Deep Learning-Based Ultrasonic Radio Frequency Data Filtering Imaging Method

Abstract

This invention discloses a deep learning-based ultrasound radio frequency data filtering imaging method. The method first proposes a novel ultrasound radio frequency data filtering algorithm for ultrasound simulation imaging—an echo signal selection algorithm based on scattering points. Then, it proposes a deep neural network to learn the target echo signal selection process based on scattering points for ultrasound radio frequency data. The trained deep neural network can effectively suppress interference signals in the radio frequency data, thereby improving ultrasound imaging quality. This data filtering deep neural network can be integrated into a time-delayed beamforming imaging framework to form a novel ultrasound imaging method, namely, a deep learning-based ultrasound radio frequency data filtering imaging method. Compared with time-delayed beamforming methods, this invention can improve the contrast and contrast-to-noise ratio of ultrasound images, exhibiting excellent imaging quality in real-world human imaging examples.

Description

Technical Field

[0001] This invention belongs to the field of ultrasound imaging, specifically relating to a deep learning-based ultrasound radio frequency data filtering imaging method. Background Technology

[0002] Ultrasound imaging involves a region composed of numerous scattering points. When emitted ultrasound waves encounter these scattering points, scattering occurs. Some of the scattered signals propagating towards the transducer are received by the transducer, becoming radio frequency (RF) data. However, in space, these numerous scattered signals interfere with each other, causing echo signals to be generated at the interference locations and received by the transducer. This means the RF data contains signals not only from the target scattering point but also from numerous interference locations. Even if only one scattering point exists in space, the interference signal can cause artifacts around that point. Effectively suppressing interference signals improves the imaging effect of RF data. Therefore, selecting RF data based on the spatial location of the scattering points can isolate a large number of signals from interfering locations in space, thereby improving the imaging performance of the RF data. However, in real-world human imaging, the location of the scattering points is unknown.

[0003] Yoon et al. proposed a deep learning ultrasound imaging method using downsampled ultrasound echo radio frequency data as input data in the paper "Yoon YH, Khan S, Huh J, et al. Efficient B-mode ultrasound image reconstruction from sub-sampled RF data using deep learning[J].IEEE Transactions on Medical Imaging, 2018, 38(2): 325-336". However, this method does not have the rich signal features of the original data.

[0004] In the paper “Luijten B, Cohen R, de Bruijn FJ, et al. Adaptive ultrasound beamforming using deep learning[J].IEEE Transactions on Medical Imaging, 2020, 39(12): 3967-3978”, Luijten et al. proposed an adaptive minimum variance beamforming algorithm based on deep learning. This method uses the minimum variance beamforming algorithm as the learning object of the neural network, but it cannot suppress echo signals from interference locations. Summary of the Invention

[0005] The main objective of this invention is to accurately select target echo signals from ultrasound radio frequency (RF) data, thereby suppressing interference signals and improving ultrasound imaging quality. This invention first proposes a target echo signal selection algorithm based on scattering points for ultrasound simulation imaging. Then, it proposes a deep neural network to learn the target echo signal selection process based on scattering points for ultrasound RF data, suppressing interference signals and improving ultrasound image quality. The imaging method of this invention improves image contrast and contrast-to-noise ratio without reducing image resolution, enabling the acquisition of high-quality ultrasound images.

[0006] The objective of this invention is achieved through at least one of the following technical solutions.

[0007] A deep learning-based ultrasonic radio frequency data filtering imaging method uses an interference signal suppression neural network to suppress interference signals in ultrasonic radio frequency data. This interference signal suppression neural network learns an echo signal selection algorithm based on scattering points for ultrasonic simulation imaging. This algorithm selects signals according to the spatial location of the scattering points, isolating a large number of signals from spatial interference locations. The method includes the following steps:

[0008] S1. The target object of ultrasonic imaging is scanned. The echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. The received real radio frequency data is converted into complex radio frequency data. After performing corresponding delay operations for different imaging points, delayed radio frequency data is obtained.

[0009] S2. Separate the real and imaginary parts of the complex radio frequency data, and use an interference signal suppression neural network to suppress the interference signal in the delayed radio frequency data obtained in step S1, so as to obtain the delayed radio frequency data after suppressing the interference signal.

[0010] S3. The delayed radio frequency data obtained in step S2 after suppressing the interference signal is recombined into complex radio frequency data, and then superimposed to complete beamforming, thereby obtaining the image pixel value of the corresponding ultrasound imaging target and thus forming a complete ultrasound image. This ultrasound image has the advantages of high contrast and high contrast-to-noise ratio.

[0011] Further, in step S1, the target object for ultrasonic imaging is scanned, and the echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. Then, the real form of radio frequency data is converted into complex form radio frequency data using Hilbert transform.

[0012] Next, the radio frequency data is delayed. That is, the delay time is calculated based on the position of the target imaging point, the position of the scan line, and the position of the receiving array element. The delay time is mapped to a signal subscript, thereby extracting the signal value corresponding to the target imaging point in the echo signal of the receiving array element to obtain the delayed radio frequency data.

[0013] The time delay operation is a routine procedure in ultrasound imaging and will not be described in detail here.

[0014] Furthermore, let the number of scan lines be L, the number of target imaging points on a scan line be P, and the number of receiving array elements be N. After the delay operation, a first delayed radio frequency data matrix M1 with a size of P×N×L is obtained. This matrix is ​​a complex matrix. Then, the real part and imaginary part of each data in the first delayed radio frequency data matrix M1 are separated to obtain a second delayed radio frequency data matrix M2 with a size of 2×P×N×L.

[0015] Further, in step S2, the second delayed radio frequency data matrix M2 is input into the interference signal suppression neural network for processing to obtain the third delayed radio frequency data matrix M3 with a size of 2×P×N×L after suppressing the interference signal;

[0016] The interference signal suppression neural network used is a three-dimensional convolutional neural network with an encoder-decoder structure. This neural network has learned a novel ultrasonic radio frequency data filtering algorithm for ultrasonic simulation imaging proposed in this invention—an echo signal selection algorithm based on scattering points.

[0017] This invention uses an encoder-decoder structure as the overall architecture of the interference signal suppression neural network, enabling the interference signal suppression neural network to utilize the strong correlation of radio frequency data to complete spatial filtering of the signal;

[0018] Since the delayed radio frequency data is a three-dimensional matrix, a neural network for suppressing interference signals is implemented based on a three-dimensional encoder-decoder network structure.

[0019] Furthermore, in the interference signal suppression neural network, since the radio frequency data is complex, the real and imaginary parts are separated and each becomes a channel, so the number of input channels of the network is 2. Due to the limitations of hardware computing resources and the large amount of input data, the initial number of channels of the intermediate feature map of the interference signal suppression neural network is set to 4. The interference signal suppression neural network includes a compression path and an expansion path. The compression path on the left increases the number of channels while reducing the size of the feature map, and the expansion path on the right decreases the number of channels while restoring the size of the feature map.

