HIFU interference removal system and method based on robust principal component analysis
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-15
- Publication Date
- 2026-08-11
Smart Images
Figure CN117562573B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of computer vision and ultrasound imaging, and relates to a system that applies machine learning methods to ultrasound imaging, aiming to remove high-intensity focused ultrasound (HIFU) interference. Background Technology
[0002] In clinical research and practice, high-intensity focused ultrasound (HIFU) has become increasingly attractive as a non-invasive technique for ablating tumors due to its ability to shorten recovery time and reduce the risk of complications. In recent years, HIFU has been successfully used to treat some soft tissue diseases, such as uterine fibroids, liver cancer, and others. During this non-invasive treatment, real-time medical imaging monitoring of the procedure is crucial for accurately ablating diseased tissue and minimizing damage to healthy tissue. However, real-time monitoring has remained a long-standing challenge in this field.
[0003] Compared to MRI, ultrasound imaging offers the advantage of real-time imaging. However, radiofrequency interference exists between HIFU systems and ultrasound imaging systems because both systems involve the transmission of acoustic signals. Even though the fundamental frequency of the acoustic signal from HIFU treatment is typically lower than that of the imaging signal, its higher harmonics can still significantly contaminate the imaging signal, leading to "overexposure" in ultrasound images. Typically, the clinical approach to address this issue is to alternate between HIFU irradiation and ultrasound imaging, which abandons real-time monitoring of HIFU treatment and prolongs the treatment process. Therefore, developing an ultrasound imaging method that can effectively eliminate HIFU acoustic interference is of significant value. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, this invention aims to propose a new method and device for high-intensity focused ultrasound (HIFU) interference removal. Therefore, the technical solution adopted by this invention is a HIFU interference removal method based on robust principal component analysis, with the following steps:
[0005] The first step is to conduct in vitro tissue and in vivo experiments. The Verasonics system is used to acquire ultrasound imaging data before and during the activation of high intensity focused ultrasound (HIFU). Each frame of data consists of 128 channels.
[0006] The second step is to construct the X matrix required for the RPCA framework from the frequency domain perspective, and decompose it into a low-rank matrix L and a sparse matrix S, and obtain the recovered spectrum from L.
[0007] The third step is to perform ultrasound imaging.
[0008] (1) The recovered time-domain channel data is obtained using inverse Fourier transform, and beamforming and angle combining are then performed;
[0009] (2) The radio frequency data is demodulated by Hilbert transform to obtain the corresponding envelope image, and after logarithmic compression, the interference-free B-mode ultrasound image is obtained.
[0010] The detailed steps for the second step are as follows:
[0011] (1) Perform a one-dimensional Fourier transform on each channel of data to obtain a one-dimensional spectrum. Combine the spectra of the 128 channels of data into a two-dimensional spectrum diagram with 128 columns;
[0012] (2) Connect the two-dimensional spectrograms end to end to form a one-dimensional sequence; combine the five interference-free spectrogram sequences with the one interference frame spectrogram sequence to form the matrix X to be decomposed;
[0013] (3) Use the RPCA matrix decomposition framework to decompose the X matrix into a low-rank matrix L and a sparse matrix S;
[0014] (4) Extract the spectrum sequence corresponding to the interference frame from the low-rank matrix L and restore it to 128-channel spectrum data.
[0015] The detailed steps of step (3) of the second step are as follows:
[0016] Robust Principal Component Analysis (RPCA) method for PCA is adopted through matrix factorization: the matrix X is divided into two parts: a low-rank matrix L and a sparse matrix S, i.e., X = L + S. The factorization procedure can be expressed as the following minimal model:
[0017]
[0018] here, It represents the rank of the matrix and serves as a sparse normalization term for the singular values of the matrix. The number of non-zero elements in the matrix S represents the sparsity of the matrix S, relaxing the non-convex optimization problem into a new convex optimization problem:
[0019]
[0020] in The kernel constant of matrix L is given by the sum of the singular values of the matrix; It is considered as the 1-norm of matrix S, given by the sum of the magnitudes of each vector in the matrix; hyperparameter , where n and m are the dimensions of X. The RPCA algorithm is solved using the principal component tracking method and iteratively calculated using the augmented Lagrange multiplier (ALM) algorithm. Specifically, the augmented Lagrange multiplier is first constructed:
[0021]
[0022] Where Y is the Lagrange multiplier matrix and υ is the hyperparameter. Denotes the F-norm;
[0023]
[0024] Then, solve and Minimize L, k to represent the number of iterations, and update the Lagrange multipliers until the formula converges.
