Ultrasonic image denoising method based on variational mode decomposition and local space sparse fusion

Abstract

The invention discloses an ultrasonic image denoising method based on variational mode decomposition and local space sparse fusion, which comprises the following steps of: firstly, adaptively optimizing key parameters of variational mode decomposition by using a grey wolf optimization algorithm to realize stable and efficient decomposition of an ultrasonic image; then, classifying the modal components according to the structural features of the modal components, and implementing differentiated denoising strategies for different types of modals to separate noise and reserve useful information; after modal reconstruction, a sparse expression method based on local space information is further introduced, according to the method, accurate boundary detection is carried out through gradient vector flow and gray scale proportion analysis, self-adaptive partitioning is carried out on an image according to boundary information, and finally sparse reconstruction is carried out through double dictionaries trained for different areas. According to the method, speckle noise in the ultrasonic image can be effectively suppressed, and meanwhile, the capability of keeping the edge and detail information of a tissue structure is remarkably improved, so that the ultrasonic image with higher quality is obtained.

Description

Technical Field

[0001] This invention belongs to the field of ultrasound medical image processing technology, and relates to an image denoising method, specifically an ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion. Background Technology

[0002] Ultrasound imaging technology has a history of several decades in the medical field and occupies an important position in medical imaging technology. Its key advantages as a primary diagnostic tool in medicine lie in its real-time performance, low cost, non-invasive nature, and lack of ionizing radiation. Ultrasound imaging works by emitting ultrasound waves that come into contact with human tissues and organs and are reflected back to the receiving end to form an image. However, due to the non-uniformity of human tissues, a large number of randomly distributed scattering particles cause speckle noise in ultrasound imaging through interference effects, severely affecting image resolution and contrast, and hindering doctors from accurately diagnosing lesions. This noise obscures subtle information in the image, increasing the difficulty of diagnosis, and cannot be completely eliminated simply by adjusting the equipment. Therefore, how to effectively remove noise from medical ultrasound images and further improve image quality has become an important issue.

[0003] Traditional image denoising methods are mainly divided into two categories: spatial domain denoising methods and transform domain denoising methods. Spatial domain denoising methods utilize the correlation between a pixel and its neighboring pixels to remove noise, operating directly on the pixel values ​​of the image, such as median filtering, mean filtering, and Gaussian filtering. However, speckle noise is a multiplicative noise, and its intensity varies with the image's grayscale value, exhibiting nonlinear characteristics related to the image itself. Spatial domain methods cannot effectively model this multiplicative relationship, often leading to over-smoothing or even loss of structure in the image while removing noise, making them unsuitable for suppressing speckle noise in ultrasound images.

[0004] With the development of artificial intelligence, denoising methods based on machine learning and deep learning have emerged, modeling noise distribution and image features through continuous training. Among them, deep learning-based denoising methods have great development potential, but they lack interpretability to some extent. Furthermore, deep learning-based denoising methods require large-scale datasets; insufficient training data or low-quality data can lead to a significant decrease in denoising performance. Although current image denoising methods include various techniques based on spatial domain, transform domain, and machine learning and deep learning, spatial domain methods struggle to handle multiplicative noise, transform domain methods are effective but the choice of transform domain is crucial, and deep learning-based methods rely on large amounts of labeled data, resulting in poor interpretability. Summary of the Invention

[0005] To address the challenge of effectively removing speckle noise in medical ultrasound image processing, this invention provides an ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion. This method overcomes the trade-off between denoising effectiveness and detail preservation inherent in traditional methods by fusing optimized variational mode decomposition with local spatial sparse representation. It effectively suppresses speckle noise in ultrasound images while significantly improving the preservation of tissue structure edges and details, thereby obtaining higher-quality ultrasound images.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] An ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion includes the following steps:

[0008] Step 1: Complete ultrasound image data acquisition, preprocess the data, and analyze noise characteristics. The specific steps are as follows:

[0009] Step 101: Acquire ultrasound images using a high-performance multi-channel programmable ultrasound imaging device and an ultrasound linear array probe with 128 channels. The ultrasound probe emits ultrasound waves and receives echo signals to generate radio frequency (RF) data. The acquired RF data undergoes signal envelope detection, logarithmic compression, noise reduction, and normalization. Bilinear interpolation is then used for resampling to convert the RF data into B-mode images. Finally, normalization is performed to ensure that the brightness range meets the requirements, thereby obtaining a visible ultrasound image of the body membrane. ;

[0010] Step 102: At the image level, a multiplicative noise model is used for description. Speckle noise will exhibit different energy distributions depending on the reflection intensity of the underlying tissue.

