A method and system for ultrasonic blood flow velocity measurement based on pulsed wave Doppler

Abstract

The application relates to the technical field of ultrasonic blood flow measurement, and discloses an ultrasonic blood flow velocity measurement method and system based on pulse wave Doppler. The method comprises the following steps: transmitting a pulse wave Doppler ultrasonic sequence and receiving a return signal, constructing a slow-time echo matrix, and quadrature demodulating to obtain a same-phase quadrature signal tensor; inputting the same-phase quadrature signal tensor into a one-dimensional convolutional neural network to extract a blood flow component and obtain a blood flow feature map; adopting a spatial attention mechanism to calculate the correlation degree of each spatial position and a blood vessel wall, generating an attention mask and attenuating a feature channel of a neighboring blood vessel wall position to obtain a pure blood flow feature; and finally inputting the pure blood flow feature into a velocity mapping network to regress to obtain the blood flow velocity in a sampling gate. The application can adaptively suppress the interference of the blood vessel wall and improve the blood flow velocity measurement precision.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic blood flow measurement technology, specifically to a method and system for measuring ultrasonic blood flow velocity based on pulse wave Doppler. Background Technology

[0002] Pulse-wave Doppler ultrasound technology is widely used in the non-invasive measurement of blood flow velocity in clinical practice. Traditional pulse-wave Doppler blood flow velocity measurement methods usually rely on Fast Fourier Transform to perform spectral analysis on the echo signal, and then extract the Doppler frequency shift and convert it into blood flow velocity.

[0003] However, in actual measurement, stationary or slowly moving tissues such as blood vessel walls produce reflected echoes with an intensity much higher than the scattering signal of red blood cells. These strong signals appear as low-frequency, high-amplitude components in the spectrum, which seriously interfere with the extraction of blood flow signals.

[0004] To suppress vessel wall signals, existing technologies generally employ high-pass filters or wall filters to preprocess echo signals. However, high-pass filters with fixed cutoff frequencies are difficult to adapt to the differences in vessel wall motion at different depths. In low-velocity blood flow measurement scenarios, they can easily filter out useful low-velocity blood flow signals, leading to distorted measurement results or failure to detect them.

[0005] Furthermore, traditional methods rely on manual setting of sampling gate positions and filtering parameters, which is complex, highly dependent on operator experience, and has poor repeatability. With the development of deep learning technology, some studies have attempted to extract blood flow information directly from echo signals using neural networks. However, existing schemes typically use the entire echo signal as input, lacking targeted suppression of interference areas in the vessel wall, resulting in limited measurement accuracy in areas adjacent to the vessel wall.

[0006] Therefore, there is an urgent need for a technical solution that can adaptively suppress interference from the blood vessel wall and improve the accuracy of blood flow velocity measurement. Summary of the Invention

[0007] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a method and system for measuring ultrasonic blood flow velocity based on pulse wave Doppler, thereby solving the problem of low measurement accuracy caused by signal interference from the blood vessel wall in the prior art.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for measuring ultrasound blood flow velocity based on pulse wave Doppler, the method comprising: Step S1: Transmit a pulse wave Doppler ultrasound sequence and receive echo signals. Arrange the echo signals along the direction of the pulse transmission sequence to construct a slow-time echo matrix. Perform orthogonal demodulation on the slow-time echo matrix to obtain an in-phase orthogonal signal tensor. Step S2: Input the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extract the blood flow component along the pulse emission sequence direction through the one-dimensional convolution kernel, and output a blood flow feature map with spatial location annotations. Step S3: Locate the region of the sampling gate adjacent to the vessel wall based on the blood flow feature map, calculate the correlation between each spatial location and the vessel wall using the spatial attention mechanism and generate an attention mask, attenuate the feature channels corresponding to the locations adjacent to the vessel wall to obtain pure blood flow features; Step S4: Input the pure blood flow characteristics into the velocity mapping network, and obtain the blood flow velocity within the pulse wave Doppler sampling gate by regression of the velocity mapping network.

[0009] Preferably, in one possible implementation of the first aspect, the construction process of the slow-time echo matrix in step S1 includes: Along the temporal dimension of the pulse emission sequence, the echo signals corresponding to each pulse emission are arranged according to the pulse emission order, forming a two-dimensional matrix structure with the pulse emission sequence number as the row index and the echo sampling point as the column index. In a two-dimensional matrix, the row direction represents the slow time dimension, and the column direction represents the fast time dimension.

