WO2019184343A1 - Gpu-based multiple mv high-definition algorithm fast medical ultrasound imaging system - Google Patents

Abstract

Disclosed in the present invention is a graphics processing unit (GPU)-based multiple MV high-definition algorithm fast medical ultrasound imaging system. The system of the present invention comprises an initial startup module, a simulation module, an algorithm imaging module, and a result display module; after the initial startup module completes system initialization configuration, the algorithm imaging module is called to perform imaging, and the simulation module is used as an auxiliary module to generate simulation data for other modules to use. The algorithm imaging module is a core module of the entire system and receives a parameter configuration from the initial startup module and data from the simulation module or a data acquisition device for imaging calculation, and the result display module performs final processing and display of the results of other modules, thereby finally achieving a rapid imaging function of multiple MV algorithms. The present invention may complete the complex calculation of high-definition medical ultrasound image algorithms in a very short period of time, and the GPU programming implementation schemes of various MV high-definition algorithms may be conveniently deployed on a computer having a GPU, which may effectively meet user requirements and which has high practicability.

Description

GPU-based multiple MV high-definition algorithm fast medical ultrasound imaging system Technical field

The invention belongs to the field of medical ultrasound imaging, and particularly relates to a GPU-based multi-MV high-definition algorithm rapid medical ultrasound imaging system.

Background technique

In recent years, medical ultrasound imaging technology has been widely used and rapidly developed. Compared with other medical imaging methods, ultrasound imaging technology has the advantages of real-time output of images, high security and low cost. As a common medical diagnostic technique, it is commonly used to observe fetal development, heart movement and blood flow. Among them, the beamforming algorithm is a key part of medical ultrasound imaging technology and has a key impact on image quality.

The ultrasound imaging algorithm widely used nowadays is a delayed superposition beamforming algorithm, which is an algorithm that can be easily implemented on different computing platforms and can realize real-time requirements quickly. Although the delayed superposition beamforming algorithm is effective for many diagnostic scenarios, it does not provide sufficient anatomical detail in some complex diagnostic scenarios. Therefore, the present invention applies a variety of higher definition adaptive minimum variance beamforming (MV) algorithms. The adaptive beamforming algorithm is adaptive to the input echo data and dynamically calculates the apodization weights at runtime. By utilizing the real-time statistical properties of the input echo signals, these algorithms effectively improve imaging quality and provide more anatomical details to improve diagnostic accuracy.

The cost of MV algorithm for high quality imaging is higher computational complexity, so its calculations take more time. Such computational complexity hinders the practical clinical application of the minimum variance beamforming algorithm. The core calculation of the traditional medical ultrasound imaging imaging system is usually implemented on the CPU of the central processing unit, but the huge imaging data of the high-definition imaging algorithm and the complicated operation process make the traditional CPU unable to meet the needs of its rapid imaging, and the current GPU ( Graphics processors are growing rapidly, and a single GPU can integrate hundreds or even thousands of computing cores with extremely powerful high-performance computing power. Therefore, the present invention implements a plurality of MV algorithms on a high-performance GPU, greatly improving the imaging speed of the algorithm, thereby completing the core processing module of the high-definition medical imaging device, and providing a rapid imaging basis for clinical diagnosis.

At present, most of the imaging algorithms used in portable medical ultrasonic detectors are traditional time-delay imaging algorithms. The algorithm is simple in operation and can meet the real-time imaging requirements of medical ultrasound, but the image quality is relatively low. Or a higher clearing algorithm is used, but its imaging speed is not ideal due to its large calculation amount.

Summary of the invention

The main object of the present invention is to solve the problem of low imaging quality and slow speed of medical ultrasound imaging devices. In order to realize the application of the high-definition imaging algorithm in the medical ultrasonic detector, the invention provides a GPU-based MV high-definition algorithm rapid medical ultrasound imaging system, and the GPU-based high-speed parallel computing capability completes imaging implementation of various MV high-definition imaging algorithms, Fast HD imaging requirements.

The object of the present invention is achieved by the following technical solutions.

