Ultrasound elastography image denoising and enhancement method

Abstract

The application relates to the technical field of medical ultrasonic imaging and digital image processing, in particular to an ultrasonic elasticity image denoising and enhancing method, which comprises the following steps: acquiring original ultrasonic elasticity imaging data to be processed, and constructing a structure tensor matrix reflecting local texture directions; solving the anisotropic coherence degree of each pixel point, obtaining a structure attribute discrimination result, and constructing a structure confidence atlas; collecting the point spread function characteristics of an ultrasonic imaging system, establishing a speckle noise statistical model; obtaining a local smoothing coefficient of the first iteration, and performing anisotropic diffusion correction on an initial elasticity distribution matrix to obtain a first modified elasticity image matrix, and updating the structure confidence atlas to obtain a final denoised and enhanced target elasticity image; the application effectively suppresses multiplicative granular speckles, eliminates the ladder effect or artifacts easily generated by conventional methods, and significantly improves the image signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of medical ultrasound imaging and digital image processing technology, specifically to a method for denoising and enhancing ultrasound elastography images. Background Technology

[0002] In the clinical application environment of ultrasound elastography, due to the coherence characteristics of ultrasound and the complexity of biological tissues, the acquired raw elastography image data inevitably contains a large amount of speckle noise that is highly coupled with tissue texture. This noise usually exhibits a multiplicative granular distribution, and its statistical characteristics show non-Gaussian forms such as Rayleigh distribution or Fisher-Tipett distribution depending on the distribution of tissue scatterers.

[0003] For the optimization processing of such images, existing technical solutions generally employ linear low-pass filtering or traditional gradient-based anisotropic diffusion algorithms. These methods detect edges by calculating the gradient magnitude of pixels and smooth the image based on a single gradient threshold. While these solutions are effective in suppressing high-frequency random noise, they fail to fully consider the specific physical and statistical laws of ultrasound speckle and lack refined characterization of anisotropic features such as local texture orientation. This makes it difficult to effectively distinguish speckle noise from weak texture structures during processing. This often results in the loss of crucial tissue boundaries and internal texture details while removing noise, leading to blurred edges or artifacts, failing to meet clinicians' high-precision requirements for accurate lesion localization and feature analysis. Therefore, how to effectively decouple speckle noise from tissue texture while adhering to the physical laws of ultrasound imaging, and fully preserve the anisotropic structural features of the image while significantly suppressing background noise, has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for denoising and enhancing ultrasonic elastography images. Specifically, the technical solution of this invention includes:

[0005] Step 1: Obtain the raw ultrasound elastography data to be processed, convert it into an initial elastic distribution matrix, and perform gradient calculation on each pixel in the initial elastic distribution matrix to construct a structure tensor matrix that reflects the local texture direction.

[0006] Step 2: Perform eigenvalue decomposition on the structure tensor matrix, extract principal eigenvalues ​​and secondary eigenvalues, calculate the anisotropic coherence of each pixel based on the ratio of principal eigenvalues ​​to secondary eigenvalues, and preset a discrimination benchmark including noise discrimination threshold and edge discrimination threshold. Compare the anisotropic coherence with the discrimination benchmark to obtain the structural attribute discrimination results and construct a structural confidence map.

[0007] Step 3: Acquire the point spread function characteristics of the ultrasound imaging system and establish a speckle noise statistical model; combine the structural confidence map and construct an elastic image evolution model using the anisotropic diffusion equation. Input the speckle noise statistical model into the elastic image evolution model as a constraint condition to construct the first... The diffusion tensor of the iteration, the edge preservation index based on the high-frequency components of the image, and the texture energy consistency factor based on the frequency domain energy features are correlated to obtain the 1st iteration. The local smoothing coefficients of the nth iteration are obtained, and the initial elastic distribution matrix is ​​corrected for anisotropic diffusion to obtain the nth iteration. The elastic image matrix is ​​corrected and the structure confidence map is updated to obtain the final denoised and enhanced target elastic image.

[0008] Preferably, step one includes:

[0009] S11. Acquire radio frequency signals of biological tissues using an ultrasonic probe, estimate tissue displacement using a cross-correlation algorithm or Doppler principle, generate raw strain data, and map it to an initial elastic distribution matrix.

[0010] S12. Perform Gaussian smoothing preprocessing on the initial elastic distribution matrix to suppress the interference of high-frequency random noise on gradient calculation;

[0011] S13, using the central difference method or The operator calculates the gradient vectors of the initial elastic distribution matrix in the horizontal and vertical directions, calculates the outer product of the gradient vectors, and performs a weighted average of the outer product results in the local neighborhood to construct the structure tensor matrix corresponding to each pixel.

[0012] Preferably, step S13 includes:

[0013] S131. Obtain the size of the neighborhood window of the current pixel in the initial elastic distribution matrix;

[0014] S132. Calculate the horizontal and vertical gradient components of each point within the neighborhood window;

[0015] S133. Construct a second-order moment matrix using the horizontal and vertical gradient components, and introduce a Gaussian weighting function to perform convolution operation on the second-order moment matrix to smooth the structure tensor field and ensure the continuity of the structure tensor matrix in space.

[0016] Preferably, step two includes:

[0017] S21. Perform eigenvalue decomposition on the structure tensor matrix of each pixel to obtain the first eigenvalue and the second eigenvalue, wherein the first eigenvalue is greater than or equal to the second eigenvalue.

