An ultrasonic signal disturbance suppression method based on covariance whitening method

Abstract

The application is an ultrasonic signal disturbance suppression method based on covariance whitening method. The application relates to the technical field of medical ultrasonic sensor signal disturbance suppression, and establishes a second-order statistical model of array elements in the channel by using original RF data collected in multiple probe states, and constructs a whitening transformation or a reference domain covariance alignment transformation in the channel dimension, so that statistical changes related to probe disturbance are normalized and suppressed. The application normalizes and aligns the second-order statistical differences caused by the probe disturbance in the original RF channel space, so as to reduce the sensitivity of subsequent signal analysis, feature extraction, classification model or regression model to non-target acquisition state factors. The corrected data still maintains the original two-dimensional RF matrix form, and when used for traditional beam forming or image reconstruction, diagonal whitening, local strip approximation, scale recovery or double branch processing mode can be adopted to give consideration to statistical stability and array element physical consistency.

Description

Technical Field

[0001] This invention relates to the field of medical ultrasound sensor signal perturbation suppression technology, and is a method for suppressing ultrasound signal perturbation based on covariance whitening. Background Technology

[0002] One-dimensional linear array ultrasound probes typically consist of multiple array elements arranged linearly. In a single acquisition, each element receives echo signals from within the tissue being measured, forming multi-channel raw RF data. For a single frame of acquisition, the data can be represented as a two-dimensional matrix, with one dimension corresponding to the array element channels and the other to the depth sampling points. This raw RF matrix preserves the amplitude, phase, correlation, and depth echo distribution information between the array element channels, serving as a crucial foundational data format for subsequent beamforming, ultrasound image reconstruction, signal analysis, and RF signal-based learning models. Under ideal acquisition conditions, the raw RF signals obtained when the same object is in the same state should have a relatively stable data distribution. However, in practical use, the relative position and contact state between the probe and the skin or tissue being measured are difficult to maintain perfectly. The probe may experience lateral movement, longitudinal movement, forward / backward offset, pitch rotation, axial rotation, changes in pressure depth, or insufficient local coupling. Even if the state of the object being measured itself does not change significantly, the aforementioned changes in probe state can still cause statistical perturbations in the raw RF signal. These perturbations typically manifest as changes in the average amplitude and variance of different array element channels, alterations in the correlation structure between adjacent or near-neighbor channels, and overall data distribution drift between different acquisition batches. For algorithms that directly use raw RF data for classification, regression, attitude estimation, or other learning tasks, the model may simultaneously learn the effective features of the object being measured and non-target features introduced by probe position, pressing state, coupling state, or acquisition batches. When the probe state is inconsistent between the training and testing phases, these non-target factors reduce the model's generalization ability and robustness. Especially in applications involving long-term repeated acquisition or acquisition across multiple days, slight probe movement or changes in contact conditions are often unavoidable. Therefore, it is necessary to suppress channel statistical changes caused by probe perturbations before the raw RF signal enters the downstream algorithm. Common correction methods in existing ultrasound signal processing include time gain compensation, channel amplitude normalization, filtering, B-mode image grayscale normalization, and image enhancement. While these methods can improve display quality or amplitude range to some extent, they still have limitations: Firstly, many methods process B-mode or envelope images, which have already undergone beamforming, envelope detection, and logarithmic compression, partially altering or losing the joint statistical relationships between the original array element channels, making it difficult to suppress the impact of probe disturbances on the original RF channel structure from the data source. Secondly, simple channel-by-channel normalization usually only considers the mean, variance, or amplitude range of a single channel, and cannot fully describe the joint variation relationships between multiple array element channels.Publicly available information also reveals schemes for ultrasound imaging based on RF data, minimum variance beamforming based on the covariance matrix, probe motion estimation based on images or markers, and compensation for amplitude variations caused by coupling inhomogeneities. These schemes primarily focus on image quality, motion estimation, echo amplitude compensation, or beamforming weight calculation. They do not establish offline-calibrated second-order statistical models for the original two-dimensional RF channel matrix under various probe perturbation states, nor do they unify channel covariance whitening, reference domain covariance alignment, online state matching, and constrained whitening for suppressing probe perturbation-related statistical distribution drift. Therefore, a probe perturbation statistical suppression method that directly acts on the original RF matrix, characterizes the second-order statistical relationships between channels, and can be rapidly applied online is needed. Summary of the Invention