GPU-based MV high-definition algorithm rapid medical ultrasound imaging system, including initial startup module, simulation simulation module, algorithm imaging module and result display module; wherein, after the initial startup module completes the system initialization setting, the algorithm imaging module is called for imaging, simulation The module acts as an auxiliary module to generate analog data for use by other modules. The algorithm imaging module is the core module of the whole system, and receives the parameter settings from the initial startup module and the data of the module simulation module or the data acquisition device for imaging calculation, and the result display module The final processing and display of the results of other modules, each module works independently and cooperates with each other, and finally realizes the rapid imaging function of various MV algorithms.

Further, the initial startup module mainly performs initialization work required during system startup, including initialization parameters, setting imaging scenes, setting enabled algorithms, evaluation indicators, and image display parameters, and most of the initialization settings are consistent with conventional ultrasound imaging systems. Option, and for the setting of imaging algorithm, the system provides a variety of high-definition imaging algorithms, including conventional MV algorithm, iterative MV algorithm, fused diagonally loaded MV algorithm, forward and backward smooth MV algorithm, fusion of generalized coherence factor MV algorithm, MV algorithm based on feature space processing and MV algorithm combining feature space and symbol coherence coefficient. As an initial startup, this module performs system check on the one hand to prevent erroneous diagnosis results caused by equipment failure. On the one hand, in addition to the relatively comprehensive initial settings built in the system, it can also be used by professionals to use initial parameters according to actual usage scenarios. The algorithm and display range are set, which can effectively shorten the running time and increase the accuracy. After the basic startup module is completed, the imaging module can be imaged by the simulation simulation module or the imaging data acquisition device.

Further, the simulation module uses the Field II analog ultrasound imaging process to obtain simulation data, first simulates the physical data required by the actual ultrasound imaging device setting, creates the transmitting and receiving array elements, and additionally needs to create a virtual object according to the actual application scenario, and then Calculate and collect echo data one by one according to the scan line. The simulation data obtained by this module can be used as the initial input data of the algorithm imaging module. On the one hand, it can be used for the warm-up start of the system software and hardware. On the one hand, it can be used as a comparison benchmark to evaluate or preview the imaging performance of each algorithm in different scenarios. Provides a reference for the algorithm selection of the actual diagnostic scenario. This module is usually initiated by the initial startup module and passes the generated data to the algorithm imaging module for further imaging calculations.

Further, the algorithm imaging module is the core part of the system. One of the basic functions that an ultrasound system needs to achieve during imaging is to create a specific sound field distribution in the imaging target area, ie beamforming. The beamforming algorithm is a key part of medical ultrasound imaging technology and has a critical impact on image quality. The algorithm imaging module of the system includes conventional MV algorithm, iterative MV algorithm, FFT algorithm with diagonally loaded loading, forward and backward smooth MV algorithm, MV algorithm combining generalized coherence factor, MV algorithm based on feature space processing and feature space and symbol coherence Coefficient fusion MV algorithm. These algorithms are high-definition adaptive beamforming algorithms. Compared with the traditional pre-fixed weight DAS algorithm, the algorithm in this module can obtain higher quality imaging. At the same time, the system implements these 7 algorithms based on GPU. Imaging, using the high-performance computing power of the GPU, solves the problem of excessive computational complexity caused by high-quality imaging.

The algorithm imaging module is started by the initial module, and the original signal data required for imaging is obtained by the data acquisition device or the simulation module, and then the pixel amplitude value is calculated according to the MV algorithm, the imaging is obtained, and the imaging result is given to the result display module. The final ultrasound image is displayed.

Further, after comprehensive research on the comparison and analysis of each MV algorithm, the main calculation process of the algorithm imaging module can be divided into calculating signal delay, calculating covariance matrix, calculating adaptive weight, calculating optimization operator and calculating pixel amplitude value 5 In part, these calculation processes are completed by multi-thread cooperation in the GPU thread block; wherein the calculation part of the optimization operator is the difference of each MV algorithm, embodies the core principles of each algorithm, and the more specific calculation principle of each algorithm and Implementations in the GPU will be explained in the examples.

The algorithm imaging module can be called after the initial startup module is set according to the actual diagnosis scenario requirement, and the MV high-definition imaging algorithm is encapsulated by the strategy mode, and the pluggable call is realized, and the addition, deletion or upgrade operation can be conveniently performed. At the same time, each MV algorithm can combine different parameters such as M and L (the number of receiving signal array elements and the number of sub-array elements) to provide more dimension choices for clinical diagnosis.