[0011]

[0012] in, For the observed image, To represent the true echo intensity, This represents the multiplicative noise intensity; when the scattering region is uniform, Modeled as a random variable following a Rayleigh distribution, speckle noise is divided into low-echo void regions and high-echo bright spot regions;

[0013] Step 2: Perform sparse representation denoising based on local spatial information, and perform boundary detection by combining gradient vector flow and grayscale ratio analysis; perform adaptive image block partitioning based on boundary information; construct a dual dictionary and use K-singular value decomposition and orthogonal matching pursuit algorithm for sparse reconstruction. The specific steps are as follows:

[0014] Step 201: Using a snake-like model, the contour curve gradually converges to the target boundary through energy minimization. The snake-like model's energy function... The definition is as follows:

[0015]

[0016] in, It is the energy inside the curve. It is the energy outside the curve;

[0017] The partial differential equation for the gradient vector flow is:

[0018]

[0019] in, The gradient vector, The edge intensity of the input image. It is the gradient of the edge map. For smoothing parameters, For Laplacian operators;

[0020] Gradient vector flow energy function Energy function as an improved external energy term:

[0021]

[0022] in, This refers to the external energy term guided by gradient vector flow;

[0023] Step 202: Perform pre-screening using grayscale ratio analysis, focusing on images containing low-echo cavity regions. Images of hyperechoic bright spots The specific steps for quickly locating potential target regions within the entire ultrasound image are as follows:

[0024] (1) Regional initial screening: radius is A solid circle is used to cover any possible hollow parts of the main body; the outer area has a radius of [missing information]. The circular area minus the inner ring, The outer ring represents the radius difference between the inner and outer rings, while the inner ring represents the background around the hole. With a given step size, the search box begins its full-image search and, after locating a suspected hole region, starts from the smallest radius. To the maximum radius Gradually increase the inner circle to find the largest circular area that still remains sufficiently dark. The criterion is:

[0025]

[0026] in, The proportion of the black area. Number of pixels Grayscale value , These are the grayscale threshold and the black area proportion threshold, respectively.

[0027] (2) Determine whether it is a low-echo void or a high-echo bright spot structure based on the initial screening results;

[0028] (3) Enter different boundary extraction paths respectively to obtain the boundary:

[0029] For low-echo cavity regions : Construct a gradient vector flow field and input the contour obtained from the initial screening. Use the internal force matrix and the normalized gradient vector flow field to guide the snake model and iteratively evolve the contour until it converges to the target edge. Use the principal axis length to judge and remove abnormal contours, and merge all contours to generate a mask. Filter out excessively small areas less than fifty pixels to obtain the final boundary trajectory.

[0030] For high echo bright spot areas The image is cropped to 50% of its width horizontally to locate the possible distribution area of ​​bright spots. The overall brightness distribution is estimated based on the proportion of black areas in this region. The threshold for bright spot extraction is adaptively calculated to extract the bright spot boundaries.

[0031] Step 203: Based on the extracted boundary information, for the low-echo cavity region image The inner region of the contour is divided into larger blocks of 16×16 to 24×24 to enhance noise reduction, while the outer region of the contour is divided into smaller blocks of 4×4 to 8×8 to preserve local details; for images with high echogenicity bright spots... The bright edge area is divided into smaller blocks of 4×4 to 8×8, while the background area outside the outline is divided into larger blocks of 16×16 to 24×24.

[0032] Step 204: Sparse representation based on dual dictionaries. During the training phase, the original image after boundary detection is processed. Simulated noise is introduced, and an adaptive block partitioning strategy is used to segment the image. For each image block, two independent dictionaries are trained using the K-Singular Value Decomposition (KSVD) algorithm. The goal is to segment the image into blocks based on a given sample set. In the case of learning the optimal dictionary This minimizes the sparse representation error of all samples:

[0033]

[0034] in, It is a sparse coefficient matrix. It is a sparsity constraint; in the reconstruction stage, the original image is... The same adaptive block partitioning is performed, and each image block is matched to its corresponding dictionary. Then, the Orthogonal Matching Pursuit (OMP) algorithm is used to solve for the sparse coefficients of each image block. OMP utilizes the redundant dictionary... Approximating the observed signal That is, to solve ,in, The vector of sparse coefficients is to be determined. Denoising image patches are obtained by multiplying the corresponding dictionary with the sparse coefficients. All denoising image patches are then restored to their original positions. A weighted average method is used to fuse the overlapping areas between the image patches to obtain the denoised image. ;

[0035] Step 3: Perform variational mode decomposition on the ultrasound images and optimize the variational mode decomposition parameters using the Grey Wolf optimization algorithm. The specific steps are as follows:

[0036] Step 301: Decompose the complex signal into several modal components with finite bandwidth using variational mode decomposition (VMD), and pre-set multiple parameters, including a penalty factor. Time step Decomposition of mode numbers Whether to retain the DC component, the initial center frequency selection method, and the convergence tolerance. and maximum number of iterations The goal of 2D-VMD decomposition, or variational mode decomposition, is to transform the original input signal... Decomposed into several modes with finite bandwidth :

[0037]

[0038] in, To different center frequencies Nearby narrowband components; Variational mode decomposition finds the optimal mode by constructing a variational optimization problem. and the corresponding center frequency The constrained variational model is as follows:

[0039]

[0040] in, For each modal function, For each center frequency;

[0041] Step 302: Reconstruct the original image by setting different parameters to obtain the reconstructed image;

[0042] Step 303: Dynamically adjust the number of decomposition modes in 2D-VMD using the Grey Wolf Optimization Algorithm (GWO). Punishment factor Time step and maximum number of iterations The minimum entropy criterion is used as the fitness function for initialization, determining the search range of each parameter. An initial position matrix of the gray wolf population is generated using a random initialization function, and the optimal wolf is initialized. Second-best wolf With the second best wolf The position and fitness score of the individual are determined, and then the parameters corresponding to the current individual position are used for 2D-VMD decomposition to obtain several modal components. The magnitude envelope of the Hilbert transform of each modal component is calculated, and after normalization, the information entropy is calculated. The formula for information entropy is:

[0043]

