Ultrasonic image processing system based on image enhancement

Abstract

The invention relates to the technical field of medical image processing, in particular to an ultrasonic image processing system based on image enhancement, which comprises a field construction step: constructing a pixel-level structure tensor matrix, and mapping original ultrasonic image data into tensor field data; a flow direction locking step: performing feature space decomposition on the tensor field data to generate a diffusion guide graph containing a main feature vector; a nonlinear evolution step: constructing an anisotropic diffusion tensor based on characteristic value information, and executing time step iterative solution based on a partial differential equation; a reunion output step: performing boundary continuity reunion on the enhanced image data; according to the method, the direction of the tissue is accurately locked through the Riemannian metric tensor field, smooth texture smoothing and inverse texture sharpening are achieved, speckle noise is greatly restrained, and meanwhile the microstructure features are reserved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an ultrasound image processing system based on image enhancement. Background Technology

[0002] In the application of ultrasound imaging in clinical diagnosis, the imaging system relies on clear tissue texture and continuous boundary features to ensure the accuracy of lesion identification. The back-end processing unit usually needs to combine the raw echo signal acquired by the probe with image enhancement algorithms to optimize visual quality in real time.

[0003] For the denoising and enhancement of ultrasound images, existing solutions generally adopt a scalar filtering architecture based on Euclidean space. This means that the image data is directly treated as a set of isolated gray values. Linear filters such as Gaussian smoothing and median filtering or conventional gradient operators are used to extract edge intensity. Noise and tissue structure are distinguished directly based on the brightness difference of pixels, and denoising is performed based on this. Although this solution is feasible for processing optical images with high signal-to-noise ratio, it is limited to scalar space and lacks the ability to describe local geometric flow. When encountering ultrasound images full of multiplicative speckle noise, the basic algorithm is very likely to misidentify high-frequency speckle noise as tissue edges, resulting in blurring of microvascular walls and fibrous tissue while smoothing noise.

[0004] Furthermore, the calculation method based directly on gray-level gradients is extremely sensitive to single-point noise interference, making it difficult to generate full-rank structural descriptions. It also lacks anisotropic diffusion constraints, which prevents the processing from performing directional enhancement along the texture direction. Moreover, it cannot solve the phenomenon of tissue boundary breakage caused by ultrasound signal attenuation or probe angle issues, making it difficult to support doctors in performing high-precision continuous observation of small lesion areas.

[0005] Therefore, how to establish a processing mechanism based on the Riemannian metric tensor field, which can effectively distinguish between speckle noise and linear texture while achieving anisotropic evolution along the texture direction and geometric repair of fracture boundaries, has become an urgent technical problem to be solved.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention discloses an ultrasound image processing system based on image enhancement. Specifically, the technical solution of this invention is as follows:

[0008] It includes a parallel computing unit, which is communicatively connected to a field construction module, a flow direction locking module, a nonlinear evolution module, and a re-aggregation output module;

[0009] The field construction module is configured to receive raw ultrasound image data and map the raw ultrasound image data into tensor field data describing local geometric features by constructing a pixel-level structure tensor matrix.

[0010] The flow direction locking module is configured to perform feature space decomposition on the tensor field data, extract the main feature vector describing the direction of tissue texture and the feature value information describing the gradient direction, and generate a diffusion guide map containing the main feature vector and the feature value information.

[0011] The nonlinear evolution module is configured to construct an anisotropic diffusion tensor based on the eigenvalue information, and use the anisotropic diffusion tensor to perform time-step iterative solution based on partial differential equations on the original ultrasound image data to generate the evolved enhanced image data.

[0012] The reconvergence output module is configured to perform boundary continuity reconstruction on the evolved enhanced image data and output the final ultrasound image result.

[0013] Preferably, the process of generating tensor field data by the field construction module includes:

[0014] Obtain the horizontal and vertical gradient components of each pixel in the raw ultrasound image data;

[0015] Calculate the outer product of the horizontal and vertical gradient components to generate the initial structure tensor matrix;

[0016] The initial structure tensor matrix is ​​convolved using a preset Gaussian smoothing kernel to obtain a local structure tensor matrix that integrates neighborhood geometric information.

[0017] Using the local structure tensor matrix as a Riemannian metric, tensor field data covering the entire image domain is constructed.

[0018] Preferably, the process of the flow direction locking module extracting the main feature vector and feature value information includes:

[0019] Perform eigenvalue decomposition on the local structure tensor matrix corresponding to each pixel in the tensor field data.

[0020] Obtain the maximum and minimum eigenvalues ​​of the local structure tensor matrix, as well as the principal eigenvectors corresponding to the maximum eigenvalues ​​and the secondary eigenvectors corresponding to the minimum eigenvalues;

[0021] The largest and smallest eigenvalues ​​are labeled as eigenvalue information, the principal eigenvectors are labeled as the main gradient direction, and the secondary eigenvectors are labeled as the texture tangential direction.

[0022] Preferably, the process by which the nonlinear evolution module constructs the anisotropic diffusion tensor includes:

[0023] Calculate the square of the difference between the largest eigenvalue and the smallest eigenvalue, and label the calculated square value as the local coherence factor;

[0024] A preset coherence threshold is obtained, and the local coherence factor is compared with the coherence threshold to determine the geometric structure type of the pixel.

[0025] The configuration is as follows: if the local coherence factor is greater than the coherence threshold, then the pixel is determined to be in a linear texture region;

[0026] If the local coherence factor is less than or equal to the coherence threshold, then the pixel is determined to be in a speckle noise region or a uniform region.

