract a series of features with different scales in the flow field from small to large. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales and degrees of abstraction. The shallow layer features represent the small-scale local flow structure in the input flow field, and the deep abstract features represent the large-scale global feature information in the input flow field. Then a set of expansion network structure composed of three-dimensional deconvolution layer is set, and the three-dimensional deconvolution layer has the functions of up-sampling and feature restoration calculation. After a series of up-sampling and feature restoration calculation, the velocity field is gradually restored from the encoded features. In the implementation process of the present application, it is also considered that the deep abstract features are directly connected with the expansion network structure, and the information will be lost due to the continuous down-sampling. Therefore, it is not enough to only restore the flow field based on the deep abstract features. In order to improve the detail restoration ability of the neural network to the flow field, a feature transfer layer is set on this basis, which is used to transfer a series of features with different scales in the feature extraction stage to the feature restoration structure and merge the features with the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration ability of the neural network to the flow field. Finally, a three-dimensional convolution layer is set as the output layer, and the output dimension of the output layer is the same as the input dimension of the input layer. The output layer is used to output the corrected flow field and the pressure field, vorticity field and other whole flow field calculated based on the velocity field. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are set after each convolution layer and fully connected layer except the output layer.

[0088] In some embodiments of the present application, the original measured velocity field includes a two-dimensional fluid velocity field, the shrinkage network has a network structure composed of two-dimensional convolution layer and pooling layer alternately, and the expansion network has a network structure composed of two-dimensional deconvolution layer.

[0089] Here, the third neural network is exemplarily given, and based on the neural network, the optimization calculation of the two-dimensional velocity field can be realized. Referring to Figure 3As shown, the third neural network has an input layer and two output layers, the input layer inputs the original measured velocity field, and then a set of shrinking network structure composed of two-dimensional convolutional layers and pooling layers is set to extract features from the input original measured velocity field, wherein the convolutional layer is used to extract features from the input velocity field, and the pooling layer is used to down-sample the input velocity field, which can reduce the learning parameters in the network and increase the calculation speed of the network, and on the other hand, the receptive field of the convolutional layer can be expanded, so that the convolutional layer can extract a series of features with different scales in the flow field. The overall effect of the shrinking network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales and degrees of abstraction, wherein the shallow features represent the small-scale local flow structure in the input flow field, and the deep abstract features represent the large-scale global feature information in the input flow field. The feature extraction module is then divided into two branches, the first branch is provided with a set of expanding network structure composed of two-dimensional deconvolutional layers, the two-dimensional deconvolutional layer has the functions of up-sampling and feature restoration calculation, and gradually restores the velocity field from the encoded features through a series of up-sampling and feature restoration calculations. In the implementation process, the present application also takes into account that the deep abstract features are directly connected with the expanding network structure, and the information will be lost due to the continuous down-sampling of the deep abstract features. Therefore, it is not enough to calculate the flow field based on the deep abstract features. In order to improve the detail restoration ability of the neural network to the flow field, a feature transfer layer is set on this basis, which is used to transfer a series of features with different scales in the feature extraction stage to the feature restoration structure and combine the features with the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration ability of the neural network to the flow field. Finally, a convolutional layer is set as the first output layer, and the output dimension of the convolutional layer is the same as the input dimension of the original velocity field input by the input layer, so as to obtain the corrected flow field and the pressure field, the vorticity field and other whole flow field calculated based on the velocity field. The second branch after the feature extraction module is provided with a flattening layer, which is used to transform the shape of the deep abstract features representing the global features of the flow field into a 1-dimensional vector, and then a fully connected layer is set to perform nonlinear transformation on the flattened feature vector, and finally a fully connected layer is set as the second output layer, which is used to output the macroscopic statistical indicators of the flow field. Since the macroscopic statistical indicators represent the global macroscopic statistical information of the flow field, the second branch can calculate the macroscopic statistical indicators of the flow field from the deep abstract features representing the global features of the flow field. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are set after each convolutional layer and fully connected layer except the output layer.