Further, the result display module further processes the calculation result of the algorithm imaging module, and then outputs the imaging result to the display for display; after the beam imaging is performed, the envelope detection is usually performed after the beam formation (using the Hilbert transform) ), logarithmic compression and grayscale range correction, the result display module outputs the image related data to the corresponding horizontal and vertical after the Hilbert transform, logarithmic compression, and grayscale range correction operations are performed by calling the corresponding function of Maltab. In the coordinate axis, the ultrasound imaging result is finally displayed on the screen; at the same time, the result display module also collects relevant information of the actual diagnosis scene such as hardware status, imaging parameters, evaluation indicators, etc., and displays them in a suitable area of the display. Provide users with comprehensive and detailed information for diagnostic reference.

The invention adopts a plurality of MV high-definition algorithms based on GPU to realize fast medical ultrasound imaging, and the imaging algorithm itself adapts and improves the GPU architecture, so that the calculation process is applicable to the GPU computing environment, thereby improving medicine. Image quality and imaging speed for ultrasound imaging. The programming implementation scheme adopted by the system is based on the GPU programming architecture and memory hierarchy for targeted parallel programming of various MV HD algorithms to realize the fast imaging function of each HD algorithm.

The present invention implements a fast medical ultrasound high-definition imaging system in a computer system with GPU hardware in accordance with the current state of the art. It is characterized in that a plurality of MV algorithms are implemented based on the GPU to achieve rapid medical ultrasound high-definition imaging purposes, and the focus is on parallelizing the calculation process of various high-definition MV algorithms, so that the operation flow is applicable to the computing environment of the GPU, thereby fully The GPU's computing power is utilized to improve the image quality and imaging frame rate of medical ultrasound imaging.

Compared with the prior art, the advantages of the present invention are mainly embodied in two aspects: on the one hand, the present invention comprehensively implements a plurality of high-definition MV imaging algorithms, which can effectively improve the image quality of medical ultrasound imaging, and the imaging effect is more than the conventional delay. The superposition algorithm is much better; on the other hand, the powerful computational power of the GPU properly solves the huge computational requirements of each high-definition MV imaging algorithm, thereby greatly increasing the output imaging frame rate of the imaging system, and can quickly output high-definition medical ultrasound images.

DRAWINGS

FIG. 1 is a schematic diagram of a workflow of a fast medical ultrasound imaging system for various MV high definition algorithms based on GPU.

Figure 2 (a) is a schematic diagram of the experimental scheme simulated in the example.

2(b) to (h) are diagrams showing an example of ultrasound imaging output by each algorithm in the example.

detailed description

The specific embodiments of the present invention are further described below in conjunction with the drawings and examples, but the implementation and protection of the present invention is not limited thereto. It should be noted that, if there is any detailed description or an existing algorithm formula symbol, etc., it can be implemented or understood by those skilled in the art with reference to the prior art, and details are not described herein again.

The system GPU programming architecture is divided into three levels: grid, block and thread. When the kernel calls GPU computing, all computational states are done in a GPU computing grid. A GPU computing grid computes a complete medical ultrasound image, and the output target image is decomposed into scan lines and pixels. Therefore, a GPU computing grid contains a matrix of computational blocks, each of which computes the pixel amplitude values according to the principles of each MV algorithm. In order to obtain a final amplitude estimate for a pixel, the computational details of each MV algorithm are implemented by some parallel computational threads (more specific algorithmic principles and their implementation have been described in the Summary of the Invention). These concurrent threads use shared memory to collaborate and perform array or matrix operations by computing independent elements. The optimal value setting for blocks and threads in the calculation hierarchy here depends on the computing platform's resources and the size of the computational problem.

In the implementation of the GPU, data is copied from the CPU memory to the GPU memory, and after being calculated by the GPU, the imaged data is copied back to the CPU memory for display. In this process, memory utilization is very important. The GPU has three basic memory levels: global memory, shared memory, and register memory. The three levels of memory access time relationship are: global memory <shared memory < register memory. Therefore, in order to give full play to the computing performance of the GPU, the global memory should be used as little as possible in the calculation process, and the register memory should be used more for calculation. However, the capacity relationship of the three memory types is: global memory > shared memory > register memory. Therefore, when the system allocates resources, it is considered that small variables are stored in registers, and large data is stored in global memory. .