[0027] Preferably, the process by which the nonlinear evolution module configures the diffusion coefficient according to the geometric structure type includes:

[0028] In response to determining that the pixel is in a linear texture region, a preset first diffusion coefficient is set in a direction parallel to the texture tangency, and a preset second diffusion coefficient is set in a direction parallel to the gradient principal direction, wherein the first diffusion coefficient is greater than the second diffusion coefficient.

[0029] In response to determining that the pixel is in a speckle noise region or a uniform region, the same preset third diffusion coefficient is set in all directions, or the diffusion coefficient in all directions is set to zero.

[0030] Based on the first diffusion coefficient, the second diffusion coefficient, or the third diffusion coefficient, the anisotropic diffusion tensor corresponding to the pixel is assembled and generated.

[0031] Preferably, the time-step iterative solution process performed by the nonlinear evolution module includes:

[0032] Step 1: Substitute the anisotropic diffusion tensor into the pre-defined divergence partial differential equation to construct the evolution operator at the current time step;

[0033] Step 2: Calculate the brightness change of the original ultrasound image data at the current time step using the evolution operator, add the brightness change to the current image data, and update the image state;

[0034] Step 3: Determine whether the cumulative number of iterations has reached the preset iteration limit. If not, return the updated image state as input data to the flow locking module to recalculate the tensor field features. If the limit has been reached, stop the iteration and output the evolved enhanced image data.

[0035] Preferably, the parallel computing unit performs an iterative solution process including:

[0036] The original ultrasound image data and the tensor field data are loaded into the GPU memory;

[0037] An independent computation thread is allocated to each pixel, and the construction of the local structure tensor matrix, eigenvalue decomposition, and assembly of the anisotropic diffusion tensor are performed in parallel within the computation thread.

[0038] By using shared memory to exchange neighborhood pixel data, the spatial derivative of divergence-type partial differential equations can be calculated in parallel.

[0039] Preferably, the process by which the reconvergence output module reconstructs the boundary continuity of the evolved enhanced image data includes:

[0040] Extract the boundaries of high-echo regions from the evolved enhanced image data;

[0041] Detect the breakpoints on the boundary and calculate the principal directions of the local structural tensors at the breakpoints;

[0042] Pixel-level interpolation is performed along the principal direction of the local structural tensor to connect the fracture points, repairing the discontinuity of tissue texture caused by signal attenuation, and generating the final ultrasound image result.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This invention constructs a tensor field mapping mechanism based on Riemannian metric, which effectively solves the technical problem that traditional Euclidean space scalar filtering is difficult to describe local geometric flow direction; unlike existing technologies that rely directly on isolated calculation of gray-level gradients, this scheme integrates neighborhood geometric information through Gaussian convolution to generate a full-rank structure tensor, which can correctly reflect the texture direction even under single-point noise pollution, and improves the robustness and structural analysis capability of the algorithm in low signal-to-noise ratio ultrasonic environment;

[0045] 2. This invention introduces a local coherence determination and adaptive threshold segmentation strategy based on eigenvalue difference, overcoming the limitation of conventional gradient magnitude method in effectively distinguishing strong speckle noise from weak tissue signals. By calculating the difference of eigenvalues ​​and combining dynamic range compression and adaptive threshold, this method can accurately identify circularly distributed speckle noise regions and linearly distributed tissue texture regions based on geometric essence, avoiding misjudgment caused by excessive noise gradient magnitude, and achieving high-precision locking of microvascular walls and fibrous tissue.

[0046] 3. This invention establishes a nonlinear anisotropic evolution model of flow along the texture, realizing dual control of smoothing and sharpening in the image enhancement process; the system intelligently configures the diffusion tensor according to the geometric structure type, performs uniform smoothing in speckle noise areas, and strictly diffuses along the tangential direction and prohibits crossing the boundary in linear texture areas, thereby effectively suppressing speckle noise while preserving clear anatomical structure features. Combined with the hardware acceleration of the parallel computing unit, it meets the stringent requirements of clinical diagnosis for real-time high frame rate imaging.

[0047] 4. This invention employs a boundary continuity reconstruction and interpolation repair technique based on tensor field guidance, which effectively solves the problem of tissue boundary breakage caused by ultrasound signal attenuation. By extracting the skeleton of the hyperechoic region and using the local tensor principal direction at the break point for geometric locking, this scheme uses an interpolation algorithm to construct a smooth curve that meets the continuity requirements to connect the break points, realizing physical-level signal reconstruction that conforms to anatomical characteristics, and providing doctors with continuous and complete visual evidence for observing small lesion areas. Attached Figure Description

[0048] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0049] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0050] 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.

[0051] Example 1:

[0052] Please see Figure 1 An ultrasound image processing system based on image enhancement includes a parallel computing unit, which is communicatively connected to a field construction module, a flow direction locking module, a nonlinear evolution module, and a refocusing output module.

[0053] The field construction module is configured to receive raw ultrasound image data and map the raw ultrasound image data into tensor field data describing local geometric features by constructing a pixel-level structure tensor matrix.

[0054] The flow direction locking module is configured to perform feature space decomposition on the tensor field data, extract the main feature vector describing the direction of tissue texture and the feature value information describing the gradient direction, and generate a diffusion guidance map containing the main feature vector and feature value information.