[0090] Here, the fourth neural network is exemplarily given, based on which the optimization calculation of the two-dimensional velocity field can be realized. Referring to Figure 4As shown, the fourth neural network has an input layer and an output layer, the input layer inputs the original measured velocity field, and then a set of shrinkage network structure composed of multiple two-dimensional convolution layers and pooling layers alternately is set to extract features from the input original measured velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer is used to down-sample the input velocity field, which can reduce the learning parameters in the network and increase the calculation speed of the network, and on the other hand, the receptive field of the convolution layer can be expanded to extract a series of features with different scales in the flow field. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales and degrees of abstraction, wherein the shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field. Then a set of expansion network structure composed of two-dimensional deconvolution layers is set, the two-dimensional deconvolution layer has the functions of up-sampling and feature restoration calculation, and the velocity field is gradually restored from the encoded features through a series of up-sampling and feature restoration calculations. In the implementation process of the present application, it is also considered that the deep abstract features are directly connected with the expansion network structure, and the deep abstract features are obtained by continuous sampling, so the information will be lost, and therefore it is not enough to restore the flow field based on the deep abstract features. In order to improve the detail restoration ability of the neural network to the flow field, a feature transfer layer is set on this basis, which is used to transfer a series of features with different scales in the feature extraction stage to the feature restoration structure and combine the features with the same scale, so that the network can restore the final flow field based on the deep abstract features and the shallow detail features, thereby enhancing the detail restoration ability of the neural network to the flow field. Finally, a convolution layer is set as the output layer, and the output dimension of the convolution layer is the same as the input dimension of the original velocity field input by the input layer, so as to output the corrected flow field and the pressure field, vorticity field and other full flow field calculated based on the velocity field. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are set after each convolution layer and fully connected layer except the output layer.

[0091] In some embodiments of the present application, before the step 103 of correcting the original measured velocity field by using the trained neural network model to process each sub-block to obtain the corrected flow field data, the training can be performed through the steps as shown in the following. Figure 5 The trained neural network model is obtained through the steps as shown in the following.

[0092] Step 301, obtaining the segmented sample measured velocity field and the segmented sample flow field data.

[0093] The segmented sample flow field data includes a corrected sample velocity field, a corrected downstream calculation field, and sample index data of a flow basin surface; and the corrected sample velocity field satisfies the following constraint conditions: the divergence inside the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity of the geometric boundary of the flow field.

[0094] In the embodiments of the application, the constraint condition at the boundary of the flow field is set, and the constraint condition at the boundary is divided into two kinds: one is a stationary boundary condition, that is, during the measurement process, the boundary of the flow field is always stationary, and according to the no-slip boundary condition of the fluid, the velocity of the fluid at the boundary is zero. That is, in this case, the fluid needs to satisfy two boundary conditions: the internal divergence is zero, and the velocity at the boundary is zero. The other case is a moving boundary condition, that is, during the measurement process, the boundary of the flow field will move, and according to the no-slip boundary condition of the fluid, the velocity of the fluid at the boundary should be equal to the movement velocity of the geometric boundary of the flow field. That is, in this case, the fluid needs to satisfy two boundary conditions: the internal divergence is zero, and the velocity at the boundary is equal to the movement velocity of the geometric boundary of the flow field. The no-slip boundary condition means that the outermost layer of the fluid is tightly attached to the geometric boundary and does not move relative to the geometric boundary, so the outermost layer of the flow field moves as the geometric boundary moves.

[0095] In the embodiments of the application, referring to FIG. 1, the data set can be constructed by the following steps: Figure 6

[0096] Step 401: Collecting original measurement velocity fields of different geometric flow basins.

[0097] Step 402: Correcting the original measurement velocity field by using a correction algorithm, and calculating a downstream field and / or a flow basin surface index.

[0098] Step 403: Standard segmentation of data.

[0099] Step 404: Constructing a data set.

