Ultra-fast ultrasonic motion artifact correction method based on VoxelMorph model

Abstract

The invention discloses a method for correcting ultrafast ultrasonic motion displacement based on a VoxelMorph model to solve motion artifacts. Comprising the following steps: acquiring ultrafast ultrasonic data, preprocessing the acquired data to obtain an initial data set, and dividing the initial data set into a training set and a test set; constructing a VoxelMorph motion correction network model based on the lightweight attention module, training the neural network model based on the training set to obtain a motion calibration network model, and testing the motion calibration network model based on the test set to obtain motion calibration data; and performing clutter filtering by using singular value decomposition to separate a blood flow signal and a tissue signal of the calibrated data, extracting the blood flow signal, and performing power Doppler blood flow imaging. According to the method, motion artifacts and random noise occurring in the ultrafast ultrasonic imaging process can be effectively restrained, blur between blood vessels is reduced, and the Doppler blood flow imaging contrast ratio is improved.

Description

Technical Field

[0001] The invention relates to the technical field of ultrafast ultrasound Doppler blood flow imaging, and in particular to a method for correcting ultrafast ultrasound motion artifacts based on a VoxelMorph model. Background Art

[0002] Ultrafast ultrasound imaging technology based on multi-angle composite plane waves can well balance imaging quality and imaging frame rate, and can measure the dynamic changes of tiny blood flows in milliseconds and hundreds of microns. It has been widely used in microvascular blood flow imaging. However, ultrafast ultrasound imaging is easily affected by motion artifacts caused by patient movement, breathing, and operator hand stability. The presence of motion artifacts can reduce image quality at best, or even affect doctors' clinical ultrasound interpretation, resulting in diagnostic errors.

[0003] The motion artifacts generated during the imaging process are mainly manifested as blur and artifacts. Common methods to reduce or eliminate these artifacts are generally: motion prevention, improving imaging acquisition methods and motion correction. Motion prevention generally requires the scanned person to keep the body still and hold his breath during the scanning process, but it will cause physiological discomfort to the patient. Improving the imaging acquisition method mainly reduces the periodic artifacts caused by physiological movements such as breathing and heartbeat by shortening the imaging time or selecting a time window with slow movement. This method cannot effectively suppress motion artifacts, and short-term sampling will also reduce image quality. Motion correction belongs to the data post-processing method, that is, after obtaining the imaging data, the image motion is estimated and the data is further integrated to achieve the effect of motion correction. However, the current data post-processing method often requires additional acquisition of a considerable number of images for processing to eliminate motion artifacts. At the same time, due to the high similarity of the overall structure of medical images, but the diversity of details, the semantic extraction precision is required to be high, which will prolong the imaging time efficiency, thereby reducing the practicality of clinical practice.

[0004] Related patents: In the Chinese patent with application number 202010790319.4, a method for ultrasonic motion calibration based on broadband pulse inversion is proposed. This method determines the displacement data by delaying two pulse echoes, and performs motion correction based on the displacement data. In the Chinese patent with application number 202011251430.2, the spatial position of the target under the influence of breathing is predicted by using the spatial position of the surface sensor at the current moment of intraoperative ultrasound data, and the static intervention path planned before surgery is corrected.

[0005] The application limitations of the above public technologies are:

[0006] 1. The registration algorithm of the deep learning model VoxelMorph has not yet been applied to the field of motion calibration of ultrafast ultrasound spinal cord images. At present, the algorithm based on VoxelMorph is mainly used for motion calibration in liver and heart datasets of MRI and PET. For deep learning algorithms, the same network structure may show different performance on different datasets. Therefore, in specific task scenarios, it is necessary to select a suitable network in combination with specific datasets. That is, the existing algorithm based on VoxelMorph needs to be innovated again to be applicable to motion calibration of ultrafast ultrasound spinal cord images.

[0007] 2. The method is less robust and requires a large amount of data sets for training and verification. Most deep learning-based algorithms rely heavily on the support of large data sets. However, due to the particularity of medical images, the amount of data is often small. It is crucial to find a model that can support both large and small data sets for training.