[0055] The nonlinear evolution module is configured to construct an anisotropic diffusion tensor based on eigenvalue information, and use the anisotropic diffusion tensor to perform time-step iterative solution based on partial differential equations on the original ultrasound image data to generate the evolved enhanced image data.

[0056] The reconvergence output module is configured to reconstruct the boundary continuity of the evolved enhanced image data and output the final ultrasound image result.

[0057] This embodiment details the overall architecture and core processing logic of the system. The system is configured to address the technical challenge of distinguishing between inherent speckle noise and tissue texture in Euclidean space during ultrasound imaging. The core of this embodiment lies in mapping image processing from conventional scalar space to Riemannian manifold space, utilizing tensor fields to describe the geometric continuity of tissues. Prior to this, to ensure the rigor of the symbol system, the following core variables are defined: It originates from gradient calculation and its physical meaning is the local structure tensor matrix; It originates from eigenvalue assembly and its physical meaning is the anisotropic diffusion tensor;

[0058] The system uses the field construction module to transform scalar pixels into geometric descriptors containing directional information. At this point, the image is no longer a collection of gray values, but a field composed of two-dimensional tensors. Its geometric shape directly reflects whether the point belongs to a flat region, an edge region, or a corner region. The flow direction locking module parses the geometric parameters of the above tensors to generate a diffusion guidance map, which is essentially a micro-navigation map, to indicate which path the subsequent diffusion process should follow.

[0059] Based on this, the nonlinear evolution module simulates the controlled heat conduction process to ensure that brightness information can only flow along the direction of tissue texture and cannot diffuse perpendicular to the direction of texture; the refocusing output module repairs tissue boundary breaks caused by ultrasound signal attenuation or probe angle issues by utilizing the continuity characteristics of the tensor field.

[0060] The system constructed in this embodiment can accurately lock the tissue orientation in a strong noise environment through the Riemann metric tensor field, achieving the dual effects of smoothing along the texture and sharpening against the texture. While greatly suppressing speckle noise, it preserves the structural features of microvascular walls and fibrous tissue, effectively solving the frequency domain aliasing problem.

[0061] Example 2:

[0062] The process of generating tensor field data by the field construction module includes:

[0063] Obtain the horizontal and vertical gradient components of each pixel in the raw ultrasound image data;

[0064] Calculate the outer product of the horizontal and vertical gradient components to generate the initial structure tensor matrix;

[0065] The initial structure tensor matrix is ​​convolved using a preset Gaussian smoothing kernel to obtain a local structure tensor matrix that integrates neighborhood geometric information.

[0066] Using the local structure tensor matrix as a Riemannian metric, tensor field data covering the entire image domain is constructed.

[0067] This embodiment specifies the process of generating tensor field data by the field construction module; the system acquires the horizontal gradient component of each pixel in the original ultrasound image data. and vertical gradient components The process preferably employs the Sobel operator or the central difference method to suppress high-frequency noise interference; the outer product of the horizontal gradient components and the vertical gradient components is calculated to generate the initial structure tensor matrix. Using a preset Gaussian smoothing kernel Convolution operations are performed on the initial structure tensor matrix to obtain a local structure tensor matrix that integrates neighborhood geometric information. The formula is as follows:

[0068]

[0069] in, : Derived from preset parameters, the physical meaning is scale parameters. A Gaussian kernel is used to incorporate neighborhood context information; to maintain the positioning accuracy of fine structures while smoothing noise, in this embodiment, the scale parameter... The preferred setting is between 1.0 and 2.0, and the Gaussian kernel window size is set to... The window parameter, with a radius of 2 pixels, is set based on a standard ultrasound image spatial resolution of 10 pixels / mm. In practical applications, the system obtains the actual spatial resolution based on the current probe frequency and depth settings, and scales the window radius proportionally to maintain a constant physical scale. Furthermore, the kernel function coefficients are normalized to ensure energy conservation. The physical meaning is convolution operator; the system converts the local structure tensor matrix... As a Riemannian metric, tensor field data covering the entire image domain is constructed;

[0070] Directly using the gradient outer product is easily affected by single-point noise because its rank is 1 and its information content is singular. This embodiment introduces Gaussian convolution to generate a full-rank matrix, so that even if a pixel itself is contaminated by noise, as long as the surrounding pixels show a consistent linear structure, the structure tensor of that point can still correctly reflect the texture direction, thereby greatly improving the robustness of the algorithm in low signal-to-noise ratio environments.

[0071] Example 3:

[0072] The process of the flow direction locking module extracting the main feature vector and feature value information includes:

[0073] Perform eigenvalue decomposition on the local structure tensor matrix corresponding to each pixel in the tensor field data.

[0074] Obtain the maximum and minimum eigenvalues ​​of the local structure tensor matrix, as well as the principal eigenvectors corresponding to the maximum eigenvalues ​​and the secondary eigenvectors corresponding to the minimum eigenvalues;

[0075] The largest and smallest eigenvalues ​​are labeled as eigenvalue information, the principal eigenvectors are labeled as the main gradient direction, and the secondary eigenvectors are labeled as the texture tangential direction.