[0020] The compression path consists of five convolutional blocks, which are connected by convolutional layers responsible for downsampling; the expansion path consists of four convolutional blocks, which are connected by convolutional layers responsible for upsampling.

[0021] Each convolutional block in the compression and expansion paths includes residual connections. The input and output of the convolutional block are superimposed before being passed to the next convolutional layer. After the first convolutional block, the feature map has 4 channels. Therefore, before the input of the interference suppression neural network is superimposed with the output of the first convolutional block, it needs to copy itself to expand the number of channels to 4. The input and output channels of other convolutional blocks are the same, so they can be directly superimposed. Downsampling is handled by a convolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which doubles the number of channels. Upsampling is handled by a deconvolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which halves the number of channels.

[0022] The output layer of the interference signal suppression neural network is a convolutional layer with a kernel size of (1, 1, 1) and a stride of 1, reducing the number of output channels to 2, equal to the number of input channels. The remaining convolutional kernels in the interference signal suppression neural network are all of size (5, 5, 5) with a stride of 1. The activation function used in the interference signal suppression neural network is the hyperbolic tangent function Tanh, instead of the commonly used ReLU series functions. This is because radio frequency data contains both positive and negative signals, and both ReLU and PReLU functions suppress negative numbers to some extent; therefore, the Tanh activation function is more suitable for processing radio frequency data. The loss function used is the MSE loss function with L1 regularization, calculated as follows:

[0023]

[0024] Where X represents a sample, Y represents the true value, f represents a neural network, W(f) represents the set of weights of the neural network, and lambda represents the L1 regularization coefficient.

[0025] Furthermore, the interference signal suppression neural network learns an echo signal selection algorithm based on scattering points for ultrasound simulation imaging, which selects the target echo signal according to the spatial location of the scattering point to isolate a large number of signals from interfering locations in space. The specific process is as follows:

[0026] In single-scattering point simulation imaging, a scattering point S is simulated within the imaging domain, and a transmitting channel t and a receiving channel r, along with the RF echo signal vector x received by the receiving channel r, are considered. r Based on the positions of the transmitting channel t, the receiving channel r, and the scattering point S, calculate the length d of the signal transmission path from the transmitting channel t to the scattering point S, and then from the scattering point S to the receiving channel r. t，r Calculate using the following formula:

[0027] d t，r =d t-S +d S-r ,t,r∈1,2,...,N (2)

[0028] Where N represents the total number of channels in the transducer, d t-S d represents the length of the signal transmission path from the transmission channel t to the scattering point S. S-r This represents the length of the signal transmission path from the scattering point S to the receiving channel r; the transmitting channel t and the receiving channel r constitute a transmission channel pair;

[0029] Next, the transmission path is calculated in the RF echo signal vector x. r The element index k in t，r :

[0030] k t，r =round(d t，r ÷c×f s ), t, r∈1, 2,..., N (3)

[0031] Where round is the floor function, c is the speed of sound, and f is the floor function. s The signal sampling frequency;

[0032] In this way, the echo signal from each scattering point is selected, thereby suppressing the echo signal from the interference point.

[0033] The center points of the transmitting channel t and the receiving channel r are defined as F1 and F2, respectively. With F1 and F2 as the two foci, the trajectory of a moving point in the plane whose sum of distances to F1 and F2 is a constant (greater than |F1F2|) forms an ellipse. Therefore, the transmission path length d of the echo signal from the scattering point located on the elliptical arc with F1 and F2 as foci is... t，r They are all the same; d lies at any point on the elliptical arc with foci F1 and F2. t，r and k t，r same.

[0034] Furthermore, the rounding function in formula (3) means that multiple different path lengths will be mapped to the same vector element index. Therefore, the elliptical arc where the scattering point is located is actually an elliptical ring, and the signal transmission path calculated at any position within the elliptical ring will be mapped to the same element index. Therefore, according to the index k t，r While selecting echo signals from the scattering point location, a portion of echo signals from the interference region on the elliptical ring are also selected; that is, the subscript k of the selected element is... t，r The echo signal sample is a composite signal, which includes not only the echo signal from the scattering point, but also the echo signal from the interference region on the elliptical ring.

[0035] For each of the remaining transmission channel pairs, there also exists an elliptical ring corresponding to the same element index k. t，rThe echo signal transmission path lengths for the two transmission channels (t1, r) and (t2, r) are equal, i.e. This indicates that the corresponding elliptical ring interference regions are different for ultrasound signals emitted from different channels; interference signals from all these interference regions are preserved along with signals from the scattering point location, while signals from other interference regions are excluded.

[0036] For imaging scenarios involving multiple scattering points, the signals generated by different scattering points interfere with each other. For example, in simulated imaging with multiple scattering points, scattering points SA, SB, and SC coexist. When calculating the echo signal from the transmitting channel t to the receiving channel r, in addition to the spatial interference signal, i.e., the elliptical ring corresponding to scattering point SA, there is also interference signal generated by scattering points SB and SC. Therefore, when applying the echo signal selection algorithm based on scattering points, it is necessary to simulate each scattering point separately, process the radio frequency echo data generated by each scattering point using the echo signal selection algorithm based on scattering points, and then superimpose the radio frequency data of all scattering points.

[0037] Furthermore, consider an N-channel transducer, with both the transmit and receive apertures encompassing all N channels; the number of scan lines is L, and the echo signal vector length recorded by each receive channel is P; since there are a total of N transmit channels, for a given scattering point S, for each receive channel of each scan line, a total of N signal samples are selected; after selecting the relevant echo signals of scattering point S using the scattering point-based echo signal selection algorithm, a radio frequency data matrix M of size P×L×N is formed. SESA In this context, for each scan line l∈{1,2,…,L} and each receiving channel n∈{1,2,…,N}, only the selected N signal samples are not zero.

[0038] After selecting the radio frequency echo data using the echo signal selection algorithm based on scattering points, the delayed superposition beamforming algorithm can be used for imaging. However, for multi-scattering point imaging scenarios, since the imaging area contains a large number of scattering points, it is necessary to simulate each scattering point separately, apply the echo signal selection algorithm based on scattering points to process the radio frequency echo data generated by each scattering point, then add up the selected radio frequency echo data of all scattering points, and finally use the delayed superposition beamforming algorithm for imaging.

[0039] However, in order to save computing resources, after processing the radio frequency echo data generated by each scattering point using the echo signal selection algorithm based on scattering points, a delay operation is first applied to the radio frequency echo data to greatly reduce the data size. Then, the selected and delayed radio frequency echo data of all scattering points are added together, and finally the radio frequency echo data of all receiving channels are superimposed to complete the delayed superimposed beamforming.