[0008] 3. The accuracy of feature extraction needs to be improved. The overall structure of medical images is highly similar, but the details are diverse, so the semantic extraction requires high precision. Although the structure of the organ itself is fixed, there are subtle differences between different people. Summary of the invention

[0009] In order to solve the technical problems existing in the above-mentioned background technology and to address the existing technical defects, the present invention provides the following solution: a method for ultrafast ultrasound motion artifact correction based on a VoxelMorph model, comprising:

[0010] S1, collecting ultrafast ultrasound data based on multi-angle composite plane waves;

[0011] S2, preprocessing the collected ultrafast ultrasound data to obtain an initial data set, and dividing the initial data set into a training set and a test set;

[0012] S3, build a VoxelMorph motion correction network model based on a lightweight attention module;

[0013] S4, training the motion correction network model based on the training set to obtain a motion calibration network model;

[0014] S5. Testing the motion calibration network model based on the test set to obtain motion calibration data;

[0015] S6, using singular value decomposition to perform clutter removal and separate the blood flow signal and tissue signal of the calibrated data;

[0016] S7. Further process the separated dynamic blood flow signal to obtain Doppler imaging results.

[0017] Preferably, the preprocessing process of the collected data in step S2 includes singular value decomposition, dimensionality conversion, and data enhancement operations;

[0018] The clutter filtering is to perform singular value decomposition on the collected ultrafast ultrasound data to separate tissue signals and blood flow signals;

[0019] The singular value decomposition is to filter out clutter on the collected ultrafast ultrasound data and separate tissue signals and blood flow signals;

[0020] The dimension conversion is to convert the collected ultrafast ultrasound data and the separated tissue signal data with an imaging depth z of 800 and a channel number x of 128 into 400 frames of data into a 2D image, that is, each of the 128 channels is divided into 8 parts, and the missing part of the imaging depth 800 compared to the size of 1024 is filled with 0 to the standard format, and the processed data is saved as 400 images with a size of 1024×1024.

[0021] The data enhancement operation is to flip the converted image vertically, horizontally, and horizontally and vertically. Horizontal mirroring is to flip the image left and right along the vertical center axis, vertical mirroring is to flip the image up and down along the horizontal center axis, and horizontal vertical mirroring is the superposition of horizontal mirroring and vertical mirroring, so as to obtain the initial data set.

[0022] Preferably, the initial data set of the data collected in step S2 includes a training set and a test set, wherein the training set accounts for 75% of the total data set, the test set accounts for 25% of the total data set, and the first image of the 400 images after dimension conversion is used as the ultrafast ultrasound reference image F, and the other images in the initial data set are used as motion images M. Preferably, the construction of the motion correction network model in step S3 includes the construction of CBAM-U-Net based on the lightweight attention module and the construction of the VoxelMorph model based on the lightweight attention module CBAM;

[0023] The construction process of the lightweight attention module CBAM-U-Net is to use the U-Net convolutional neural network as a basis and embed the lightweight attention module CBAM into the U-Net structure;

[0024] The VoxelMorph model based on the lightweight attention module CBAM includes CBAM-U-Net, spatial transformer network and local variance weighted loss function.

[0025] Preferably, the working process of the lightweight attention module CBAM-U-Net is as follows: based on the preprocessed ultrafast ultrasound reference image F and the motion image M, the CBAM-U-Net convolutional neural network g θThe encoder path of (F,M)=u is downsampled, where θ can be learned and optimized as a parameter. During the downsampling process, the number of channels increases and the feature map shrinks; a lightweight attention module CBAM is added at the end of the encoder. After the channel attention is first performed, the maximum pooling and average pooling are used, and the feature map is input into the multi-layer perceptron MLP with shared weights. After 1×1 convolution to fuse the channel information, the output features are added element by element according to their corresponding positions, and then the channel feature map M is obtained through the activation function sigmoid. c (F * ). Input feature map F * And channel feature map M c (F * ) correspond to each other and multiply them to generate the input features F' that subsequent operations depend on. The channel attention mechanism can be expressed as:

[0026]

[0027] Where σ represents the sigmoid function, W1 and W0 are the weights of the MLP shared input feature map, represents the average pooling feature, Represents the maximum pooling feature. AvgPool represents average pooling, and MaxPool represents maximum pooling.