[0076] This embodiment specifies the process for the flow direction locking module to extract geometric features; the system defines the local structure tensor matrix corresponding to each pixel in the tensor field data. Eigenvalue decomposition is performed. To ensure computational efficiency, especially for subsequent GPU parallel processing, this embodiment abandons the computationally intensive iterative solver and instead uses a closed-form analytical solution to directly calculate the eigenpairs of the 2x2 symmetric matrix. Let the local structure tensor matrix be... ,in, These correspond to the smoothed local structure tensor components;

[0077] The system directly calculates the eigenvalues ​​using the following algebraic formula. :

[0078]

[0079] in, Take the value corresponding to the plus sign. Take the value corresponding to the minus sign; in physical terms, The energy representing the direction of the most drastic local gradient change. Energy representing the direction of minimum gradient change;

[0080] For the feature vector, the system calculates the principal direction angle. :

[0081]

[0082] Directly construct the normalized eigenvector basis using trigonometric functions:

[0083]

[0084] The system performs semantic tagging, and and Marked as feature value information, Marked as the principal gradient direction, perpendicular to the texture, and... Marked as the texture tangential direction, along the texture direction; this analytical solution-based processing method ensures the direct portability of the algorithm in the shader language and avoids dependence on external linear algebra libraries.

[0085] Example 4:

[0086] The process of constructing anisotropic diffusion tensors using the nonlinear evolution module includes:

[0087] Calculate the square of the difference between the largest and smallest eigenvalues, and label the calculated square value as the local coherence factor;

[0088] Obtain a preset coherence threshold, compare the local coherence factor with the coherence threshold, and determine the geometric structure type of the pixel.

[0089] The configuration is as follows: if the local coherence factor is greater than the coherence threshold, then the pixel is determined to be in a linear texture region.

[0090] If the local coherence factor is less than or equal to the coherence threshold, the pixel is determined to be in a speckle noise region or a uniform region.

[0091] This embodiment specifies the logic for determining the geometric structure type; the nonlinear evolution module calculates the maximum eigenvalue. with minimum eigenvalue The square of the difference generates the local coherence factor. The formula is as follows:

[0092]

[0093] in, Derived from eigenvalue difference calculations, its physical meaning is an index that quantifies the degree of anisotropy in local geometric structure; it should be noted that, due to eigenvalues... The original value is proportional to the square of the pixel gradient. The values ​​typically have a very large dynamic range, potentially reaching [value missing] in 8-bit images. Directly comparing the magnitude of the grayscale value with a standard grayscale threshold can lead to numerical overflow or logical failure. Before comparison, the system must perform dynamic range compression and normalization. Specifically, the system uses the following normalization formula to... Mapped to the [0, 255] interval:

[0094]

[0095] in, To adjust the dynamic range using an adaptive saturation constant based on global image statistical properties, For gamma correction coefficients; for parameters The determination of this value is not based on randomly selected empirical values, but on physical constants obtained from standard ultrasonic phantom experiments; specifically, due to the local coherence factor... Defined as the square of the difference in eigenvalues, i.e. the fourth power of the gradient energy, its numerical distribution spans a wide range; in order to map it to the linear receptive domain that conforms to the Weber-Fechner law of the human eye, nonlinear compression is required.

[0096] Multiple comparative tests were conducted on the CIRS040GSE standard phantom to determine the optimal contrast-to-noise ratio (CNR) between the cyst area and the background tissue, with the goal of maximizing this ratio. The value is 0.5; this value physically corresponds to nonlinear compression, which aims to map high-order energy signals to a linear response range suitable for human visual perception; therefore, in practical applications, The value is fixed at 0.5 and does not change with the image content; at the same time, in order to strictly meet the reproducibility requirements of the algorithm on different gain ultrasound devices and avoid using vague empirical values, this embodiment clearly defines... The calculation formula is for all pixels in the current frame image. Arithmetic mean of the values:

[0097]

[0098] in, and These represent the width and height of the currently processed image frame, respectively, in pixels. For continuous video stream processing scenarios, to prevent parameter flickering caused by abrupt changes in the statistical characteristics of a single frame, The calculation can preferably employ a moving average strategy in the time domain to enhance display stability; this adaptive calculation mechanism ensures... Always positioned at the energy center of the histogram, it plays a crucial role in providing a soft threshold and preventing... The screen displays either pure black or pure white saturation; the aforementioned preset coherence threshold... Based on this normalized feature space Set up;

[0099] The system obtains the preset coherence threshold. The algorithm is then compared and judged; to further improve its adaptability to ultrasound images with different gains, the Otsu's maximum inter-class variance method is used based on the current frame. Histogram automatically calculates adaptive threshold To replace fixed empirical values; in response to Greater than The system determines that the pixel is located in a linear tissue texture region, such as a blood vessel wall or muscle fiber, indicating a clear directionality; in response to Less than or equal to The system determines that the pixel is in a speckle noise region or a uniform region, indicating that the local structure is isotropic.

[0100] Introducing local coherence factor This method effectively overcomes the limitations of conventional gradient magnitude methods. This embodiment uses the square of the eigenvalue difference as a criterion, cleverly utilizing the geometric nature that speckle noise is distributed in a circular pattern while tissue texture is distributed in a linear pattern. This enables effective differentiation between strong speckle noise and weak tissue signals, avoiding misjudgment caused by excessively large noise gradient magnitudes.

[0101] Example 5:

[0102] The process by which the nonlinear evolution module configures the diffusion coefficient based on the geometry type includes:

[0103] In response to determining that a pixel is in a linear texture region, a preset first diffusion coefficient is set in the direction parallel to the texture tangency, and a preset second diffusion coefficient is set in the direction parallel to the gradient principal direction, wherein the first diffusion coefficient is greater than the second diffusion coefficient.

[0104] In response to determining whether a pixel is in a speckle noise region or a uniform region, set the same preset third diffusion coefficient in all directions, or set the diffusion coefficient in all directions to zero.