[0003] This invention does not reconstruct RF data into B-mode images before performing grayscale correction, nor is it limited to performing amplitude normalization independently for each channel. Instead, it utilizes the array element channel covariance structure of the original RF data under various probe disturbance states to establish a second-order statistical model for the channels. Through whitening or reference domain covariance alignment transformation, it suppresses channel scale differences, inter-channel redundancy correlation differences, and multi-state acquisition distribution drift caused by probe disturbance. This invention discloses an ultrasound signal disturbance suppression method based on covariance whitening.

[0004] This invention provides the following technical solutions: A method for suppressing ultrasonic signal perturbation based on covariance whitening, the method comprising the following steps: Step 1: Collect raw RF calibration data covering the perturbation states of various ultrasonic sensor probes, and decompose each frame of the RF matrix into C-dimensional channel vector samples along the depth direction; Step 2: Estimate the global channel mean and global channel covariance matrix based on all channel vector samples; perform regularization on the covariance matrix to obtain the regularized covariance matrix, and construct the global symmetric whitening matrix through eigenvalue decomposition. Step 3: Acquire the original RF matrix of the current frame X t Perform basic preprocessing consistent with the offline phase; center the preprocessed RF matrix using the global channel mean, and multiply it by the global whitening matrix in the element channel dimension to obtain the corrected RF matrix. .

[0005] Preferably, when collecting calibration data, the object under test is kept in a stable state so that the difference in calibration data mainly comes from changes in probe state rather than changes in the state of the object under test itself; the calibration data covers all typical probe disturbance states in practical applications, and state labels are added to the data of different probe states.

[0006] Preferably, the number of array element channels of the one-dimensional linear array ultrasonic probe is set to... C The number of depth sampling points for each channel is D Any frame of raw RF data is represented as:

[0007] in, X The c The line represents the first c RF signals of each array element channel, X The d The column indicates the first d The channel vector formed by all channels at each depth sampling point; Record No. d The channel vector corresponding to each depth sampling point is:

[0008] A frame of RF matrix is ​​considered as being composed of D indivual C Composition of 3D channel vectors:

[0010] Preferably, multiple probe disturbance states, multiple acquisition frames, and channel vectors at multiple depth positions are used as statistical samples, with the total number of samples set to be... M Then the global channel mean and global channel covariance can be written as:

[0012] Preferably, to improve numerical stability, a regularization form related to the covariance scaling is adopted:

[0013] in, I C for C × C The identity matrix. This regularization term avoids numerical amplification caused by small eigenvalues ​​and makes the regularization strength adaptively change with the overall energy scale of the RF data. Eigenvalue decomposition is performed on the regularization covariance matrix, and a symmetric whitening matrix is ​​constructed:

[0014] During online processing, the current input RF matrix is ​​processed. X t By centering the channels and left-multiplying by the global whitening matrix along the channel dimension, we obtain:

[0015] Among them, 1 D Let D be a vector of length D consisting entirely of 1s. It is important to emphasize that this whitening matrix strictly whitens the regularized mixture covariance matrix, i.e.:

[0016] For the original covariance matrix without a regularization term, when the regularization term is small and the covariance estimate is sufficient, we have:

[0018] Preferably, for the first s For each probe disturbance state, estimate the mean and covariance under that state:

[0019] The standard fit state or the multi-state average state is selected as the reference domain, and its mean and covariance are denoted as . μ r With Σ r In the case of Σ s and Σ r After performing similar regularization, the square root of a symmetric positive definite matrix is ​​used to construct the state. s Covariance alignment matrix to the reference domain:

[0020] For the current state is s The input RF matrix, the correction result is:

[0021] When the current state is correctly matched, the square root of the matrix is ​​taken as the principal symmetric square root, and the small deviation introduced by the regularization term is ignored, this transformation can make the corrected channel vector satisfy:

[0023] Preferably, when the current probe status is unknown during online data acquisition, the current mean is estimated using a short period of RF data. and current covariance The data is then matched with an offline perturbation state template. To avoid interference from the amplitude scale on the correlation comparison, the covariance matrix is ​​converted into a normalized correlation matrix, and the following distance metric is used:

[0024] in, α >0 and β >0 is used to balance the difference in mean and the difference in correlation structure, respectively.