[0028] The feature value F' generated by the channel attention is then pooled twice through spatial attention. The two pooling operations aggregate the channel information of a feature map to generate two two-dimensional feature maps: the average pool feature based on the channel axis and max pooling features Then, the two feature maps are combined and sent to the standard convolution to transform the learning parameters, and finally a two-dimensional weighted spatial attention feature map M is output. s (F'). The expression of the spatial attention module is:

[0029]

[0030] Where σ represents the sigmoid function, f 7×7 Represents a convolution operation with a filter size of 7×7. AvgPool represents average pooling, and MaxPool represents maximum pooling. This method can improve the feature extraction accuracy to achieve multi-scale feature extraction;

[0031] The decoder path includes upsampling the feature image, concatenating the feature map in the upsampling process with the feature map processed by the lightweight attention mechanism from the encoder path in the channel dimension, halving the image channels in the process of restoring the image resolution until the output image resolution is consistent with the input image resolution, and finally outputting the predicted deformation field φ=u+id, where φ is the same size as F and M, and id is a constant representing the part without displacement or deformation.

[0032] Preferably, the working process of the VoxelMorph model based on the lightweight attention module CBAM is to obtain deformation field parameters by passing the moving image and the reference image through the CBAM-U-Net convolutional neural network, and perform linear interpolation transformation on the moving image based on the deformation field parameters using the spatial transformation network, and the conversion method is as follows:

[0033]

[0034] p'=p+u(p);

[0035] Where p represents a pixel point, q represents p's neighboring pixels, p' represents pixel p weighted by pixels q in eight adjacent directions, and Z(p') represents the transformation parameters of the neighboring positions of the design point p.

[0036] Spatial transformation to obtain the deformed image The local weighted variance loss and deformation field smoothness loss are used as loss functions to measure the similarity between the reference image and the deformed image, and the training is completed when the evaluation indicators are met to obtain the final calibrated image.

[0037] Preferably, the loss function includes: deformation field smoothing loss and local weighted variance loss.

[0038] The deformation field smoothing loss calculation method is as follows:

[0039]

[0040]

[0041] In the formula, φ is used as the deformation field parameter. The loss can suppress the local sudden changes of the vector by setting the gradient size of u as the loss.

[0042] The local weighted variance loss is composed of a local variance function and a weight function. The calculation method of the local variance is as follows:

[0043]

[0044] in and Represents n centered on p 3The average grayscale value of the reference image of the neighborhood block and the average grayscale value of the motion image. LV After normalization, we get The σ parameter is introduced to adjust the weight factor coefficient and construct the weight function:

[0045]

[0046] Then the local weighted variance loss can be expressed as:

[0047]

[0048] Preferably, the training process in step S4 is to build a training model using the Pytorch deep learning framework, use the local weighted variance and deformation field smoothing loss function and the Adam optimizer to train the model, and after multiple iterations, adjust the model parameters to obtain a motion calibration network model.

[0049] Preferably, in step S5, selecting the test data set to test the motion calibration network model includes completing the performance evaluation of the model by inputting the test set into the motion calibration network model based on the mean square error MSE indicator, the peak signal-to-noise ratio PSNR indicator and the structural similarity SSIM indicator.

[0050] Preferably, in step S6, the ultrafast ultrasound data generally includes echo signals of static tissues, echo signals of blood flow and noise, and the image data after motion calibration needs to be filtered to remove the noise and data of static tissue signals;

[0051] Preferably, in step S7, Doppler blood flow imaging is performed on the data after clutter filtering. Doppler blood flow imaging is to express the blood flow signal in the form of intensity value. The Doppler average signal intensity is expressed as:

[0052]

[0053] Where x represents the direction parallel to the probe array, z represents the imaging depth, and S f (x,z,t i ) represents the signal after noise filtering, and N represents the number of sampling frames.

[0054] The present invention discloses the following technical effects:

[0055] The present invention provides a method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model. First, the preprocessed ultrafast ultrasound image is input into the CBAM-U-Net network to obtain the deformation field parameters. Subsequently, the motion image is spatially transformed based on the deformation field parameters to obtain the deformed image. The local weighted variance loss and the deformation field smoothing loss are used as loss functions to measure the similarity between the calibrated image and the reference image. The training is completed under the condition of meeting the evaluation index to obtain the calibrated image. The invention can greatly reduce the cost and complexity of the system, making it possible for the system to be widely used. At the same time, there is no need to obtain a large number of repeated images to eliminate artifacts. Better results can be obtained on a smaller training set, which solves the problem of small sample size of medical imaging data sets. At the same time, the lightweight attention module can also improve the accuracy of semantic feature extraction and improve the calibration effect. In summary, the invention has the characteristics of good calibration effect, simple structure and wide application range. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the specific embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0058] Figure 2 It is a network structure diagram of CBAM-U-Net according to an embodiment of the present invention;