[0010] Preferably, in one possible implementation of the first aspect, the process of acquiring the in-phase orthogonal signal tensor includes: Perform quadrature demodulation based on the transmission center frequency on each column of the slow-time echo matrix to generate in-phase component matrix and quadrature component matrix; By stacking the in-phase component matrix and the orthogonal component matrix along the channel dimension, a three-dimensional in-phase orthogonal signal tensor with channel dimension, slow time dimension and fast time dimension is constructed.

[0011] Preferably, in one possible implementation of the first aspect, the output process of the blood flow feature map in step S2 includes: One-dimensional convolutional neural networks take in-phase orthogonal signal tensors as input, use one-dimensional convolution kernels to slide in the slow time dimension, perform local convolution operations on the echo signals at each slow time position, extract blood flow component features in the slow time dimension, and output blood flow feature maps with spatial location labels through the channel dimension. The channel dimension of the blood flow feature map corresponds to the blood flow feature expression at different levels of abstraction.

[0012] Preferably, in one possible implementation of the first aspect, the calculation process of the correlation between each spatial location and the blood vessel wall in step S3 includes: Based on the spatial location markings carried by the blood flow feature map, the sampling gate is located in the region adjacent to the vessel wall in both the slow and fast time dimensions; The feature vectors of each spatial location in the adjacent blood vessel wall region are extracted from the blood flow feature map and the mean value is calculated to obtain the feature vector of the blood vessel wall region. The feature vector of the blood vessel wall region is used as the query vector, and the feature vectors of each spatial location of the blood flow feature map are used as the key vectors. By calculating the cosine similarity between the query vector and the key vector, the correlation between each spatial location and the blood vessel wall is obtained.

[0013] Preferably, in one possible implementation of the first aspect, the attention mask generation process includes: After negative values ​​are taken for the correlation between each spatial location and the blood vessel wall, exponential normalization is performed to generate an inhibition weight map with values ​​between 0 and 1. In the inhibition weight map, spatial locations with values ​​close to 0 correspond to locations near the vessel wall, while spatial locations with values ​​close to 1 correspond to locations far from the vessel wall. This inhibition weight map is used as an attention mask.

[0014] Preferably, in one possible implementation of the first aspect, the process of obtaining the pure blood flow characteristics includes: The attention mask is multiplied element-wise with the blood flow feature map along the channel dimension. The feature channel response corresponding to the location near the blood vessel wall is attenuated, while the feature channel response corresponding to the location far from the blood vessel wall is retained. The result of the multiplication is the pure blood flow feature.

[0015] Preferably, in one possible implementation of the first aspect, the specific structure of the velocity mapping network in step S4 includes: The velocity mapping network consists of a sequentially connected one-dimensional convolutional layer, a feature flattening layer, a fully connected hidden layer, and a velocity regression output layer. One-dimensional convolutional layers receive pure blood flow features and perform convolution operations; The feature flattening layer reshapes the output of the convolution operation into a one-dimensional feature vector; Fully connected hidden layers perform nonlinear transformations on one-dimensional feature vectors; The velocity regression output layer outputs the scalar value of blood flow velocity within the sampling gate.

[0016] Preferably, in one possible implementation of the first aspect, the process of obtaining the blood flow velocity includes: Pure blood flow features are input into a velocity mapping network, which then sequentially undergoes convolution operations in a one-dimensional convolutional layer, vector reshaping in a feature flattening layer, nonlinear transformation in a fully connected hidden layer, and mapping in a velocity regression output layer. The velocity regression output layer directly regresses and outputs the blood flow velocity values ​​within the pulse wave Doppler sampling gate.