[0100] Here, as many original measurement velocity fields as possible of various geometric flow basins under various working conditions such as stationary geometric boundary and moving geometric boundary are collected; then the original velocity field is corrected by using a no-divergence smoothing processing algorithm and a boundary no-slip boundary condition to obtain a corrected velocity field, and the optimized velocity field needs to satisfy two constraint conditions: the internal divergence of the flow field is zero, and the velocity at the boundary is consistent with the velocity of the geometric boundary of the flow field. If the geometric boundary of the flow field is stationary, the boundary velocity of the flow field is zero; if the geometric boundary of the flow field moves, the velocity at the boundary of the flow field is equal to the movement velocity at the boundary of the flow field. The data set constructed by using different boundary conditions can train the trained stretching-in network model to process flow field optimization problems under different boundary conditions. Finally, the required pressure field, vorticity field, wall shear stress, and other downstream flow basin fields or surface indexes are calculated based on the corrected velocity field according to the actual needs of specific scenes.​

[0101] Since the input dimension of the neural network needs to be set in advance, such as in the two-dimensional case, the input dimension is 256x256, and the actual measured velocity field size varies, it is necessary to standardize the original velocity field and the corrected velocity field and the downstream flow field or surface indicators calculated based on the corrected velocity field. The segmentation method is shown in Figure 7 , Figure 8 The two-dimensional segmentation case is shown, where H and W are consistent with the dimensions of the input layer of the network model, and a is the overlap when segmenting, that is, the starting position of the next part is moved forward by size a from the end position of the previous part segmentation. When segmenting, the same size overlap is kept in each segmentation direction. According to Figure 9 The segmentation method shown is used to segment the original velocity field and the corrected velocity field and the pressure field, vorticity field, wall shear stress and other downstream flow field calculated based on the corrected velocity field. The segmented original measured velocity field is used as input, and the corrected velocity field and the pressure field, vorticity field, wall shear stress and other downstream flow field calculated based on the corrected velocity field or macroscopic statistical indicators are used as labels to construct a data set for training.

[0102] Step 302, using the segmented sample measurement velocity field as the input data of the neural network model, using the corrected sample velocity field and the corrected downstream calculation field as the first output label, and using the sample indicator data of the flow field surface as the second output label, the neural network model is trained to obtain the trained neural network model.

[0103] In the embodiments of the present application, the constructed neural network model can be trained using an adaptive matrix estimation (Adam) optimizer. The final total loss function is composed of the sum of the first loss function L mse and the second loss function L gradient After training and optimization, the required neural network model can be obtained.

[0104] For example, the construction process of the first neural network is as follows: the first neural network comprises an input layer and an output layer, the dimension of the input layer is (64, 64, 64, 3) for inputting the original measured velocity field, and a shrinkage network structure composed of four two-dimensional convolution layers and pooling layers is arranged after the input layer to extract features from the input original velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer has a step of 2 and is used to downsample the input velocity field. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of scales such as (64, 64, 64), (32, 32, 32), (16, 16, 16), (8, 8, 8) and different abstract features, wherein the shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field; the feature extraction module is divided into two branches, the first branch is a dilatation network structure composed of three three-dimensional deconvolution layers after the feature extraction module, wherein the step of deconvolution is 2, and different scales such as (16, 16, 16), (32, 32, 32), ((64, 64, 64) are gradually restored from the features. In addition, in order to enhance the neural network's ability to restore the details of the flow field, a feature transfer layer is arranged to transfer the features of different scales (16, 16, 16), (32, 32, 32), (64, 64, 64) extracted in the feature extraction stage to the feature restoration structure, and the features of the same scale are combined. Finally, a convolution layer is arranged as the final output layer, and the output layer has a dimension of (64, 64, 64, 4) to output the velocity field u, v, w and the pressure field p. The second branch after the feature extraction module is arranged with a flatten layer to flatten the features extracted by the feature extraction module into a feature vector, and then a fully connected layer is arranged to perform nonlinear transformation on the feature vector. Finally, a fully connected layer is arranged as the second output layer to output the macroscopic statistical indicators of the flow field. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are arranged after each convolution layer and fully connected layer except the output layer.