[0018] S22. Using the ratio of the difference between the first eigenvalue and the second eigenvalue to the sum, calculate the anisotropic coherence, which characterizes the degree of consistency of texture direction in a local region.

[0019] S23. Preset noise discrimination threshold and edge discrimination threshold, and limit the noise discrimination threshold to be less than the edge discrimination threshold;

[0020] S24. Compare the anisotropic coherence with the noise discrimination threshold and the edge discrimination threshold to obtain the structural attribute discrimination result, including: if the anisotropic coherence is less than or equal to the noise discrimination threshold, the pixel is determined to be located in an isotropic noise region and marked as a flat region; if the anisotropic coherence is greater than the noise discrimination threshold and less than or equal to the edge discrimination threshold, the pixel is determined to be located in a weak texture region and marked as a transition region; if the anisotropic coherence is greater than the edge discrimination threshold, the pixel is determined to be located in a strong edge or texture region and marked as a feature region.

[0021] S25. Based on the structural attribute discrimination results, assign a corresponding confidence weight to each pixel and draw a grayscale map to obtain the structural confidence map.

[0022] Preferably, step three includes:

[0023] S31. Obtain the Rayleigh distribution or Fisher-Tipett distribution parameters of ultrasonic speckle noise from historical databases or phantom experimental data, and establish a statistical model of speckle noise.

[0024] S32. Construct an anisotropic diffusion partial differential equation as an elastic image evolution model, which includes a diffusion tensor term and a time step term;

[0025] S33. Based on the feature vector directions in the structural confidence map, construct the first... The diffusion tensor of the next iteration is made so that the diffusion direction is parallel to the texture direction and perpendicular to the gradient direction, in order to protect edge features;

[0026] S34. Calculate the ratio of the local variance of the current image to the variance of the pre-selected reference uniform region, and construct the first... The equivalent number of views in each iteration is used as an evaluation index for the degree of noise suppression.

[0027] Preferably, step three also includes:

[0028] S35. Utilize the Laplacian operator to detect changes in high-frequency components in an image, and construct the first... The edge preservation index of the next iteration;

[0029] S36. Perform image processing Filtering transformation is used to extract the energy features of subbands in different directions, and the first subband is constructed. Texture energy consistency factor in the next iteration;

[0030] S37. Extract the diffusion tensor, equivalent number of views, edge preservation index, and texture energy consistency factor obtained in S33-S36, perform weighted fusion, and obtain the... The local smoothing coefficient of the second iteration is used to dynamically adjust the diffusion intensity of the diffusion equation based on this coefficient. The image is updated through each iteration.

[0031] Preferably, step three also includes:

[0032] S38, Calculate the first The elastic image matrix after the second correction and the first Root mean square error between sub-image matrices;

[0033] S39. Preset a convergence threshold and compare the root mean square error (RMSE) with the convergence threshold: if the RMS error is greater than the convergence threshold, then let... The structure confidence map is corrected using the updated image data, and the process returns to step S33 for the next iteration. If the root mean square error is less than or equal to the convergence threshold, the iteration stops, and the current iteration is output. The modified elastic image matrix is ​​used as the final target elastic image.

[0034] Preferably, step three also includes:

[0035] S40. Perform quality assessment on the final output target elastic image to obtain image quality grading results, including: calculating the structural similarity index (SSIM) between the target elastic image and the initial elastic distribution matrix; if SSIM is greater than 0.9, the quality grading result is labeled as excellent, and a storage or display instruction is executed; if SSIM is greater than 0.75 and less than or equal to 0.9, the quality grading result is labeled as good, an edge sharpening compensation instruction is generated, and the target elastic image is post-processed for sharpening; if SSIM is less than or equal to 0.75, the quality grading result is labeled as poor, a parameter callback instruction is generated, the noise discrimination threshold in step S23 is reduced, and the process returns to step two to update the structural confidence map, and then step three is re-executed.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This invention utilizes the structural tensor matrix and anisotropic coherence to construct a structural confidence map, breaking through the limitation of traditional gradient algorithms that only focus on amplitude; by quantifying the consistency of texture direction, the system can accurately divide the noise region, transition region and feature region, thereby achieving effective decoupling of speckle noise and tissue texture, ensuring that while smoothing background noise, the texture and boundary information of tiny lesions that are crucial for clinical diagnosis are fully preserved.

[0038] 2. This invention introduces a speckle noise statistical model as a physical constraint into the evolution equation to model the non-Gaussian distribution characteristics of ultrasonic echo signals. This denoising strategy based on imaging physics can more accurately adapt to the noise morphology of Rayleigh or Fisher-Tipett distribution compared with traditional linear filtering, effectively suppressing multiplicative granular speckle, eliminating the step effect or artifacts that are easily generated by conventional methods, and significantly improving the image signal-to-noise ratio.

[0039] 3. This invention constructs a multi-dimensional fusion mechanism that includes diffusion tensor, edge preservation index and texture energy consistency factor to dynamically generate local smoothing coefficients. The mechanism adjusts the diffusion intensity and direction in real time according to the local features of the image. It performs isotropic diffusion in flat areas to thoroughly remove noise, and performs anisotropic diffusion in edge and texture areas to lock in details, thus achieving the best balance between noise suppression and anatomical structure preservation.