[0017] Secondly, the present invention provides an ultrasonic blood flow velocity measurement system based on pulse-wave Doppler, the system being used to implement the ultrasonic blood flow velocity measurement method based on pulse-wave Doppler as described in the first aspect, comprising: The echo acquisition and demodulation module transmits pulse wave Doppler ultrasound sequences and receives echo signals. It arranges the echo signals along the direction of the pulse transmission sequence to construct a slow-time echo matrix and performs orthogonal demodulation on the slow-time echo matrix to obtain an in-phase orthogonal signal tensor. The blood flow feature extraction module inputs the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extracts the blood flow components along the pulse emission sequence direction through the one-dimensional convolution kernel, and outputs a blood flow feature map with spatial location annotations. The vessel wall attenuation module locates the region of the sampling gate adjacent to the vessel wall based on the blood flow feature map, uses a spatial attention mechanism to calculate the correlation between each spatial location and the vessel wall and generates an attention mask, attenuates the feature channels corresponding to the locations adjacent to the vessel wall, and obtains pure blood flow features. The velocity regression module inputs the pure blood flow characteristics into the velocity mapping network, and the velocity mapping network regresses the blood flow velocity within the pulse wave Doppler sampling gate.

[0018] The beneficial effects of this invention are as follows: This invention extracts blood flow components along the pulse emission sequence direction using a one-dimensional convolutional neural network, making full use of the differences in the temporal variation patterns of blood flow signals and blood vessel wall signals in the slow time dimension, thereby achieving adaptive learning and extraction of blood flow features and avoiding the signal distortion problem caused by the fixed cutoff frequency of traditional high-pass filters.

[0019] A spatial attention mechanism is introduced to locate the region of the sampling gate near the blood vessel wall based on the blood flow feature map. By calculating the correlation between each spatial location and the blood vessel wall and generating an attention mask, the feature channels corresponding to the locations near the blood vessel wall are attenuated, thereby accurately suppressing the interference of blood vessel wall signals at the feature level and preserving pure blood flow features.

[0020] By directly regressing blood flow velocity from pure blood flow features through a velocity mapping network, an end-to-end mapping from echo signals to velocity values ​​is constructed. This simplifies multiple intermediate steps such as spectrum analysis and peak detection in traditional methods, reduces the complexity of manual parameter setting, and improves the automation and repeatability of measurements.

[0021] The system adopts a deep learning architecture and uses labeled data for supervised training, enabling the network to learn the optimal feature extraction and velocity mapping strategies. Compared with traditional methods, it has significant improvements in measurement accuracy, anti-interference ability, and ease of operation, providing a more reliable technical means for clinical ultrasound hemodynamic assessment. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This application provides a flowchart of an ultrasonic blood flow velocity measurement method based on pulse wave Doppler. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: As Figure 1 As shown, this invention provides a method for measuring ultrasonic blood flow velocity based on pulse wave Doppler, comprising: Step S1: Transmit a pulse wave Doppler ultrasound sequence and receive the echo signal. Arrange the echo signal along the direction of the pulse transmission sequence to construct a slow-time echo matrix. Perform orthogonal demodulation on the slow-time echo matrix to obtain the in-phase orthogonal signal tensor.

[0026] In this embodiment, the ultrasound probe first emits a pulsed-wave Doppler ultrasound sequence toward the target blood vessel region. This sequence consists of multiple pulse repetition cycles, with each cycle emitting a short-duration, high-frequency ultrasound pulse. The center frequency of the pulse is set to [value missing]. The pulse repetition frequency is set to After each pulse transmission, the ultrasound probe immediately enters receiving mode, acquiring echo signals scattered by red blood cells within the blood vessel and reflected from the vessel wall. These echo signals are then pre-amplified, bandpass filtered, and converted from analog to digital to form a digitized echo signal sequence.

[0027] Along the direction of the pulse transmission sequence, the echo signals corresponding to all pulse transmissions are arranged in the order of pulse transmission. Specifically, the echo signal corresponding to the first pulse transmission is placed in the first row, the echo signal corresponding to the second pulse transmission is placed in the second row, and so on, until the [number missing]th pulse transmission. The echo signal corresponding to the next pulse transmission is used as the first Rows. The number of elements in each row equals the number of sampling points for a single echo signal, i.e., the number of sampling points along the depth direction. This constructs a two-dimensional matrix where the rows represent the slow-time dimension, i.e., the temporal variation of the pulse emission; and the columns represent the fast-time dimension, i.e., the depth sampling points along the ultrasound propagation path. This two-dimensional matrix is ​​the slow-time echo matrix, denoted as [matrix name missing]. ,in For slow time index (value range 1 to...) , (Total number of pulses emitted). Fast time index (value range 1 to...) , (Number of sampling points per echo).