[0040] After processing the radio frequency data of all scattering points using the echo signal selection algorithm based on scattering points, the data is first delayed to reduce the size of the radio frequency data from P×N×L to C×N×L, where C represents the number of vertical pixels, and C is much smaller than P. Then, all the delayed radio frequency data are added together to obtain the complete delayed radio frequency data of the complex scene with multiple scattering points. Finally, the channel dimensions are superimposed to obtain the delayed superimposed radio frequency data of size C×L.

[0041] In real-world ultrasound imaging scenarios, the locations of scattering points in the imaging area are unknown beforehand. Therefore, echo signal selection algorithms based on scattering points cannot be directly applied to real-world imaging scenarios. In simulated ultrasound imaging scenarios, the tissue structure and scattering point locations of the imaging area are known, allowing for the effective selection of echo signals generated by scattering points, thereby suppressing interference signals. Previous literature has shown that neural networks trained using only single-scattering-point radio frequency data can generalize to simulated cyst imaging and real-world imaging. Moreover, neural networks can approximate any mathematical function. Therefore, this invention proposes a deep neural network to implement an echo signal selection algorithm based on scattering points, making it applicable to real-world imaging scenarios. The question then becomes: how to learn single-point radio frequency data M that has not been processed by an echo signal selection algorithm based on scattering points? original And single-point radio frequency data M processed by the echo signal selection algorithm based on scattering points SESA The transformation relationship between them is a regression task, therefore it uses a large number of {M} original M SESA Deep neural networks trained on datasets composed of tuples are a suitable tool for generalizing echo signal selection algorithms based on scattering points to real-world imaging.

[0042] Due to the large volume of radio frequency (RF) data, a delay operation is required to reduce GPU memory usage. Therefore, an interference signal suppression neural network that has learned the echo signal selection process based on scattering points is integrated into the delay-superimposed beamforming framework.

[0043] Furthermore, the training process of the interference signal suppression neural network is as follows: Ultrasonic radio frequency data of a single scattering point is generated using Field II simulation software. The radio frequency data is processed using an echo signal selection algorithm based on the scattering point to obtain the training ground truth. A single scattering point's radio frequency data is considered a sample. The training set and validation set each contain multiple samples, with a sample size ratio of 10:1. The deep neural network is trained using the Adam optimizer. The Adam optimizer uses both the mean and variance of the gradient to calculate the update step size of the network weight parameters. Training set data is input into the deep neural network batch by batch, and the Adam optimizer updates the parameters of the deep neural network. Processing all training set data constitutes one training cycle. After each training cycle, the validation error is calculated using the validation set data, and the network weight parameters corresponding to the minimum validation error are saved. When the training cycle exceeds a set threshold, the network weights for each round are saved, and the optimal network weights are determined during testing.

[0044] Further, in step S3, the third delayed radio frequency data M3 obtained in S2 after suppressing the interference signal is recombined into a complex delayed radio frequency data matrix M4 of size P×N×L. The complex delayed radio frequency data matrix M4 has the same size as the delayed radio frequency data matrix M1 before the deep neural network processing. Then, the complex delayed radio frequency data matrix M4 is superimposed to obtain the pixel values ​​of all target imaging points.

[0045] After the above steps, the pixel values ​​of all target imaging points on all scan lines can be obtained. Subsequent envelope detection, logarithmic compression, and dynamic range display operations will then yield a complete ultrasound image. Envelope detection and logarithmic compression are routine operations in ultrasound imaging and will not be described in detail here.

[0046] The deep learning-based ultrasound radio frequency data filtering imaging method consists of two main components: a deep neural network for suppressing interference signals and beamforming. Based on the traditional time-delay superposition beamforming framework, an interference signal suppression neural network is added to suppress interference signals in ultrasound radio frequency data, thereby improving the quality of ultrasound images.

[0047] Compared with the prior art, the advantages of the present invention are mainly reflected in:

[0048] This invention uses deep learning technology to suppress interference signals in ultrasound radio frequency data. It integrates a deep neural network that can suppress interference signals into a traditional time-delay superposition beamforming framework, which can improve the contrast and contrast signal-to-noise ratio of ultrasound images while maintaining resolution. Attached Figure Description

[0049] Figure 1This is a schematic diagram of the imaging process of the ultrasound radio frequency data filtering imaging method based on deep learning in an embodiment of the present invention;

[0050] Figure 2 This is a topology diagram of the interference signal suppression neural network used in the deep learning-based ultrasound radio frequency data filtering imaging method in this embodiment of the invention.

[0051] Figure 3 This is a schematic diagram illustrating the single-scattering point imaging principle of the echo signal selection algorithm based on scattering points in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the interference area for different transmission channels of the echo signal selection algorithm based on scattering points in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram illustrating the principle of multi-scattering point imaging based on the echo signal selection algorithm of scattering points in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram illustrating the process of calculating the echo signal for a specific receiving channel using a single scattering point in the echo signal selection algorithm based on scattering points in an embodiment of the present invention.

[0055] Figure 7 This is a schematic diagram of multi-scattering point beamforming based on the echo signal selection algorithm of scattering points in an embodiment of the present invention;

[0056] Figure 8 This is an example of a single-point target image output by a deep learning-based ultrasonic radio frequency data filtering imaging method in this invention.

[0057] Figure 9 This is an example of a cyst image output by the time-delay superposition imaging method, the echo signal selection algorithm based on scattering points, and the ultrasound radio frequency data filtering imaging method based on deep learning in this invention.

[0058] Figure 10 This is an example of a human thyroid image output by the time-lapse imaging method and the deep learning-based ultrasound radio frequency data filtering imaging method in this invention. Detailed Implementation

[0059] The specific implementation of the present invention will be further described below with reference to the accompanying drawings and embodiments, but the implementation and protection of the present invention are not limited thereto. It should be noted that any details not specifically described below are implementations that can be achieved by those skilled in the art with reference to existing technology.

[0060] Example:

[0061] The ultrasound radio frequency data filtering imaging method based on deep learning proposed in this invention will ultimately be applied to ultrasound imaging. In one embodiment, an ultrasound imaging system with three modules is constructed, namely a data acquisition module, a core computing module, and an image display module. The advantages of the ultrasound radio frequency data filtering imaging method based on deep learning proposed in this invention are illustrated based on certain quantitative indicators and visual effect evaluation.