[0040] 4. This invention designs a closed-loop feedback system that includes iterative convergence determination and output quality evaluation, and has parameter adaptive callback and post-processing sharpening functions. When the image structure similarity does not meet the standard, the system can automatically reduce the discrimination threshold and re-evolve, avoiding image distortion or loss of details caused by parameter solidification, which greatly enhances the intelligence level and reliability of the algorithm in complex clinical application scenarios. Attached Figure Description

[0041] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0042] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0044] Example 1:

[0045] Please see Figure 1 A method for denoising and enhancing ultrasound elastography images, the specific steps of which include:

[0046] Step 1: Obtain the raw ultrasound elastography data to be processed, convert it into an initial elastic distribution matrix, and perform gradient calculation on each pixel in the initial elastic distribution matrix to construct a structure tensor matrix that reflects the local texture direction.

[0047] Step 2: Perform eigenvalue decomposition on the structure tensor matrix, extract principal eigenvalues ​​and secondary eigenvalues, calculate the anisotropic coherence of each pixel based on the ratio of principal eigenvalues ​​to secondary eigenvalues, and preset a discrimination benchmark including noise discrimination threshold and edge discrimination threshold. Compare the anisotropic coherence with the discrimination benchmark to obtain the structural attribute discrimination results and construct a structural confidence map.

[0048] Step 3: Acquire the point spread function characteristics of the ultrasound imaging system and establish a speckle noise statistical model; combine the structural confidence map and construct an elastic image evolution model using the anisotropic diffusion equation. Input the speckle noise statistical model into the elastic image evolution model as a constraint condition, construct the diffusion tensor of the k-th iteration, the edge preservation index based on the high-frequency components of the image, and the texture energy consistency factor based on the frequency domain energy characteristics. After correlating the diffusion tensor, the edge preservation index, and the texture energy consistency factor, obtain the local smoothing coefficient of the k-th iteration, and perform anisotropic diffusion correction on the initial elastic distribution matrix to obtain the elastic image matrix after the k-th correction. Update the structural confidence map to obtain the final denoised and enhanced target elastic image.

[0049] This embodiment details the core execution logic of an ultrasound elastography image denoising and enhancement method. This method aims to solve the technical problem of decoupling speckle noise and tissue texture in existing technologies. The system executes step one: acquiring the raw ultrasound elastography data to be processed and obtaining an initial elastic distribution matrix through mapping transformation; calculating the gradient for each pixel in this matrix. Unlike conventional methods that only utilize gradient magnitude, this step constructs a structural tensor matrix, which not only contains gradient magnitude information but also preserves gradient direction information through gradient outer product smoothing; the system executes step two: performing eigenvalue decomposition on the acquired structural tensor matrix to extract principal and secondary eigenvalues; based on this, calculating the anisotropic coherence, which serves as a quantitative index characterizing the consistency of texture direction within a local region; the system retrieves a preset discrimination benchmark, which includes a noise discrimination threshold and an edge discrimination threshold, and compares the calculated anisotropic coherence with this benchmark point by point to obtain a structural attribute discrimination result determining whether the point belongs to noise, weak texture, or strong edge.

[0050] Based on the discrimination result, a structural confidence map is constructed. This map is a weighted map of the same size as the original image, and the higher the value, the greater the probability that the pixel belongs to the real tissue structure. The system executes step three to construct an evolutionary model under physical constraints. The point spread function characteristics of the ultrasound imaging system are collected to establish a speckle noise statistical model describing the noise distribution law. Combining the aforementioned structural confidence map, an evolutionary model is constructed using the anisotropic diffusion equation, and the speckle noise statistical model is used as the input physical constraint condition. In the k-th iteration, the system constructs three key parameters: diffusion tensor, edge preservation index, and texture energy consistency factor. After correlating the above three parameters, the local smoothing coefficient of the k-th iteration is obtained. The initial elastic distribution matrix is ​​anisotropically diffused and corrected using the local smoothing coefficient to obtain the elastic image matrix after the k-th correction. The structural confidence map is updated synchronously. After multiple iterations until convergence, the final denoised and enhanced target elastic image is obtained.

[0051] This embodiment achieves the technical effect of fundamentally distinguishing speckle noise from biological tissue texture by introducing structural tensors and anisotropic coherence. By introducing the speckle statistical model as a constraint into the diffusion equation, the denoising process conforms to the physical laws of ultrasound imaging. The final output image not only significantly suppresses background noise but also fully preserves the tissue boundaries and internal texture details that are crucial for clinical diagnosis, effectively solving the problem that traditional filtering methods easily lose anisotropic structural features.

[0052] Example 2:

[0053] Step one includes:

[0054] S11. Acquire radio frequency signals of biological tissues using an ultrasonic probe, estimate tissue displacement using a cross-correlation algorithm or Doppler principle, generate raw strain data, and map it to an initial elastic distribution matrix.

[0055] S12. Perform Gaussian smoothing preprocessing on the initial elastic distribution matrix to suppress the interference of high-frequency random noise on gradient calculation;

[0056] S13, using the central difference method or The operator calculates the gradient vectors of the initial elastic distribution matrix in the horizontal and vertical directions, calculates the outer product of the gradient vectors, and performs a weighted average of the outer product results in the local neighborhood to construct the structure tensor matrix corresponding to each pixel.

