Optimized ultrasonic imaging method based on fast singular value decomposition

Abstract

An optimized ultrasonic imaging method based on fast singular value decomposition comprises the steps that after original ultrasonic signals are collected and subjected to beam forming processing, local low-order singular value decomposition (SVD) is achieved through eigenvalue decomposition, a sparse matrix optimization method is combined to reconstruct low-order signals related to background tissue noise, and the low-order signals related to background tissue noise are obtained. And finally, extracting a blood flow signal through low-order truncation and reconstructing an ultrasonic image. According to the method, the calculation amount of singular value decomposition is remarkably reduced by utilizing a low-order truncation and sparse matrix optimization method, and rapid ultrasonic image processing is realized.

Description

Technical Field

[0001] The present invention relates to a technology in the field of medical image processing, specifically an optimized ultrasonic imaging method based on fast singular value decomposition. Background Art

[0002] Ultrasonic imaging is a medical imaging technology that uses ultrasonic waves to measure and visualize internal structures in the human body, and is widely used in the fields of cardiovascular disease diagnosis, tumor localization and monitoring, pregnancy and childbirth detection, musculoskeletal assessment, visceral disease diagnosis, and surgical guidance. Improving the accuracy of ultrasonic imaging can significantly enhance the accuracy and reliability of these medical examinations, thereby more effectively guiding clinical decision-making and treatment planning. Summary of the Invention

[0003] Aiming at the above deficiencies of the existing technology, the present invention proposes an optimized ultrasonic imaging method based on fast singular value decomposition, which significantly reduces the computational complexity of singular value decomposition by using low-order truncation and sparse matrix optimization methods, and realizes fast ultrasonic image processing.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to an optimized ultrasonic imaging method based on fast singular value decomposition. After collecting the original ultrasonic signal and performing beamforming processing, local low-order singular value decomposition (SVD) is realized through eigenvalue decomposition, and then the low-order signal related to background tissue noise is reconstructed by combining the sparse matrix optimization method. Finally, the blood flow signal is extracted through low-order truncation and the ultrasonic image is reconstructed. Brief Description of the Drawings

[0006] Figure 1 is the flowchart of the present invention;

[0007] Figure 2 is the ultrasonic imaging diagram of the human kidney;

[0008] In the figure: (a) The B-mode imaging diagram after only removing noise by filtering; (b) The B-mode imaging diagram after extracting the blood flow-related signal through the fast SVD algorithm;

[0009] Figure 3 is the schematic diagram of the spatio-temporal matrix;

[0010] Figure 4 is the flowchart of the singular value decomposition in the embodiment;

[0011] Figure 5 is the flowchart of the eigenvalue decomposition combining power iteration and orthogonalization in the embodiment;

[0012] Figure 6 is the effect diagram of the embodiment;

[0013] In the figure: Matlab refers to the Matlab function library, MKL refers to the Intel Math Kernel Library, GTX280 is the GPU model, SLR-SVD is the fast SVD algorithm proposed by the present invention, and 8K is the abbreviated form of 8000). Detailed implementation manners

[0014] As Figure 1 shown, this embodiment relates to an optimized ultrasonic imaging method based on fast singular value decomposition. After collecting the original ultrasonic signal and performing beamforming processing, local low-order singular value decomposition (SVD) is achieved through eigenvalue decomposition, and then a sparse matrix optimization method is combined to reconstruct the low-order signal related to the background tissue noise. Finally, the blood flow signal is extracted through low-order truncation and the ultrasonic image is reconstructed.

[0015] As Figure 4 shown, the realization of local low-order singular value decomposition (SVD) through eigenvalue decomposition specifically includes:[[]]

[0016] Step 1: Rearrange the two-dimensional ultrasonic image with a size of x×y arranged in t frames of time after beamforming processing into a two-dimensional spatio-temporal matrix S with a size of t×u as shown, that is, all pixels of one frame of two-dimensional ultrasonic image are arranged in sequence, and u = x×y. Figure 3 shown, that is, all pixels of one frame of two-dimensional ultrasonic image are arranged in sequence, and u = x×y.