[0105] Based on the first diffusion coefficient, the second diffusion coefficient, or the third diffusion coefficient, the anisotropic diffusion tensor corresponding to the pixel is assembled and generated.

[0106] This embodiment elaborates on the diffusion tensor. The system's construction logic; in response to determining that a pixel is located in a linearly organized texture region, the system operates in a direction parallel to the texture tangent. Set a larger first diffusion coefficient. And parallel to the principal gradient direction A very small second diffusion coefficient is set on top. ,make sure Significantly greater than To ensure the stability and anisotropic effect of the diffusion process, the diffusion coefficient must satisfy specific numerical constraints; specifically, the first diffusion coefficient... A value of 1.0 is recommended to maximize the smoothing effect; second diffusion coefficient A value between 0.01 and 0.05 is recommended to minimize cross-boundary diffusion. This configuration aims to ensure that smoothing operations are performed along the vessel wall, while prohibiting smoothing across the vessel wall. Specifically, in response to determining whether a pixel is in a speckle noise region or a uniform region, the system sets the same preset third diffusion coefficient in all directions. It is recommended that the value be between 0.5 and 0.8, or the coefficient be set to zero to preserve the original details; based on the above coefficients, an anisotropic diffusion tensor is generated. The formula is as follows:

[0107]

[0108] in, : Derived from local structure tensor decomposition, its physical meaning is eigenvector basis; : Derived from logical judgment configuration, its physical meaning is diffusion characteristic value;

[0109] By constructing this intelligent fluid-like diffusion tensor, this embodiment achieves adaptive image evolution; when encountering a blood vessel wall, the diffusion flow direction automatically changes to flow along the vessel wall, while it is uniformly smoothed when encountering speckle noise; this Riemannian-based guiding mechanism solves the problem of blurred boundaries in traditional Gaussian filtering, ensuring the edge sharpness of the enhanced image.

[0110] Regarding the key diffusion coefficient involved in the embodiments, this embodiment further clarifies its numerical source and acquisition method: the coefficient is not an arbitrarily selected empirical value, but is obtained based on the calibration experiment of the CIRS040GSE standard ultrasound phantom; specifically:

[0111] First diffusion coefficient Normalized to 1.0 as the baseline flow rate;

[0112] Second diffusion coefficient The range of values ​​is determined based on the mathematical constraint that the anisotropy ratio must be greater than 20:1 to ensure that the lateral diffusion flux approaches zero; the preferred value of 0.02 is the optimal calibration solution that makes the contrast noise ratio (CNR) of the vascular simulation structure reach its peak in the phantom test.

[0113] Third diffusion coefficient This is achieved by calculating the speckle autocorrelation length in a uniform background region. This is based on statistical analysis; specifically, the system selects the image with the smallest local variance. The pixel region is treated as a uniform background region. The normalized autocorrelation function of this region is calculated, and the autocorrelation coefficient is reduced to [value missing]. The spatial lag distance corresponding to the time is defined as The specific mapping calculation formula is as follows:

[0114]

[0115] in, This is a preset calibration constant, with its physical unit being pixels. In this embodiment, it is preferably set to 1.0 to ensure that when the autocorrelation length is large, i.e. the speckle particles are large, a larger diffusion coefficient is used to enhance the smoothing force, thereby effectively smoothing out coarse speckle noise; conversely, when the autocorrelation length is small, a smaller diffusion coefficient is used, thereby effectively suppressing fine speckle noise; this formula aims to match the natural spatial correlation of background noise.

[0116] Example 6:

[0117] According to the image enhancement-based ultrasound image processing system of Example 5, the process of the nonlinear evolution module performing time-step iterative solution includes:

[0118] Step 1: Substitute the anisotropic diffusion tensor into the pre-defined divergence partial differential equation to construct the evolution operator at the current time step;

[0119] Step 2: Calculate the brightness change of the original ultrasound image data at the current time step using the evolution operator, add the brightness change to the current image data, and update the image state;

[0120] Step 3: Determine whether the cumulative number of iterations has reached the preset iteration limit. If not, return the updated image state as input data to the flow locking module to recalculate the tensor field features. If the limit has been reached, stop the iteration and output the evolved enhanced image data.

[0121] This embodiment specifies the detailed execution process of the time-step iterative solution; the module uses the anisotropic diffusion tensor Substituting into the divergence-type partial differential equation, an evolution operator is constructed; an explicit difference scheme is used to calculate the brightness change and update the image state, as shown in the following formula:

[0122]

[0123] in, : Derived from iterative calculation, its physical meaning is the first Image brightness after the next iteration; Derived from iterative control, its physical meaning is the current iteration number in the evolution process, and it is dimensionless; Derived from vector calculus, its physical meaning is the gradient operator, which acts on a scalar image. To obtain the first-order derivative vector in space; The physical meaning is divergence operator, used to calculate the degree of divergence of a flux field; Derived from preset parameters, its physical meaning is the time step in the virtual evolution process, dimensionless; for the numerical stability of explicit difference solutions, the time step... It must be less than 0.25 to satisfy the stability criterion; in this embodiment, The preferred setting is 0.15;

[0124] To ensure the accuracy of the numerical solution and avoid mesh artifacts, in specific implementations, the aforementioned divergence term... The discretization employs a nine-point difference scheme; this is because the anisotropic diffusion tensor Including non-zero mixed derivative terms, i.e., off-diagonal elements, the standard five-point difference scheme is insufficient to accurately describe the anisotropic characteristics after rotation. To meet the requirements of code-level reproducibility and avoid making the calculation of mixed derivatives a black box description, this embodiment explicitly defines a specific discretized flux calculation formula: using the eastern interface of the computational grid, i.e., the coordinate... flux at For example, the calculation formula is as follows:

[0125]

[0126] in, For pixels and The arithmetic mean of the diffusion tensor at that point; that is This operation involves summing and averaging the elements at corresponding positions in the matrix.