[0057] Step S13 includes:

[0058] S131. Obtain the size of the neighborhood window of the current pixel in the initial elastic distribution matrix;

[0059] S132. Calculate the horizontal and vertical gradient components of each point within the neighborhood window;

[0060] S133. Construct a second-order moment matrix using the horizontal and vertical gradient components, and introduce a Gaussian weighting function to perform convolution operation on the second-order moment matrix to smooth the structure tensor field and ensure the continuity of the structure tensor matrix in space.

[0061] This embodiment further specifies step one of Embodiment 1, detailing the data acquisition and tensor construction process; in step S11, ultrasound waves are emitted to biological tissue and radio frequency signals are received through an ultrasound probe; using a cross-correlation algorithm or Doppler principle, the displacement of the tissue before and after compression is calculated, thereby generating the original strain data; according to Hooke's law or the generalized elastic modulus formula, the strain data is mapped to an initial elastic distribution matrix. ;

[0062] The source is a mapping of the original strain data, with the physical meaning being the spatial distribution of the initial elastic modulus of biological tissue, in kilopascals; Step S12 is executed to prevent spurious extrema from arising in the gradient calculation due to random noise. Perform Gaussian smoothing preprocessing to generate a preprocessed matrix; execute step S13 and its sub-steps S131 to S133 to capture local geometric structures; utilize... The operator calculates the horizontal gradient component for each pixel in the image. and vertical gradient components

[0063] For the current pixel, its neighborhood window is sampled, for example, a 5×5 pixel window, and a second-order moment matrix, i.e., the structure tensor matrix, is constructed. ;

[0064] The Gaussian convolution, derived from the outer product of gradient components, physically represents the second moment describing the gradient direction distribution within a local neighborhood of an image; it is dimensionless. However, directly using the outer product of gradients to obtain the matrix is ​​susceptible to noise interference; therefore, a Gaussian weighting function is introduced. Convolution operation is performed on a second-order moment matrix, where the Gaussian weighting function is used. Standard deviation The value range is [1.0, 3.0], and in this embodiment, 1.5 is preferred. The size of the convolution kernel window is determined according to... The nearest odd integer is used to determine the structure of the tensor field, thus smoothing the structure.

[0065] This embodiment effectively suppresses the interference of high-frequency random noise on structural feature extraction through Gaussian preprocessing and weighted averaging of tensor fields, ensuring the spatial continuity of the structural tensor matrix and making the calculated texture direction more robust and less susceptible to single-point noise, thus providing a stable data foundation for subsequent feature decomposition.

[0066] Example 3:

[0067] Step two includes:

[0068] S21. Perform eigenvalue decomposition on the structure tensor matrix of each pixel to obtain the first eigenvalue and the second eigenvalue, wherein the first eigenvalue is greater than or equal to the second eigenvalue.

[0069] S22. Using the ratio of the difference between the first eigenvalue and the second eigenvalue to the sum, calculate the anisotropic coherence, which characterizes the degree of consistency of texture direction in a local region.

[0070] S23. Preset noise discrimination threshold and edge discrimination threshold, and limit the noise discrimination threshold to be less than the edge discrimination threshold;

[0071] S24. Compare the anisotropic coherence with the noise discrimination threshold and the edge discrimination threshold to obtain the structural attribute discrimination result, including: if the anisotropic coherence is less than or equal to the noise discrimination threshold, the pixel is determined to be located in an isotropic noise region and marked as a flat region; if the anisotropic coherence is greater than the noise discrimination threshold and less than or equal to the edge discrimination threshold, the pixel is determined to be located in a weak texture region and marked as a transition region; if the anisotropic coherence is greater than the edge discrimination threshold, the pixel is determined to be located in a strong edge or texture region and marked as a feature region.

[0072] S25. Based on the structural attribute discrimination results, assign a corresponding confidence weight to each pixel and draw a grayscale map to obtain the structural confidence map.

[0073] This embodiment further specifies step two in embodiment 1, detailing how to use the structure tensor for feature classification and confidence map construction; executing step S21, for each pixel's structure tensor matrix Perform eigenvalue decomposition to obtain two non-negative eigenvalues: the first eigenvalue... Second eigenvalue And satisfy ;

[0074] The source is the eigenvalue decomposition of the tensor matrix, and its physical meaning is the intensity along the direction of maximum gradient change, which is dimensionless.

[0075] The source is tensor matrix eigenvalue decomposition, which physically represents the intensity of change along the direction perpendicular to the gradient, i.e., the texture direction, and is dimensionless; execute step S22 to calculate the anisotropic coherence. The calculation formula is the square of the ratio of the difference between the first and second eigenvalues ​​to the sum of the eigenvalues. The specific mathematical expression is:

[0076]

[0077] in, To prevent tiny positive numbers with a denominator of zero from being counted, a sign is used here. To distinguish it from the convergence threshold in subsequent steps Its value is usually in to In this embodiment, the preferred embodiment is... The coherence is used to quantitatively describe the directionality of local textures; steps S23 to S24 are executed to preset the noise discrimination threshold. and edge discrimination threshold ;

[0078] Specifically, and The setup logic is as follows: Calculate the gray-level histogram of the anisotropic coherence spectrum, and then use... The Otsu's method calculates the global optimal threshold as... ; Take anisotropic coherence less than The average or median of a subset of the data is calculated and set as follows: Alternatively, based on empirical statistics from a large number of ultrasound soft tissue elastography images, a setting can be established. The range of values ​​is This corresponds to the low coherence of homogeneous structures; The range of values ​​is The corresponding high coherence of the fascia or lesion boundary; in response to the anisotropic coherence being less than or equal to the noise discrimination threshold, the system determines that the pixel is located in the isotropic noise region and marks it as a flat region.