[0017] Step 2: Transpose the spatio-temporal matrix S, and perform eigenvalue decomposition on the matrix M = SS T using an improved power iteration method, which specifically includes:[[]]

[0018] 2.1) As Figure 5 shown, perform eigenvalue decomposition of the matrix M by combining power iteration and orthogonalization to obtain the right singular vector matrix and the singular value array, which specifically includes:[[]]

[0019] 2.1.1) Perform QR decomposition on the matrix M using the modified Gram-Schmidt algorithm (MGS) to obtain the orthogonal matrix Q and the norm of each column, which specifically includes:[[]]

[0020] 2.1.1.1) Divide M into several column vectors m1, m2,..., m u and then normalize each column to obtain the first column q1 of the orthogonal matrix Q and calculate the norm of the vector q1 as an element p1 of the eigenvalue candidate array.

[0021] 2.1.1.2) For the remaining column vectors of M, subtract the projections of all previous columns on this column in sequence, and then obtain the subsequent columns of Q, namely q2, q3,..., q u while the array composed of the norms of its columns is the eigenvalue candidate array P λ , that is, the diagonal elements of the upper triangular matrix R.

[0022] 2.1.2) Arrange each column in the orthogonal matrix Q in descending order according to the modulus length of each column.

[0023] 2.1.3) Update each column vector in turn through iterative multiplication, and re-orthogonalize with the previous vectors in each step of iteration until the convergence condition is met. Until the k-th vector is updated, at this time, only the k-order effective right singular vector matrix Q and the singular value array λ arranged in descending order are obtained, and the values higher than the K-th order are all set to zero.

[0024] 2.2) Calculate the reciprocal singular value matrix, specifically including:

[0025] 2.2.1) Calculate the square root of each element in the singular value array λ as the singular value;

[0026] 2.2.2) Construct the reciprocal singular value diagonal matrix R △ .

[0027] Reconstructing the low-order signal related to the background tissue noise specifically includes:

[0028] 1) Calculate the intermediate matrix, specifically: U = R △ Q T S, where: Q is the matrix obtained by eigenvalue decomposition, R △ is the reciprocal singular value diagonal matrix, T represents the transpose, and S is the two-dimensional spatio-temporal matrix to be processed.

[0029] 2) Calculate the reconstruction matrix, specifically: X = QR'U, where: R' is the singular value diagonal matrix, and U is the left singular vector matrix.

[0030] The low-order truncation specifically includes: Output = S - X, where: X is the low-order signal reconstructed related to the background tissue noise, and S is the two-dimensional spatio-temporal matrix.

[0031] After specific experiments, an convex array probe with a radius of 60 mm, 128 array elements, an element pitch of 0.5 mm, and a center frequency of 3.5 MHz is used for imaging, the acquisition frequency is 2.5 MHz, and the acquisition frame rate is 300 FPS. As Figure 2 shown in a, it is the B-mode ultrasound image only after high-pass filtering. It can be seen that simple filtering cannot completely eliminate the influence of surrounding tissues. As Figure 2 shown in b, after fast SVD processing, the bright spots caused by surrounding tissues are basically eliminated, so that it can better focus on the morphology of microvessels.

[0032] As Figure 6As shown, in order to compare the operation speed, the operation speed results of the traditional SVD algorithm in the central processing unit (CPU) environment and the GPU environment in "Singular value decomposition on GPU using CUDA" (2009 IEEE International Symposium on Parallel & Distributed Processing, 2009, doi: 10.1109 / IPDPS.2009.5161058) are referred to, and compared with the test results of the present invention on the CPU (model i7-11800H, single-core operation). Taking the case of n×m=32×8000 as an example, the operation speed of the present invention is 4.2 times that of the literature, while the actual calculation time is only 1 / 1528 of it, wherein the operation speed brought by the present invention is increased by up to 364 times. In addition, although the GPU can accelerate the calculation, the overall calculation speed is still lower than the result of the present invention because the traditional SVD algorithm is not optimized. Taking the case of n×m=32×8000 as an example, the operation speed of the low-order truncation and sparse matrix optimization SVD method based on CPU is at least 313 times that of the traditional SVD method based on GPU.