[0025] A robust principal component analysis-based HIFU interference removal system comprises a computer and the Verasonics ultrasound imaging system. Verasonics acquires ultrasound imaging data before and during HIFU activation, with each frame consisting of 128 channels.
[0026] In computers:
[0027] 1. Construct the X matrix required for the RPCA framework from the frequency domain perspective, and decompose it into a low-rank matrix L and a sparse matrix S, and obtain the recovered spectrum from L;
[0028] II. Performing ultrasound imaging
[0029] (1) Use inverse Fourier transform to obtain the recovered time-domain channel data, and perform beamforming and angle combining;
[0030] (2) The radio frequency data is demodulated by Hilbert transform to obtain the corresponding envelope image, and after logarithmic compression, an interference-free B-mode image is obtained.
[0031] The features and beneficial effects of this invention are:
[0032] This invention experimentally demonstrates the hypothesis that matrix factorization can be used to eliminate HIFU acoustic interference in B-mode guided images, enabling real-time monitoring of HIFU treatment. In vitro and in vivo experiments in a clinical setting confirmed the feasibility of the proposed frequency-domain RPCA method. This novel method offers advantages in dynamically observing HIFU-induced tissue damage and comprehensively evaluates the time series before, during, and after HIFU. This new technology has great potential to overcome the difficulties of current real-time monitoring of HIFU ablation. Attached Figure Description
[0033] Figure 1 Experimental system diagram.
[0034] Figure 2 Two-dimensional spectrum.
[0035] Figure 3 Construction of the X matrix and RPCA decomposition diagram.
[0036] Figure 4 B-mode image and image after RPCA interference removal. Detailed Implementation
[0037] This invention proposes a HIFU interference removal system based on robust principal component analysis. It assumes that the contaminated ultrasound image consists of two parts: the HIFU signal and the ultrasound imaging signal, and establishes a frequency domain robust principal component analysis (RPCA) imaging processing method to separate the ultrasound imaging signal. The technical solution of this invention is as follows:
[0038] A HIFU interference removal system based on robust principal component analysis includes a memory, a processor, and a computer program running on the processor. When the processor executes the program, it performs the steps of the following method:
[0039] The first step is to conduct in vitro tissue and in vivo experiments. The Verasonics system is used to acquire ultrasound imaging data before and during HIFU activation. Each frame of data consists of 128 channels.
[0040] The second step is to construct the X matrix required for the RPCA framework from the frequency domain perspective, and decompose it into a low-rank matrix L and a sparse matrix S, and obtain the recovered spectrum from L.
[0041] (1) Perform a one-dimensional Fourier transform on each channel of data to obtain a one-dimensional spectrum. Combine the spectra of the 128 channels of data into a two-dimensional spectrum diagram with 128 columns.
[0042] (2) Connect the two-dimensional spectrograms end to end to form a one-dimensional sequence; combine the five interference-free spectrogram sequences with the one interference frame spectrogram sequence to form the matrix X to be decomposed.
[0043] (3) Use the RPCA matrix decomposition framework to decompose the X matrix into a low-rank matrix L and a sparse matrix S.
[0044] (4) Extract the spectrum sequence corresponding to the interference frame from the low-rank matrix L and restore it to 128-channel spectrum data.
[0045] The third step is to perform ultrasound imaging.