[0025] Preferably, the state with the smallest distance is selected and the corresponding alignment matrix is ​​used, or soft weights are used to fuse the correction results of multiple states:

[0026]

[0028] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement an ultrasonic signal perturbation suppression method based on covariance whitening.

[0029] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement an ultrasonic signal perturbation suppression method based on covariance whitening.

[0030] The present invention has the following beneficial effects: This invention does not rely on B-mode images, envelope images, log-compressed images, or other image domain post-processing results. Addressing the channel energy differences, inter-channel correlation changes, and overall statistical distribution drift caused by lateral translation, longitudinal translation, pitch rotation, axial rotation, pressure depth changes, and coupling state changes relative to the measured tissue during actual acquisition, this invention utilizes raw RF data acquired under various probe states to establish a second-order statistical model of the array element channels. A whitening transformation or reference domain covariance alignment transformation is constructed in the channel dimension to normalize and suppress statistical changes related to probe perturbation. In the basic implementation, the RF matrix under various probe perturbation states is decomposed into channel vector samples along the depth direction. The global channel mean and global channel covariance matrix are estimated offline, and a global whitening matrix is ​​constructed through regularized eigenvalue decomposition. During online processing, the current input RF matrix is ​​channel-centered, and the whitening matrix is ​​left-multiplied in the array element channel dimension, making the mixed calibration distribution more normalized in a second-order statistical sense. In the enhanced implementation, the state mean and state covariance are estimated for different probe disturbance states respectively. The standard fitting state or the multi-state average state is used as the reference domain to construct a covariance alignment matrix from the disturbance state to the reference domain, so that the RF channel distribution under the current disturbance state is aligned with the reference domain in terms of mean and covariance.

[0031] The disturbance suppression described in this invention does not represent the restoration of the complete physical sound field when the probe has not shifted, nor does it represent the elimination of all tissue cross-sectional differences, scatterer position differences, or sound path changes caused by probe position changes. Its technical effect is limited to normalizing and aligning the second-order statistical differences caused by probe disturbances in the original RF channel space, thereby reducing the sensitivity of subsequent signal analysis, feature extraction, classification models, or regression models to non-target acquisition state factors. The corrected data still retains the original two-dimensional RF matrix form, facilitating direct integration into subsequent algorithms. When used for traditional beamforming or image reconstruction, diagonal whitening, local strip approximation, scale restoration, or bi-branch processing methods can be employed to balance statistical stability and element physical consistency. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 The flowchart shown is that of the present invention; Figure 2 The flowchart shown is a specific embodiment of the present invention. Figure 3 The original RF matrix is ​​displayed under different probe offset states; Figure 4 The display shows the channel correlation matrix under different probe offset states before correction; Figure 5 Displayed as a comparison of statistical distances relative to the correct position before and after correction; Figure 6 Displayed as the channel correlation matrix after reference domain covariance alignment. Detailed Implementation

[0034] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] The present invention will be described in detail below with reference to specific embodiments. Specific Implementation Example 1: according to Figures 1 to 6As shown, the specific optimized technical solution adopted by the present invention to solve the above-mentioned technical problems is: The present invention relates to an ultrasonic signal perturbation suppression method based on covariance whitening method.

[0037] This invention provides a method for suppressing ultrasonic signal perturbations based on covariance whitening. The overall processing approach of the invention is as follows: Figure 1 As shown. This method does not reconstruct the RF data into a B-mode image before performing grayscale correction, nor is it limited to performing amplitude normalization independently for each channel. Instead, it utilizes the array element channel covariance structure of the original RF data under various probe disturbance states to establish a second-order statistical model for the channels. Through whitening or reference domain covariance alignment transformation, it suppresses channel scale differences, inter-channel redundancy correlation differences, and multi-state acquisition distribution drift caused by probe disturbance.