[0028] Next, orthogonal demodulation is performed on the slow-time echo matrix. Orthogonal demodulation is based on the center frequency of the transmitted pulse. For each column of the slow-time echo matrix... The slow time series corresponding to this column It is considered as a time-varying signal, and its sampling interval is equal to the pulse repetition period. Multiplying the sequence by both a cosine and a sine reference signal, the expression for the cosine reference signal is: The expression for the sinusoidal reference signal is: ,in The multiplication operation yields two mixed signals:

[0029]

[0030] Subsequently, a low-pass filter is applied to each of the two mixed signals in each column to filter out high-frequency components (i.e., sum-frequency components near the center frequency) and retain the baseband signal. The cutoff frequency of the low-pass filter is set to... The filter is of order 32 and uses a finite impulse response (FIR) structure. The filtered output consists of in-phase and quadrature components:

[0031]

[0032] For all columns After performing the above operations, two two-dimensional matrices are obtained: the in-phase component matrix and the matrix of in-phase components. and orthogonal component matrix Both are the same size .

[0033] Finally, the in-phase component matrix and the orthogonal component matrix are stacked along a new dimension. This new dimension is defined as the channel dimension, where the first channel stores the data of the in-phase component matrix and the second channel stores the data of the orthogonal component matrix. After stacking, a three-dimensional tensor is formed, denoted as . ,in For channel index ( Corresponding in-phase components, (corresponding orthogonal components) For the slow time dimension, It is a fast time dimension. This three-dimensional tensor is the in-phase orthogonal signal tensor, which fully preserves the amplitude and phase information of the original echo signal.

[0034] Step S2: Input the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extract the blood flow component along the pulse emission sequence direction through the one-dimensional convolution kernel, and output a blood flow feature map with spatial location annotation.

[0035] In this embodiment, the one-dimensional convolutional neural network is designed to extract features along the slow time dimension. The network consists of five sequentially connected one-dimensional convolutional blocks, each containing a one-dimensional convolutional layer, a batch normalization layer, and a... Activation layers. The first convolutional block has 2 input channels, corresponding to the two channels of the in-phase orthogonal signal tensor. The kernel size is set to 7, the stride to 1, and the padding to 3 to ensure the slow time dimension remains unchanged. It has 16 output channels. The second convolutional block has a kernel size of 5, a stride of 1, and padding to 2, resulting in 32 output channels. The third convolutional block has a kernel size of 5, a stride of 1, and padding to 2, resulting in 64 output channels. The fourth convolutional block has a kernel size of 3, a stride of 1, and padding to 1, resulting in 128 output channels. The fifth convolutional block has a kernel size of 3, a stride of 1, and padding to 1, resulting in 256 output channels.

[0036] All convolutional operations slide along the slow-time dimension, i.e., for each fixed fast-time index. One-dimensional convolution kernel along slow-time index The system performs local convolution operations in the direction of the input to extract blood flow signal features that change over time at that depth location. A batch normalization layer normalizes the feature map after each convolutional layer, accelerating training convergence and stabilizing the gradient. The activation layer introduces a non-linear transformation to enhance the network's expressive power. After processing through five convolutional blocks, a three-dimensional feature map is output, with a size of [size missing]. ,in The length of the slow time dimension is the total number of pulses emitted. The fast time dimension is the length of the sampling point per echo, and 256 represents the channel dimension, with each channel corresponding to a different level of blood flow feature representation. Each spatial location in this feature map corresponds to a slow time index. and fast time index Therefore, it naturally carries spatial location annotations. The network training employs supervised learning. The training dataset contains a large number of pulse-wave Doppler ultrasound echo signals. The in-phase orthogonal signal tensor of each sample serves as input, and the corresponding ideal blood flow component feature map serves as the label. The ideal blood flow component feature map is extracted from the echo signals using a traditional high-pass filtering method and verified by experts. The training loss function uses mean squared error, the optimizer is Adam, the initial learning rate is set to 0.001, the batch size is 32, and the training epochs are 100. During training, the network automatically learns to suppress low-frequency strong signals generated by stationary tissues such as the blood vessel wall and highlights high-frequency weak signals generated by red blood cell movement, achieving effective extraction of blood flow components.

[0037] Step S3: Based on the blood flow feature map, locate the region of the sampling gate adjacent to the blood vessel wall, use the spatial attention mechanism to calculate the correlation between each spatial location and the blood vessel wall and generate an attention mask, attenuate the feature channels corresponding to the locations adjacent to the blood vessel wall to obtain pure blood flow features.