[0079] In response to anisotropic coherence greater than the noise discrimination threshold and less than or equal to the edge discrimination threshold, the system determines that the pixel is located in a weak texture region and marks it as a transition region; in response to anisotropic coherence greater than the edge discrimination threshold, the system determines that the pixel is located in a strong edge or texture region and marks it as a feature region; proceed to step S25, and assign a confidence weight to each pixel based on the above structural attribute discrimination results. For example, high weights, such as 1.0, are assigned to feature regions; low weights, such as 0.1, are assigned to flat regions; and linear interpolation is used for transition regions, with the following calculation formula:

[0080]

[0081] in, For flat regions, the weight is 0.1. For the feature region weights, such as 1.0, To obtain the structure confidence map, plot the anisotropic coherence of the current pixel as a grayscale map. ;

[0082] This embodiment uses anisotropic coherence calculation based on eigenvalues ​​and a multi-threshold discrimination strategy to accurately identify flat areas, transition areas and feature areas in an image, providing clear spatial guidance for subsequent adaptive diffusion and avoiding edge blurring or noise residue problems caused by the one-size-fits-all processing in traditional filtering algorithms.

[0083] Example 4:

[0084] Step three includes:

[0085] S31. Obtain the Rayleigh distribution or Fisher-Tipett distribution parameters of ultrasonic speckle noise from historical databases or phantom experimental data, and establish a statistical model of speckle noise.

[0086] S32. Construct an anisotropic diffusion partial differential equation as an elastic image evolution model, which includes a diffusion tensor term and a time step term;

[0087] S33. Based on the feature vector direction in the structural confidence map, construct the diffusion tensor of the k-th iteration, so that the diffusion direction is parallel to the texture direction and perpendicular to the gradient direction, in order to protect the edge features.

[0088] S34. Calculate the ratio of the local variance of the current image to the variance of the pre-selected reference uniform region, and construct the equivalent number of views for the k-th iteration as an evaluation index of the degree of noise suppression.

[0089] S35. Detect high-frequency component changes in an image using the Laplacian operator and construct the edge-preserving index for the kth iteration.

[0090] S36. Perform image processing The filtering transformation extracts the energy features of sub-bands in different directions, and constructs the texture energy consistency factor for the k-th iteration.

[0091] S37. Extract the diffusion tensor, equivalent number of views, edge preservation index and texture energy consistency factor obtained from S33-S36, perform weighted fusion to obtain the local smoothing coefficient of the k-th iteration, and dynamically adjust the diffusion intensity of the diffusion equation according to the coefficient to evolve and update the image of the k-th iteration.

[0092] This embodiment is a further development of step three in embodiment 1, detailing the construction of the evolutionary model and the multi-factor fusion mechanism; step S31 is executed to obtain ultrasonic speckle noise parameters from the historical database. For fully developed speckles, a Rayleigh distribution model is adopted, with its probability density function being... For underdeveloped speckle patterns, the Fisher-Tipett distribution parameter, i.e., the Gumbel distribution, is used, and its probability density function is:

[0093]

[0094] in, The statistical parameters are fitted from the data, and are denoted here by [symbol]. Specifically refers to noise distribution parameters, to distinguish them from subsequent parameters. The scale parameter in the filter; the speckle noise statistical model is established; step S32 is executed to construct the anisotropic diffusion partial differential equation (PDE) as the elastic image evolution model. This model includes a diffusion tensor term and a time step term, and the specific equation form is as follows:

[0095]

[0096] in For divergence operators, For gradient operators, For virtual time variables, For the diffusion tensor, For local smoothing coefficients, The coefficient for the data fidelity term is adaptively determined based on the speckle noise statistical parameters obtained in step S31: for the Rayleigh distribution, set... For the Fisher-Tipping distribution, set ,in, This is the intensity normalization factor; it is introduced here. That is, the initial elasticity distribution matrix The global root mean square (RMS squared) is used to ensure the consistency of the physical dimensions of the evolution equations, so that... The weights are converted into dimensionless weighting coefficients to balance the magnitudes of the data fidelity term and the diffusion tensor term; this introduces the statistical properties of noise as a constraint into the evolution equation; this is used to prevent the image from becoming overly smoothed and deviating from the original data during iteration. Step S33 is then executed. In each iteration, a diffusion tensor is constructed based on the eigenvector directions associated with the structural confidence map. Its mathematical expression is:

[0097]

[0098] in, The eigenvector matrix, Let be a diagonal matrix of diffusion eigenvalues, where and These represent the diffusion coefficients along the main diffusion direction and the secondary diffusion direction, respectively, with values ​​ranging from [value range missing]. Specifically, in order to integrate the structure discrimination results from steps S23-S25 into the evolutionary model, a diffusion eigenvalue and structure confidence map is established. Medium weight Function mapping relationship:

[0099] ,

[0100] in, The gain coefficient, used to control the degree of anisotropy, typically ranges from [value range missing]. For example, take 10; The larger the value, the stronger the suppression of diffusion in non-feature regions; therefore, when Higher, that is, from and When determining the feature region, Approaching 0, it suppresses diffusion perpendicular to the texture direction to protect the edges; when When the temperature is lower, i.e., in a flat region, Approaching 1, isotropic diffusion is performed to remove noise; and it is usually set To enhance smoothness along the texture direction; to protect the edges, the design... The eigenvector directions are such that the main diffusion direction is parallel to the texture direction (i.e., perpendicular to the gradient direction), and the secondary diffusion direction is perpendicular to the texture direction.

