A method for improving the time resolution of the doppler spectrum of an ultrasound system

Abstract

The present application relates to a kind of methods for improving the time resolution of Doppler spectrum of ultrasonic system, comprising the following steps: step 1: first, the complex Doppler IQ data of size M×N is obtained;Step 2: input Doppler time resolution enhancement network, and the time resolution of Doppler IQ data is enhanced;Step 3: wall filtering, for filtering out the noise or the clutter of uninterested in data;Step 4: time domain windowing, for inhibiting the sidelobe in spectrum, improves the contrast of spectrum display;Step 5: tail zero padding, data is padded into the integer power of 2, to facilitate subsequent FFT calculation, also play the role of improving spectral resolution;Step 6: using FFT, the data after zero padding is carried out fast Fourier transform, and Doppler spectrum is obtained and is displayed.The present application improves the time resolution of Doppler spectrum imaging under the premise of maintaining the frequency resolution of Doppler spectrum.

Description

Technical Field

[0001] This invention relates to the field of ultrasound, and more particularly to a method for improving the time resolution of the Doppler spectrum in an ultrasound system. Background Technology

[0002] Doppler spectral imaging is an important application mode in conventional ultrasound imaging, widely used in imaging various organs of the human body and small animals. Because Doppler spectral imaging is often performed simultaneously with B-mode imaging (two-synchronous mode) or B+C mode imaging (three-synchronous mode), the scanning time of the Doppler spectrum needs to be shared with the scanning time of B-mode or C-mode. Generally, time-division multiplexing is used for scanning, which poses a significant technical challenge to acquiring high spatiotemporal resolution Doppler spectral images.

[0003] Existing technologies generally use the traditional Welch algorithm to calculate the Doppler spectrum, and obtain Doppler IQ data after orthogonal demodulation of the ultrasonic echo signal (see reference). Figure 1 The Doppler spectrum is obtained by performing an FFT (Fast Fourier Transform). The Doppler IQ data is organized into a Doppler scanning window and sampled in two directions. Sampling is performed in Fast Time according to the sampling frequency of the analog-to-digital converter (ADC) used in the system, sampling M points to form a sampling line. Then, sampling is performed in Slow Time at intervals of PRT. Generally, the sampling frequency in Slow Time, PRF = 1 / PRT, is much lower than the sampling frequency in Fast Time. N lines are obtained in Slow Time, with M points on each line, resulting in a total of M*N data points, stored as an M x N matrix. Each data point is a complex number containing both real and imaginary components. Data point (m, n) represents the m-th data point on the n-th line, where 1 ≤ m ≤ M, 1 ≤ n ≤ N. The Welch algorithm takes one line of data (N data points) each time, performs an FFT, repeats this process M times, and finally averages the resulting M FFTs to output the final Doppler spectrum. To obtain good FFT spectral calculation results, a relatively large N is generally required. However, when N is large, the time required for a single Doppler scan window (N*PRT) is also large, thus reducing the temporal resolution of the Doppler spectrum. This necessitates the use of a longer Doppler scan window, resulting in poor temporal resolution. Conversely, using a shorter Doppler scan window leads to spectral broadening, resulting in poor frequency resolution. Especially when the PRF is low, current techniques for imaging low-velocity blood flow or low-velocity tissue Doppler can miss useful information, potentially leading to misdiagnosis. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a method for improving the temporal resolution of the Doppler spectrum in an ultrasound system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for improving the temporal resolution of the Doppler spectrum in an ultrasound system includes the following steps:

[0007] Step 1: First, acquire complex Doppler IQ data of size M×N;

[0008] Step 2: Input the Doppler time resolution enhancement network to perform time resolution enhancement on the Doppler IQ data. The steps are as follows:

[0009] Step 21: First, split and rearrange the real and imaginary parts of the input Doppler IQ data to obtain a tensor of size [M, N, 2].

[0010] Where M represents the rows of the tensor, N represents the columns of the tensor, and 2 represents the number of channels of the tensor (the size of tensors in the following text is arranged in this order);

[0011] Step 22: Then, the tensor obtained in Step 21 is input into the Doppler encoder network for encoding to obtain an encoded output tensor, which is used to better represent the various potential features in the input data;

[0012] Step 23: Finally, the output tensor encoded in Step 22 is input into the Doppler decoder network for decoding. Based on various latent features in the input data, the time-resolution enhanced Doppler IQ data is obtained, which is a tensor of size [M, KN, 2]. By increasing the number of lines from N to KN, the Doppler time-resolution enhancement network increases the length of the Doppler scanning window, thereby improving the temporal resolution of the spectrum while preserving the frequency resolution after subsequent calculations.