[0105] For example, the construction process of the second neural network is as follows: the second neural network comprises an input layer and an output layer, the dimension of the input layer is (64, 64, 64, 3) for inputting the original measured velocity field, and a shrinkage network structure composed of four two-dimensional convolution layers and pooling layers is arranged after the input layer to extract features from the input velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer has a step of 2 and is used to down-sample the input velocity field, which can reduce the learning parameters in the network, increase the calculation speed of the network, and expand the receptive field of the convolution layer to extract a series of features with different scales in the flow field. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales such as (64, 64, 64), (32, 32, 32), (16, 16, 16), (8, 8, 8) and different degrees of abstraction, wherein the shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field; then a dilatation network structure composed of three three-dimensional deconvolution layers is arranged, wherein the step of deconvolution is 2, and different scales such as (16, 16, 16), (32, 32, 32), (64, 64, 64) are gradually restored from the features; in addition, in order to enhance the calculation ability of the neural network for restoring the details of the flow field, a feature transfer layer is also arranged, which is used to transfer the features with different scales (16, 16, 16), (32, 32, 32), (64, 64, 64) extracted in the feature extraction stage to the feature restoration structure and combine the features with the same scale; finally, a convolution layer is arranged as the final output layer, and the output layer has a dimension of (64, 64, 64, 3) to output the velocity field u, v and w. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are arranged after each convolution layer and fully connected layer except the output layer.

[0106] For example, the third neural network is constructed as follows: the third neural network comprises an input layer and an output layer, the dimension of the input layer is (256, 256, 2) for inputting the original measured velocity field, and a shrinkage network structure composed of five two-dimensional convolution layers and pooling layers is arranged after the input layer to extract features from the input velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer with a step of 2 is used to down-sample the input velocity field. On the one hand, the down-sampling can reduce the learning parameters in the network and increase the calculation speed of the network, and on the other hand, the receptive field of the convolution layer can be expanded to extract a series of features with different scales from small to large in the flow field. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales such as (256, 256), (128, 128), (64, 64), (32, 32), (16, 16) and different degrees of abstraction, wherein the shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field; then a dilatation network structure composed of four two-dimensional deconvolution layers is arranged, wherein the step of the deconvolution is 2, and different scales such as (32, 32), (64, 64), (128, 128), (256, 256) are gradually restored from the features. In addition, in order to enhance the calculation ability of the neural network for restoring the details of the flow field, a feature transfer layer is arranged to transfer the features with different scales such as (32, 32), (64, 64), (128, 128), (256, 256) extracted in the feature extraction stage to the feature restoration structure and combine the features with the same scale. Finally, a convolution layer is arranged as the final output layer, and the output layer has a dimension of (256, 256, 3) to output the velocity field u, v and the pressure field p. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are arranged after each convolution layer and fully connected layer except the output layer.

[0107] For example, the fourth neural network is constructed as follows: the fourth neural network comprises an input layer and an output layer, the input layer has a dimension of (256, 256, 2) for inputting the original measured velocity field, and a shrinkage network structure composed of five two-dimensional convolution layers and pooling layers alternately arranged after the input layer is arranged to extract features from the input original velocity field, wherein the convolution layer is used to extract features from the input velocity field, and the pooling layer has a step of 2 and is used to downsample the input velocity field, which can reduce the learning parameters in the network, increase the calculation speed of the network, and also enable the receptive field of the convolution layer to expand continuously, so that the convolution layer can extract a series of features with scales from small to large in the flow field. The overall effect of the shrinkage network structure is equivalent to a hierarchical encoder, which encodes the input flow field into a series of features with different scales such as (256, 256), (128, 128), (64, 64), (32, 32), (16, 16) and different degrees of abstraction, wherein the shallow features represent small-scale local flow structures in the input flow field, and the deep abstract features represent large-scale global feature information in the input flow field; then a dilatation network structure composed of four two-dimensional deconvolution layers is arranged, wherein the deconvolution layers have a step of 2, and gradually restore features with different scales such as (32, 32), (64, 64), (128, 128), (256, 256) from the features. In addition, in order to enhance the calculation ability of the neural network for restoring the details of the flow field, a feature transfer layer is arranged to transfer the features with different scales such as (32, 32), (64, 64), (128, 128), (256, 256) extracted in the feature extraction stage to the feature restoration structure and combine the features with the same scale. Finally, a convolution layer is arranged as the final output layer, and the output layer has a dimension of (256, 256, 3) to output the velocity fields u, v and the pressure field p. In order to increase the nonlinear fitting ability of the neural network and facilitate training, an activation layer and a batch normalization layer are arranged after each convolution layer and fully connected layer except the output layer.