[0059] Figure 3 A structural diagram of a VoxelMorph model according to an embodiment of the present invention;

[0060] Figure 4 The B-mode image and the residual image of the 256th frame image before and after calibration according to the embodiment of the present invention;

[0061] Figure 5 1 is a power Doppler image of rat spinal cord microblood flow before and after calibration according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] Ultrasound imaging motion artifacts are used to describe the phenomenon of blurring, artifacts, etc. in the image that cannot accurately represent the anatomical or physiological structure of the imaged object due to motion. The presence of motion artifacts may cause image blur or distortion at the least, or may cover up the lesion area and produce false images at the worst, making it impossible for doctors to accurately observe the lesion area or misjudge the lesion area, thus leading to misdiagnosis, missed diagnosis and other problems. The acquisition speed of ultrafast ultrasound images based on multi-angle composite plane waves can be within milliseconds, but the separation and processing of hemodynamic signals still requires the accumulation of dozens to hundreds of frames of images. During this measurement time window, the object being measured will produce relative displacement or morphological changes due to breathing and body organ movement, resulting in motion artifacts in the original image or processed microblood flow image, which in turn affects the quality of microblood flow images and also brings huge challenges to ultrafast ultrasound imaging research. Although a series of research progress has been made in the research of ultrasound image motion calibration, its application still has the following limitations:

[0065] (1) The feature extraction accuracy of ultrafast ultrasound images obtained in the spinal cord is difficult to meet the needs of motion calibration. Due to the complex structure of the spinal cord and the imaging mechanism of spinal microvessels, ultrafast ultrasound spinal cord blood flow imaging is difficult and has low spatiotemporal resolution and blurred microvessel images. Improving the feature extraction accuracy of reference images and motion images plays an important role in motion calibration applications based on deep learning algorithms.

[0066] (2) There are few studies on ultrafast ultrasound motion calibration of the spinal cord based on deep learning algorithms, and the motion calibration of ultrasound images is mainly focused on the interpretation of the liver and heart. Non-rigid motion correction methods based on deep learning have not been widely studied in ultrafast ultrasound spinal cord images.

[0067] In view of the status of VoxelMorph network structure in MRI image registration tasks and its excellent performance, this paper will solve the problem of ultrafast ultrasound motion artifact correction by building a VoxelMorph motion correction network model based on a lightweight attention module based on the VoxelMorph framework. The singular value decomposition method is used to perform clutter removal to separate the blood flow signal and tissue signal of the calibrated data, extract the blood flow signal, and perform power Doppler blood flow imaging.

[0068] Figure 1 The flowchart proposed by the present invention is a flowchart of ultrafast ultrasound motion artifact correction based on the VoxelMorph model, which is also applicable to this embodiment. This example uses a method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model provided by the present invention to perform calibration, mainly for ultrafast ultrasound images of rat spinal cord for motion calibration, but is not limited to motion calibration of rat spinal cord. It includes the following steps:

[0069] 1) Data acquisition: Data acquisition was performed by conducting in vivo experiments on Sprague-Dawley rats and using the multi-channel emission acquisition Vantage256 phased ultrasound experimental platform to perform ultrafast ultrasound imaging on the rat spinal cord data obtained by scanning. The Vantage256 phased ultrasound experimental platform is an experimental platform for ultrasound imaging and phased array ultrasound technology research. It provides a complete set of hardware and software tools for controlling and operating phased array ultrasound probes to acquire and process ultrasound images. The probe model used in this experimental platform is L22-14vX, which includes 128 array elements, a center frequency of 15.625MHz, and an array element spacing of 0.1mm. The imaging frame rate used for ultrafast ultrasound imaging is 400 frames / second, and each subsequence emits 15 tilted plane waves at uniform intervals in the range of -10° to 10°, with an imaging depth of 5.9mm. The acquired waveform data is input into the computer for processing to obtain image information at the corresponding position.