[0013] Where K is a positive integer greater than 1, and the appropriate value can be selected according to different applications;

[0014] Step 3: Wall filtering, used to filter out noise or unwanted clutter in the data;

[0015] Step 4: Time-domain windowing is used to suppress sidelobes in the spectrum and improve the contrast of the spectrum display;

[0016] Step 5: Pad the data with zeros to make it a power of 2, which facilitates the subsequent FFT calculation and also improves the spectral resolution.

[0017] Step 6: FFT, perform Fast Fourier Transform on the zero-padding data, and display the Doppler spectrum.

[0018] Preferably, in the method for improving the temporal resolution of the Doppler spectrum of an ultrasound system, the wall filtering in step 3 is either low-pass filtering (for tissue imaging) or high-pass filtering (for blood flow imaging).

[0019] Preferably, in the method for improving the time resolution of the Doppler spectrum of an ultrasound system, the time-domain windowing in step 4 includes Hamming window, Hanning window, rectangular window, triangular window, Gaussian window, Blackman window, Kaiser window, Bessel window, and Chebyshev window.

[0020] Preferably, in the method for improving the Doppler spectral temporal resolution of an ultrasound system, the specific steps of the Doppler encoder network in step 22 are as follows:

[0021] First, the input Doppler IQ data is processed through a channel upscaling network to increase the number of channels from 2 to 2K, where K is a positive integer greater than 1. The appropriate value can be selected according to different applications to obtain an upscaling feature map of size [M, N, 2K].

[0022] Then, after processing by C downsampled 1D CNNs, where 1D means one-dimensional and CNN is a convolutional neural network, the final result is of size [M, N / 2]. C 2K×2 C The downsampled feature map of ] is used as the output tensor of the encoded tensor, where C is a positive integer greater than or equal to 1, and the appropriate value can be selected according to different applications and computing resources.

[0023] Preferably, in the method for improving the temporal resolution of the Doppler spectrum of an ultrasound system, the channel upscaling network consists of a CNN, which is composed of a 3*3 convolutional layer with a stride of 1. Therefore, the size of rows M and columns N is not changed, but the number of channels is increased from 2 to 2K through 2*2K different convolution coefficients.

[0024] Preferably, in the method for improving the temporal resolution of the Doppler spectrum of an ultrasound system, the downsampling 1DCNN is implemented by a 1*3 1D convolutional layer followed by a linear rectified (ReLU) layer, wherein the stride of the 1*3 1D convolutional layer is set to 1 in the M direction and 2 in the N direction.

[0025] Each downsampled 1D CNN generates a downsampled feature map, which serves as the input to the next module. After processing by C downsampled 1D CNNs, the number of columns in the C downsampled feature maps becomes N / 2. CTo compensate for the information loss caused by downsampling the number of columns to half its original value each time, each downsampled 1D CNN doubles the number of channels in the input data using a 1D convolutional layer. Therefore, after processing by C downsampled 1D CNNs, the number of channels in the downsampled feature map C becomes 2K×2. C The product of the number of channels and the number of columns remains constant at 2K×N, thus ensuring that the overall information content does not decrease.

[0026] Preferably, in the method for improving the Doppler spectral temporal resolution of an ultrasound system, the purpose of the Doppler decoder network is to decode the result encoded by the Doppler encoder network; therefore, its structure has a certain degree of symmetry with 310. The specific steps are as follows:

[0027] First, the downsampled feature map C output by the Doppler encoder network is input into C consecutive upsampled 1D CNNs for processing, where 1D means one-dimensional and CNN means convolutional neural network, to obtain an upsampled feature map C of size [M, N, 2K]. Then, it is processed by a channel dimensionality reduction network to reduce the number of channels from 2K to 2, resulting in Doppler IQ data with enhanced temporal resolution of size [M, KN, 2]. This achieves the effect of increasing the number of columns N of the input Doppler IQ data of the Doppler encoder network to KN.