[0044] in, For the first Normalized amplitude probability distribution for each mode To minimize convergence tolerance, the minimum value among all modal entropy values ​​is taken as the fitness value of the current individual, and the fitness function is dynamically updated based on the fitness function value. , , The wolf's position and score are used to maintain the three best individuals in the population. Then, the male individuals are updated according to the gray wolf optimization algorithm. , , The three-headed wolf guides its own position change, and records the current position after the iteration ends. The wolf's position determines the optimal number of modalities. Optimal penalty factor Optimal time step and the optimal maximum number of iterations ;

[0045] Step 304: After performing variable mode decomposition parameter optimization and decomposition on multiple images, the corresponding intrinsic mode components are obtained. ;

[0046] Step 4: Classify the decomposed modalities and implement a differentiated denoising strategy; reconstruct the processed modalities into an image, and apply the local spatial sparse representation method again for fusion denoising to output the final denoised image. The specific steps are as follows:

[0047] Step 401: Decompose each image to obtain all intrinsic mode components. Enhancement processing is performed to obtain the main modes. Secondary modes Other modes ;

[0048] Step 402: Based on multiple evaluation indicators and through multiple experiments, design modal classification criteria: Judgment condition 1: equivalent visual effect < 8 and peak signal-to-noise ratio > 9; Judgment condition 2: equivalent visual effect > 64 or feature similarity > 1.43; Judgment condition 3: |principal direction angle -90| < 10; if judgment condition 1 is met, it is classified as the primary mode; if judgment condition 1 is not met but judgment conditions 2 and 3 are met, it is classified as other modes, and the rest are secondary modes. Separate the modes one by one, and finally output the mode type.

[0049] Step 403: Design differentiated denoising strategies based on the different information content and noise intensity in different types of modes: For the main mode For secondary modes, simply retain their original state; A strategy combining local region denoising and edge enhancement is adopted; for other modalities... Denoising is performed on local areas, and median filtering is used to smooth the entire image.

[0050] Step 404: The processed primary mode, secondary mode, and other modes are directly weighted and superimposed to reconstruct the image, resulting in a differentiated denoised image. After reconstruction, a sparse representation denoising method is used to obtain the final denoised image. .

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. This invention addresses the shortcomings of traditional compressed sensing denoising methods, such as insufficient denoising performance in certain regions and susceptibility of dictionary learning to speckle noise interference. It proposes a sparse representation denoising method based on local spatial information. Utilizing gradient vector flow and grayscale ratio analysis, a boundary extraction method for local regions of ultrasound images is designed to automatically acquire the boundary contours of target regions. Based on boundary information, an adaptive block partitioning strategy for the image is improved, automatically dividing image blocks in different regions for sparse representation. Furthermore, the dictionary training method based on the K-singular value decomposition algorithm is improved, training dual dictionaries for image blocks in different regions to achieve targeted denoising of different regions of the ultrasound image.

[0053] 2. This invention addresses the high sensitivity of variational mode decomposition (VM) denoising methods to parameter selection by proposing a VM optimization method based on the Grey Wolf optimization algorithm. The Grey Wolf optimization algorithm optimizes the parameter selection process of VM, obtaining modes with stable decomposition results. The characteristics of different modes are analyzed, and a classification method is designed to facilitate denoising. Differentiated denoising methods are designed for different classifications of modes, achieving targeted denoising of different modes in ultrasound images. After differential denoising and reconstruction using VM, a sparse representation method based on local spatial information is further used to process the ultrasound image. The improved fusion method retains the advantages of the unfused method, achieving targeted denoising of the local space of the ultrasound image. Attached Figure Description

[0054] Figure 1 This is a flowchart of ultrasound image denoising based on variational mode decomposition and local spatial sparse fusion;

[0055] Figure 2 This is a schematic diagram of the sparse representation denoising process;

[0056] Figure 3 This is a diagram illustrating the dual dictionaries established in different regions;

[0057] Figure 4 This is a schematic diagram of the fitness function value change curve;

[0058] Figure 5 This is a flowchart of modal classification;

[0059] Figure 6 This is a schematic diagram of differentiated noise reduction;

[0060] Figure 7 It is a denoised image with sparse representation after reconstruction of the hollow region;

[0061] Figure 8 This is an improved method for denoising bright areas in images. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0063] This invention provides an ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion. First, the Grey Wolf optimization algorithm is used to adaptively optimize the key parameters of variational mode decomposition, achieving stable and efficient decomposition of the ultrasound image. Then, the modal components are classified according to their structural characteristics, and differentiated denoising strategies are implemented for different modal categories to separate noise and retain useful information. After modal reconstruction, a sparse representation method based on local spatial information is further introduced. This method performs accurate boundary detection through gradient vector flow and grayscale ratio analysis, and adaptively divides the image into blocks based on the boundary information. Finally, sparse reconstruction is performed using dual dictionaries trained for different regions. Figure 1 As shown, the specific steps are as follows:

[0064] Step 1: Complete ultrasound image data acquisition, preprocess the data, and analyze noise characteristics. Specific steps are as follows:

[0065] Step 101: Acquire ultrasound images using a high-performance multi-channel programmable ultrasound imaging device and an ultrasound linear array probe with 128 channels. The ultrasound probe emits ultrasound waves and receives echo signals to generate radio frequency (RF) data. Preprocess the acquired RF data, including signal envelope detection, logarithmic compression, noise reduction, and normalization. Then, resampling is performed using bilinear interpolation to convert the RF data into B-mode images. Finally, normalization is performed to ensure that the brightness range meets the requirements, thereby obtaining a visible ultrasound image of the body membrane. .