[0062] For ultrasound simulation experiments, Field II simulation software was used to simulate the propagation process of ultrasound waves in ultrasound imaging and obtain simulation data. First, based on the corresponding configuration of the actual ultrasound imaging equipment, the corresponding simulated physical data was simulated, transmitting and receiving array elements were created, and simulated detection objects were created. Then, radio frequency data was simulated and received line by line according to the scan lines. In the simulation experiments of the data acquisition module, simulated single-point target objects and cyst objects were created respectively, and their imaging effects were observed. For the human thyroid imaging experiment, the radio frequency data was saved in a .mat file, and the human thyroid ultrasound radio frequency data could be obtained by reading the .mat file.

[0063] The core computing module includes a delay calculation module and an interference signal suppression neural network processing module. The interference signal suppression neural network implements a deep learning-based ultrasonic radio frequency data filtering imaging method.

[0064] In an ultrasound imaging system, after the core calculation module obtains the pixel data, the display image module uses the corresponding decoding program to perform Hilbert transformation, logarithmic compression, grayscale range correction, and image depth and width calculations and image display operations on the data. Finally, the image-related data is output to the corresponding coordinate system, and the ultrasound image is displayed on the screen.

[0065] A deep learning-based ultrasonic radio frequency data filtering imaging method uses an interference signal suppression neural network to suppress interference signals in ultrasonic radio frequency data. This interference signal suppression neural network learns a scattering point-based echo signal selection algorithm for ultrasonic simulation imaging. This algorithm selects signals based on the spatial location of the scattering points, isolating a large number of signals from spatially interfering locations. Figure 1 As shown, it includes the following steps:

[0066] S1. The target object of ultrasonic imaging is scanned. The echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. The received real radio frequency data is converted into complex radio frequency data. After performing corresponding delay operations for different imaging points, delayed radio frequency data is obtained.

[0067] The target object is scanned by ultrasound imaging. The echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. Then, the Hilbert transform is used to convert the real radio frequency data into complex radio frequency data.

[0068] Next, the radio frequency data is delayed. That is, the delay time is calculated based on the position of the target imaging point, the position of the scan line, and the position of the receiving array element. The delay time is mapped to a signal subscript, thereby extracting the signal value corresponding to the target imaging point in the echo signal of the receiving array element to obtain the delayed radio frequency data.

[0069] The time delay operation is a routine procedure in ultrasound imaging and will not be described in detail here.

[0070] Let the number of scan lines be L, the number of target imaging points on a scan line be P, and the number of receiving array elements be N. After a delay operation, a first delayed radio frequency data matrix M1 with a size of P×N×L is obtained. This matrix is ​​a complex matrix. Then, the real part and imaginary part of each data in the first delayed radio frequency data matrix M1 are separated to obtain a second delayed radio frequency data matrix M2 with a size of 2×P×N×L.

[0071] S2. Separate the real and imaginary parts of the complex radio frequency data, and use an interference signal suppression neural network to suppress the interference signal in the delayed radio frequency data obtained in step S1, so as to obtain the delayed radio frequency data after suppressing the interference signal.

[0072] The second delayed radio frequency data matrix M2 is input into the interference signal suppression neural network for processing to obtain the third delayed radio frequency data matrix M3 with a size of 2×P×N×L after suppressing the interference signal;

[0073] The interference signal suppression neural network used is a three-dimensional convolutional neural network with an encoder-decoder structure. This neural network has learned a novel ultrasonic radio frequency data filtering algorithm for ultrasonic simulation imaging proposed in this invention—an echo signal selection algorithm based on scattering points.

[0074] This invention uses an encoder-decoder structure as the overall architecture of the interference signal suppression neural network, enabling the interference signal suppression neural network to utilize the strong correlation of radio frequency data to complete spatial filtering of the signal;

[0075] Since the delayed radio frequency data is a three-dimensional matrix, a neural network for suppressing interference signals is implemented based on a three-dimensional encoder-decoder network structure.

[0076] like Figure 2As shown, in the interference signal suppression neural network, since the radio frequency data is complex, the real and imaginary parts are separated and each becomes a channel, so the number of input channels of the network is 2. Due to the limitations of hardware computing resources and the large amount of input data, the initial number of channels of the intermediate feature map of the interference signal suppression neural network is set to 4. The interference signal suppression neural network includes a compression path and an expansion path. The compression path on the left increases the number of channels while reducing the size of the feature map, and the expansion path on the right decreases the number of channels while restoring the size of the feature map.

[0077] The compression path consists of five convolutional blocks, which are connected by convolutional layers responsible for downsampling; the expansion path consists of four convolutional blocks, which are connected by convolutional layers responsible for upsampling.

[0078] Each convolutional block in the compression and expansion paths includes residual connections. The input and output of the convolutional block are superimposed before being passed to the next convolutional layer. After the first convolutional block, the feature map has 4 channels. Therefore, before the input of the interference suppression neural network is superimposed with the output of the first convolutional block, it needs to copy itself to expand the number of channels to 4. The input and output channels of other convolutional blocks are the same, so they can be directly superimposed. Downsampling is handled by a convolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which doubles the number of channels. Upsampling is handled by a deconvolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which halves the number of channels.

[0079] The output layer of the interference signal suppression neural network is a convolutional layer with a kernel size of (1, 1, 1) and a stride of 1, reducing the number of output channels to 2, equal to the number of input channels. The remaining convolutional kernels in the interference signal suppression neural network are all of size (5, 5, 5) with a stride of 1. The activation function used in the interference signal suppression neural network is the hyperbolic tangent function Tanh, instead of the commonly used ReLU series functions. This is because radio frequency data contains both positive and negative signals, and both ReLU and PReLU functions suppress negative numbers to some extent; therefore, the Tanh activation function is more suitable for processing radio frequency data. The loss function used is the MSE loss function with L1 regularization, calculated as follows:

[0080]

[0081] Where X represents a sample, Y represents the true value, f represents a neural network, W(f) represents the set of weights of the neural network, and lambda represents the L1 regularization coefficient.

[0082] The interference signal suppression neural network learns an echo signal selection algorithm based on scattering points for ultrasound simulation imaging. This algorithm selects the target echo signal based on the spatial location of the scattering points to isolate a large number of signals from interfering locations in space. The specific process is as follows:

[0083] like Figure 3 As shown, in single-scattering point simulation imaging, a scattering point S is simulated and set in the imaging domain. Consider a transmitting channel t, a receiving channel r, and the radio frequency echo signal vector x received by the receiving channel r. r Based on the positions of the transmitting channel t, the receiving channel r, and the scattering point S, calculate the length d of the signal transmission path from the transmitting channel t to the scattering point S, and then from the scattering point S to the receiving channel r. t，r Calculate using the following formula:

[0084] d t，r =d t-S +d S-r ,t,r∈1,2,...,N (2)

[0085] Where N represents the total number of channels in the transducer, d t-S d represents the length of the signal transmission path from the transmission channel t to the scattering point S. S-r This represents the length of the signal transmission path from the scattering point S to the receiving channel r; the transmitting channel t and the receiving channel r constitute a transmission channel pair;

[0086] Next, the transmission path is calculated in the RF echo signal vector x. r The element index k in t，r :

[0087] k t，r =round(d t，r ÷c×f s ), t, r∈1, 2,..., N (3)

[0088] Where round is the floor function, c is the speed of sound, and f is the floor function. s The signal sampling frequency;

[0089] In this way, the echo signal from each scattering point is selected, thereby suppressing the echo signal from the interference point.