[0033] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. An optimized ultrasonic imaging method based on fast singular value decomposition, characterized in that, After collecting the original ultrasonic signal and performing beamforming processing, local low-order singular value decomposition is achieved through eigenvalue decomposition, and then combined with a sparse matrix optimization method to reconstruct the low-order signal related to background tissue noise. Finally, the blood flow signal is extracted through low-order truncation and the ultrasonic image is reconstructed.

2. The optimized ultrasonic imaging method based on fast singular value decomposition according to claim 1, characterized in that, The realization of local low-order singular value decomposition through eigenvalue decomposition specifically includes: Step 1: Rearrange the two-dimensional ultrasonic image with a size of x×y arranged in t frames of time after beamforming processing into a two-dimensional spatio-temporal matrix S with a size of t×u, that is, all pixels of one frame of two-dimensional ultrasonic image are arranged in sequence, and u = x×y; Step 2: Transpose the spatio-temporal matrix S, and for the matrix M = SS T Perform eigenvalue decomposition using the improved power iteration method.

3. The optimized ultrasonic imaging method based on fast singular value decomposition according to claim 2, characterized in that The specific content of step 2 includes: 2.1) Perform eigenvalue decomposition by combining power iteration and orthogonalization on matrix M to obtain the right singular vector matrix and the singular value array, specifically including: 2.1.1) Perform QR decomposition on matrix M using the modified Gram-Schmidt algorithm (MGS) to obtain the orthogonal matrix Q and the norm of each column; 2.1.2) Arrange each column in the orthogonal matrix Q in descending order according to the norm of each column; 2.1.3) Update each column vector through iterative multiplication in sequence, and re-orthogonalize with the previous vector at each step of iteration until the convergence condition is satisfied. Until the kth vector is updated, at this time, only the k-order effective right singular vector matrix Q and the descending-order singular value array λ are obtained, and the values higher than the Kth order are all cleared; 2.2) Calculate the singular value reciprocal matrix, specifically including: 2.2.1) Calculate the square root of each element in the singular value array λ as the singular value; 2.2.2) Construct the reciprocal singular value diagonal matrix R △ .

4. The optimized ultrasonic imaging method based on fast singular value decomposition according to claim 3, characterized in that The specific content of step 2.1.1 includes: 2.1.1.1) Divide M into a number of column vectors m1, m2, …, m u After that, normalize each column to obtain the first column q1 of the orthogonal matrix Q and calculate the norm of the vector q1 as an element p1 of the eigenvalue candidate array; 2.1.1.2) For the remaining column vectors of M, after successively subtracting the projections of all the previous columns on this column and then performing normalization processing, the subsequent columns of Q are obtained, namely q2, q3, …, q u At the same time, the array composed of the norms of its columns is the candidate eigenvalue array P λ , that is, the diagonal elements of the upper triangular matrix R.

5. The optimized ultrasonic imaging method based on fast singular value decomposition according to claim 2, wherein The reconstruction of the low-order signal related to background tissue noise specifically includes: 1) Calculate the intermediate matrix, specifically: U = R △ Q T S, where: Q is the matrix obtained by eigenvalue decomposition, R △ is the diagonal matrix of reciprocal singular values, T represents transpose, and S is the two-dimensional spatio-temporal matrix to be processed; 2) Calculate the reconstruction matrix, specifically: X = QR'U, where: R' is the singular value diagonal matrix, and U is the left singular vector matrix.

6. The optimized ultrasonic imaging method based on fast singular value decomposition according to claim 2, characterized in that, The low-order truncation specifically includes: Output = S - X, where: X is the low-order signal related to the reconstruction of background tissue noise, and S is the two-dimensional spatio-temporal matrix.

Images