[0028] Preferably, in the method for improving the temporal resolution of the Doppler spectrum of an ultrasound system, the upsampling 1D CNN is implemented by a linear interpolation upsampling layer, a 1*3 1D convolutional layer, followed by a linear rectified (ReLU) layer. The purpose of the linear interpolation upsampling layer is to upsample the number of columns of the input data to twice the original value. For any input x of length L, it generates an output y of length 2L, where L is a positive integer greater than 1. The calculation formula is as follows:

[0029]

[0030] In this system, a 1*3 1D convolutional layer is set to a stride of 1. Each upsampled 1D CNN generates an upsampled feature map, which serves as the input to the next module. After processing by C upsampled 1D CNNs, the number of columns in the upsampled feature map C becomes N. To balance the increase in data volume and computational cost caused by doubling the number of columns each time, each upsampled 1D CNN uses a 1D convolutional layer to halve the number of channels in the input data. Therefore, after processing by C upsampled 1D CNNs, the number of channels in the upsampled feature map C becomes 2K. The multiplication of the number of channels and the number of columns remains constant at 2K×N, thus ensuring that the overall data volume and computational cost do not increase.

[0031] Preferably, in the method for improving the temporal resolution of the Doppler spectrum of an ultrasound system, the channel dimensionality reduction network includes a CNN, a channel rearrangement layer, a 3*3 convolutional layer, and finally a hyperbolic tangent activation layer (tanh). The channel rearrangement layer simply rearranges the input tensor, changing its size from [M, N, 2K] to [M, KN, 2]. Assuming the input is x and the output is y, the formula for calculating the channel rearrangement is:

[0032]

[0033] In the above formula, mod represents the modulo operation, and ceil represents the floor operation.

[0034] The stride of the 3*3 convolutional layer is set to 1, so the size of row M and column KN is not changed. By using 2*2 different convolutional coefficients, the number of channels is kept constant at 2.

[0035] tanh takes an input x and produces an output y, and its mathematical expression is as follows:

[0036]

[0037] By means of the above-described solution, the present invention has at least the following advantages:

[0038] 1. This invention improves the temporal resolution of Doppler spectral imaging. By using two deep convolutional networks: a Doppler encoder network and a Doppler decoder network, the temporal resolution of the Doppler IQ data received by the ultrasound system is enhanced, resulting in a longer Doppler scanning window. This improves the spectral broadening phenomenon that is easily caused by using a shorter Doppler scanning window, and greatly improves the temporal resolution of Doppler spectral imaging while maintaining the frequency resolution of the spectrum.

[0039] 2. The technology of this invention will significantly improve the image quality of Doppler spectrum imaging of ultrasound systems, especially for low-speed blood flow or low-speed tissue Doppler imaging, thereby providing doctors with more accurate diagnostic information and reducing possible misdiagnosis.

[0040] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a diagram illustrating Doppler IQ data;

[0043] Figure 2 This is a flowchart of the Doppler spectrum calculation process of the present invention;

[0044] Figure 3 This is a flowchart of the Doppler temporal resolution enhancement network of the present invention;

[0045] Figure 4 This is a flowchart of the Doppler encoder network of the present invention;

[0046] Figure 5 This is a flowchart of the Doppler decoder network of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0048] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0049] Example

[0050] like Figures 2 to 5 As shown, a method for improving the temporal resolution of the Doppler spectrum of an ultrasound system includes the following steps:

[0051] Step 1: First, acquire complex Doppler IQ data of size M×N;

[0052] Step 2: Input the Doppler time resolution enhancement network to perform time resolution enhancement on the Doppler IQ data. The steps are as follows:

[0053] Step 21: First, split and rearrange the real and imaginary parts of the input Doppler IQ data to obtain a tensor of size [M, N, 2].

[0054] Where M represents the rows of the tensor, N represents the columns of the tensor, and 2 represents the number of channels of the tensor (the size of tensors in the following text is arranged in this order);

[0055] Step 22: Then, the tensor obtained in Step 21 is input into the Doppler encoder network for encoding to obtain an encoded output tensor, which is used to better represent the various potential features in the input data;

[0056] Step 23: Finally, the output tensor encoded in Step 22 is input into the Doppler decoder network for decoding. Based on various latent features in the input data, the time-resolution enhanced Doppler IQ data is obtained, which is a tensor of size [M, KN, 2]. By increasing the number of lines from N to KN, the Doppler time-resolution enhancement network increases the length of the Doppler scanning window, thereby improving the temporal resolution of the spectrum while preserving the frequency resolution after subsequent calculations.

[0057] Where K is a positive integer greater than 1, and the appropriate value can be selected according to different applications;

[0058] Step 3: Wall filtering, used to filter out noise or unwanted clutter in the data, employing low-pass filtering (for tissue imaging) or high-pass filtering (for blood flow imaging).