[0066] Step 102: At the image level, a multiplicative noise model is used for description. Speckle noise will exhibit different energy distributions depending on the reflection intensity of the underlying tissue.

[0067]

[0068] in, For the observed image, To represent the true echo intensity, This represents the multiplicative noise intensity. When the scattering region is uniform, It can be modeled as a random variable following a Rayleigh distribution. Speckle noise is easily affected by various factors, and its manifestation and spatial distribution patterns differ in different regions. It is divided into low-echo cavity regions and high-echo bright spot regions. Twenty images of low-echo cavity regions were selected. and 20 images of hyperechoic bright spot regions .

[0069] Step 2: Perform sparse representation denoising based on local spatial information, and perform boundary detection by combining gradient vector flow and grayscale ratio analysis; perform adaptive image block partitioning based on boundary information; construct a dual dictionary and use K-singular value decomposition and orthogonal matching pursuit algorithm for sparse reconstruction. The specific steps are as follows:

[0070] Step 201: Using a snake-like model to minimize energy, this addresses the problem that direct sparse reconstruction of the entire image has insufficient denoising effect in local areas, affecting the observation effect. The snake-like model gradually converges the contour curve to the target boundary. The definition is as follows:

[0071]

[0072] in, It is the energy inside the curve. It is the external energy of the curve, combined with gradient vector flow (GVF), which extends the force field based on the gradient field, and can guide the evolution of the profile in the weak gradient region. The partial differential equation of gradient vector flow is:

[0073]

[0074] in, The gradient vector, The edge intensity of the input image. It is the gradient of the edge map. This is a smoothing parameter used to control the smoothness of the vector field. Using the Laplacian operator, iterative solutions can smooth the gradient field and provide a stronger force field in regions far from the edges. The gradient vector flow energy function... Energy function as an improved external energy term:

[0075]

[0076] in, This is the external energy term guided by gradient vector flow.

[0077] Step 202: Perform pre-screening using grayscale ratio analysis, including... and This method rapidly locates potential target regions within the entire ultrasound image, addressing the issue of unclear gradient information in the gradient vector flow field inside cavities or at the edges of bright spots. This can sometimes lead to misleading vector flows, causing serpentine contours to deviate from the true boundaries. Initial region screening is performed first, with a radius of [missing information]. A solid circle is used to cover any possible hollow parts of the main body; the outer area has a radius of [missing information]. The circular area minus the inner ring, The outer ring represents the radius difference between the inner and outer rings, while the inner ring represents the background around the hole. With a given step size, the search box begins its full-image search and, after locating a suspected hole region, starts from the smallest radius. To the maximum radius Gradually increase the inner circle to find the largest circular area that still remains sufficiently dark. The criterion is:

[0078]

[0079] in, The proportion of the black area. Number of pixels Grayscale value , These are the grayscale threshold and the black area proportion threshold, respectively. Based on the initial screening results, it can be determined whether it belongs to a low-echo cavity or a high-echo bright spot structure. Then, different boundary extraction paths are used for each, and finally, boundary acquisition is performed for the low-echo cavity region. A gradient vector flow field is constructed and input into the contour obtained from the initial screening. The internal force matrix and the normalized gradient vector flow field guide the serpentine model, iteratively evolving the contour until it converges to the target edge. Then, abnormal contours are removed based on the principal axis length, and all contours are merged to generate a mask. Regions smaller than fifty pixels are filtered out to obtain the final boundary trajectory. For high-echo bright spot regions... The process involves cropping a 50% horizontal section from the center of the image to locate potential bright spot distribution areas. The overall brightness distribution is estimated based on the proportion of black areas within this region. An adaptive threshold for bright spot extraction is then calculated, and finally, the bright spot boundaries are extracted.

[0080] Step 203: Based on the extracted boundary information, for the low-echo cavity region image The structure within the contour area is relatively uniform, and noise is relatively significant; therefore, larger blocks of 16×16 to 24×24 are used to enhance noise reduction. The outer contour area is mostly background noise, so smaller blocks of 4×4 to 8×8 are used to preserve local details. For images with high echogenic bright spots... The highlight structure is prominent, and the grayscale changes in the edge area of ​​the highlight are obvious. Therefore, smaller blocks of 4×4 to 8×8 are used for the division; the background area outside the outline is divided into larger blocks of 16×16 to 24×24.

[0081] Step 204: Sparse representation based on dual dictionaries. During the training phase, the original image after boundary detection is processed. Simulated noise is introduced, and an adaptive block partitioning strategy is used for image segmentation. For each image block class, two independent dictionaries are trained using the K-Singular Value Decomposition (KSVD) algorithm, with the goal of segmenting the image into blocks based on a given sample set. In the case of learning the optimal dictionary This minimizes the sparse representation error of all samples:

[0082]

[0083] in, It is a sparse coefficient matrix. It is a sparsity constraint; during the reconstruction stage, the original image... The same adaptive block partitioning is performed, and each image block is matched to its corresponding dictionary. Then, the Orthogonal Matching Pursuit (OMP) algorithm is used to solve for the sparse coefficients of each image block. OMP utilizes the redundant dictionary... Approximating the observed signal That is, to solve ,in, The vector of sparse coefficients is used to obtain denoised image patches by multiplying the corresponding dictionary with the sparse coefficients. Finally, all denoised image patches are restored to their original positions, and overlapping areas between the image patches are fused using a weighted average method to obtain the denoised image. .