[0038] Let the number of array element channels of a one-dimensional linear ultrasound probe be... C The number of depth sampling points for each channel is D Any frame of raw RF data is represented as:

[0039] in, X The c The line represents the first c RF signals of each array element channel, X The d The column indicates the first d The channel vector formed by all channels at the nth depth sampling point. Let the nth depth sampling point be... d The channel vector corresponding to each depth sampling point is:

[0040] A frame of RF matrix can be viewed as composed of D indivual C Composition of 3D channel vectors:

[0041] This invention uses channel vectors from multiple probe disturbance states, multiple acquisition frames, and multiple depth locations as statistical samples. Let the total number of samples be... M Then the global channel mean and global channel covariance can be written as:

[0042] The diagonal elements of this covariance matrix reflect the variance of each array element channel individually, while the off-diagonal elements reflect the covariance between different array element channels. Compared to simply performing channel-by-channel amplitude normalization, this matrix can describe the joint variation relationship between array element channels, and is therefore more suitable for characterizing channel statistical structure changes caused by probe perturbations. To improve numerical stability, this invention preferably employs a regularization form related to the covariance scale:

[0043] in, I C for C × C The identity matrix. This regularization term avoids numerical amplification caused by small eigenvalues ​​and makes the regularization strength adaptively change with the overall energy scale of the RF data. Eigenvalue decomposition is performed on the regularization covariance matrix, and a symmetric whitening matrix is ​​constructed:

[0044] During online processing, the current input RF matrix is ​​processed. X t By centering the channels and left-multiplying by the global whitening matrix along the channel dimension, we obtain:

[0045] Among them, 1 D Let D be a vector of length D consisting entirely of 1s. It is important to emphasize that this whitening matrix strictly whitens the regularized mixture covariance matrix, i.e.:

[0046] For the original covariance matrix without a regularization term, when the regularization term is small and the covariance estimate is sufficient, we have:

[0047] Therefore, the reasonable technical meaning of the basic implementation is to perform channel second-order statistical normalization on the multi-perturbation hybrid calibration distribution, rather than to perform strict physical state recovery for each specific probe perturbation state.

[0048] To further enhance the statistical consistency between different probe states, this invention also provides a reference domain covariance alignment method. For the... s For each probe disturbance state, estimate the mean and covariance under that state:

[0049] The standard fit state or the multi-state average state is selected as the reference domain, and its mean and covariance are denoted as . μ r With Σ rIn the case of Σ s and Σ r After performing similar regularization, the square root of a symmetric positive definite matrix is ​​used to construct the state. s Covariance alignment matrix to the reference domain:

[0050] For the current state is s The input RF matrix, the correction result is:

[0051] When the current state is correctly matched, the square root of the matrix is ​​taken as the principal symmetric square root, and the small deviation introduced by the regularization term is ignored, this transformation can make the corrected channel vector satisfy:

[0052] More precisely, this transformation can regularize the state covariance Σ s,ε Align to regularized reference covariance Σ r,ε When the regularization term is small and the estimate is sufficient, the alignment of the original state covariance to the reference covariance can be approximately achieved. Therefore, the enhanced implementation has a more explicit meaning of state-conditional second-order statistical alignment compared to single global whitening.

[0053] When the current probe status is unknown during online data acquisition, the current mean can be estimated using a short period of RF data. and current covariance This is then matched against an offline perturbation state template. To avoid interference from the amplitude scale on the correlation comparison, the covariance matrix can be converted into a normalized correlation matrix, and the following distance metric can be used:

[0054] in, α >0 and β >0 is used to balance mean differences and correlation structure differences. The state with the smallest distance can be selected and its corresponding alignment matrix applied, or soft weights can be used to fuse the correction results of multiple states.

[0055] Soft weighted fusion can reduce discontinuities caused by hard switching when probe states change continuously, but its strict second-order covariance alignment property is weaker than that of single-state matching. Therefore, this invention uses hard matching as a strict state alignment implementation method and soft fusion as a smooth online processing method in continuous perturbation scenarios. The beneficial effects of this invention are: First, the method directly acts on the original RF signal domain, which can suppress channel statistical changes caused by probe perturbation at a position closer to the data source; Second, the method uses the channel covariance structure to describe the joint statistical relationship between array elements, which is more suitable for handling the overall channel distribution drift caused by probe state changes than channel-by-channel normalization; Third, the basic whitening parameters can be established offline, and the online stage mainly performs mean centering and matrix multiplication, with low computational overhead, making it suitable for insertion into real-time or near-real-time RF processing flows; Fourth, reference domain covariance alignment can provide a more explicit mean and covariance alignment effect when the state is known or the match is correct; Fifth, the corrected RF data still retains C × D The two-dimensional matrix form facilitates direct use in subsequent signal analysis, feature extraction, classification models, or regression models.