[0070] 2) Data preprocessing: The collected ultrafast ultrasound data is cluttered and the tissue signal and blood flow signal are separated; at the same time, the collected ultrafast ultrasound data and the separated tissue signal data are dimensional converted, and the 400 frames of data with an imaging depth z of 800 and a channel number x of 128 are converted into 2D images, that is, each of the 128 channels is divided into 8 parts, and the missing part of the imaging depth 800 compared to 1024 is filled with 0 to the standard format, and the processed data is saved as 400 images of size 1024×1024. The image obtained after the dimension conversion is flipped vertically, horizontally, and horizontally and vertically by using the data enhancement method, that is, the sample is rotated 90 degrees, 180 degrees, and 270 degrees, and the number of samples becomes 4 times the original. In this way, the diversity of samples can be increased while increasing the number of samples, which is conducive to improving the robustness of the model. After obtaining the initial data set, the collected ultrafast ultrasound data set is divided into a training set and a test set according to a certain ratio, where the training set accounts for 75% of the total data set and the test set accounts for 25% of the total data set. In this example, 1600 images are used as the experimental data set, of which 1200 images are used as the training set and 400 images are used as the test set. The first image of the 400 images after dimension conversion is used as the ultrafast ultrasound reference image F, and the other images in the initial data set are used as motion images M.

[0071] 3) Model construction:

[0072] 3.1) CBAM-U-Net network construction: Figure 2 As shown in the figure, the U-Net network structure is used as the basic network structure, and the lightweight attention module CBAM is embedded into the U-Net structure. The lightweight attention module can be used to infer the attention map from two independent dimensions, channel and space, in sequence, to capture the detailed features of the image that are difficult to distinguish along the channel dimension, and to focus on the area with more effective information along the spatial dimension, retaining the precise location information. Then, the generated attention map is multiplied with the input feature map and adaptive feature optimization is performed to form a pair of detail-aware and position-sensitive feature maps, which can be input into the feature map to enhance the representation of the target of interest.

[0073] The CBAM module and the local variance weighted loss function are embedded in the U-Net network structure.

[0074] 3.2) VoxelMorph model network construction: Figure 3As shown in the figure, the VoxelMorph model network model can be divided into three parts. The first is the convolutional network module based on CBAM-U-Net. Secondly, the motion artifacts are corrected through the spatial transformation network based on the deformation field parameters. Finally, the calibrated image and the reference image are measured by similarity measurement based on the local variance weighted loss function to measure the calibration result.

[0075] The most basic VoxelMorph model, MSE+UNET, is used to process images, but the problem of signal holes occurs. In the training process, poor feature extraction accuracy and over-parameterization often occur, resulting in extremely poor calibration results. This application makes optimization adjustments based on this method.

[0076] Specifically, the CBAM-U-Net convolutional neural network g is used to generate a pre-processed ultrafast ultrasound reference image F and a motion image M. θ The encoder path of (F,M)=u is downsampled, where θ can be learned and optimized as a parameter. During the downsampling process, the number of channels increases and the feature map is reduced to the original 116. A lightweight attention module CBAM is added at the end of each layer of the encoder. After channel attention, maximum pooling and average pooling are used, and the feature map is input into the multi-layer perceptron MLP with shared weights. After 1×1 convolution to fuse the channel information, the output features are added element by element according to their corresponding positions, and then the channel feature map M is obtained through the activation function sigmoid. c (F * ). Input feature map F * And channel feature map M c (F * ) correspond to each other and multiply them to generate the input features F' that subsequent operations depend on. The channel attention mechanism can be expressed as:

[0077]

[0078] Where σ represents the sigmoid function, W1 and W0 are the weights of the MLP shared input feature map, represents the average pooling feature, Represents the maximum pooling feature. AvgPool represents average pooling, and MaxPool represents maximum pooling.

[0079] The feature value F' generated by the channel attention is then pooled twice through spatial attention. The two pooling operations aggregate the channel information of a feature map to generate two two-dimensional feature maps: the average pool feature based on the channel axis and max pooling features Then, the two feature maps are combined and sent to the standard convolution to transform the learning parameters, and finally a two-dimensional weighted spatial attention feature map M is output. s (F'). The expression of the spatial attention module is:

[0080]

[0081] Where σ represents the sigmoid function, f 7×7 Represents a convolution operation with a filter size of 7×7. AvgPool represents average pooling, and MaxPool represents maximum pooling. This method can improve the feature extraction accuracy to achieve multi-scale feature extraction.

[0082] Finally, a standard convolution is performed to output a weighted feature map to extract the target. The addition of a lightweight attention module CBAM can improve the feature extraction accuracy to achieve multi-scale feature extraction.