[0090] The center points of the transmitting channel t and the receiving channel r are defined as F1 and F2, respectively. With F1 and F2 as the two foci, the trajectory of a moving point in the plane whose sum of distances to F1 and F2 is a constant (greater than |F1F2|) forms an ellipse. Therefore, the transmission path length d of the echo signal from the scattering point located on the elliptical arc with F1 and F2 as foci is... t，r They are all the same; d lies at any point on the elliptical arc with foci F1 and F2.t，r and k t，r same.

[0091] The rounding function in formula (3) means that multiple different path lengths will be mapped to the same vector element index. Therefore, the elliptical arc where the scattering point is located is actually an elliptical ring, and the signal transmission path calculated at any position within the elliptical ring will be mapped to the same element index. Therefore, according to the index k t，r While selecting echo signals from the scattering point location, a portion of echo signals from the interference region on the elliptical ring are also selected; that is, the subscript k of the selected element is... t，r The echo signal sample is a composite signal, which includes not only the echo signal from the scattering point, but also the echo signal from the interference region on the elliptical ring.

[0092] For each of the remaining transmission channel pairs, there also exists an elliptical ring corresponding to the same element index k. t，r ,like Figure 4 As shown, the echo signal transmission path lengths for the two transmission channels (t1, r) and (t2, r) are equal, i.e. This indicates that the corresponding elliptical ring interference regions are different for ultrasound signals emitted from different channels; interference signals from all these interference regions are preserved along with signals from the scattering point location, while signals from other interference regions are excluded.

[0093] In imaging scenarios involving multiple scattering points, signals generated by different scattering points interfere with each other; for example, ... Figure 5 As shown, in multi-scattering point simulation imaging, scattering points SA, SB, and SC coexist. When calculating the echo signal from the transmitting channel t to the receiving channel r, in addition to the spatial interference signal, i.e., the elliptical ring corresponding to scattering point SA, there is also interference signal generated by scattering points SB and SC. Therefore, when applying the echo signal selection algorithm based on scattering points, it is necessary to simulate each scattering point separately, process the radio frequency echo data generated by each scattering point using the echo signal selection algorithm based on scattering points, and then superimpose the radio frequency data of all scattering points.

[0094] Consider an N-channel transducer, with both the emitter and receiver apertures encompassing all N channels; the number of scan lines is L, and the length of the echo signal vector recorded by each receiver channel is P; since there are a total of N emitter channels, for a given scattering point S, for each receiver channel along each scan line, a total of N signal samples are selected, such as... Figure 6 As shown; after selecting the correlated echo signal of scattering point S using the echo signal selection algorithm based on scattering point, the resulting RF data matrix M is of size P×L×N. SESAIn this context, for each scan line 1∈{1,2,…,L} and each receiving channel n∈{1,2,…,N}, only the selected N signal samples are not zero.

[0095] After selecting the radio frequency echo data using the echo signal selection algorithm based on scattering points, the delayed superposition beamforming algorithm can be used for imaging. However, for multi-scattering point imaging scenarios (such as cyst simulation imaging), since the imaging area contains a large number of scattering points, it is necessary to simulate each scattering point separately. The echo signal selection algorithm based on scattering points is applied to process the radio frequency echo data generated by each scattering point. Then, the selected radio frequency echo data of all scattering points are added together, and finally, the delayed superposition beamforming algorithm is used for imaging.

[0096] However, in order to save computing resources, after processing the radio frequency echo data generated by each scattering point using the echo signal selection algorithm based on scattering points, a delay operation is first applied to the radio frequency echo data to greatly reduce the data size. Then, the selected and delayed radio frequency echo data of all scattering points are added together, and finally the radio frequency echo data of all receiving channels are superimposed to complete the delayed superimposed beamforming.

[0097] like Figure 7 As shown, after processing the radio frequency data of all scattering points using the echo signal selection algorithm based on scattering points, the data is first delayed to reduce the size of the radio frequency data from P×N×L to C×N×L, where C represents the number of vertical pixels, and C is much smaller than P. Then, all the delayed radio frequency data are added together to obtain the complete delayed radio frequency data of the complex scene with multiple scattering points. Finally, the channel dimensions are superimposed to obtain the delayed superimposed radio frequency data of size C×L.

[0098] In real-world ultrasound imaging scenarios, the locations of scattering points in the imaging area are unknown beforehand. Therefore, echo signal selection algorithms based on scattering points cannot be directly applied to real-world imaging scenarios. In simulated ultrasound imaging scenarios, the tissue structure and scattering point locations of the imaging area are known, allowing for the effective selection of echo signals generated by scattering points, thereby suppressing interference signals. Previous literature has shown that neural networks trained using only single-scattering-point radio frequency data can generalize to simulated cyst imaging and real-world imaging. Moreover, neural networks can approximate any mathematical function. Therefore, this invention proposes a deep neural network to implement an echo signal selection algorithm based on scattering points, making it applicable to real-world imaging scenarios. The question then becomes: how to learn single-point radio frequency data M that has not been processed by an echo signal selection algorithm based on scattering points? original And single-point radio frequency data M processed by the echo signal selection algorithm based on scattering points SESA The transformation relationship between them is a regression task, therefore it uses a large number of {M} original MSESA Deep neural networks trained on datasets composed of tuples are a suitable tool for generalizing echo signal selection algorithms based on scattering points to real-world imaging.

[0099] Due to the large volume of radio frequency (RF) data, a delay operation is required to reduce GPU memory usage. Therefore, an interference signal suppression neural network, which has learned the echo signal selection process based on scattering points, is integrated into the delay-superimposed beamforming framework. Figure 1 As shown.

[0100] The training process for an interference signal suppression neural network is as follows: using Field The simulation software II generates ultrasound radio frequency data from a single scattering point. An echo signal selection algorithm based on the scattering point is used to process the radio frequency data to obtain the training ground truth. A single scattering point's radio frequency data is considered a sample. The training and validation sets each contain multiple samples, with a sample size ratio of 10:1. The training set consists of 12,000 samples, and the validation set consists of 1,200 samples. The deep neural network is trained using the Adam optimizer. The Adam optimizer uses both the mean and variance of the gradient to calculate the update step size of the network weight parameters. Training data is input into the deep neural network in batches, and the Adam optimizer updates the parameters of the deep neural network. Processing all training data constitutes one training cycle. After each training cycle, the validation error is calculated using the validation set data, and the network weight parameters corresponding to the minimum validation error are saved. Experiments show that the validation error decreases slowly after 33 training cycles. Therefore, after 33 cycles, the network weights for each round are saved, and the optimal network weights are determined during testing. The training process is time-consuming, but once training is complete, the deep neural network can process data quickly.