[0046] (1) Use inverse Fourier transform to obtain the recovered time-domain channel data, and perform beamforming and angle combining.
[0047] (2) The radio frequency data is demodulated by Hilbert transform to obtain the corresponding envelope image, and after logarithmic compression, an interference-free B-mode image is obtained.
[0048] To make the technical solution of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. It includes a memory, a processor, and a computer program running on the processor, as shown in the flowchart below. Figure 1 As shown, the specific steps are as follows:
[0049] The first step is to conduct in vitro tissue and in vivo experiments, using the Verasonics system to acquire ultrasound imaging data before and during HIFU activation.
[0050] This invention combines a HIFU treatment system (JC-200, Haifu, Chongqing, China) and an ultrasound imaging system (Vantage Verasonics 256, Kirkland, WA, USA) in experiments to obtain experimental data, such as... Figure 1 As shown. For HIFU treatment, a continuous waveform with a center frequency of 0.97 MHz from a signal generator (RIGOL, DG4202, Jiangsu, China) is amplified by a power amplifier to drive the HIFU transducer (outer diameter: 210 mm, inner diameter: 58 mm, focal length: 175 mm, center frequency: 0.97 MHz; -6 dB bandwidth: 0.88 MHz - 1.08 MHz; manufactured by Chongqing HIFU Co., Ltd., China). An electrical matching box is required between the power amplifier and the HIFU transducer, and the driving acoustic power is calibrated by an acoustic radiation force scale. The ultrasound imaging system uses a 128-element convex array transducer C5-2 (center frequency 3.13 MHz) for ultrasound imaging. Raw channel data can be acquired by this research ultrasound system for customized post-processing. The imaging transducer C5-2 is coaxially mounted with the HIFU transducer, and its imaging plane is calibrated to intersect the focal center of the HIFU. The signal generator is used to establish synchronization between HIFU treatment and ultrasound imaging, and to control the timing sequence between the two modules. The ambient temperature in the water tank is maintained at room temperature.
[0051] The second step is to construct the X matrix required for the RPCA framework from the frequency domain perspective, and decompose it into a low-rank matrix L and a sparse matrix S, and obtain the recovered spectrum from L.
[0052] (1) such as Figure 2 As shown, a one-dimensional Fourier transform is performed on each channel of data to obtain a one-dimensional spectrum. The spectra of the 128 channels of data can then be used to form a two-dimensional spectrum diagram with 128 columns.
[0053] (2) Connect the first and last two-dimensional spectrograms to form a one-dimensional sequence; combine the five interference-free spectrogram sequences with the one interference frame spectrogram sequence to form the matrix X to be decomposed, as shown below. Figure 3 As shown.
[0054] (3) Use the RPCA matrix decomposition framework to decompose the X matrix into a low-rank matrix L and a sparse matrix S.
[0055] Robust Principal Component Analysis (RPCA) is a technique that evolved from the classical statistical procedure of Principal Component Analysis (PCA). PCA is typically used to find the dominant subspace of a two-dimensional matrix X ∈ Rm xn. This subspace has a low dimension and can approximate the original dataset. The biggest drawback of PCA is its sensitivity to outliers due to the use of the Frobenius criterion. To make PCA more robust, various methods have been proposed, called robust PCA, among which matrix factorization has recently attracted considerable attention. In this paper, RPCA refers only to PCA implemented through matrix factorization. This RPCA does not derive the dominant subspace through a statistical procedure, but instead assumes that the matrix X can be divided into two parts: a low-rank matrix L and a sparse matrix S, i.e., X = L + S. The factorization procedure can be expressed as the following minimal model:
[0056]
[0057] here, It represents the rank of the matrix and serves as a sparse normalization term for the singular values of the matrix. The rank function represents the number of non-zero elements in the matrix S, measuring its sparsity. Due to the non-convexity of the rank function and the zero norm, solving this minimization problem is a complex NP-hard problem. To obtain an approximate solution, the original non-convex optimization problem can be relaxed into a new convex optimization problem:
[0058] ,
[0059] in The kernel constant of matrix L is given by the sum of the singular values of the matrix; It is considered as the 1-norm of the matrix, given by the sum of the magnitudes of each vector in the matrix; hyperparameter , where n and m are the dimensions of X. The RPCA algorithm is solved using the principal component tracking method and iteratively computed using the enhanced Lagrange multiplier (ALM) algorithm. Specifically, we first construct the amplified Lagrange multiplier:
[0060]
[0061] Where Y is the Lagrange multiplier matrix and υ is the hyperparameter. This represents the F-norm.