[0083] Then the feature image is upsampled. The feature map in the upsampling process is spliced ​​with the feature map processed by the lightweight attention mechanism from the encoder path in the channel dimension. In the process of restoring the image resolution, the image channel is halved until the output image resolution is consistent with the input image resolution. Finally, the predicted deformation field is output to obtain the deformation field parameters. Based on the deformation field parameters, the spatial transformation network is used to perform linear interpolation transformation on the motion image. The conversion method is as follows:

[0084]

[0085] p'=p+u(p);

[0086] Where p represents a pixel point, q represents p's neighboring pixels, p' represents pixel p weighted by pixels q in eight adjacent directions, and Z(p') represents the transformation parameters of the neighboring positions of the design point p.

[0087] After the deformed image is obtained, the local weighted variance loss and deformation field smoothness loss are used as loss functions for similarity measurement, and the training is completed when the evaluation indicators are met to obtain the calibrated image.

[0088] In the application, since the rat spinal cord data is affected by respiratory movement, some areas have large local differences. Therefore, in order to achieve a high calibration effect, a local variance weighting operation based on MSE is selected, which can fully consider the grayscale information elements of the surrounding neighborhood feature blocks, accurately screen out areas with poor matching effects, and reduce the probability of mismatching of the calibrated feature areas. At the same time, this method can also delete some noise points to prevent spots and excessive smoothing. This method has good robustness and noise resistance, and is not easily affected by uneven grayscale distribution and poor contrast of the image.

[0089] 4) Model training: The computing devices used in the experiment are NVIDIA GeForceGTX2080ti GPU with 32GB video memory and Intel I7 9700 CPU. The software environment uses Python 3.9, and the calibration implementation architecture is Pytorch. The original learning rate is 0.001, the initial momentum of the Adam optimizer is set to 0.95, and the number of iterations is 40,000. The deformation field smoothing and local weighted variance loss functions are used.

[0090] The calculation method of deformation field smoothing loss is as follows:

[0091]

[0092]

[0093] In the formula, φ is used as the deformation field parameter. The loss can suppress the local sudden changes of the vector by setting the gradient size of u as the loss.

[0094] The local weighted variance loss is composed of the local variance function and the weight function. The calculation method of the local variance is as follows:

[0095]

[0096] in and Represents n centered on p 3 The average grayscale value of the reference image of the neighborhood block and the average grayscale value of the motion image. LV After normalization, we get The σ parameter is introduced to adjust the weight factor coefficient and construct the weight function:

[0097]

[0098] Then the local weighted variance loss can be expressed as:

[0099]

[0100] The Pytorch deep learning framework was used to build the training model, and the local weighted variance and deformation field smoothing loss function and Adam optimizer were used for model training. After multiple iterations, the model parameters were adjusted to obtain the motion calibration network model.

[0101] 5) Model evaluation: Draw the absolute grayscale subtraction map of the image before and after calibration and the reference image, and compare the residual results, such as Figure 4 As shown in the figure, mean square error (MSE), peak signal-to-noise ratio (PNSR) and structural similarity (SSIM) indicators are used to evaluate the performance of the model on the test set after training.

[0102] The mean square error (MSE) indicates the degree to which the measured data deviates from the true value. The smaller the value, the higher the similarity between the motion calibration image and the reference image. The calculation method is as follows:

[0103]

[0104] Among them, N represents the number of frames of data, x i and i represent the original data and the calibrated data, respectively.

[0105] The peak signal-to-noise ratio (PNSR) mainly indicates the ratio of the maximum signal power to the noise power that affects the image accuracy. The larger the value, the lower the distortion rate of the calibration image and the higher the image quality. The calculation method is as follows:

[0106]

[0107] Where L represents the maximum grayscale value of the image.

[0108] Structural similarity SSIM measures the brightness, contrast and structure of the image to compare the image similarity. It is symmetrical and the value range is (0,1). The larger the value, the smaller the difference between the corrected image and the reference image, the higher the similarity and the better the image quality. The calculation method is as follows:

[0109] SSIM(x,y)=[I(x,y)] α [C(x,y)] β [S(x,y)] γ ;

[0110] Among them α>0, β>0, γ>0>0.