The system is designed according to the actual use scenario. The overall module design is shown in Figure 1. It can be divided into four modules: initial startup module, simulation module, algorithm imaging module and result display module. Each module works independently and cooperates with each other, and finally realizes the rapid imaging function of various MV algorithms. The specific design of each module is explained below.

(1) Initial startup module

The initial startup module mainly performs the initialization work required during system startup, including initialization parameters, setting the imaging scene, setting the enabled algorithm and image display parameters. As an initial startup, this module performs system check on the one hand to prevent erroneous diagnosis results caused by equipment failure. On the one hand, in addition to the relatively comprehensive initial settings built in the system, it can also be used by professionals to use initial parameters according to actual usage scenarios. The algorithm and display range are set, which can effectively shorten the running time and increase the accuracy. After the basic setup is completed, the simulation module obtains the data required by the imaging simulation module to perform imaging.

(2) Simulation module

The simulation module uses the Field ii analog ultrasound imaging process to obtain simulation data. First, simulate the physical data required by the actual ultrasound imaging device setup, create the transmitting and receiving array elements, and additionally create a virtual object according to the actual application scenario, and then press the scan. Lines calculate and collect echo data one by one. The simulation data obtained by this module can be used as the initial input data of the algorithm imaging module to evaluate the imaging performance of each algorithm in this scenario, and provide reference for the selection algorithm.

(3) Algorithm imaging module

The algorithm imaging module contains a variety of medical ultrasound MV imaging algorithms and is a core part of the system. The module is started by the initial module, and the original signal data required for imaging is obtained by combining the simulation module, and then the pixel amplitude value calculation is performed according to the MV algorithm, the imaging is obtained, and the imaging result is given to the result display module to display the final ultrasound image. . After comprehensive research such as comparative analysis of each MV algorithm, the main calculation process of this module can be divided into five parts, namely, calculating signal delay, calculating covariance matrix, calculating adaptive weight, other improved calculation and calculating pixel amplitude value. The calculation process is done in a GPU thread block by multi-threaded collaboration. Among them, the improved calculation part is the difference of each MV algorithm, which embodies the core principle of each algorithm. The more specific calculation principle of each algorithm and its implementation in GPU will be explained below.

In order to facilitate the upgrade and expansion of the system algorithm, this module implements the policy mode, which is called by the initial startup module according to the needs. Each MV algorithm is independently implemented and pluggable, which is convenient to add, delete or upgrade. And each MV algorithm can combine different parameters such as M and L (the number of receiving signal array elements and the number of sub-array elements) to provide more choices for clinical diagnosis.

(4) Result display module

After the imaging module obtains the imaging data, the result is displayed by the result display module. The module performs the Hilbert transform, the logarithmic compression, the grayscale range correction, etc. after calling the Maltab function. The data is output to the corresponding horizontal and vertical axis, and finally, the ultrasound imaging result is displayed on the screen.

The following is a description of the various medical ultrasound MV imaging algorithms of the present example. The following is only an example. In view of the space, some of the contents, such as the meaning of the formula, are not described in detail, and those skilled in the art can implement or understand with reference to the prior art.

1, the conventional MV algorithm

(1) Introduction to the algorithm

According to the development of the minimum variance beamforming algorithm in medical ultrasound imaging, the most common adaptive beamforming algorithm is based on the adaptive minimum variance (MV) beamformer designed by Capon in 1969. The conventional MV algorithm is based on a delayed superposition beamforming algorithm that has the same input and output data streams and the same delay superposition process. The main difference is that the minimum variance beamforming algorithm uses apodized weight adaptive input of ultrasound data, while the delayed superposition beamforming algorithm cannot adapt to the fixed apodization weight of the input data. It is this primary difference that makes the quality of the output image of the minimum variance beamforming algorithm higher than that of the delay stacking algorithm.

The MV beamforming algorithm uses a subaperture averaging method. A receive aperture consists of M consecutive input data channels and is divided into a set of sub-apertures consisting of L consecutive input channels. Therefore, one receiving aperture consists of (M‐L+1) sub-apertures. By the subaperture averaging method, the covariance matrix R _MV (p) of the pixel point p can be calculated by the following equation.