[0101] S3. The delayed radio frequency data obtained in step S2 after suppressing the interference signal is recombined into complex radio frequency data, and then superimposed to complete beamforming, obtain the image pixel value of the corresponding ultrasound imaging target, and thus form a complete ultrasound image. This ultrasound image has the advantages of high contrast and high contrast-to-noise ratio.

[0102] The third delayed radio frequency data M3 obtained in S2 after suppressing the interference signal is recombined into a complex delayed radio frequency data matrix M4 of size P×N×L. The complex delayed radio frequency data matrix M4 has the same size as the delayed radio frequency data matrix M1 before the deep neural network processing. Then, the complex delayed radio frequency data matrix M4 is superimposed to obtain the pixel values ​​of all target imaging points.

[0103] In one embodiment, in an ultrasound imaging system, the same radio frequency data is calculated using a traditional time-delay superposition imaging method, an echo signal selection algorithm based on scattering points, and a deep learning-based ultrasound radio frequency data filtering imaging method, and the imaging effects are compared.

[0104] In the single-point target simulation experiment, the only scattering point was located at a focusing depth of 15mm. After obtaining the radio frequency data by calling the data acquisition module, the core calculation module was called to perform image calculations using the traditional time-delay superposition imaging method, the echo signal selection algorithm based on the scattering point, and the ultrasound radio frequency data filtering imaging method based on deep learning. The calculated image is shown below. Figure 8 As shown, within a dynamic range of 60 dB, both the echo signal selection algorithm based on scattering points and the ultrasound radio frequency data filtering imaging method based on deep learning can effectively suppress sidelobe artifacts. The echo signal selection algorithm based on scattering points suppresses the most sidelobe artifacts, while the ultrasound radio frequency data filtering imaging method based on deep learning still retains some sidelobe artifacts.

[0105] After simulating imaging a single scattering point, the normalized envelope curve of the row containing the scattering point is plotted, called the point spread function. The peak in the middle of the curve is the main lobe, and the width of the peak can be used to measure the width of the main lobe. The smaller the main lobe width, the higher the resolution. The sides of the main lobe are the side lobes. A lower side lobe level indicates better suppression of the side lobes, fewer artifacts on both sides of the scattering point, and higher contrast between the cyst image and the real human image. Simultaneously, the distance between positions -6 dB to the left and right of the scattering point (where the lateral distance is 0) is calculated, called the full width at half maximum (FWHM). This value serves as a measure of lateral resolution; the smaller the value, the higher the lateral resolution.

[0106] The point spread function (PSF) results from the single-point target simulation experiment show that the echo signal selection algorithm based on scattering points reduces sidelobes by 10 to 20 dB, while the overall PSF of the deep learning-based ultrasound radio frequency data filtering imaging method falls between that of time-delay stacking and the echo signal selection algorithm based on scattering points. Table 1 shows the resolution performance of traditional time-delay stacking, the echo signal selection algorithm based on scattering points, and the deep learning-based ultrasound radio frequency data filtering imaging method. Numerically, the echo signal selection algorithm based on scattering points improves resolution compared to time-delay stacking, while the resolution of the deep learning-based ultrasound radio frequency data filtering imaging method is basically the same as that of time-delay stacking.

[0107] Table 1. Comparison of full width at half maximum (FWHM) values ​​for traditional time-delay stacking, echo signal selection based on scattering points, and ultrasound radio frequency data filtering imaging methods based on deep learning.

[0108]

[0109] In the cyst imaging simulation experiment, the cyst's center was located at a focal depth of 15 mm, and its radius was 2.5 mm. The scattering point density within the imaging region was 100 scattering points per cubic millimeter. The amplitude of the scattering points inside the cyst region was set to zero. After obtaining radio frequency data from the data acquisition module, the core calculation module was invoked to perform image calculations using a traditional time-delay superposition imaging method, a scattering point-based echo signal selection algorithm, and a deep learning-based ultrasound radio frequency data filtering imaging method. The resulting image is shown below. Figure 9 As shown, compared with time-delay superposition, the cyst region in the cyst image obtained by the echo signal selection algorithm based on scattering points and the ultrasound radio frequency data filtering imaging method based on deep learning is deeper, and the speckle image is not damaged. This indicates that the deep neural network effectively suppresses the interference signal while preserving the target signal.

[0110] Contrast, contrast-to-noise ratio, and speckle signal-to-noise ratio are used to measure the quality of cyst images. When calculating contrast and contrast-to-noise ratio, the region of interest (ROI) of the cyst area is the area enclosed by a circle with a radius of 0.8 times that of the cyst, and the ROI of the background area is the area enclosed by a circle with a radius of 0.8 times that of the cyst on the left and right sides. When calculating speckle signal-to-noise ratio, the ROI of the speckle area is the area enclosed by a circle with a radius of 0.8 times that of the cyst on the left and right sides.

[0111] The formula for calculating contrast ratio (CR) is as follows:

[0112]

[0113] Where μ cyst This represents the mean uncompressed envelope within the cyst region, in μ. background This represents the mean of the uncompressed envelope within the background region.

[0114] The formula for calculating the contrast-to-noise ratio (CNR) is as follows:

[0115]

[0116] Where σ background σ represents the standard deviation of the uncompressed envelope within the background region. cyst This represents the standard deviation of the uncompressed envelope within the cyst area.

[0117] The formula for calculating the speckle signal-to-noise ratio (SNR) is as follows:

[0118]

[0119] Where μspeckle σ represents the mean of the uncompressed envelope within the speckle region. speckle This represents the standard deviation of the uncompressed envelope within the speckle area.

[0120] In one embodiment, Table 2 shows a comparison of cyst evaluation metrics for traditional time-delay stacking, echo signal selection based on scattering points, and ultrasound radiofrequency data filtering imaging methods based on deep learning. Table 2 shows that the echo signal selection based on scattering points effectively improves CR and CNR, while the speckle SNR remains essentially unchanged. The ultrasound radiofrequency data filtering imaging method based on deep learning also effectively improves CR and CNR, but the speckle SNR is slightly lower than that of the time-delay stacking imaging method. Nevertheless, in the cyst images generated by the ultrasound radiofrequency data filtering imaging method based on deep learning, the speckle structure is not destroyed, and there are no visual deficiencies.