[0108] For example, in a scenario of generating a data set, a data set for training can be obtained by the following process: collecting as many MRI original measured velocity fields of aortic blood vessels of different patients as possible, and then using a velocity field smoothing algorithm and a boundary non-slip boundary condition to correct the original velocity field to obtain a corrected velocity field, wherein the optimized divergence field satisfies two constraint conditions that the velocity divergence of the internal flow field is zero and the velocity of the boundary flow field is consistent with the velocity of the geometric boundary of the flow field. Then, the corrected velocity field is divided into a plurality of training samples according to the following steps: Figure 10The shown segmentation manner segments the original measurement velocity field and the corrected velocity field, and the sizes W, H and D of the segmentation are 64, 64 and 64 respectively, and the overlapping size a between adjacent two sub-blocks is 8. Finally, the original velocity field after segmentation is taken as input, and the corrected velocity field and the pressure field are used to construct a data set required for model training.

[0109] For example, when calculating the statistical index, such as the flow field uniformity index CV, on each sub-block, the following formula can be used for calculation:

[0110] wherein,

[0111] wherein, is the average velocity, V j is the sampling point at the j-th position, and the smaller the CV value is, the more uniform the flow field is. The above-mentioned original velocity field after segmentation is taken as input, the corrected velocity field and the pressure field are taken as the first output label, and the flow field uniformity index CV is taken as the second output label, so as to construct a data set for model training. Wherein, j and n are positive integers, and the maximum value of n is the total number of all sub-blocks.

[0112] The embodiment of the present application provides a correction device for flow field data, which can be used for realizing Figure 10 The corresponding embodiment provides a correction method for flow field data, which refers to Figure 10 As shown in the figure, the flow field data correction device 500 comprises:

[0113] The obtaining module 501 is configured to obtain an original measurement velocity field of a measurement region of a blood vessel image;

[0114] The segmentation module 502 is configured to perform image segmentation on the original measurement velocity field to obtain each sub-block;

[0115] The processing module 503 is configured to process each sub-block by using the trained neural network model, and correct the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following data:

[0116] The corrected velocity field;

[0117] The corrected downstream calculation field associated with the corrected velocity field;

[0118] The index data of the flow domain surface.

[0119] In some embodiments of the present application, the corrected velocity field satisfies the following constraint condition: the divergence of the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity at the geometric boundary of the flow field.

[0120] In some embodiments of the present application, the model loss function of the trained neural network model comprises a first loss function and a second loss function, wherein the first loss function is a loss function constructed by the mean square error between the predicted value calculated by inputting the segmented sample measurement velocity field into the neural network model and the segmented sample flow field data, and the second loss function is a loss function constructed by the mean square error between the gradient of the predicted value calculated by inputting the segmented sample measurement velocity field into the neural network model and the segmented sample flow field data.

[0121] The segmented sample flow field data comprises data obtained by image segmentation on the corrected sample velocity field, the downstream calculation field calculated based on the corrected sample velocity field, and the sample index data of the flow field surface.