[0056] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement an ultrasonic signal perturbation suppression method based on covariance whitening.

[0057] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an ultrasonic signal perturbation suppression method based on covariance whitening. Specific Implementation Example 2: The only difference between Embodiment 2 and Embodiment 1 of this application is that: The specific implementation plan is as follows: Figure 2 As shown. First, calibration data is collected to establish statistical parameters. The calibration data should cover typical probe states that may occur in practical applications, such as standard fit, leftward movement, rightward movement, forward movement, backward movement, slight pitch, axial rotation, light pressure, heavy pressure, and changes in coupling state. During data collection, the subject should be kept in a relatively stable state as much as possible, so that the main differences between different data points come from changes in probe position, posture, pressure, or coupling conditions, rather than from significant changes in the subject's own state. In forearm tissue or muscle data collection scenarios, the subject can be asked to maintain the same hand gesture or muscle state, and the probe contact position and contact method can be changed accordingly. For example... Figure 3As shown, the original RF matrix under different probe offset states exhibits different amplitude distributions and local texture differences in the element channel dimension and depth sampling dimension. This figure shows the original RF matrix itself, not the B-mode image or the displayed image after beamforming. Further... Figure 4 As shown, the correlation matrices of array element channels differ under different probe offset states, indicating that probe perturbation not only manifests as amplitude changes in a single channel but also causes changes in the second-order correlation structure between channels. If the changes in the state of the measured object in the calibration data are highly coupled with probe perturbation, the estimated covariance may simultaneously include differences in the target's physiological state and probe state, leading to whitening transformation weakening some effective task features. Therefore, it is preferable to use multi-probe state calibration data under the same measured state and establish statistical parameters separately under different measured states. Each frame of data acquired is saved as a two-dimensional RF matrix. If the number of array element channels is 128 and the number of depth sampling points is 1664, then each frame of RF data is a 128×1664 matrix. Data under different probe states can be separately created into folders or state labels for subsequent statistical analysis and parameter estimation. The calibration process of this method does not require manual annotation of tissue boundaries, image contours, or downstream task labels; it only requires the inclusion of raw RF data under multiple typical probe states. Therefore, it can be used as a general preprocessing module before the acquisition system or downstream algorithms.

[0059] Before calculating statistical parameters, basic preprocessing can be performed on the raw RF data. The purpose of preprocessing is not to change the basic matrix structure of the RF signal, but to remove obvious invalid components and outliers, making the covariance estimation more stable. Preprocessing may include removing DC bias, cropping invalid depth regions, cropping outlier amplitudes, unifying data types, amplitude scaling, and necessary bandpass filtering. If some array element channels have obvious anomalies, such as all-zero channels, channel saturation, excessive random noise, or no effective echoes for a long period, abnormal channel detection can be performed before calculating the covariance, and processed using methods such as rejection, interpolation, weight reduction, or hardware maintenance marking. The preprocessed data still retains... C × D The matrix form is used to ensure consistency between subsequent whitening correction and the original data interface. When the calibration data volume is large, to avoid loading the entire RF matrix at once, this invention preferably uses a batch iterative method to calculate the global mean and covariance. Let the first... b Each batch contains M b Each channel vector sample has a batch mean of 1. μ b The batch scatter matrix is Q b If the sample size has already been accumulated before being added to this batch. n old mean μ old and scatter matrix Q old The update following the addition of this batch can be written as:

[0060] This iterative approach is mathematically equivalent to directly calculating the mean and covariance after summarizing all channel vectors at once, but it significantly reduces memory usage and is suitable for large-scale RF datasets. After estimating the global mean and global covariance, a global whitening matrix is ​​constructed based on the regularized covariance matrix, and the offline-obtained mean is used... μ and whitening matrix P Save as calibration parameters. During online acquisition or inference, it is not necessary to reread all calibration data or recalculate the covariance matrix; only the offline parameters need to be loaded to complete the processing. In online processing, the system first acquires the original RF matrix of the current frame and performs the same basic preprocessing as in the offline stage. Then, channel centering is performed using the offline mean, and the result is left-multiplied by the global whitening matrix to obtain the whitened RF matrix. Since the whitening matrix only affects the channel dimension and not the depth sampling dimension, the input is... C × D The RF matrix, the output is still C × DThe RF matrix is ​​used. For example, when the input is a 128×1664 RF matrix, the output is also a 128×1664 matrix. If the whitened RF data is inconsistent with the expected range of the subsequent model or display module in terms of amplitude scale, global scale recovery or channel-by-channel scale recovery can be added to make the numerical range more stable. When practical applications require more explicit suppression of several known probe disturbance states, a reference domain covariance alignment method can be used. In the offline stage, the mean and covariance of each disturbance state are estimated separately, and a standard fitting state is selected as the reference domain, or multiple states are weighted according to the number of samples or the actual occurrence probability to form an average reference domain. For each disturbance state, the corresponding alignment matrix is ​​constructed and saved. In the online stage, if the current acquisition state is known, the mean and alignment matrix corresponding to the state are directly called to complete the correction; if the current state is unknown, the current statistic is estimated using short-time RF data, the distance is calculated with the offline template, the closest state is selected, or soft weight fusion is used. This method, when the current state is correctly matched, can more clearly align the mean and covariance of the perturbation state to the reference domain, making it suitable for applications requiring reduced statistical differences between different probe contact states. In some ultrasound scenarios, the statistical properties of RF signals in different depth regions may differ significantly. Shallow regions are more susceptible to probe contact states, skin surface reflections, and coupling conditions; mid-depth regions may contain major tissue structures; and deep regions are more susceptible to attenuation and noise. To improve local adaptability, this invention divides the depth sampling index into several windows and estimates the mean, covariance, and whitening matrix within each window. During online processing, channel whitening is performed on each depth window, and then the data is stitched together along the depth direction to obtain the complete output. This method allows shallow, mid, and deep regions to use correction parameters that better match their local statistical characteristics without altering the basic RF data structure. When subsequent tasks have a strong dependence on the physical location of the array elements, a constrained whitening method can be used. Diagonal whitening uses only the diagonal elements of the covariance matrix to correct the scale of each channel, without mixing different array elements, thus maintaining the one-to-one correspondence between the output channels and the real array elements to the greatest extent. Local strip approximation only allows a finite number of linear combinations between adjacent or nearest-neighbor array elements, reducing the impact of mixing of distant channels on the array geometry. Its form can be written as:

[0061] Where ⊙ represents Hadamard element-wise multiplication, kThe nearest neighbor channel radius is used to allow for mixing. The banded constraint yields an approximate channel correlation suppression matrix, which generally no longer strictly satisfies the complete whitening property. Therefore, output variance normalization or scale recovery steps can be added in practical implementations. For traditional imaging tasks, it is preferable to use a parallel approach of the original RF imaging branch and the whitened RF learning branch; for learning tasks such as classification, regression, or pose estimation, the globally whitened or reference-domain aligned RF matrix can be directly used as input to improve the statistical stability of the model under different probe states. To verify the implementation effect of this invention, RF data of various probe states such as standard fit, translation, rotation, pressing, and coupling changes can be collected under the same test state, and the channel covariance heatmaps, mean distances between states, normalized correlation matrix distances, and downstream task errors before and after correction can be compared. Figure 5 The display shows the decrease in the statistical distance of each offset state relative to the correct position after correction. Figure 6 The channel correlation structure after covariance alignment with the reference domain converges towards the reference domain. After correction, the covariance structures between different probe states are more similar, the statistical distance between states decreases, and the downstream model's error or stability under perturbation conditions decreases, indicating that this invention can effectively suppress second-order statistical differences in RF channels caused by probe perturbation. The above verification method does not require proof that the complete physical sound field after probe offset is recovered, but rather serves to demonstrate that the original RF channel statistical structure becomes more stable and consistent after processing by this method.