[0062]
[0063] Then, solve and Minimize L, where k represents the number of iterations, and update the Lagrange multiplier until the formula converges.
[0064] (4) Extract the spectrum sequence corresponding to the interference frame from the low-rank matrix L and restore it to 128-channel spectrum data.
[0065] The third step is to perform ultrasound imaging.
[0066] (1) Use inverse Fourier transform to obtain the recovered time-domain channel data, and perform beamforming and angle combining.
[0067] (2) Demodulate the radio frequency data using Hilbert transform to obtain the corresponding envelope image. After logarithmic compression, obtain an interference-free B-mode image, such as... Figure 4 As shown, the RPCA method can effectively eliminate HIFU interference and reveal tissue information.
Claims
1. A HIFU interference removal system based on robust principal component analysis, characterized in that, Verasonics, which includes computers and ultrasound imaging systems, has the following workflow: The first step is to conduct in vitro tissue and in vivo experiments. The Verasonics system is used to acquire ultrasound imaging data before and during the activation of high intensity focused ultrasound (HIFU). Each frame of data consists of 128 channels. The second step is to construct the X matrix required for the RPCA framework from the frequency domain perspective, and decompose it into a low-rank matrix L and a sparse matrix S. The recovered spectrum is then obtained from L. The detailed steps are as follows: (1) Perform a one-dimensional Fourier transform on each channel of data to obtain a one-dimensional spectrum, and combine the spectra of the 128 channels of data into a two-dimensional spectrum diagram with 128 columns. (2) Connect the two-dimensional spectrograms end to end to form a one-dimensional sequence; combine the five interference-free spectrogram sequences with the one interference frame spectrogram sequence to form the matrix X to be decomposed; (3) Use the RPCA matrix decomposition framework to decompose the X matrix into a low-rank matrix L and a sparse matrix S; (4) Extract the spectrum sequence corresponding to the interference frame from the low-rank matrix L and restore it to 128-channel spectrum data; In step (3): Robust Principal Component Analysis (RPCA) method for PCA is adopted through matrix factorization: the matrix X is divided into two parts: a low-rank matrix L and a sparse matrix S, i.e., X = L + S. The factorization procedure can be expressed as the following minimal model: It represents the rank of the matrix and serves as a sparse normalization term for the singular values of the matrix. The number of non-zero elements in the matrix S represents the sparsity of the matrix S, relaxing the non-convex optimization problem into a new convex optimization problem: in The kernel constant of matrix L is given by the sum of the singular values of the matrix; It is considered as the 1-norm of matrix S, given by the sum of the magnitudes of each vector in the matrix; hyperparameter Where n and m are the dimensions of X, the RPCA algorithm is solved using the principal component tracking method, and the augmented Lagrange multiplier (ALM) algorithm is used for iterative calculation. Specifically, the augmented Lagrange multiplier is first constructed: Where Y is the Lagrange multiplier matrix and υ is the hyperparameter. Denotes the F-norm; Then, solve and Minimize L, k to represent the number of iterations, and update the Lagrange multipliers until the formula converges. The third step is to perform ultrasound imaging. (1) The recovered time-domain channel data is obtained using inverse Fourier transform, and beamforming and angle combining are then performed; (2) The radio frequency data is demodulated by Hilbert transform to obtain the corresponding envelope image, and after logarithmic compression, the interference-free B-mode ultrasound image is obtained.