Where x _i (p) is a (L × 1) vector consisting of echo signal data input in the i ^th subarray, eg x _i (p) represents the i ^{th th} to (i) in x(p) +L-1) A set of ^th elements, and x(p) is a (L × 1) vector composed of delayed signal values corresponding to pixel points. After the calculation of the covariance matrix R _MV (p) is completed, the adaptive apodization weight can be estimated by the following formula:

Since the data in the input channel has been delayed, here a is a simple direction vector that is all one. Finally, the amplitude value of pixel p can be estimated by:

(2) GPU implementation

The GPU implementation of the conventional MV algorithm is the basis for other improved MV algorithms. The calculation grid and the thread block size are organized by a two-dimensional subscript, wherein the X, Y size of the calculation grid corresponds to the number of scanned lines and the number of pixels on each scan line; X≥L in the thread block. Y is obtained according to the set total number of threads per thread block and the size of X. Before starting the GPU calculation, the space of the signal data, the covariance matrix, the weight vector value and the intermediate value after the delay is allocated in the shared memory, and the original signal data is stored in the GPU global memory through the CPU copy.

The delay calculation part obtains valid start and end array elements according to the scan line position, and then calculates the delay in parallel by the thread in the thread block to obtain the corresponding signal and stores the corresponding address in the shared memory. In this process, the position of the array element is symmetric. Sexuality, part of the signal can be obtained by only one calculation process to obtain the corresponding signal, thereby reducing the amount of calculation.

Since the covariance matrix of the MV algorithm has symmetry, only the values of the diagonal and upper triangular portions can be calculated, thereby achieving the purpose of saving computation. The calculation process performs parallel computation through multiple threads in one thread block.

In the weight vector calculation part, the system directly calculates the product of the inverse of the covariance matrix and the direction vector by Cholesky decomposition, thereby eliminating the inversion process and parallel computing through multi-thread cooperation in a thread block.

The final superimposed part is small in magnitude because of the actual array element, so it can be directly calculated by a specific thread. Through the above steps, the powerful computing performance of the GPU can be fully utilized, and the rapid imaging process of the high-definition MV algorithm can be realized.

2. Iterative MV algorithm

(1) Introduction to the algorithm

The core of the conventional MV beamforming algorithm is the estimation of the noise covariance matrix, and the main cause of the MV error is the result of using the sample covariance matrix instead of the noise covariance matrix. Therefore, Eng Nai Wee, Ser et al. proposed an iterative minimum variance beamforming (IMV) algorithm in 2011, hoping to improve the estimation accuracy of the noise covariance matrix by multiple iterations.

In this algorithm, the output of the conventional MV beamforming algorithm is used as an estimate of the expected signal amplitude:

The desired signal is separated from the received signal, and the rest is the interference and the acoustic component, as shown in equation (2-5):

Then just use

Instead of x _i (p), a new beamforming weight vector is calculated according to the formula of the conventional MV algorithm, and substituted into the equation (3-3) to finally obtain the amplitude estimation of the pixel. Since this step is equivalent to repeating the MV optimization process for the interference noise component, we call it the two-step iterative minimum variance (2StgMV) algorithm. And this step can be repeated as needed.

The IMV algorithm can significantly improve the contrast of the MV algorithm, but the inverse calculation in the iterative process will cause a large computational overhead, which is not conducive to real-time performance of the algorithm.

(2) GPU implementation

The computational process of the iterative MV algorithm is essentially the same as that of the conventional MV algorithm, except that the process of weighting in the conventional MV algorithm is repeated, and the amplitude estimate of the previous iteration is subtracted in the calculation process. Therefore, the amplitude value obtained by the last calculation is recorded by adding a variable, and is subtracted every time the original signal is read. The rest of the process uses the parallel calculation scheme of the conventional MV algorithm, so that iteration can be obtained by further iterative calculation. The pixel amplitude value of the MV algorithm. Its imaging speed is compared to the conventional MV algorithm depending on the number of iterations desired.

3. Fusion diagonal algorithm for MV loading

(1) Introduction to the algorithm

The performance of conventional MV beamforming algorithms will degrade rapidly due to inaccurate estimates of frequency or direction of arrival. Li et al. use a diagonal loading (DL) approach to improve the robustness of the MV beamforming algorithm. In 2005, Sasso et al. applied it to the field of medical ultrasound imaging, which also improved the imaging resolution to some extent.