[0084] Step 3: Perform variational mode decomposition on the ultrasound images and optimize the variational mode decomposition parameters using the Grey Wolf optimization algorithm. The specific steps are as follows:

[0085] Step 301: Decompose the complex signal into several modal components with finite bandwidth using variational mode decomposition (VMD), and pre-set multiple parameters, including a penalty factor. Time step Decomposition of mode numbers Whether to retain the DC component, the initial center frequency selection method, and the convergence tolerance. and maximum number of iterations The goal of variational mode decomposition is to perform 2D-VMD decomposition on the original input signal. Decomposed into several modes with finite bandwidth :

[0086]

[0087] in, To different center frequencies The nearby narrowband components. Variational mode decomposition seeks the optimal mode by constructing a variational optimization problem. and the corresponding center frequency The constrained variational model is as follows:

[0088]

[0089] in, For each modal function, For each center frequency.

[0090] Step 302: The original image is reconstructed by setting different parameters to obtain the reconstructed image. The reconstructed image under different parameters all showed different degrees of striped noise and blurring. Improper parameter selection led to many problems in the decomposition results. An optimization algorithm was introduced to achieve adaptive selection of parameters.

[0091] Step 303: Dynamically adjust the number of decomposition modes in 2D-VMD using the Grey Wolf Optimization Algorithm (GWO). Punishment factor Time step and maximum number of iterations Using the minimum entropy criterion as the fitness function, the initial settings are as follows: the gray wolf population size is 10, the problem dimension is 4, and the maximum number of iterations is 10. The search range for each parameter is determined. An initial position matrix for the gray wolf population is generated using a random initialization function, and the positions and fitness scores of the best, second-best, and next-second-best wolves are initialized. Then, the parameters corresponding to the current individual position are used for 2D-VMD decomposition to obtain several modal components. The magnitude envelope of the Hilbert transform of each modal component is calculated, and after normalization, the information entropy is calculated. The information entropy formula is:

[0092]

[0093] in, For the first Normalized amplitude probability distribution for each mode To mitigate convergence tolerance, the minimum value among all modal entropy values ​​is taken as the fitness value of the current individual, and the fitness function is dynamically updated based on the fitness function value. , , The wolf's position and score are used to maintain the three best individuals in the population. Then, the position is updated according to the gray wolf optimization algorithm formula, as follows:

[0094]

[0095] in, , , This represents the positions of the individuals ranked first, second, and third in fitness within the population. , , , , , For independently generated coefficient vectors, all individuals are based on , , The three-headed wolf guides its own position change, and records the current position after the iteration ends. The wolf's position determines the optimal number of modalities. Optimal penalty factor Optimal time step and the optimal maximum number of iterations ;

[0096] Step 304: After performing variable mode decomposition parameter optimization and decomposition on multiple images, the corresponding intrinsic mode components are obtained. .

[0097] Step 4: Classify the decomposed modalities and implement a differential denoising strategy; reconstruct the processed modalities into an image, and apply the local spatial sparse representation method again for fusion denoising to output the final denoised image. The specific steps are as follows:

[0098] Step 401: Decompose each image to obtain all intrinsic mode components. Enhancement processing is performed, and the differences between different modalities are significant. One type of modality usually has stronger structural features and plays a dominant role in overall image understanding and subsequent tasks; therefore, it is called the dominant modality. Another type of mode, although containing less information than the primary mode, can provide auxiliary information and is therefore called a secondary mode. The remaining modes mostly have low information density and are referred to as other modes. .

[0099] Step 402: Based on multiple evaluation indicators and through numerous experiments, a modality classification criterion was designed. Evaluation indicators were calculated, including equivalent visual count, peak signal-to-noise ratio (PSNR), feature similarity, and the principal direction angle (ANGLE) determined using a gradient direction histogram. Criterion 1: Equivalent visual count < 8 and PSNR > 9; Criterion 2: Equivalent visual count > 64 or feature similarity > 1.43; Criterion 3: |Principal direction angle - 90| < 10. If criterion 1 is met, the modality is classified as the primary mode; if criterion 1 is not met but criterion 2 and criterion 3 are met, the modality is classified as another mode, and the rest are secondary modes. The modalities are separated sequentially, and the final modality category is output.

[0100] Step 403: Design differentiated denoising strategies based on the different information content and noise intensity in different types of modes: For the main mode To avoid information loss, the original state is directly preserved; for secondary modes... A strategy combining local region denoising and edge enhancement is adopted to strengthen key boundary features while suppressing local noise; for other modalities... In addition to denoising local areas, median filtering is used to smooth the entire image, improving the overall image quality.

[0101] Step 404: After processing, different modalities each have their advantages in different regions. The primary modality serves as the overall structural skeleton, the secondary modality enhances the edges, and other modalities provide details. First, we attempt to directly weight and superimpose the processed primary modality, secondary modality, and other modalities to reconstruct the image, removing some multiplicative noise and obtaining the differentiated denoising image. Even in the background area, some edge blurring and texture disorder still exist. After reconstruction, a sparse representation denoising method is used to achieve better denoising results in local areas and suppress background multiplicative noise, ultimately resulting in a more natural image. .