[0127] The explicit definitions of parameter subscripts in the formula are as follows: Definition It is a second-order symmetric matrix ,but The diagonal element corresponding to the first row and first column of the matrix has the physical meaning of the diffusion coefficient along the x-axis. The off-diagonal element in the first row and second column of the corresponding matrix physically represents the cross-coupling diffusion coefficient between the x-axis and y-axis; the second term in this formula precisely quantifies the mixing derivative. In terms of interface contribution, this is achieved by selecting the four adjacent corner pixels on both sides of the interface. This is achieved by using the average difference.

[0128] Similarly, to ensure the precise conservation of anisotropic diffusion in the direction of the mixed derivative, this embodiment explicitly defines the north interface, i.e., the coordinate... flux The calculation formula is as follows:

[0129]

[0130] in, For pixels and The arithmetic mean of the diffusion tensor at that point, the second term of the formula utilizes The average gradient in the direction is used to calculate the contribution of the mixed derivative; the final divergence update is... Let the spatial step size be 1. Here, the spatial step size refers to the pixel spacing, that is, the Euclidean distance between the centers of adjacent pixels, which is normalized to 1 during the discretization process. This processing method ensures that the diffusion flow is transmitted not only in the axial direction, but also in the diagonal direction, thereby avoiding the checkerboard effect commonly found in anisotropic filtering.

[0131] Furthermore, to address the array out-of-bounds issue that may occur when calculating the neighborhood difference of image edge pixels, the system performs calculations... Previously, strict boundary processing was implemented; the image domain was defined as... When index When using the following piecewise mapping function... Read data:

[0132]

[0133] in, For example, for , mapped to ,Right now Although the above clamping operation effectively prevents program crashes and makes the normal derivative zero, in anisotropic diffusion, the mixed derivative term... Non-zero flux may still be introduced at the boundary; in order to strictly satisfy the Neumann boundary conditions, i.e. To ensure the conservation of total system energy, the system performs a forced adiabatic truncation after calculating the initial flux: explicitly forcing the flux values ​​on the four borders of the image to 0, for example, if the current calculation is... West-side flux Then force assignment This physically eliminates the possibility of heat flowing out of the image boundary, preventing energy leakage or artifacts from appearing at the image edge during the iteration process.

[0134] The system determines whether the cumulative number of iterations has reached the preset limit. Meanwhile, in order to strike a balance between noise reduction and detail preservation, the iteration limit is set. The number of iterations is typically set to 10 to 30, with 15 being preferred. This parameter combination ensures that the algorithm's execution time on the GPU is kept within 30ms, meeting the real-time imaging requirement of 30fps. If this requirement is not met, the updated image state is returned to the flow-locking module, and the tensor field features are recalculated. This nonlinear feedback mechanism ensures that the estimation of the structure tensor becomes more accurate as the iteration progresses. If the upper limit is reached, the iteration stops and the result is output.

[0135] This iterative process simulates the physical evolution of anisotropic diffusion. As the time step progresses, speckle noise is rapidly smoothed due to its lack of coherence, while the tissue structure, due to its high coherence and the fact that its diffusion direction is locked in the tangential direction, is not only not blurred, but becomes more continuous and clearer due to the influx of surrounding pixels, thus achieving simultaneous noise suppression and detail enhancement.

[0136] Regarding the time step and number of iterations involved in the embodiments, this embodiment further clarifies the physical definition basis:

[0137] Time step Derivation: The setting of this value strictly follows the Courant-Friedrichs-Lewy stability criterion (CFL) for numerical solutions of parabolic partial differential equations; in the two-dimensional explicit difference scheme, the theoretical stability limit is... After normalization; in order to offset the discretization truncation error while ensuring convergence, this embodiment introduces an engineering safety factor of 0.6, i.e. This eliminates the risk of numerical divergence in mathematics.

[0138] Iteration limit The derivation is as follows: This range is derived from quantitative calculations based on real-time time budgets; the system must meet a frame rate of 30fps, i.e., a total processing window of 33ms; actual measurements show that on a standard computing unit, such as a 1024-core GPU, the average time for a single iteration is approximately 2.0ms; therefore, the physical limit is... The optimal value of 15 iterations represents the maximum iteration depth achievable while reserving 3ms for data transmission and display, and also satisfies the requirement of residual convergence. The minimum number of steps required to scale up to that magnitude.

[0139] Example 7:

[0140] The process of parallel computing units performing iterative solutions includes:

[0141] The raw ultrasound image data and tensor field data are loaded into the GPU memory;

[0142] Each pixel is assigned an independent computation thread, and the construction of the local structure tensor matrix, eigenvalue decomposition, and assembly of the anisotropic diffusion tensor are performed in parallel within the computation thread.

[0143] By using shared memory to exchange neighborhood pixel data, the spatial derivative of divergence-type partial differential equations can be calculated in parallel.