The diagonal loading process is to superimpose a certain proportion of the unit matrix on the original matrix, as shown in the following equation:

Where δ is the diagonal load, trace{} represents the trace of the matrix, and I represents the unit array. Then use

Instead of R _MV (p), the calculation is performed in the same manner as in the equation (3‐2) and the equation (3‐3) to obtain the amplitude estimation of the pixel.

The meaning of diagonal loading is to superimpose Gaussian white noise proportional to the signal power on the original signal, increasing the independence between signals. However, while improving the robustness, it will reduce the adaptability of the algorithm, resulting in lower resolution, and the imaging performance tends to the traditional delay superposition algorithm. Therefore, the diagonal loading should not be too large, and should satisfy: δ<< 1/L, and usually takes: δ = 1 / (100L).

(2) GPU implementation

The MV algorithm combined with diagonal loading mainly adds a certain proportion of Gaussian white noise to the conventional MV algorithm after the covariance matrix is obtained, so the GPU parallel implementation scheme of the conventional MV algorithm can still be used. Then, based on this On the other hand, after the calculation of the covariance matrix is completed, the part of the thread is calculated to calculate the value to be added, and then added to the diagonal of the matrix. Since there is no access conflict in the overlay process, parallel operations can be performed through L threads in the thread block to complete the diagonal loading process faster, and the purpose of fast imaging is achieved.

4, forward and backward smooth MV algorithm

(1) Introduction to the algorithm

The forward and backward smoothed minimum variance (FBMV) beamforming algorithm is an improved algorithm of AsL et al. for the MV algorithm proposed to improve contrast. The core idea of the algorithm is to estimate the more accurate covariance matrix needed to calculate the apodization weights. It uses a forward-backward smoothing technique instead of the traditional forward-only spatial smoothing, resulting in more accurate background speckle statistics and thus improved contrast.

The estimation method given by equation (3‐1) can be considered as an estimate of forward spatial smoothing, which is recorded as

Estimation of backward spatial smoothing by calculation of equation (3‐8)

Where J represents the exchange matrix, that is, the unit matrix is flipped left and right.

Furthermore, a more accurate estimate can be obtained by combining the forward and backward spatial smoothing estimates. The expression is as follows:

Then use

Instead of R _MV (p), the calculation is performed in the same manner as in the equation (3‐2) and the equation (3‐3) to obtain the amplitude estimation of the pixel. This smoothing process also makes it more robust than conventional MV algorithms.

(2) GPU implementation

It can be seen from the algorithm formula that the forward-and-forward smoothing MV algorithm corresponds to the forward spatial smoothing estimate obtained by the conventional MV algorithm, and the backward spatial smoothing value can be obtained in the forward estimated relative position. Therefore, after the conventional MV algorithm completes its covariance matrix calculation, the parallel scheme when calculating the covariance matrix is used to obtain the backward smoothed value stored in the temporary variable, and at the same time, the value of the joint estimation is forward and backward. The estimated averages are thus numerically equal, and the average calculation can be performed directly in the corresponding position in the matrix after the values are obtained. This scheme can achieve good parallel computing without increasing memory overhead, which greatly saves the computation time of the joint covariance matrix.

6. MV algorithm combining generalized coherence factors

(1) Introduction to the algorithm

In order to obtain high resolution and correct phase error caused by uneven sound velocity in tissue propagation, Wu Wenzhao et al. proposed a high-resolution medical ultrasound imaging method combining generalized coherence factor (GCF) and MV beamforming algorithm in 2010.

The calculation of GCF is based on multi-channel data after focus delay. First, the array element domain data is transformed from the array element domain to the beam domain by Fourier transform:

Where y(m) is the data after transforming into the beam domain, and then calculating the energy of each beam direction, and obtaining the ratio of the energy of the coherent direction to the total energy as the value of GCF:

Where K is the energy ratio that controls the low frequency component of GCF. By changing the value of K, the performance of the algorithm can be changed.

In order to calculate the amplitude value of the pixel, the calculation process of the conventional MV algorithm should be completed once, and then combined with the GCF by the equation (3-12) to obtain the final amplitude estimate of the pixel.

v _{MV_GCF} (p)=c _GCF (p)·v _MV (p). (3-12)

This method combines the robustness of the generalized coherence coefficient with phase error and the high resolution capability of the MV algorithm.