[0062] The above description is merely a preferred embodiment of an ultrasonic signal perturbation suppression method based on covariance whitening. The scope of protection for such a method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the scope of protection of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the scope of protection of this invention.

Claims

1. A method for suppressing ultrasonic signal perturbation based on covariance whitening, characterized in that: The method includes the following steps: Step 1: Collect raw RF calibration data covering the perturbation states of various ultrasonic sensor probes, and decompose each frame of the RF matrix into C-dimensional channel vector samples along the depth direction; Step 2: Estimate the global channel mean and global channel covariance matrix based on all channel vector samples; perform regularization on the covariance matrix to obtain the regularized covariance matrix, and construct the global symmetric whitening matrix through eigenvalue decomposition. Step 3: Acquire the original RF matrix of the current frame. X t Perform basic preprocessing consistent with the offline phase; center the preprocessed RF matrix using the global channel mean, and multiply it by the global whitening matrix in the element channel dimension to obtain the corrected RF matrix. .

2. The method according to claim 1, characterized in that: When collecting calibration data, the object under test is kept in a stable state so that the difference in calibration data mainly comes from changes in probe state rather than changes in the state of the object under test itself; the calibration data covers all typical probe disturbance states in practical applications, and state labels are added to the data of different probe states.

3. The method according to claim 2, characterized in that: set up The number of array element channels of a one-dimensional linear ultrasound probe is C The number of depth sampling points for each channel is D Any frame of raw RF data is represented as: in, X The c The line represents the first c RF signals of each array element channel, X The d The column indicates the first d The channel vector formed by all channels at each depth sampling point; Record No. d The channel vector corresponding to each depth sampling point is: A frame of RF matrix is ​​considered as being composed of D indivual C Composition of 3D channel vectors: 。 4. The method according to claim 3, characterized in that: Using multiple probe disturbance states, multiple acquisition frames, and channel vectors at multiple depth locations as statistical samples, let the total number of samples be... M Then the global channel mean and global channel covariance can be written as: 。 5. The method according to claim 4, characterized in that: To improve numerical stability, a regularization form related to the covariance scaling is adopted: in, I C for C × C The identity matrix, this regularization term avoids numerical amplification caused by small eigenvalues ​​and makes the regularization strength adaptively change with the overall energy scale of the RF data. Eigenvalue decomposition is performed on the regularization covariance matrix, and a symmetric whitening matrix is ​​constructed: During online processing, the current input RF matrix is ​​processed. X t By centering the channels and left-multiplying by the global whitening matrix along the channel dimension, we obtain: Among them, 1 D Let D be a vector of all ones. This whitening matrix strictly whitens the regularized mixture covariance matrix, that is: For the original covariance matrix without a regularization term, when the regularization term is small and the covariance estimate is sufficient, we have: 。 6. The method according to claim 5, characterized in that: For the s For each probe disturbance state, estimate the mean and covariance under that state: The standard fit state or the multi-state average state is selected as the reference domain, and its mean and covariance are denoted as . μ r With Σ r In the case of Σ s and Σ r After performing similar regularization, the square root of a symmetric positive definite matrix is ​​used to construct the state. s Covariance alignment matrix to the reference domain: For the current state is s The input RF matrix, the correction result is: When the current state matches correctly, the square root of the matrix is ​​taken as the principal symmetric square root, and the small deviation introduced by the regularization term is ignored, this transformation makes the corrected channel vector satisfy: 。 7. The method according to claim 6, characterized in that: When the current probe status is unknown during online data acquisition, the current mean is estimated using a short period of RF data. and current covariance The data is then matched with an offline perturbation state template. To avoid interference from the amplitude scale on the correlation comparison, the covariance matrix is ​​converted into a normalized correlation matrix, and the following distance metric is used: in, α >0 and β >0 is used to balance the difference in mean and the difference in correlation structure, respectively.

8. The method according to claim 7, characterized in that: Alternatively, select the state with the smallest distance and apply the corresponding alignment matrix, or use soft weights to fuse the correction results of multiple states: 。 9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method as claimed in any one of claims 1-8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the method of any one of claims 1-8.

Images