[0102] Example

[0103] This embodiment utilizes a high-performance multi-channel programmable ultrasound imaging device and an ultrasound linear array probe to acquire ultrasound images with 128 channels. To verify the reliability of the ultrasound denoising algorithm, this embodiment uses body membranes for data acquisition.

[0104] Step 1: Analyze the noise characteristics of the input ultrasound image. Speckle noise is easily affected by various factors, and its manifestation and spatial distribution differ in different regions. Hypoechoic cavity regions are usually distributed within liquid structures or homogeneous soft tissues. In these regions, the scattering ability of ultrasound waves is weak, and the echo signal intensity is low, appearing as areas with low grayscale in the image. The noise mainly consists of sparsely distributed small bright spots or linear bright areas within the cavity regions. Hyperechoic bright spot regions are mostly distributed near calcified tissue, bone tissue, and metal implants. In these regions, ultrasound waves are strongly reflected near the tissue, and the interference between scattering sources forms dense bright spots.

[0105] To verify the effectiveness of the denoising algorithm, experimental results were analyzed from both subjective and objective evaluation perspectives. For ultrasound images, since there is no reference image, a method for evaluation within a local region of the ultrasound image is proposed. Pixel values ​​of the region of interest and the surrounding background region are averaged within a small local area to serve as an ideal noise-free reference image. A multi-evaluation index system is constructed by combining various objective evaluation metrics.

[0106] Step 2: First, add noise to the image. Then, divide the image into blocks, moving one pixel at a time. Next, use the orthogonal matching pursuit algorithm to calculate the sparse coefficients of all image blocks according to a pre-generated dictionary, and then reconstruct the image blocks. The dictionary is continuously adjusted using the K-singular value decomposition algorithm until the mean square error difference between two iterations is less than 0.0001, at which point the iteration stops. Then, divide the original image into blocks again, calculate the sparse coefficients using the generated dictionary, and reconstruct the denoised image blocks. Finally, weighted summaries of all image blocks according to their positions are used to obtain the denoised image. Figure 2As shown, a snake-like model boundary detection strategy combining gradient vector flow and grayscale ratio analysis is introduced before sparse representation to accurately locate the boundary contour of the region of interest. Sparse representation reconstruction is performed by dividing different regions with different block sizes, which improves the traditional block division method. By utilizing structural prior information, an image adaptive block division method is proposed. Based on the extracted boundary information, two regions in a single image are first divided, and then the corresponding block division method is adopted according to the features of the regions.

[0107] A dual dictionary is constructed for the region of interest (ROI) and background region of the ultrasound image for sparse representation reconstruction. During the training phase, simulated noise is added to the original image after boundary detection, and then an adaptive block partitioning strategy is used to divide the image into blocks. For each type of image block, two independent dictionaries are trained using the K-singular value decomposition algorithm. The dictionaries for different regions are as follows: Figure 4 As shown, in the reconstruction stage, the original image is divided into identical adaptive blocks, and each image block is matched to its corresponding dictionary. Then, the orthogonal matching pursuit algorithm is used to solve for the sparse coefficients of each image block. The denoised image block is obtained by multiplying the corresponding dictionary with the sparse coefficients. Finally, all denoised image blocks are restored according to their original positions, and the overlapping areas between image blocks are fused using a weighted average method. The evaluation metrics for the denoising effect of the improved method are shown in Table 1.

[0108]

[0109] As shown in Table 1, the equivalent number of views, horizontal half-width at half-maximum (HWHM), and vertical half-width at half-maximum (HWHM) are improved by 9.0%, 1.8%, and 1.6% respectively compared to the original image, indicating that the denoising method helps improve the overall uniformity, but the effect on improving horizontal and vertical resolution is not significant. Since the improved sparse representation denoising method has shortcomings in edge preservation, further optimization is needed.

[0110] Step 3: Variational Mode Decomposition (VMD) decomposes a complex signal into several mode components with finite bandwidth. Performing 2D-VMD decomposition requires pre-setting several parameters, including a penalty factor. Time step Decomposition of mode numbers Whether to retain the DC component, the initial center frequency selection method, and the convergence tolerance. and maximum number of iterations The settings of these parameters directly affect the decomposition effect and speed. Improper parameter selection will lead to numerous problems in the decomposition results. Traditional parameter setting relies on experience or multiple trials, which is highly subjective and inefficient. Therefore, optimization algorithms are needed to achieve adaptive parameter selection. Using the minimum entropy criterion as the fitness function, the Grey Wolf optimization algorithm is used to dynamically adjust the parameters of 2D-VMD. Figure 4Several sets of fitness function values ​​are shown to change with the number of iterations. In the initial few iterations, the fitness function value decreases rapidly, and most of the optimization process approaches convergence around the 6th iteration. By observing the trend of the fitness function value, it is found that in most cases, only 10 iterations are needed to achieve a satisfactory optimization effect.

[0111] Step four: Enhancement processing is performed on all intrinsic modal components obtained from the decomposition of each image. After processing, the features of different modes are significantly enhanced. After comparing and analyzing the differences between the processed image and the original image, a modality classification criterion is designed based on multiple evaluation indicators and through multiple experiments, such as... Figure 5 As shown, the evaluation indicators include equivalent number of views, peak signal-to-noise ratio (PSNR), feature similarity, and principal direction angle (ANGLE) determined using the gradient direction histogram. The first criterion is equivalent view count < 8 and peak SNR > 9; the second criterion is equivalent view count > 64 or feature similarity > 1.43; and the third criterion is |principal direction angle -90| < 10. By repeatedly determining whether the conditions are met, the modes are separated step by step, ultimately outputting the mode types. After classifying the modes, differentiated denoising is performed based on the different information content and noise intensity in different types of modes, such as... Figure 6 As shown, different types of modalities have their own advantages in different regions after processing: the primary modality serves as the overall structural skeleton, the secondary modality strengthens the edges, and other modalities provide details. Applying a sparse representation reconstruction denoising method based on local spatial information to the reconstructed image not only enables better denoising in local regions but also suppresses background multiplicative noise, resulting in a more natural image.