[0111]

[0112]

[0113]

[0114] Where I(x,y) is the brightness comparison, C(x,y) is the contrast comparison, and S(x,y) is the structure comparison. y and μ x represent the mean values ​​of x and y, σ xy represents the covariance of x and y, σ x and σ y represents the standard deviation of x and y. c1, c2, c3 are constants to avoid system errors with denominators of 0. Let c3 = c2 / 2, α = β = γ = 1, then SSIM can be simplified to:

[0115]

[0116] 6) Clutter filtering:

[0117] The data of 400 consecutive frames of composite B-mode images before and after motion calibration are constructed into a two-dimensional matrix S f (x,z,t i ); perform singular value decomposition on the two-dimensional matrix to extract the blood flow signal, and the blood flow signal threshold of the singular value decomposition is set to (30,380).

[0118] 7) Power Doppler blood flow imaging: The extracted blood flow map is further processed to obtain a power Doppler blood flow map, such as Figure 5 Doppler blood flow imaging is to express the blood flow signal in the form of intensity value. The Doppler average signal intensity is expressed as:

[0119]

[0120] Where x represents the direction parallel to the probe array, z represents the imaging depth, and S f (x,z,t i ) represents the signal after noise filtering, and N represents the number of sampling frames.

[0121] Functions and effects of this embodiment:

[0122] In this embodiment, a method for ultrafast ultrasound motion artifact correction based on a VoxelMorph model proposed by the present invention is applied. This method can reduce residual artifacts of B-mode ultrasound images, such as Figure 4 As shown in the figure, the white residual traces in the residual artifact image are reduced, indicating that the model has a certain calibration effect. At the same time, the motion calibration image after clutter filtering is subjected to power Doppler blood flow imaging. It can be seen that the background noise of Doppler blood flow imaging after motion correction is significantly reduced, the contrast of Doppler blood flow imaging is improved, the vascular contour is clearer, and it shows a sharper edge and thinner blood vessels. In addition, by comparing the MSE, SSIM, PNSR and other data indicators of the images before and after correction, the MSE, SSIM, and PNSR before calibration are 126.10, 0.90, and 27.42, respectively, and the MSE, SSIM, and PNSR after calibration are 94.03, 0.92, and 28.77, respectively. The data optimization rates (expressed in absolute values) of each indicator are 25.43%, 2.22%, and 4.69%, respectively. It can be seen that the data indicators after motion artifact algorithm correction are well improved.

[0123] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. Without departing from the spirit of the present invention, various modifications and improvements made by ordinary technicians in this field to the technical solution of the present invention should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for ultrafast ultrasound motion artifact correction based on VoxelMorph model, characterized in that: The following steps are involved: S1, collecting ultrafast ultrasound data based on multi-angle composite plane waves; S2, preprocessing the collected ultrafast ultrasound data to obtain an initial data set, and dividing the initial data set into a training set and a test set; S3, build a VoxelMorph motion correction network model based on a lightweight attention module; S4, training the motion correction network model based on the training set to obtain a motion calibration network model; S5. Testing the motion calibration network model based on the test set to obtain motion calibration data; S6, using singular value decomposition to perform clutter removal and separate the blood flow signal and tissue signal of the calibrated data; S7. Further process the separated dynamic blood flow signal to obtain Doppler imaging results.

2. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 1, characterized in that: The preprocessing process of the collected data in step S2 includes clutter filtering, dimensionality conversion, and data enhancement operations; The clutter filtering is to perform singular value decomposition on the collected ultrafast ultrasound data to separate tissue signals and blood flow signals; The dimension conversion is to convert the collected ultrafast ultrasound data and the separated tissue signal data with an imaging depth z of 800 and a channel number x of 128 into 2D images: each of the 128 channels is divided into 8 parts, and the missing part of the imaging depth 800 compared to the size of 1024 is filled with 0 to a standard format, and the processed data is saved as 400 images with a size of 1024×1024; The data enhancement operation is to flip the converted image vertically, horizontally, and horizontally and vertically. Horizontal flipping is to flip the image left and right along the vertical center axis, vertical flipping is to flip the image up and down along the horizontal center axis, and horizontal and vertical flipping is the superposition of horizontal mirroring and vertical mirroring, so as to obtain the initial data set.

3. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 1, characterized in that: The initial data set for collecting data in step S2 includes a training set and a test set, wherein the training set accounts for 75% of the total data set, and the test set accounts for 25% of the total data set. The first image of the 400 images after dimension conversion is used as the ultrafast ultrasound reference image F, and the other images in the initial data set are used as motion images M.

4. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 1, characterized in that: The motion correction network model constructed in step S3 includes the construction of CBAM-U-Net based on the lightweight attention module and the construction of VoxelMorph model based on the lightweight attention module CBAM; The construction process of the lightweight attention module CBAM-U-Net is to use the U-Net convolutional neural network as a basis and embed the lightweight attention module CBAM into the U-Net structure; The VoxelMorph model based on the lightweight attention module CBAM includes CBAM-U-Net, spatial transformer network and local variance weighted loss function.

5. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 4, characterized in that: The working process of the lightweight attention module CBAM-U-Net is as follows: based on the preprocessed ultrafast ultrasound reference image F and the motion image M, the CBAM-U-Net convolutional neural network g is used to θ The encoder path of (F,M)=u is downsampled, where θ can be learned and optimized as a parameter. During the downsampling process, the number of channels increases and the feature map shrinks. A lightweight attention module CBAM is added at the end of the encoder. After channel attention, maximum pooling and average pooling are used, and the feature map is input into a multi-layer perceptron with shared weights. The channel information is fused through 1×1 convolution, and then the above two poolings are performed through spatial attention. The obtained feature maps are spliced, and finally a standard convolution is performed to output a weighted feature map to extract the target. The addition of the lightweight attention module CBAM can improve the feature extraction accuracy to achieve multi-scale feature extraction. The decoder path includes upsampling the feature image, concatenating the feature map in the upsampling process with the feature map processed by the lightweight attention mechanism from the encoder path in the channel dimension, halving the image channels in the process of restoring the image resolution until the output image resolution is consistent with the input image resolution, and finally outputting the predicted deformation field φ=u+id, where φ is the same size as F and M, and id is a constant representing the part without displacement or deformation.

6. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 4, characterized in that: The working process of the VoxelMorph model based on the lightweight attention module CBAM is to obtain deformation field parameters by passing the moving image and the reference image through the CBAM-U-Net convolutional neural network, and perform linear interpolation conversion on the moving image based on the deformation field parameters using the spatial transformation network. The conversion method is as follows: p'=p+u(p); Where p represents a pixel, q represents the adjacent pixel of p, p' represents the weight of pixel p by the adjacent pixel q in eight directions, z(p') represents the transformation parameter of the adjacent position of p, Spatial transformation to obtain the deformed image The local weighted variance loss and deformation field smoothness loss are used as loss functions to measure the similarity between the deformed image and the reference image, and the training is completed when the evaluation indicators are met to obtain the final calibrated image.

7. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 6, characterized in that: The loss function includes: deformation field smoothing loss and local weighted variance loss. The deformation field smoothing loss calculation method is as follows: In the formula, φ is used as the deformation field parameter. This loss can suppress the local sudden changes of the vector by setting the gradient size of u as the loss. The local weighted variance loss is composed of a local variance function and a weight function. The calculation method of the local variance is as follows: in and Represents n centered on p 3 The average gray value of the reference image of the neighborhood block and the average gray value of the motion image are L LV After normalization, we get The σ parameter is introduced to adjust the weight factor coefficient and construct the weight function: Then the local weighted variance loss can be expressed as:

8. The method for ultrafast ultrasound motion artifact correction based on VoxelMorph model according to claim 1, characterized in that: The training process in step S4 is to build a training model using the Pytorch deep learning framework, use the local weighted variance and deformation field smoothing loss function and the Adam optimizer for model training, and after multiple iterations, adjust the model parameters to obtain a motion calibration network model.

9. The method for ultrafast ultrasound motion artifact correction based on the VoxelMorph model according to claim 1, characterized in that: In the step S5, the test data set is selected to test the motion calibration network model, including based on the mean square error MSE indicator, the peak signal-to-noise ratio PSNR indicator and the structural similarity SSIM indicator, and the performance evaluation of the model is completed by inputting the test set into the motion calibration network model.

10. The method for ultrafast ultrasound motion artifact correction based on VoxelMorph model according to claim 1, characterized in that: In step S7, Doppler blood flow imaging is performed on the data after clutter filtering, and the blood flow signal is expressed in the form of intensity value. The Doppler average signal intensity is expressed as: Where x represents the direction parallel to the probe array, z represents the imaging depth, and S f (x,z,t i ) represents the signal after noise filtering, and N represents the number of sampling frames.

Images