[0144] This embodiment specifies the hardware acceleration implementation of the parallel computing unit; the system loads the original ultrasound image data and the tensor field data generated by intermediate calculations into the GPU memory; an independent computing thread is allocated for each pixel, and the construction of the local structure tensor matrix, the closed-form solution eigenvalue decomposition, and the assembly of the anisotropic diffusion tensor are performed in parallel using GPU registers; during this period, in order to optimize the calculation of the spatial derivative of the divergence partial differential equation, the system uses shared memory to exchange neighborhood pixel data and preloads image blocks into shared memory to reduce the latency of repeated access to the video memory;

[0145] This embodiment ensures the real-time performance of the algorithm through GPU parallelization and shared memory optimization. Faced with the massive floating-point operations generated by pixel-by-pixel tensor decomposition and PDE solving, this system can keep the processing latency within an extremely low range, meeting the real-time display requirements of high frame rate ultrasound equipment, and enabling doctors to obtain a latency-free enhanced imaging experience during the scanning process.

[0146] Example 8:

[0147] The reconstructing output module performs boundary continuity reconstruction on the evolved enhanced image data, including the following steps:

[0148] Extract the boundaries of high-echo regions from the evolved enhanced image data;

[0149] Detect breakpoints on the boundary and calculate the principal directions of the local structural tensor at the breakpoints;

[0150] Pixel-level interpolation is performed along the principal direction of the local structural tensor to connect the fracture points, repairing the discontinuity of tissue texture caused by signal attenuation and generating the final ultrasound image result.

[0151] This embodiment specifically defines the boundary continuity reconstruction process; the module extracts the boundaries of high-echo regions in the evolved image; specifically, the system performs adaptive binarization processing on the enhanced image data, using the Otsu algorithm or setting a brightness threshold, where the high-echo region is defined as the region in the image whose gray value is higher than a specific threshold, such as the first 15% of the bright area, and marks the pixels as foreground or background; then, morphological closing operations are performed on the foreground region to fill internal holes, and the Zhang-Suen thinning algorithm is used to extract the skeleton of a single pixel width, and morphological pruning operations are performed to remove burr interference with a length of less than 3 pixels, thereby obtaining the boundaries of the high-echo regions; breakpoints on the boundaries are detected, and here, to eliminate algorithmic ambiguity, the specific detection operator is clarified: calculate the skeleton pixels. The total number of non-zero pixels in the 8-neighborhood ,like If so, then that point is determined to be the break point. ;

[0152] The system calculates the principal directions of the local structural tensor at the break point. That is, texture tangent ; for the inherent features of the feature vector directional ambiguity, i.e. and Representing the same axis, the system performs a geometric orientation locking step: obtaining the relationship between the skeleton and the break point. Directly connected preceding pixels That is, the unique non-zero pixel within the above 8 neighborhoods, constructing the geometric extension vector. and correct The symbol ensures This ensures that the search direction points to the external region of the fractured skeleton; based on this, the system follows the corrected... The direction searches for the nearest neighbor of the same type of breakpoint within the preset sector. In order to convert the preset sector search into executable code logic, this embodiment sets a search radius. A pixel corresponds to a physical scale of approximately 1.5 mm. This parameter is also dynamically adjusted with the image DPI, and the angle threshold... System traversal Centered Candidate points within the neighborhood Construct the displacement vector If candidate points Simultaneously satisfying the following two geometric constraints: Euclidean distance determination: ; Directional consistency determination: Then mark it as a valid candidate point;

[0153] To overcome the potential for erroneous connections based solely on distance matching, such as misconnecting two parallel blood vessels, this embodiment introduces a bidirectional tangent consistency constraint and a curvature cost function; for any valid candidate point The system obtains the texture tangent at that point. And calculate the connection vector. and The cosine value of the included angle; only when At that time, a candidate point is retained only if the connecting line smoothly cuts into the target point; calculate the following structural continuity cost function. :

[0154]

[0155] in, This is the curvature penalty coefficient, with physical units of pixels per radian, and a preferred value of 5.0 pixels per radian; system selection The point with the smallest number of connections (1) is selected as the target breakpoint. Furthermore, to prevent erroneous connections at blood vessel bifurcation points, the system further calculates the cost difference between the optimal candidate point and the second-best candidate point. ,like If the value is less than the preset ambiguity threshold, it is determined that there is structural ambiguity, and the automatic connection operation for that breakpoint is abandoned; if no candidate point that meets the conditions is found or there is structural ambiguity, the breakpoint is skipped to avoid incorrect connection.

[0156] To address the mathematical singularity of conventional interpolation and ensure perfect alignment of the curve with the tangents of the textures on both sides, this embodiment explicitly employs Hermite interpolation to construct a parameterized cubic spline curve; to determine the coefficients of the interpolation polynomial, the system defines the magnitude parameter of the tangent vector. The value is the Euclidean distance between the two points. The unit is pixels, thus constructing the control tangent vector. and Correction is required. The direction is set so that the dot product of the displacement vector is positive;

[0157] Final interpolation curve It is synthesized from the following explicit basis functions:

[0158]

[0159] in, System traversal The value ranges from 0 to 1, with a step size set to... To ensure the sampling density meets the Nyquist criterion and prevent connection drops, calculate the corresponding coordinate points. And fill in pixels; in order to achieve physical reconstruction of the echo signal, fill in brightness values. Two-endpoint linear interpolation is used:

[0160]

[0161] The calculated Write the result image At the coordinates; this formula ensures that the generated connection curve is at... The tangent at that point is ,exist The tangent at that point is Furthermore, the brightness transitions smoothly, thus mathematically ensuring that the repaired boundary conditions are met. Continuity ensures a smooth closure that conforms to anatomical features.