[0059] Step 4: Time-domain windowing is used to suppress side lobes in the spectrum and improve the contrast of the spectrum display; choose any window function commonly found in signal processing textbooks, such as Hamming window, Hanning window, rectangular window, triangular window, Gaussian window, Blackman window, Kaiser window, Bessel window, and Chebyshev window;

[0060] Step 5: Padding the data with zeros at the end is to pad the data to powers of 2, which facilitates the subsequent FFT calculation and also improves the spectral resolution.

[0061] Step 6: FFT (Fast Fourier Transform) is performed on the zero-padding data to obtain the Doppler spectrum for display.

[0062] The specific steps of the Doppler encoder network in step 22 of this invention are as follows:

[0063] First, the input Doppler IQ data is processed through a channel upscaling network to increase the number of channels from 2 to 2K, where K is a positive integer greater than 1. The appropriate value can be selected according to different applications to obtain an upscaling feature map of size [M, N, 2K].

[0064] Then, after processing by C downsampled 1D CNNs, where 1D means one-dimensional and CNN is a convolutional neural network, the final result is of size [M, N / 2]. C 2K×2 C The downsampled feature map of ] is used as the output tensor of the encoded tensor, where C is a positive integer greater than or equal to 1, and the appropriate value can be selected according to different applications and computing resources.

[0065] The larger the value of C, the higher the computational resource requirements, and vice versa. However, an excessively small C may cause the encoded output tensor to fail to fully extract the hidden potential features in the input Doppler IQ data, thus causing difficulties for the subsequent Doppler decoder. Therefore, an appropriate value should be selected according to different applications and computational resources. C is set by those skilled in the art.

[0066] The channel upscaling network described in this invention consists of a CNN. Since CNN is common knowledge among those skilled in the art, its basic knowledge will not be introduced in detail. The CNN consists of a 3*3 convolutional layer with a stride of 1. Therefore, the size of rows M and columns N is not changed, but the number of channels is increased from 2 to 2K through 2*2K different convolution coefficients.

[0067] The downsampled 1D CNN described in this invention consists of a 1*3 1D convolutional layer followed by a Rectified Linear Array (ReLU) layer. Although the Doppler encoder network has C downsampled 1D CNNs, their structures are identical, differing only in the coefficients of each convolutional layer. Therefore, only the structure of the first downsampled 1D CNN is described here. The 1*3 1D convolutional layer has a stride of 1 in the M direction and 2 in the N direction. This does not change the size of row M, but the size of column N is downsampled to half its original size. Furthermore, to save computational resources, convolution operations are performed only in the N direction, not in the M direction, making it a one-dimensional convolutional layer with very high computational efficiency.

[0068] Each downsampled 1D CNN generates a downsampled feature map, which serves as the input to the next module. After processing by C downsampled 1D CNNs, the number of columns in the C downsampled feature maps becomes N / 2. C To compensate for the information loss caused by downsampling the number of columns to half its original value each time, each downsampled 1D CNN doubles the number of channels in the input data using a 1D convolutional layer. Therefore, after processing by C downsampled 1D CNNs, the number of channels in the downsampled feature map C becomes 2K×2. C The product of the number of channels and the number of columns remains constant at 2K×N, thus ensuring that the overall information content does not decrease.

[0069] The ReLU used in downsampling 1D CNNs is common knowledge to those skilled in the art, and will only be briefly explained here. ReLU takes an input x and generates an output y, and its mathematical expression is as follows:

[0070]

[0071] The purpose of the Doppler decoder network described in this invention is to decode the result encoded by the Doppler encoder network. Therefore, its structure has a certain degree of symmetry with the Doppler encoder network. The specific steps are as follows:

[0072] First, the downsampled feature map C output by the Doppler encoder network is input into C consecutive upsampled 1D CNNs for processing, where 1D means one-dimensional and CNN means convolutional neural network, to obtain an upsampled feature map C of size [M, N, 2K]. Then, it is processed by a channel dimensionality reduction network to reduce the number of channels from 2K to 2, resulting in Doppler IQ data with enhanced temporal resolution of size [M, KN, 2]. This achieves the effect of increasing the number of columns N of the input Doppler IQ data of the Doppler encoder network to KN.