[0101] Based on this, steps S34 to S36 are executed to construct a multidimensional evaluation index; the ratio of the local variance of the current image to the variance of the pre-selected reference uniform region is calculated to construct the equivalent number of views for the k-th iteration.

[0102]

[0103] in, For a local window centered on the current pixel, such as the grayscale variance within a 7x7 area, The gray-level variance of a pre-selected uniform region with no texture is used; the Laplacian operator is used to detect changes in high-frequency components in the image, and the first... The edge preservation index of the next iteration ; Perform on the image The filtering transformation extracts the energy features of sub-bands in different directions, and a texture energy consistency factor is constructed. The calculation formula is as follows:

[0104]

[0105] in, For the direction set, the symbol Represents the convolution operation; Two-dimensional using complex form The filter kernel function, used to obtain phase-independent texture energy, has the following spatial domain expression:

[0106]

[0107]

[0108] in, The imaginary unit is used. In actual calculation, the real part and the imaginary part are convolved with the image respectively, and the square of the magnitude of the convolution result is taken as the energy feature.

[0109] : The spatial scaling parameter of the filter typically takes values ​​in the range of... In this embodiment, it is possible to take ;

[0110] : The typical range of wavelength parameters for the filter is as follows: In this embodiment, it is possible to take Note that subscripts are used here to distinguish them from eigenvalues. and noise parameters Confusion;

[0111] The aspect ratio of the space, for example, is set to 0.5 to control the ellipticity of the kernel function. This indicates a phase shift, for example, a value of 0;

[0112] The source is the calculation of the local variance ratio, and its physical meaning is a quantitative indicator of the degree of noise suppression, which is dimensionless;

[0113] The source is Laplacian operator detection, and its physical meaning is the maintenance of edge sharpness, which is dimensionless;

[0114] Source: Filtered energy extraction, physically meaning the retention rate of texture direction performance quantities, is dimensionless; the diffusion tensor, equivalent number of views, edge preservation index, and texture energy consistency factor obtained from S33-S36 are extracted, and then subjected to max-min normalization based on global data, as shown in the formula. ,in and Let be the maximum and minimum values ​​of this index for the entire image in the current iteration. Map the three indices to the interval [0,1], and denote them as follows: , and To eliminate the dimensional differences between different indicators; weighted fusion is performed to obtain the first... The local smoothing coefficient of the next iteration is calculated using the following formula:

[0115]

[0116] in, For normalized weight coefficients, satisfying Specifically, the weighting coefficients can be determined using the entropy weighting method, which calculates the information entropy of ENL, EPI, and TEC in the current iteration image. The smaller the information entropy, the greater the corresponding index weight, thus achieving adaptive adjustment. Alternatively, they can be set based on prior knowledge, for example, all of them can be set to 1 / 3. The diffusion intensity of the diffusion equation is dynamically adjusted according to this coefficient, and the image of the kth iteration is updated. In this embodiment, by integrating the three dimensions of ENL noise suppression, EPI edge protection, and TEC texture protection, a dynamically adjusted local smoothing coefficient is constructed. This achieves powerful denoising while maximizing the preservation of edge and texture information reflecting lesion features, overcoming the limitation of traditional diffusion models that rely solely on gradient information.

[0117] Example 5:

[0118] Step three also includes:

[0119] S38. Calculate the root mean square error between the elastic image matrix after the kth correction and the image matrix after the (k-1)th correction.

[0120] S39. Preset a convergence threshold and compare the root mean square error with the convergence threshold: If the root mean square error is greater than the convergence threshold, let k=k+1, use the updated image data to correct the structure confidence map, and return to execute step S33 for the next round of iteration; If the root mean square error is less than or equal to the convergence threshold, stop the iteration and output the current k-th corrected elastic image matrix as the final target elastic image.

[0121] This embodiment describes the convergence control mechanism of the iterative process, which aims to ensure the stability of the evolution process; step S38 is executed to calculate the image matrix after the k-th correction. Compared to the previous image The root mean square error (RMSE) between them is calculated using the following formula:

[0122]

[0123] in, and These represent the number of rows and columns of the image matrix, respectively.