[0162] 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 image processing system based on image enhancement, characterized in that, It includes a parallel computing unit, which is communicatively connected to a field construction module, a flow direction locking module, a nonlinear evolution module, and a re-aggregation output module; The field construction module is configured to receive raw ultrasound image data and map the raw ultrasound image data into tensor field data describing local geometric features by constructing a pixel-level structure tensor matrix. The flow direction locking module is configured to perform feature space decomposition on the tensor field data, extract the main feature vector describing the direction of tissue texture and the feature value information describing the gradient direction, and generate a diffusion guide map containing the main feature vector and the feature value information. The nonlinear evolution module is configured to construct an anisotropic diffusion tensor based on the eigenvalue information, and use the anisotropic diffusion tensor to perform time-step iterative solution based on partial differential equations on the original ultrasound image data to generate the evolved enhanced image data. The reconvergence output module is configured to perform boundary continuity reconstruction on the evolved enhanced image data and output the final ultrasound image result.

2. The ultrasound image processing system based on image enhancement according to claim 1, characterized in that, The process of generating tensor field data by the field construction module includes: Obtain the horizontal and vertical gradient components of each pixel in the raw ultrasound image data; Calculate the outer product of the horizontal and vertical gradient components to generate the initial structure tensor matrix; The initial structure tensor matrix is ​​convolved using a preset Gaussian smoothing kernel to obtain a local structure tensor matrix that integrates neighborhood geometric information. Using the local structure tensor matrix as a Riemannian metric, tensor field data covering the entire image domain is constructed.

3. The ultrasound image processing system based on image enhancement according to claim 2, characterized in that, The process of the flow direction locking module extracting the main feature vector and feature value information includes: Perform eigenvalue decomposition on the local structure tensor matrix corresponding to each pixel in the tensor field data. Obtain the maximum and minimum eigenvalues ​​of the local structure tensor matrix, as well as the principal eigenvectors corresponding to the maximum eigenvalues ​​and the secondary eigenvectors corresponding to the minimum eigenvalues; The largest and smallest eigenvalues ​​are labeled as eigenvalue information, the principal eigenvectors are labeled as the main gradient direction, and the secondary eigenvectors are labeled as the texture tangential direction.

4. The ultrasound image processing system based on image enhancement according to claim 3, characterized in that, The process by which the nonlinear evolution module constructs the anisotropic diffusion tensor includes: Calculate the square of the difference between the largest eigenvalue and the smallest eigenvalue, and label the calculated square value as the local coherence factor; A preset coherence threshold is obtained, and the local coherence factor is compared with the coherence threshold to determine the geometric structure type of the pixel. The configuration is as follows: if the local coherence factor is greater than the coherence threshold, then the pixel is determined to be in a linear texture region; If the local coherence factor is less than or equal to the coherence threshold, then the pixel is determined to be in a speckle noise region or a uniform region.

5. The ultrasound image processing system based on image enhancement according to claim 4, characterized in that, The process by which the nonlinear evolution module configures the diffusion coefficient according to the geometric structure type includes: In response to determining that the pixel is in a linear texture region, a preset first diffusion coefficient is set in a direction parallel to the texture tangency, and a preset second diffusion coefficient is set in a direction parallel to the gradient principal direction, wherein the first diffusion coefficient is greater than the second diffusion coefficient. In response to determining that the pixel is in a speckle noise region or a uniform region, the same preset third diffusion coefficient is set in all directions, or the diffusion coefficient in all directions is set to zero. Based on the first diffusion coefficient, the second diffusion coefficient, or the third diffusion coefficient, the anisotropic diffusion tensor corresponding to the pixel is assembled and generated.

6. The ultrasound image processing system based on image enhancement according to claim 5, characterized in that, The time-step iterative solution process performed by the nonlinear evolution module includes: Step 1: Substitute the anisotropic diffusion tensor into the pre-defined divergence partial differential equation to construct the evolution operator at the current time step; Step 2: Calculate the brightness change of the original ultrasound image data at the current time step using the evolution operator, add the brightness change to the current image data, and update the image state; Step 3: Determine whether the cumulative number of iterations has reached the preset iteration limit. If not, return the updated image state as input data to the flow locking module to recalculate the tensor field features. If the limit has been reached, stop the iteration and output the evolved enhanced image data.

7. The ultrasound image processing system based on image enhancement according to claim 6, characterized in that, The parallel computing unit performs an iterative solution process, which includes: The original ultrasound image data and the tensor field data are loaded into the GPU memory; An independent computation thread is allocated to each pixel, and the construction of the local structure tensor matrix, eigenvalue decomposition, and assembly of the anisotropic diffusion tensor are performed in parallel within the computation thread. By using shared memory to exchange neighborhood pixel data, the spatial derivative of divergence-type partial differential equations can be calculated in parallel.

8. The ultrasound image processing system based on image enhancement according to claim 6, characterized in that, The process by which the reconvergence output module reconstructs the boundary continuity of the evolved enhanced image data includes: Extract the boundaries of high-echo regions from the evolved enhanced image data; Detect the breakpoints on the boundary and calculate the principal directions of the local structural tensors at the breakpoints; Pixel-level interpolation is performed along the principal direction of the local structural tensor to connect the fracture points, repairing the discontinuity of tissue texture caused by signal attenuation, and generating the final ultrasound image result.

Images