[0112] The denoising effect of low-echo cavity region images is compared to... Figure 7 As shown. From a visual perspective: the method based on sparse representation of local spatial information removes some noise in local areas, but some large bright linear noise remains. However, this method has a better smoothing effect in the background area. The differential denoising and reconstruction method has limited effect on noise suppression, but it removes large bright linear noise and retains boundary information completely, while the smoothing effect in the background area is average. The denoising method using adaptive sparse representation after differential denoising and reconstruction not only removes all bright linear noise in the hole area, but also has a strong ability to preserve boundary details. At the same time, the smoothing effect in the background area is better, and the overall visual effect is the best. For images with high echo bright spots, the improved method also has strong denoising ability and structure preservation ability. Figure 8As shown, the adaptive sparse representation denoising image suppresses background noise to a certain extent, but the edges of the bright spot region are overly smoothed; the differential denoising reconstruction method improves the edge clarity, but there is still a lot of noise in its background region; the fused sparse representation denoising image removes artifact noise while better preserving the details of the structural edges, and the overall visual effect of the image is more natural.

Claims

1. A method for denoising ultrasound images based on variational mode decomposition and local spatial sparse fusion, characterized in that... The method includes the following steps: Step 1: Complete the acquisition of ultrasound image data, preprocess the data, and analyze the noise characteristics; Step 2: Perform sparse representation denoising based on local spatial information, and perform boundary detection by combining gradient vector flow and gray-level ratio analysis; perform adaptive image block division based on boundary information; Construct a dual dictionary and perform sparse reconstruction using K-singular value decomposition and orthogonal matching pursuit algorithm; Step 3: Perform variational mode decomposition on the ultrasound images and optimize the variational mode decomposition parameters using the Grey Wolf optimization algorithm; Step 4: Classify the decomposed modalities and implement a differentiated denoising strategy; reconstruct the processed modalities into an image, and apply the local spatial sparse representation method again for fusion denoising to output the final denoised image.

2. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 1, characterized in that... The specific steps of step 1 are as follows: Step 101: Acquire ultrasound images using a high-performance multi-channel programmable ultrasound imaging device and an ultrasound linear array probe with 128 channels. The ultrasound probe emits ultrasound waves and receives echo signals to generate radio frequency (RF) data. The acquired RF data undergoes signal envelope detection, logarithmic compression, noise reduction, and normalization. Bilinear interpolation is then used for resampling to convert the RF data into B-mode images. Finally, normalization is performed to ensure that the brightness range meets the requirements, thereby obtaining a visible ultrasound image of the body membrane. ; Step 102: At the image level, a multiplicative noise model is used for description. Speckle noise will exhibit different energy distributions depending on the reflection intensity of the underlying tissue. in, For the observed image, To represent the true echo intensity, This represents the multiplicative noise intensity; when the scattering region is uniform, Modeled as a random variable following a Rayleigh distribution, speckle noise is divided into low-echo void regions and high-echo bright spot regions.

3. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 1, characterized in that... The specific steps of step 2 are as follows: Step 201: Using a snake-like model, the contour curve gradually converges to the target boundary through energy minimization. The snake-like model's energy function... The definition is as follows: in, It is the energy inside the curve. It is the energy outside the curve; The partial differential equation for the gradient vector flow is: in, The gradient vector, The edge intensity of the input image. It is the gradient of the edge map. For smoothing parameters, For Laplacian operators; Gradient vector flow energy function Energy function as an improved external energy term: in, This refers to the external energy term guided by gradient vector flow; Step 202: Perform pre-screening using grayscale ratio analysis, focusing on images containing low-echo cavity regions. Images of hyperechoic bright spots Rapidly locate potential target areas within the entire ultrasound image; Step 203: Based on the extracted boundary information, for the low-echo cavity region image The inner region of the contour is divided into larger blocks of 16×16 to 24×24 to enhance noise reduction, while the outer region of the contour is divided into smaller blocks of 4×4 to 8×8 to preserve local details; for images with high echogenicity bright spots... The bright edge area is divided into smaller blocks of 4×4 to 8×8, while the background area outside the outline is divided into larger blocks of 16×16 to 24×24. Step 204: Sparse representation based on dual dictionaries. During the training phase, the original image after boundary detection is processed. Simulated noise is introduced, and an adaptive block partitioning strategy is used to segment the image into blocks. For each type of image block, two independent dictionaries are trained using the K-singular value decomposition algorithm. The goal is to achieve the desired result for a given sample set. In the case of learning the optimal dictionary This minimizes the sparse representation error of all samples: in, It is a sparse coefficient matrix. It is a sparsity constraint; in the reconstruction stage, the original image is... The same adaptive block partitioning is performed, and each image block is matched to its corresponding dictionary. Then, the orthogonal matching pursuit algorithm is used to solve for the sparse coefficients of each image block. Orthogonal matching pursuit utilizes the redundant dictionary. Approximating the observed signal That is, to solve ,in, The vector of sparse coefficients is to be determined. Denoising image patches are obtained by multiplying the corresponding dictionary with the sparse coefficients. All denoising image patches are then restored to their original positions. A weighted average method is used to fuse the overlapping areas between the image patches to obtain the denoised image. .

4. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 3, characterized in that... The specific steps of step 202 are as follows: (1) Regional initial screening: radius is A solid circle is used to cover any possible hollow parts of the main body; the outer area has a radius of [missing information]. The circular area minus the inner ring, The outer ring represents the radius difference between the inner and outer rings, while the inner ring represents the background around the hole. With a given step size, the search box begins its full-image search and, after locating a suspected hole region, starts from the smallest radius. To the maximum radius Gradually increase the inner circle to find the largest circular area that still remains sufficiently dark. The criterion is: in, The proportion of the black area. Number of pixels Grayscale value , These are the grayscale threshold and the black area proportion threshold, respectively. (2) Determine whether it is a low-echo void or a high-echo bright spot structure based on the initial screening results; (3) Enter different boundary extraction paths respectively to obtain the boundary: For low-echo cavity regions : Construct a gradient vector flow field and input the contour obtained from the initial screening. Use the internal force matrix and the normalized gradient vector flow field to guide the snake model and iteratively evolve the contour until it converges to the target edge. Use the principal axis length to judge and remove abnormal contours, and merge all contours to generate a mask. Filter out excessively small areas less than fifty pixels to obtain the final boundary trajectory. For high echo bright spot areas The algorithm extracts a 50% horizontal section from the middle of the image to locate the possible distribution area of ​​bright spots. It estimates the overall brightness distribution based on the proportion of black areas in this region, adaptively calculates the threshold for bright spot extraction, and extracts the boundaries of the bright spots.

5. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 1, characterized in that... The specific steps of step 3 are as follows: Step 301: Decompose the complex signal into several modal components with finite bandwidth through variational mode decomposition, and pre-set multiple parameters, including a penalty factor. Time step Decomposition of mode numbers Whether to retain the DC component, the initial center frequency selection method, and the convergence tolerance. and maximum number of iterations The goal of 2D-VMD decomposition, or variational mode decomposition, is to transform the original input signal... Decomposed into several modes with finite bandwidth : in, To different center frequencies Nearby narrowband components; Variational mode decomposition finds the optimal mode by constructing a variational optimization problem. and the corresponding center frequency The constrained variational model is as follows: in, For each modal function, For each center frequency; Step 302: Reconstruct the original image by setting different parameters to obtain the reconstructed image; Step 303: Dynamically adjust the number of decomposition modes in 2D-VMD using the Grey Wolf optimization algorithm. Punishment factor Time step and maximum number of iterations The minimum entropy criterion is used as the fitness function for initialization, determining the search range of each parameter. An initial position matrix of the gray wolf population is generated using a random initialization function, and the optimal wolf is initialized. Second-best wolf With the second best wolf The position and fitness score of the current individual are determined, and then the parameters corresponding to the current individual's position are used for 2D-VMD decomposition to obtain several modal components. For each modal component, the magnitude envelope of its Hilbert transform is calculated, and after normalization, the information entropy is calculated. The minimum value among all modal entropy values ​​is taken as the fitness value of the current individual, and the fitness function is dynamically updated based on the fitness function value. , , The wolf's position and score are used to maintain the three best individuals in the population. Then, the male individuals are updated according to the gray wolf optimization algorithm. , , The three-headed wolf guides its own position change, and records the current position after the iteration ends. The wolf's position is used to obtain the optimal number of modalities. Optimal penalty factor Optimal time step and the optimal maximum number of iterations ; Step 304: After performing variable mode decomposition parameter optimization and decomposition on multiple images, the corresponding intrinsic mode components are obtained. .

6. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 5, characterized in that... The formula for information entropy is: in, For the first Normalized amplitude probability distribution for each mode To achieve convergence tolerance.

7. The ultrasound image denoising method based on variational mode decomposition and local spatial sparse fusion according to claim 1, characterized in that... The specific steps of step 4 are as follows: Step 401: Decompose each image to obtain all intrinsic mode components. Enhancement processing is performed to obtain the main modes. Secondary modes Other modes ; Step 402: Based on multiple evaluation indicators and through multiple experiments, design modal classification criteria: Judgment condition 1: equivalent visual effect < 8 and peak signal-to-noise ratio > 9; Judgment condition 2: equivalent visual effect > 64 or feature similarity > 1.43; Judgment condition 3: |principal direction angle -90| < 10; if judgment condition 1 is met, it is classified as the primary mode; if judgment condition 1 is not met but judgment conditions 2 and 3 are met, it is classified as other modes, and the rest are secondary modes. Separate the modes one by one, and finally output the mode type. Step 403: Design differentiated denoising strategies based on the different information content and noise intensity in different types of modes: For the main mode For secondary modes, simply retain their original state; A strategy combining local region denoising and edge enhancement is adopted; for other modalities... Denoising is performed on local areas, and median filtering is used to smooth the entire image. Step 404: The processed primary mode, secondary mode, and other modes are directly weighted and superimposed to reconstruct the image, resulting in a differentiated denoised image. After reconstruction, a sparse representation denoising method is used to obtain the final denoised image. .

Images