[0124] The source is the root mean square of the difference between adjacent iterative images, which physically represents the magnitude of change during image evolution, and the unit is the same as pixel gray level; execute step S39, preset convergence threshold. ,For example The calculated root mean square error is compared with a convergence threshold; in response to a root mean square error greater than the convergence threshold, the system determines that the image is still evolving, and sets... The elastic image matrix output from the previous round The input data for step S13 is used to update the structure confidence map. Steps S13 to S25 are then re-executed to recalculate the structure tensor matrix and anisotropic coherence, thereby generating a new structure confidence map to adapt to texture changes during image evolution. The process then returns to step S33 for the next iteration. In response to the root mean square error being less than or equal to the convergence threshold, the system determines that the evolution has stabilized, stops the iteration, and outputs the current iteration. The modified elastic image matrix is ​​used as the final target elastic image;

[0125] This embodiment utilizes RMSE as an adaptive stopping criterion to implement a dynamic convergence control mechanism. This mechanism avoids incomplete denoising caused by insufficient iterations and prevents waste of computational resources and image blurring caused by excessive iterations, thus ensuring the algorithm's adaptability and computational efficiency under different noise levels.

[0126] Example 6:

[0127] Step three also includes:

[0128] S40. Perform quality assessment on the final output target elastic image to obtain image quality grading results, including:

[0129] Calculate the structural similarity index (SSIM) between the target elasticity image and the initial elasticity distribution matrix;

[0130] If SSIM is greater than 0.9, the quality grading result is labeled as excellent, and the storage or display instruction is executed;

[0131] If SSIM is greater than 0.75 and less than or equal to 0.9, the quality grading result is labeled as good, an edge sharpening compensation instruction is generated, and the target elastic image is post-processed for sharpening.

[0132] If SSIM is less than or equal to 0.75, the quality grading result is labeled as poor, a parameter callback instruction is generated, the noise discrimination threshold in step S23 is reduced, and the execution of step two is returned to update the structure confidence map, and step three is executed again.

[0133] This embodiment details the quality closed-loop control mechanism after image output, aiming to construct an intelligent processing flow with self-correction capabilities; step S40 is executed to process the final output target elastic image. Conduct quality assessment; calculate With the initial matrix The Structural Similarity Index (SSIM) between them is calculated using the following formula:

[0134]

[0135] in, Images and The local mean, These are the local standard deviations. For covariance, and It is a constant. For pixel dynamic range, such as 255. ; The source is image structure similarity calculation, which physically measures the similarity between two images in terms of brightness, contrast, and structure, with a value ranging from 0 to 1. The system performs graded processing based on the SSIM value. When SSIM is greater than 0.9, the system determines that the denoising effect is ideal and the structure is well preserved, and the quality grade result is labeled as excellent, and the storage or display instruction is executed directly. When SSIM is greater than 0.75 and less than or equal to 0.9, the system determines that the denoising is acceptable but the edges are slightly blurred, and the quality grade result is labeled as good.

[0136] At this point, an edge sharpening compensation instruction is generated, and the sharpening filter is invoked. Post-processing is performed, specifically using an unsharpened masking algorithm, with the following formula:

[0137]

[0138] in, for The image after Gaussian smoothing with a standard deviation of 0.5. A preset sharpening intensity coefficient, for example, 0.5, is used to enhance edge contrast. When SSIM is less than or equal to 0.75, the system determines severe structural loss and labels the quality grading result as poor. At this point, a parameter callback instruction is generated to automatically lower the noise discrimination threshold in step S23. The specific parameter update formula is as follows:

[0139]

[0140] in, The step size coefficient is used to control the magnitude of parameter callback; its value range is typically [range missing]. For example, take 0.5 and set a lower limit protection. According to this formula, the lower the SSIM, the greater the reduction in threshold, thus more aggressively expanding the area considered as texture, and returning to step two to reconstruct the structure confidence map and re-perform the diffusion iteration;

[0141] This embodiment introduces an automated quality control and feedback mechanism, especially a parameter callback strategy for low-quality ratings, which enables the algorithm to have self-correction capabilities and adaptively adjust parameters for raw data of different qualities, greatly improving the robustness and reliability of the algorithm in complex clinical application scenarios.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An ultrasound elastography image denoising and enhancement method, characterized in that, The specific steps include: Step 1: Obtain the raw ultrasound elastography data to be processed, convert it into an initial elastic distribution matrix, and perform gradient calculation on each pixel in the initial elastic distribution matrix to construct a structure tensor matrix that reflects the local texture direction. Step 2: Perform eigenvalue decomposition on the structure tensor matrix, extract principal eigenvalues ​​and secondary eigenvalues, calculate the anisotropic coherence of each pixel based on the ratio of principal eigenvalues ​​to secondary eigenvalues, and preset a discrimination benchmark including noise discrimination threshold and edge discrimination threshold. Compare the anisotropic coherence with the discrimination benchmark to obtain the structural attribute discrimination results and construct a structural confidence map. Step 3: Acquire the point spread function characteristics of the ultrasound imaging system and establish a speckle noise statistical model; combine the structural confidence map and construct an elastic image evolution model using the anisotropic diffusion equation. Input the speckle noise statistical model into the elastic image evolution model as a constraint condition to construct the first... The diffusion tensor of the iteration, the edge preservation index based on the high-frequency components of the image, and the texture energy consistency factor based on the frequency domain energy features are correlated to obtain the 1st iteration. The local smoothing coefficients of the nth iteration are obtained, and the initial elastic distribution matrix is ​​corrected for anisotropic diffusion to obtain the nth iteration. The elastic image matrix is ​​corrected and the structure confidence map is updated to obtain the final denoised and enhanced target elastic image.