[0038] In this embodiment, the depth range of the sampling gate in the fast time dimension is first determined based on the spatial location markers carried by the blood flow feature map. The fast time index interval of the sampling gate is given by preset parameters of the ultrasound system, with a lower limit of [missing information]. The upper limit is The coordinates of the blood vessel wall in the fast time dimension. The sampling gate is obtained by averaging the blood flow feature map along the slow time dimension and then detecting abrupt amplitude changes. The region adjacent to the vessel wall within the sampling gate is defined as all areas inside the sampling gate that meet the conditions. The fast time position, where 3 is the preset number of extended sampling points, and all indices in the slow time dimension. All of these belong to this region. The collection of these spatial locations constitutes the region adjacent to the blood vessel wall.

[0039] Feature vectors for all spatial locations within the blood flow feature map are extracted. Each feature vector is a 256-dimensional vector composed of the values ​​across all 256 channels at that location. The mean of all extracted feature vectors at their spatial locations is then calculated to obtain a 256-dimensional feature vector representing the blood vessel wall region. .

[0040] Feature vector of blood vessel wall region As a query vector, each spatial location in the blood flow feature map eigenvectors As key-value vectors, calculate the cosine similarity between the two:

[0041] in Represents the vector dot product. This represents the L2 norm of a vector. The value ranges from -1 to 1, with a larger value indicating a greater similarity between the spatial location's features and the vessel wall's features. After calculating for all spatial locations, a correlation map with the same spatial dimensions as the blood flow feature map is obtained, where each element corresponds to the correlation degree between a spatial location and the vessel wall.

[0042] By taking the negative value for each correlation degree in the correlation graph, a negative correlation graph is obtained, whose element values ​​are... The negative correlation graph is subjected to exponential normalization, specifically by calculating the following for each spatial location:

[0043] Then find all spatial locations The maximum value is denoted as Divide the exponent value at each spatial location by The normalized suppression weight values ​​are obtained as follows:

[0044] The inhibition weight values ​​approach 0 near the vessel wall and approach 1 away from the vessel wall. Arranging all inhibition weight values ​​spatially creates an inhibition weight map with values ​​between 0 and 1; this inhibition weight map is the attention mask. .

[0045] Attention mask The blood flow feature map is multiplied element-wise along the channel dimension. Specifically, for each spatial location... and each channel The blood flow feature map values ​​at this location and channel Multiply by the value of the attention mask at that position. The updated feature values ​​are obtained as follows:

[0046] This process significantly attenuates the responses of all feature channels located near the vessel wall, while preserving the responses of feature channels located far from the vessel wall. The resulting three-dimensional feature tensor after multiplication is the pure blood flow feature, with the same size as the input blood flow feature map. However, its characteristic response has eliminated the interference of blood vessel wall signals, retaining only the motion signal components from blood flow.

[0047] Step S4: Input the pure blood flow characteristics into the velocity mapping network, and obtain the blood flow velocity within the pulse wave Doppler sampling gate by regression of the velocity mapping network.

[0048] In this embodiment, the velocity mapping network consists of a one-dimensional convolutional layer, a feature flattening layer, a fully connected hidden layer, and a velocity regression output layer connected in sequence. The one-dimensional convolutional layer receives pure blood flow features and performs convolution operations. This one-dimensional convolutional layer has 128 kernels, a kernel size of 3, a stride of 1, padding of 1, 256 input channels, and 128 output channels. The one-dimensional convolutional layer performs sliding convolution on the pure blood flow features in the slow-time dimension, for each fixed fast-time index... One-dimensional convolution kernel along slow-time index The direction performs local convolution operations to further extract high-order motion features of pure blood flow in the slow time dimension. ,in The length of the slow time dimension is the total number of pulses emitted. The length of the fast time dimension is the number of sampling points per echo.

[0049] The feature flattening layer reshapes the 3D feature map output by the 1D convolutional layer into a 1D feature vector. Specifically, it arranges the values ​​of all dimensions sequentially to form a vector of length [length missing]. The one-dimensional feature vector is used. A fully connected hidden layer receives this one-dimensional feature vector and performs a non-linear transformation. The fully connected hidden layer contains 512 neurons and employs... The activation function and the weight matrix have the following dimensions: The bias vector has a dimension of 512.