[0073] The upsampled 1D CNN described in this invention consists of a linear interpolation upsampling layer, a 1*3 1D convolutional layer, and a rectified linear LU (ReLU) layer. Although the Doppler decoder network has C upsampled 1D CNNs, their structures are identical; the only difference lies in the coefficients of each convolutional layer. Therefore, only the structure of the first upsampled 1D CNN is described here. The aforementioned linear interpolation upsampling layer is common knowledge to those skilled in the art and will only be briefly explained here. The purpose of the linear interpolation upsampling layer is to upsample the number of columns of the input data to twice its original value. For any input x of length L, it generates an output y of length 2L, where L is a positive integer greater than 1. The calculation formula is as follows:

[0074]

[0075] The 1*3 1D convolutional layer has a stride of 1, thus not changing the row and column sizes. Furthermore, to conserve computational resources, convolution operations are performed only in the N direction, not the M direction, making it a one-dimensional convolutional layer with very high computational efficiency. Each upsampled 1D CNN generates an upsampled feature map, which serves as the input to the next module. After processing by C upsampled 1D CNNs, the number of columns in the upsampled feature map C becomes N. To balance the increased data volume and computational cost caused by doubling the number of columns each time, each upsampled 1D CNN uses a 1D convolutional layer to halve the number of channels in the input data. Therefore, after processing by C upsampled 1D CNNs, the number of channels in the upsampled feature map C becomes 2K. The product of the number of channels and the number of columns remains constant at 2K×N, ensuring that the overall data volume and computational cost do not increase.

[0076] The ReLU used in upsampling 1D CNNs is the same as the ReLU used in downsampling 1D CNNs, so it will not be repeated here.

[0077] The channel dimensionality reduction network described in this invention includes a CNN, a channel rearrangement layer, a 3*3 convolutional layer, and finally a hyperbolic tangent activation layer (tanh). The channel rearrangement layer simply rearranges the input tensor, changing its size from [M, N, 2K] to [M, KN, 2]. Assuming the input is x and the output is y, the formula for calculating the channel rearrangement is:

[0078]

[0079] In the above formula, mod represents the modulo operation, and ceil represents the floor operation.

[0080] In this design, the stride of the 3x3 convolutional layer is set to 1, thus maintaining the size of rows M and columns KN without altering their values. By using 2x2 different convolutional coefficients, the number of channels remains constant at 2. The tanh parameter mentioned above is common knowledge to those skilled in the art; it is only briefly explained here. tanh accepts an input x and generates an output y, and its mathematical expression is as follows:

[0081]

[0082] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0083] In the description of this application, it should be noted that the terms "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0084] Furthermore, terms such as "horizontal" and "vertical" do not imply that components must be absolutely horizontal or vertical, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0085] In the description of this application, it should also be noted that, unless otherwise expressly specified and limited, the terms "set up," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for improving the temporal resolution of the Doppler spectrum in an ultrasound system. Its features are, Includes the following steps: Step 1: First, acquire complex Doppler IQ data of size M×N; Step 2: Input the Doppler time resolution enhancement network to perform time resolution enhancement on the Doppler IQ data. The steps are as follows: Step 21: First, split and rearrange the real and imaginary parts of the input Doppler IQ data to obtain a tensor of size [M, N, 2]. Where M represents the rows of the tensor, N represents the columns of the tensor, and 2 represents the number of channels of the tensor; Step 22: Then, the tensor obtained in Step 21 is input into the Doppler encoder network for encoding to obtain an encoded output tensor, which is used to better represent the various potential features in the input data; Step 23: Finally, the output tensor encoded in Step 22 is input into the Doppler decoder network for decoding. Based on various latent features in the input data, the time-resolution enhanced Doppler IQ data is obtained, which is a tensor of size [M, KN, 2]. By increasing the number of lines from N to KN, the Doppler time-resolution enhancement network increases the length of the Doppler scanning window, thereby improving the temporal resolution of the spectrum while preserving the frequency resolution after subsequent calculations. Where K is a positive integer greater than 1, and the appropriate value can be selected according to different applications; Step 3: Wall filtering, used to filter out noise or unwanted clutter in the data; Step 4: Time-domain windowing is used to suppress sidelobes in the spectrum and improve the contrast of the spectrum display; Step 5: Pad the data with zeros to make it a power of 2, which facilitates the subsequent FFT calculation and also improves the spectral resolution. Step 6: FFT, perform Fast Fourier Transform on the zero-padding data, and display the Doppler spectrum.

2. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 1, characterized in that: The wall filtering in step 3 is either low-pass filtering or high-pass filtering.

3. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 1, characterized in that: The time-domain windowing in step 4 includes Hamming window, Hanning window, rectangular window, triangular window, Gaussian window, Blackman window, Kaiser window, Bessel window, and Chebyshev window.

4. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 1, characterized in that: The specific steps of the Doppler encoder network in step 22 are as follows: First, the input Doppler IQ data is processed through a channel upscaling network to increase the number of channels from 2 to 2K, where K is a positive integer greater than 1. The appropriate value can be selected according to different applications to obtain an upscaling feature map of size [M, N, 2K]. Then, after processing by C downsampled 1D CNNs, where 1D means one-dimensional and CNN is a convolutional neural network, the final result is of size [M, N / 2]. C 2K×2 C The downsampled feature map of ] is used as the output tensor of the encoded tensor, where C is a positive integer greater than or equal to 1, and the appropriate value can be selected according to different applications and computing resources.

5. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 4, characterized in that: The channel-upgrading network consists of a CNN, and the CNN consists of a 3D... The system consists of 3 convolutional layers with a stride of 1, thus not changing the size of row M and column N, but through 2... With 2K different convolution coefficients, the number of channels is increased from 2 to 2K.

6. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 4, characterized in that: The downsampled 1D CNN consists of a 1 A 3D convolutional layer followed by a linear rectified layer is used to implement this, where 1 A 3-dimensional 1D convolutional layer with a stride of 1 in the M direction and 2 in the N direction; Each downsampled 1D CNN generates a downsampled feature map, which serves as the input to the next module. After processing by C downsampled 1D CNNs, the number of columns in the C downsampled feature maps becomes N / 2. C To compensate for the information loss caused by downsampling the number of columns to half its original value each time, each downsampled 1D CNN doubles the number of channels in the input data using a 1D convolutional layer. Therefore, after processing by C downsampled 1D CNNs, the number of channels in the downsampled feature map C becomes 2K×2. C The product of the number of channels and the number of columns remains constant at 2K×N, thus ensuring that the overall information content does not decrease.

7. The method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 1, characterized in that: The purpose of the Doppler decoder network is to decode the result encoded by the Doppler encoder network. The specific steps are as follows: First, the downsampled feature map C output by the Doppler encoder network is input into C consecutive upsampled 1D CNNs for processing, where 1D means one-dimensional and CNN means convolutional neural network, to obtain an upsampled feature map C of size [M, N, 2K]. Then, it is processed by a channel dimensionality reduction network to reduce the number of channels from 2K to 2, resulting in Doppler IQ data with enhanced temporal resolution of size [M, KN, 2]. This achieves the effect of increasing the number of columns N of the input Doppler IQ data of the Doppler encoder network to KN.

8. A method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 7, characterized in that: The upsampled 1D CNN consists of a linear interpolation upsampling layer and a 1 The system uses a 3x1D convolutional layer followed by a linear rectified layer. The purpose of the linear interpolation upsampling layer is to upsample the number of columns of the input data to twice its original value. For any input x of length L, it generates an output y of length 2L, where L is a positive integer greater than 1. The calculation formula is as follows: Among them, 1 The 3-layer 1D convolutional layer has a stride of 1. Each upsampled 1D CNN generates an upsampled feature map, which serves as the input to the next module. After processing by C upsampled 1D CNNs, the number of columns in the upsampled feature map C becomes N. To balance the increase in data volume and computation caused by doubling the number of columns each time, each upsampled 1D CNN uses a 1D convolutional layer to halve the number of channels in the input data. Therefore, after processing by C upsampled 1D CNNs, the number of channels in the upsampled feature map C becomes 2K. The multiplication of the number of channels and the number of columns remains unchanged at 2K×N, thus ensuring that the overall data volume and computation do not increase.

9. A method for improving the time resolution of the Doppler spectrum of an ultrasound system according to claim 7, characterized in that: The channel dimensionality reduction network includes a CNN, a channel rearrangement layer, and a 3D layer. The system consists of three convolutional layers, followed by a hyperbolic tangent activation layer. The channel rearrangement layer simply rearranges the input tensor, changing its size from [M, N, 2K] to [M, KN, 2]. Assuming the input is x and the output is y, the formula for calculating the channel rearrangement is: In the above formula, mod is the modulo operation and ceil is the floor operation; Among them, 3 The stride of the 3rd convolutional layer is set to 1, so the size of row M and column KN is not changed, by using 2 Two different convolution coefficients are used, while keeping the number of channels constant at 2.

Images