[0121] Table 2. Cyst evaluation indicators of traditional time-delay superposition, echo signal selection algorithm based on scattering points, and ultrasound radio frequency data filtering imaging method based on deep learning.

[0122]

[0123] In one embodiment, three different focal depths (15mm, 20mm, and 25mm) were used to image the thyroid gland in human thyroid imaging, while the training data for the neural network was generated with a focal depth of 15mm. Figure 10 Images of the thyroid gland obtained using three different time-lapse stacking imaging methods at varying focal depths, as well as images obtained using a deep learning-based ultrasound radiofrequency data filtering imaging method, are shown. Since the echo signal selection algorithm based on scattering points cannot be directly applied to human imaging, corresponding imaging results are not available. Figure 10 As shown in Figures a and b, at a focusing depth of 15mm, the deep learning-based ultrasound radiofrequency data filtering imaging method visually improves the quality of the thyroid image. In images created using the traditional time-lapse overlay imaging method, the thyroid gland is obscured by a layer of blur, resulting in a relatively dark overall image. By processing the radiofrequency data using the deep learning-based ultrasound radiofrequency data filtering imaging method and suppressing interference signals, the thyroid gland becomes clearer and brighter, with significant noise suppression, making it easily identifiable. Furthermore, the cross-section of the left carotid artery of the thyroid gland is also clearer. Since the depth range of the thyroid image is 5mm to 30mm, which is the same as the depth coverage range of the training dataset, therefore… Figure 10 The thyroid images in Figures a and b, obtained using a deep learning-based ultrasound radiofrequency data filtering imaging method, show improved image quality across the entire depth range. Figure 10 Images c and d in the image compare images of the thyroid gland obtained using the traditional time-lapse imaging method and the deep learning-based ultrasound radiofrequency data filtering imaging method at a focusing depth of 20 mm. Figure 10 Figures e and f in the image compare images of the thyroid gland obtained using the traditional time-lapse imaging method and the deep learning-based ultrasound radiofrequency data filtering imaging method at a focal depth of 25 mm. At both focal depths, the deep learning-based ultrasound radiofrequency data filtering imaging method also effectively suppresses noise and improves image brightness and contrast. This demonstrates that the deep learning-based ultrasound radiofrequency data filtering imaging method acquires the ability to suppress interference signals through training, and the learned interference signal patterns are independent of the focal depth value, indicating the generalization ability of the deep learning-based ultrasound radiofrequency data filtering imaging method.

[0124] The data above shows that, compared with the traditional time-delay superposition imaging method, the deep learning-based ultrasound radio frequency data filtering imaging method in this invention improves the contrast and contrast-to-noise ratio of the image without affecting the image resolution, and obtains high-quality ultrasound images of the human thyroid gland.

[0125] This embodiment describes the design and evaluation of a deep learning-based ultrasound radio frequency data filtering imaging method to improve ultrasound image quality in simulated ultrasound imaging and human thyroid imaging. Experimental results show that the deep learning-based ultrasound radio frequency data filtering imaging method improves the contrast and contrast-to-noise ratio of ultrasound images compared to the traditional time-delay stacking imaging method, and obtains high-quality human thyroid images.

[0126] Therefore, compared with the deep learning ultrasound imaging method proposed by Yoon et al. in the paper "Yoon YH, Khan S, Huh J, et al. Efficient B-mode ultrasound image reconstruction from sub-sampled RF data using deep learning[J].IEEE Transactions on Medical Imaging, 2018, 38(2): 325-336", which uses downsampled ultrasound echo radio frequency data as input data, the method proposed in this invention uses the original ultrasound echo radio frequency data as input data, thus preserving the rich signal features in the original data. Compared to the adaptive minimum variance beamforming algorithm based on deep learning proposed by Luijten et al. in the paper "Luijten B, Cohen R, de Bruijn FJ, et al. Adaptive ultrasound beamforming using deep learning[J]. IEEE Transactions on Medical Imaging, 2020, 39(12): 3967-3978", the deep neural network model designed in this invention takes the proposed echo signal selection algorithm based on scattering points as the learning object. Its purpose is to select each echo signal from the scattering point location, thereby suppressing the echo signal from the interference location. It can be flexibly combined with any beamforming imaging framework.