[0122] In some embodiments of the present application, the processing module 503 is configured to use each sub-block as input data of the trained neural network model, extract features of each sub-block by a contraction network of the trained neural network model to obtain deep abstract features, wherein the deep abstract features comprise first scale global features in the input data, transfer shallow detail features obtained in the feature extraction process of each sub-block by the contraction network to an expansion network of the trained neural network model through a feature transfer layer of the trained neural network model, wherein the shallow detail features comprise second scale local flow features in the input data, and the first scale is greater than the second scale, calculate at least one of the corrected velocity field and the downstream calculation field based on the deep abstract features and the shallow detail features through the expansion network, and / or perform shape transformation on the deep abstract features through a flattening layer of the trained neural network model to obtain flattened feature vectors, and perform nonlinear transformation on the flattened feature vectors through a fully connected layer of the trained neural network model to obtain the index data of the flow field surface.

[0123] In some embodiments of the present application, the original measurement velocity field comprises a three-dimensional fluid velocity field, the contraction network has a network structure composed of three-dimensional convolution layers and pooling layers alternately, and the expansion network has a network structure composed of three-dimensional deconvolution layers.

[0124] In some embodiments of the present application, the original measurement velocity field comprises a two-dimensional fluid velocity field, the contraction network has a network structure composed of two-dimensional convolution layers and pooling layers alternately, and the expansion network has a network structure composed of two-dimensional deconvolution layers.

[0125] In some embodiments of the present application, the obtaining module 501 is configured to obtain the segmented sample measurement velocity field and the segmented sample flow field data, wherein the segmented sample flow field data comprises the corrected sample velocity field, the corrected downstream calculation field and the sample index data of the drainage surface; the corrected sample velocity field satisfies the following constraint condition: the divergence inside the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity of the geometric boundary of the flow field;

[0126] The processing module 503 is configured to take the segmented sample measurement velocity field as the input data of the neural network model, take the corrected sample velocity field and the corrected downstream calculation field as the first output label, and take the sample index data of the drainage surface as the second output label, train the neural network model, and obtain the trained neural network model.

[0127] The flow field data correction device provided by the embodiments of the present application obtains the original measurement velocity field of the measurement region of the blood vessel image; performs image segmentation on the original measurement velocity field to obtain each sub-block; processes each sub-block by using the trained neural network model, and corrects the original measurement velocity field to obtain the corrected flow field data, wherein the corrected flow field data comprises at least one of the following data: the corrected velocity field; the downstream calculation field associated with the corrected velocity field; and the index data of the drainage surface. That is, the flow field data correction method provided by the present application segments the original measurement velocity field to obtain a series of to-be-calculated sub-blocks, then inputs each sub-block into the trained neural network model for calculation, corrects the original measurement velocity field, and thus obtains the corrected flow field data, thereby achieving the purpose of correcting the original measurement velocity field by using the neural network model for flow field optimization calculation, and improving the calculation efficiency.

[0128] The embodiments of the present application provide a data processing device, Figure 7 The data processing device 600 shown in the figure comprises a processor 601, a memory 602 and a communication bus 603, wherein:

[0129] The communication bus 603 is configured to realize the communication connection between the processor 601 and the memory 602.

[0130] The processor 601 is configured to execute the flow field data correction program stored in the memory 602 to realize Figure 1 The corresponding embodiments provide a flow field data correction method.

[0131] The processor can be an integrated circuit chip with a processing capability of signals, such as a general purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., and the general purpose processor can be a microprocessor or any conventional processor.

[0132] The data processing device provided in the embodiment of the present application obtains the original measurement velocity field of the measurement region of the blood vessel image, performs image segmentation on the original measurement velocity field to obtain each sub-block, processes each sub-block by using the trained neural network model, and corrects the original measurement velocity field to obtain the corrected flow field data, wherein the corrected flow field data includes at least one of the following data: the corrected velocity field; the downstream calculation field associated with the corrected velocity field; and the index data of the flow region surface. That is, the method for correcting the flow field data provided in the present application segments the original measurement velocity field to obtain a series of to-be-calculated sub-blocks, then inputs each sub-block into the trained neural network model for calculation, corrects the original measurement velocity field, and thus obtains the corrected flow field data, so as to achieve the purpose of correcting the original measurement velocity field by using the neural network model for flow field optimization calculation, and improve the calculation efficiency.