[0050] The fully connected hidden layer performs a linear weighted summation on the one-dimensional feature vector and adds a bias, then passes through... The activation function outputs a 512-dimensional hidden feature vector. The velocity regression output layer receives this hidden feature vector and outputs a scalar value of the blood flow velocity within the sampling gate. The velocity regression output layer contains one neuron, does not use an activation function, and the weight matrix has a dimension of 1. The bias vector has a dimension of 1. The velocity regression output layer performs a linear weighted summation on the hidden feature vectors and adds the bias, directly outputting a floating-point number, which is the blood flow velocity value within the pulse wave Doppler sampling gate.

[0051] The velocity mapping network is trained using supervised learning. The training dataset contains a large number of pulse-wave Doppler ultrasound echo signals. The pure blood flow features of each sample are used as input, and the corresponding real blood flow velocity is used as the label. The real blood flow velocity is extracted from the raw echo signals using high-precision Doppler spectrum analysis and verified by experts. The training loss function uses mean squared error, and the optimizer is selected from... The initial learning rate was set to 0.0001, the batch size to 64, and the training epochs to 200. During training, the network automatically learned the nonlinear mapping relationship from pure blood flow features to blood flow velocity, achieving accurate regression of blood flow velocity within the sampling gate.

[0052] Example 2: This invention provides an ultrasonic blood flow velocity measurement system based on pulse wave Doppler, comprising: The echo acquisition and demodulation module transmits pulse wave Doppler ultrasound sequences and receives echo signals. It arranges the echo signals along the direction of the pulse transmission sequence to construct a slow-time echo matrix and performs orthogonal demodulation on the slow-time echo matrix to obtain an in-phase orthogonal signal tensor. The blood flow feature extraction module inputs the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extracts the blood flow components along the pulse emission sequence direction through the one-dimensional convolution kernel, and outputs a blood flow feature map with spatial location annotations. The vessel wall attenuation module locates the region of the sampling gate adjacent to the vessel wall based on the blood flow feature map, uses a spatial attention mechanism to calculate the correlation between each spatial location and the vessel wall and generates an attention mask, attenuates the feature channels corresponding to the locations adjacent to the vessel wall, and obtains pure blood flow features. The velocity regression module inputs the pure blood flow characteristics into the velocity mapping network, and the velocity mapping network regresses the blood flow velocity within the pulse wave Doppler sampling gate.

[0053] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for measuring ultrasonic blood flow velocity based on pulse wave Doppler, characterized in that, The method includes: Step S1: Transmit a pulse wave Doppler ultrasound sequence and receive echo signals. Arrange the echo signals along the direction of the pulse transmission sequence to construct a slow-time echo matrix. Perform orthogonal demodulation on the slow-time echo matrix to obtain an in-phase orthogonal signal tensor. Step S2: Input the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extract the blood flow component along the pulse emission sequence direction through the one-dimensional convolution kernel, and output a blood flow feature map with spatial location annotations. Step S3: Locate the region of the sampling gate adjacent to the vessel wall based on the blood flow feature map, calculate the correlation between each spatial location and the vessel wall using the spatial attention mechanism and generate an attention mask, attenuate the feature channels corresponding to the locations adjacent to the vessel wall to obtain pure blood flow features; Step S4: Input the pure blood flow characteristics into the velocity mapping network, and obtain the blood flow velocity within the pulse wave Doppler sampling gate by regression of the velocity mapping network.

2. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 1, characterized in that, The construction process of the slow-time echo matrix in step S1 includes: Along the temporal dimension of the pulse emission sequence, the echo signals corresponding to each pulse emission are arranged according to the pulse emission order, forming a two-dimensional matrix structure with the pulse emission sequence number as the row index and the echo sampling point as the column index. In a two-dimensional matrix, the row direction represents the slow time dimension, and the column direction represents the fast time dimension.

3. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 2, characterized in that, The process of obtaining the in-phase orthogonal signal tensor includes: Perform quadrature demodulation based on the transmission center frequency on each column of the slow-time echo matrix to generate in-phase component matrix and quadrature component matrix; By stacking the in-phase component matrix and the orthogonal component matrix along the channel dimension, a three-dimensional in-phase orthogonal signal tensor with channel dimension, slow time dimension and fast time dimension is constructed.

4. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 1, characterized in that, The output process of the blood flow feature map in step S2 includes: One-dimensional convolutional neural networks take in-phase orthogonal signal tensors as input, use one-dimensional convolution kernels to slide in the slow time dimension, perform local convolution operations on the echo signals at each slow time position, extract blood flow component features in the slow time dimension, and output blood flow feature maps with spatial location labels through the channel dimension. The channel dimension of the blood flow feature map corresponds to the blood flow feature expression at different levels of abstraction.

5. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 1, characterized in that, The calculation process for the correlation between each spatial location and the blood vessel wall in step S3 includes: Based on the spatial location markings carried by the blood flow feature map, the sampling gate is located in the region adjacent to the vessel wall in both the slow and fast time dimensions; The feature vectors of each spatial location in the adjacent blood vessel wall region are extracted from the blood flow feature map and the mean value is calculated to obtain the feature vector of the blood vessel wall region. The feature vector of the blood vessel wall region is used as the query vector, and the feature vectors of each spatial location of the blood flow feature map are used as the key vectors. By calculating the cosine similarity between the query vector and the key vector, the correlation between each spatial location and the blood vessel wall is obtained.

6. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 5, characterized in that, The attention mask generation process includes: After negative values ​​are taken for the correlation between each spatial location and the blood vessel wall, exponential normalization is performed to generate an inhibition weight map with values ​​between 0 and 1. In the inhibition weight map, spatial locations with values ​​close to 0 correspond to locations near the vessel wall, while spatial locations with values ​​close to 1 correspond to locations far from the vessel wall. This inhibition weight map is used as an attention mask.

7. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 6, characterized in that, The process of obtaining the pure blood flow characteristics includes: The attention mask is multiplied element-wise with the blood flow feature map along the channel dimension. The feature channel response corresponding to the location near the blood vessel wall is attenuated, while the feature channel response corresponding to the location far from the blood vessel wall is retained. The result of the multiplication is the pure blood flow feature.

8. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 1, characterized in that, The specific structure of the velocity mapping network in step S4 includes: The velocity mapping network consists of a sequentially connected one-dimensional convolutional layer, a feature flattening layer, a fully connected hidden layer, and a velocity regression output layer. One-dimensional convolutional layers receive pure blood flow features and perform convolution operations; The feature flattening layer reshapes the output of the convolution operation into a one-dimensional feature vector; Fully connected hidden layers perform nonlinear transformations on one-dimensional feature vectors; The velocity regression output layer outputs the scalar value of blood flow velocity within the sampling gate.

9. The method for measuring ultrasonic blood flow velocity based on pulse wave Doppler as described in claim 8, characterized in that, The process of obtaining the blood flow velocity includes: Pure blood flow features are input into a velocity mapping network, which then sequentially undergoes convolution operations in a one-dimensional convolutional layer, vector reshaping in a feature flattening layer, nonlinear transformation in a fully connected hidden layer, and mapping in a velocity regression output layer. The velocity regression output layer directly regresses and outputs the blood flow velocity values ​​within the pulse wave Doppler sampling gate.

10. A pulse-wave Doppler-based ultrasonic blood flow velocity measurement system, characterized in that, The system is used to implement a pulse-wave Doppler-based ultrasonic blood flow velocity measurement method as described in any one of claims 1 to 9, comprising: The echo acquisition and demodulation module transmits pulse wave Doppler ultrasound sequences and receives echo signals. It arranges the echo signals along the direction of the pulse transmission sequence to construct a slow-time echo matrix and performs orthogonal demodulation on the slow-time echo matrix to obtain an in-phase orthogonal signal tensor. The blood flow feature extraction module inputs the in-phase orthogonal signal tensor into a one-dimensional convolutional neural network, extracts the blood flow components along the pulse emission sequence direction through the one-dimensional convolution kernel, and outputs a blood flow feature map with spatial location annotations. The vessel wall attenuation module locates the region of the sampling gate adjacent to the vessel wall based on the blood flow feature map, uses a spatial attention mechanism to calculate the correlation between each spatial location and the vessel wall and generates an attention mask, attenuates the feature channels corresponding to the locations adjacent to the vessel wall, and obtains pure blood flow features. The velocity regression module inputs the pure blood flow characteristics into the velocity mapping network, and the velocity mapping network regresses the blood flow velocity within the pulse wave Doppler sampling gate.

Images