2. The method for denoising and enhancing ultrasonic elastography images according to claim 1, characterized in that: Step one includes: S11. Acquire radio frequency signals of biological tissues using an ultrasonic probe, estimate tissue displacement using a cross-correlation algorithm or Doppler principle, generate raw strain data, and map it to an initial elastic distribution matrix. S12. Perform Gaussian smoothing preprocessing on the initial elastic distribution matrix to suppress the interference of high-frequency random noise on gradient calculation; S13, using central difference method or An operator is used to calculate the gradient vectors of the initial elastic distribution matrix in horizontal and vertical directions respectively, to calculate the outer product of the gradient vectors, and to construct a structure tensor matrix corresponding to each pixel point by weighted averaging the outer product results in the local neighborhood.

3. The method of ultrasound elastography image denoising and enhancement of claim 2, wherein: Step S13 includes: S131. Obtain the size of the neighborhood window of the current pixel in the initial elastic distribution matrix; S132. Calculate the horizontal and vertical gradient components of each point within the neighborhood window; S133. Construct a second-order moment matrix using the horizontal and vertical gradient components, and introduce a Gaussian weighting function to perform convolution operation on the second-order moment matrix to smooth the structure tensor field and ensure the continuity of the structure tensor matrix in space.

4. The method of ultrasound elastography image denoising and enhancement of claim 3, wherein: Step two includes: S21. Perform eigenvalue decomposition on the structure tensor matrix of each pixel to obtain the first eigenvalue and the second eigenvalue, wherein the first eigenvalue is greater than or equal to the second eigenvalue. S22. Using the ratio of the difference between the first eigenvalue and the second eigenvalue to the sum, calculate the anisotropic coherence, which characterizes the degree of consistency of texture direction in a local region. S23. Preset noise discrimination threshold and edge discrimination threshold, and limit the noise discrimination threshold to be less than the edge discrimination threshold; S24. Compare the anisotropic coherence with the noise discrimination threshold and the edge discrimination threshold to obtain the structural attribute discrimination result, including: if the anisotropic coherence is less than or equal to the noise discrimination threshold, the pixel is determined to be located in an isotropic noise region and marked as a flat region; if the anisotropic coherence is greater than the noise discrimination threshold and less than or equal to the edge discrimination threshold, the pixel is determined to be located in a weak texture region and marked as a transition region; if the anisotropic coherence is greater than the edge discrimination threshold, the pixel is determined to be located in a strong edge or texture region and marked as a feature region. S25. Based on the structural attribute discrimination results, assign a corresponding confidence weight to each pixel and draw a grayscale map to obtain the structural confidence map.

5. The method of ultrasound elastography image denoising and enhancement of claim 4, wherein: Step three includes: S31. Obtain the Rayleigh distribution or Fisher-Tipett distribution parameters of ultrasonic speckle noise from historical databases or phantom experimental data, and establish a statistical model of speckle noise. S32. Construct an anisotropic diffusion partial differential equation as an elastic image evolution model, which includes a diffusion tensor term and a time step term; S33. Based on the feature vector directions in the structural confidence map, construct the first... The diffusion tensor of the next iteration is made so that the diffusion direction is parallel to the texture direction and perpendicular to the gradient direction, in order to protect edge features; S34. Calculate the ratio of the local variance of the current image to the variance of the pre-selected reference uniform region, and construct the first... The equivalent number of views in each iteration is used as an evaluation index for the degree of noise suppression.

6. The method of ultrasound elastography image denoising and enhancement of claim 5, wherein: Step three also includes: S35. Utilize the Laplacian operator to detect changes in high-frequency components in an image, and construct the first... The edge preservation index of the next iteration; S36. Process the image Filtering transformation is used to extract the energy features of subbands in different directions, and the first subband is constructed. Texture energy consistency factor in the next iteration; S37, extracting the diffusion tensor, the equivalent view number, the edge preserving index and the texture energy consistency factor obtained in S33-S36, performing weighted fusion to obtain the first iteration local smoothing coefficient, and dynamically adjusting the diffusion strength of the diffusion equation according to the coefficient to perform evolution update on the image of the second iteration.

7. The method of ultrasound elastography image denoising and enhancement of claim 6, wherein: Step three also includes: S38, Calculate the first The elastic image matrix after the second correction and the first Root mean square error between sub-image matrices; S39. Preset a convergence threshold and compare the root mean square error (RMSE) with the convergence threshold: if the RMS error is greater than the convergence threshold, then let... The structure confidence map is corrected using the updated image data, and the process returns to step S33 for the next iteration. If the root mean square error is less than or equal to the convergence threshold, the iteration stops, and the current iteration is output. The modified elastic image matrix is ​​used as the final target elastic image.

8. The method of ultrasound elastography image denoising and enhancement of claim 7, wherein: Step three also includes: S40. Perform quality assessment on the final output target elastic image to obtain image quality grading results, including: calculating the structural similarity index (SSIM) between the target elastic image and the initial elastic distribution matrix; if SSIM is greater than 0.9, the quality grading result is labeled as excellent, and a storage or display instruction is executed; if SSIM is greater than 0.75 and less than or equal to 0.9, the quality grading result is labeled as good, an edge sharpening compensation instruction is generated, and the target elastic image is post-processed for sharpening; if SSIM is less than or equal to 0.75, the quality grading result is labeled as poor, a parameter callback instruction is generated, the noise discrimination threshold in step S23 is reduced, and the process returns to step two to update the structural confidence map, and then step three is re-executed.

Images