[0133] It should be noted that the specific implementation process of the steps performed by the processor in the embodiment can be referred to Figure 11 The implementation process in the method for correcting the flow field data provided in the corresponding embodiment will not be described herein.

[0134] The description of the device in the embodiment of the present application is similar to the description of the method embodiment described above, and has similar beneficial effects to the method embodiment, and thus will not be described herein. For technical details not disclosed in the device embodiment, please refer to the description of the method embodiment of the present application for understanding.

[0135] The embodiment of the present application provides a storage medium storing executable instructions, wherein the storage medium stores executable instructions, and when the executable instructions are executed by a processor, the processor will execute the method provided in the embodiment of the present application, for example, as Figure 12 Figure 1 Figure 1 Figure 1 The method shown in the figure.

[0136] The storage medium provided in the embodiments of the present application obtains an original measurement velocity field of a measurement region of a blood vessel image, performs image segmentation on the original measurement velocity field to obtain each sub-block, processes each sub-block by using a trained neural network model, and corrects the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data includes at least one of the following data: a corrected velocity field; a downstream calculation field associated with the corrected velocity field; and index data of a flow region surface. That is, the correction method of the flow field data provided in the present application segments the original measurement velocity field to obtain a series of to-be-calculated sub-blocks, then inputs each sub-block into the trained neural network model for calculation, corrects the original measurement velocity field, and thus obtains the corrected flow field data, so as to achieve the purpose of correcting the original measurement velocity field by using the neural network model for flow field optimization calculation, and improve the calculation efficiency.

[0137] In some embodiments, the storage medium can be a computer-readable storage medium, for example, a ferroelectric memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM), and the like. The storage medium can also be various devices including one or any combination of the above storage mediums.

[0138] In some embodiments, the executable instructions can be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and can be deployed in any form, including being deployed as independent programs or being deployed as modules, components, subroutines or other units suitable for use in a computing environment.

[0139] By way of example, an executable instruction can be, but is not limited to, a file, a part of a file, executing on a storage, a library, a service, a procedure, etc. Whether such an instruction is stored in an object, a library, a procedure, or a file (e.g., in a file that stores other programs or data), the instruction need not be stored in its own file, or in a file dedicated to such an instruction. By way of example, the executable files can be, but need not be, stored on memory devices, such as single or multiple USBs and / or hard drives, or in RAM or ROM. The executable files can be executed by various virtual machines, such as the Java Virtual Machine, or an emulator for the machine on which the files runs. The files carrying instructions can be executed by using virtual machines.

[0140] The above merely provides example embodiments of the present application, but is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application shall fall within the protection scope of the present application.

Claims

1. A method of modifying flow field data, characterized by, The method comprises: obtaining an original measurement velocity field of a measurement region of a blood vessel image; performing image segmentation on the original measurement velocity field to obtain each sub-block; processing each sub-block by using a trained neural network model to correct the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following: a corrected velocity field; a downstream calculation field associated with the corrected velocity field; index data of a flow region surface. The processing of each sub-block by using the trained neural network model to correct the original measurement velocity field to obtain corrected flow field data comprises: taking each sub-block as input data of the trained neural network model; extracting features of each sub-block by a contraction network of the trained neural network model to obtain deep abstract features, wherein the deep abstract features comprise first scale global features in the input data; transferring shallow detail features obtained in the feature extraction process of each sub-block by the contraction network to a dilatation network of the trained neural network model through a feature transfer layer of the trained neural network model, wherein the shallow detail features comprise second scale local flow features in the input data, and the first scale is greater than the second scale; calculating at least one of the corrected velocity field and the downstream calculation field based on the deep abstract features and the shallow detail features through the dilatation network; and / or performing shape transformation on the deep abstract features through a flattening layer of the trained neural network model to obtain flattened feature vectors, and performing nonlinear transformation on the flattened feature vectors through a fully connected layer of the trained neural network model to obtain the index data of the flow region surface.