Claims

1. A deep learning-based ultrasonic radio frequency data filtering imaging method, characterized in that, Includes the following steps: S1. The ultrasonic imaging target object is scanned. The echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. The received real radio frequency data is converted into complex radio frequency data. After performing corresponding delay operations for different imaging points, delayed radio frequency data is obtained. The ultrasonic imaging target object is scanned. The echo signal received by the ultrasonic transducer receiving aperture constitutes radio frequency data. Then, Hilbert transform is used to convert the real radio frequency data into complex radio frequency data. Next, a delay operation is performed on the radio frequency data. That is, the delay time is calculated based on the position of the target imaging point, the position of the scan line, and the position of the receiving array element. The delay time is mapped to a signal subscript, thereby extracting the signal value corresponding to the target imaging point in the echo signal of the receiving array element, and obtaining the delayed radio frequency data. Let the number of scan lines be L, the number of target imaging points on a scan line be P, and the number of receiving array elements be N. After the delay operation, a first delayed radio frequency data matrix M1 with a size of P×N×L is obtained. This matrix is ​​a complex matrix. Then, the real part and imaginary part of each data in the first delayed radio frequency data matrix M1 are separated to obtain a second delayed radio frequency data matrix M2 with a size of 2×P×N×L. S2. Separate the real and imaginary parts of the complex radio frequency data, and use an interference signal suppression neural network to suppress the interference signal in the delayed radio frequency data obtained in step S1, so as to obtain the delayed radio frequency data after suppressing the interference signal. The second delayed radio frequency data matrix M2 is input into the interference signal suppression neural network for processing to obtain the third delayed radio frequency data matrix M3 with a size of 2×P×N×L after the interference signal is suppressed. The interference signal suppression neural network is a three-dimensional convolutional neural network with an encoder-decoder structure. This neural network learns an echo signal selection algorithm based on scattering points. In the interference signal suppression neural network, since the radio frequency data is complex, the real part and the imaginary part are separated and each becomes a channel. Therefore, the number of input channels of the network is 2. Due to limitations in hardware computing resources and the large amount of input data, the initial number of channels in the intermediate feature map of the interference signal suppression neural network is set to 4. The interference signal suppression neural network includes a compression path and an expansion path. The compression path on the left increases the number of channels while reducing the size of the feature map, and the expansion path on the right reduces the number of channels while restoring the size of the feature map. The compression path consists of five convolutional blocks, which are connected by convolutional layers responsible for downsampling; the expansion path consists of four convolutional blocks, which are connected by convolutional layers responsible for upsampling. Each convolutional block in the compression and expansion paths includes residual connections. The input and output of a convolutional block are superimposed before being passed to the next convolutional layer. After the first convolutional block, the feature map has 4 channels. Therefore, before the input of the interference suppression neural network is superimposed with the output of the first convolutional block, it needs to copy itself to expand the number of channels to 4. The input and output channels of other convolutional blocks are the same, so they can be directly superimposed. Downsampling is handled by a convolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which doubles the number of channels. Upsampling is handled by a deconvolutional layer with a kernel size of (2, 2, 2) and a stride of 2, which halves the number of channels. The output layer of the interference signal suppression neural network is a convolutional layer with a kernel size of (1, 1, 1) and a stride of 1, reducing the number of output channels to 2, equal to the number of input channels. The remaining convolutional kernels in the interference signal suppression neural network are all of size (5, 5, 5) with a stride of 1. The activation function used in the interference signal suppression neural network is the hyperbolic tangent function (Tanh). The loss function used is the MSE loss function with L1 regularization, calculated as follows: Where X represents a sample, Y represents the true value, f represents a neural network, W(f) represents the set of weights of the neural network, and lambda represents the L1 regularization coefficient. The interference signal suppression neural network learns an echo signal selection algorithm based on scattering points for ultrasound simulation imaging. This algorithm selects the target echo signal based on the spatial location of the scattering points to isolate a large number of signals from interfering locations in space. The specific process is as follows: In single-scattering point simulation imaging, a scattering point S is simulated within the imaging domain, and a transmitting channel t and a receiving channel r are considered, along with the RF echo signal vector received by the receiving channel r. Based on the positions of the transmitting channel t, the receiving channel r, and the scattering point S, calculate the length of the signal transmission path from the transmitting channel t to the scattering point S, and then from the scattering point S to the receiving channel r. Calculate using the following formula: Where N represents the total number of channels in the transducer. This represents the length of the signal transmission path from the transmission channel t to the scattering point S. This represents the length of the signal transmission path from the scattering point S to the receiving channel r; the transmitting channel t and the receiving channel r constitute a transmission channel pair; Next, the RF echo signal vector of this transmission path is calculated. element subscript : Where round is the floor function and c is the speed of sound. The signal sampling frequency; In this way, the echo signal from each scattering point is selected, thereby suppressing the echo signal from the interference point. The center points of the transmitting channel t and the receiving channel r are defined as F1 and F2, respectively. With F1 and F2 as the two foci, the trajectory of a moving point in the plane whose sum of distances to F1 and F2 is a constant (greater than |F1F2|) forms an ellipse. Therefore, the path length of the echo signal transmission from the scattering point located on the elliptical arc with F1 and F2 as foci is... They are all the same; located at any point on the elliptical arc with foci F1 and F2. and same. S3. The delayed radio frequency data obtained in step S2 after suppressing the interference signal is recombined into complex radio frequency data, and then superimposed to complete beamforming, thereby obtaining the image pixel value of the corresponding ultrasound imaging target and thus forming a complete ultrasound image. This ultrasound image has the advantages of high contrast and high contrast-to-noise ratio.

2. The ultrasound radio frequency data filtering imaging method based on deep learning according to claim 1, characterized in that, The rounding function in formula (3) means that multiple different path lengths will be mapped to the same vector element index. Therefore, the elliptical arc where the scattering point is located is actually an elliptical ring. The signal transmission path calculated at any position in the elliptical ring will be mapped to the same element index. Therefore, according to the subscript While selecting echo signals from the scattering point location, a portion of echo signals from the interference region on the elliptical ring were also selected; that is, the subscript of the selected element... The echo signal sample is a composite signal, which includes not only the echo signal from the scattering point, but also the echo signal from the interference region on the elliptical ring. For each of the remaining transmission channel pairs, there also exists an elliptical ring corresponding to the same element index. The echo signal transmission path lengths for the two transmission channels (t1,r) and (t2,r) are equal, that is... This indicates that the corresponding elliptical ring interference regions are different for ultrasonic signals emitted from different channels; interference signals from all these interference regions are preserved along with signals from the scattering point location, while signals from other interference regions are excluded. For imaging scenarios involving multiple scattering points, the signals generated by different scattering points interfere with each other. In simulated imaging with multiple scattering points, scattering points SA, SB, and SC coexist. When calculating the echo signal from the transmitting channel t to the receiving channel r, in addition to the spatial interference signal, i.e., the elliptical ring corresponding to scattering point SA, there are also interference signals generated by scattering points SB and SC. Therefore, when applying the echo signal selection algorithm based on scattering points, it is necessary to simulate each scattering point separately, process the RF echo data generated by each scattering point using the echo signal selection algorithm based on scattering points, and then superimpose the RF data of all scattering points.

3. The ultrasound radio frequency data filtering and imaging method based on deep learning according to claim 2, characterized in that, Consider an N-channel transducer, with transmit and receive apertures encompassing all N channels; the number of scan lines is L, and the length of the echo signal vector recorded by each receive channel is P; since there are a total of N transmit channels, for a given scattering point S, for each receive channel of each scan line, a total of N signal samples are selected; after selecting the relevant echo signals of scattering point S using a scattering point-based echo signal selection algorithm, a radio frequency data matrix of size P×L×N is generated. In this context, for each scan line l∈{1,2,…,L} and each receiving channel n∈{1,2,…,N}, only the selected N signal samples are not zero.

4. The ultrasound radio frequency data filtering and imaging method based on deep learning according to claim 2, characterized in that, The training process of the interference signal suppression neural network is as follows: use Field II simulation software to simulate and generate ultrasonic radio frequency data of a single scattering point, use the echo signal selection algorithm based on the scattering point to process the radio frequency data, and obtain the training ground truth. The radio frequency data of a single scattering point constitutes one sample. The training set and validation set each contain multiple samples, with a sample size ratio of 10:

1. The deep neural network is trained using the Adam optimizer, which combines the mean and variance of the gradient to calculate the update step size of the network weight parameters. The training set data is input into the deep neural network in batches, and the Adam optimizer updates the parameters of the deep neural network. Processing all the training set data constitutes one training cycle. After each training cycle, the validation error is calculated using the validation set data, and the network weight parameters corresponding to the minimum validation error are saved. When the training cycle exceeds a set threshold, the network weights of each round are saved, and the optimal network weights are determined during testing.

5. The ultrasound radio frequency data filtering imaging method based on deep learning according to claim 1, characterized in that, In step S3, the third delayed radio frequency data M3 obtained in S2 after suppressing the interference signal is recombined into a complex delayed radio frequency data matrix M4 of size P×N×L. The complex delayed radio frequency data matrix M4 has the same size as the delayed radio frequency data matrix M1 before the deep neural network processing. Then, the complex delayed radio frequency data matrix M4 is superimposed to obtain the pixel values ​​of all target imaging points.

Images