2. The method of claim 1, wherein, The corrected velocity field satisfies the following constraint condition: the divergence inside the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity at the geometric boundary of the flow field.

3. The method of claim 1, wherein, The model loss function of the trained neural network model comprises a first loss function and a second loss function, wherein the first loss function is a loss function constructed by the mean square error between the predicted value calculated by inputting the segmented sample measurement velocity field into the neural network model and the segmented sample flow field data, and the second loss function is a loss function constructed by the mean square error between the gradient of the predicted value calculated by inputting the segmented sample measurement velocity field into the neural network model and the gradient of the segmented sample flow field data. The segmented sample flow field data comprises data obtained by performing image segmentation on the corrected sample velocity field, the downstream calculation field calculated based on the corrected sample velocity field, and the sample index data of the flow region surface, respectively.

4. The method of claim 1, wherein, The original measurement velocity field comprises a three-dimensional fluid velocity field, the contraction network has a network structure composed of three-dimensional convolution layers and pooling layers in an alternating manner, and the dilatation network has a network structure composed of three-dimensional deconvolution layers.

5. The method of claim 1, wherein, The original measurement velocity field comprises a two-dimensional fluid velocity field, the contraction network has a network structure of two-dimensional convolution layers and pooling layers alternately arranged, and the expansion network has a network structure of two-dimensional deconvolution layers.

6. The method according to any one of claims 1 to 5, characterized in that, Before the processing of the each sub-block by the trained neural network model and the correction of the original measurement velocity field to obtain the corrected flow field data, the method comprises: obtaining the segmented sample measurement velocity field and the segmented sample flow field data, wherein the segmented sample flow field data comprises a corrected sample velocity field, a corrected downstream calculation field and sample index data of a flow region surface; the corrected sample velocity field satisfies the following constraint condition: the divergence of the flow field is zero, and the velocity at the boundary of the flow field is the same as the velocity at the geometric boundary of the flow field; training the neural network model by taking the segmented sample measurement velocity field as input data of the neural network model, taking the corrected sample velocity field and the corrected downstream calculation field as first output labels, and taking the sample index data of the flow region surface as second output labels, to obtain the trained neural network model.

7. A flow field data correction device characterized by comprising: The device comprises: an obtaining module configured to obtain an original measurement velocity field of a measurement region of a blood vessel image; a segmentation module configured to perform image segmentation on the original measurement velocity field to obtain each sub-block; a processing module configured to process the each sub-block by using a trained neural network model, and correct the original measurement velocity field to obtain corrected flow field data, wherein the corrected flow field data comprises at least one of the following data: a corrected velocity field; a downstream calculation field associated with the corrected velocity field; index data of a flow region surface; The processing of the each sub-block by the trained neural network model and the correction of the original measurement velocity field to obtain the corrected flow field data comprises: taking the each sub-block as input data of the trained neural network model; extracting features of the each sub-block by a contraction network of the trained neural network model to obtain deep abstract features, wherein the deep abstract features comprise first scale global features in the input data; transferring shallow detail features obtained in the feature extraction of the each sub-block by the contraction network to an expansion network of the trained neural network model through a feature transfer layer of the trained neural network model, wherein the shallow detail features comprise second scale local flow features in the input data, and the first scale is greater than the second scale; calculating at least one of the corrected velocity field and the downstream calculation field based on the deep abstract features and the shallow detail features by the expansion network; and / or performing shape transformation on the deep abstract features by a flattening layer of the trained neural network model to obtain flattened feature vectors, and performing nonlinear transformation on the flattened feature vectors by a fully connected layer of the trained neural network model to obtain the index data of the flow region surface.

8. A data processing device, characterized by The data processing device comprises: a memory for storing executable instructions; a processor for executing the executable instructions stored in the memory to implement the method of modifying flow field data according to any one of claims 1 to 6.

9. A storage medium, characterized by executable instructions stored in the memory, when executed, to cause the processor to perform the method of modifying flow field data according to any one of claims 1 to 6.

Images