(2) GPU implementation

The MV algorithm combining the generalized coherence factor has a calculation process for calculating the pixel amplitude value consistent with the conventional MV algorithm. Therefore, the GPU implementation scheme of the conventional MV algorithm is still applied thereto. Then, after obtaining the pixel amplitude value, Fourier transform is performed on the array metadata stored in the shared memory, and the coherence factor is calculated according to the above algorithm, and finally the amplitude value obtained by the conventional MV algorithm is multiplied. In the calculation part of the algorithm improvement, the part without read-write cross-linking can still be performed in parallel in the same thread block through multiple threads, thereby speeding up the calculation process. In addition, in the implementation process, the system also tries to use less equivalent calculation of equivalent replacement calculation or unnecessary calculation to further accelerate the imaging speed.

7. MV algorithm based on feature space processing

(1) Introduction to the algorithm

The minimum variance beamforming algorithm can improve the resolution of the image to a certain extent, but can not significantly improve the contrast of the image. To this end, Asl et al. proposed a minimum variance (ESBMV) beamforming algorithm based on feature space processing. The ESBMV algorithm uses the feature structure of the covariance matrix to enhance the performance of the conventional MV algorithm, which can reduce the main lobe signal while reducing the side. The amplitude of the flap improves imaging quality.

The steps of the ESBMV algorithm before calculating the apodization weight are the same as the conventional MV algorithm. Then, the eigendecomposition of the covariance matrix can be expressed as:

Here Λ=diag(λ ₁ ,λ ₂ ,...,λ _L ), where λ ₁ ≥λ ₂ ≥...≥λ _L is the eigenvalue of the covariance matrix; U=[v ₁ ,v ₂ ,...,v _L ], v _i is a feature vector corresponding to λ _i .

a signal subspace consisting of N _sig eigenvectors corresponding to larger eigenvalues, and

Then it is the noise subspace. The value of N _sig determines the ability of the beamforming method to maintain the main lobe signal and reduce the side lobe signal, usually greater than the maximum eigenvalue α (0.1 ≥ α ≥ 5.5) times or greater than the minimum eigenvalue β (0.1 ≥ β ≥. 5) The number of feature vectors is determined by multiples.

The weighted value obtained by the conventional MV algorithm is projected to the signal subspace obtained by the feature space method, and the weighted value of the ESBMV is obtained:

The resulting output of the resulting ESBMV beamforming is:

Compared to conventional MV beamforming, ESBMV is capable of effective contrast and signal to noise ratio. However, its drawback is that when the noise exceeds the strength of the desired signal, the algorithm will recognize the noise as the desired signal, which may result in a completely erroneous result.

(2) GPU implementation

The MV algorithm based on feature space processing has one more computational feature space decomposition process than the conventional MV algorithm. Based on the consideration of precision and parallelism, the Jocobi decomposition method is used for feature decomposition. The part that still has no access conflict is paralleled by multi-thread in the thread block as much as possible. In order to perform parallel cooperation between threads, it is necessary to add shared memory for storing the intermediate value of the feature decomposition process based on the original MV algorithm. Here, it is mainly used to store feature vectors, and the feature values can be stored in the last calculation that has not been used yet. Allocated in shared memory. Because the size of the shared memory is limited, the number of thread blocks that can be run in parallel in the actual GPU is limited. Therefore, in order to save memory, in the implementation process, according to the symmetry of the calculation process, it is only necessary to increase the capacity of the matrix by half, and when the data needs to be used. Then the corresponding data is obtained by subscript conversion calculation, which not only saves the use of shared memory, increases the number of parallel thread blocks, but also reduces the calculation amount by nearly half, so as to achieve the purpose of imaging as fast as possible.

8. MV algorithm for feature space and symbol coherence coefficient fusion

(1) Introduction to the algorithm

The resolution of the image is mainly determined by the width of the main lobe, and the resolution of the image obtained by ESBMV beamforming is almost the same as that of the conventional MV algorithm. In order to improve the image contrast while further improving the image resolution, Liu Tingping people cohered the symbols in 2015. The coefficients are introduced into the ESBMV algorithm, and a beamforming algorithm combining feature space and symbol coherence coefficient (SCF) is proposed.

The SCF can be considered as a nonlinear function of the beamforming device. It requires less hardware resources and can be quickly integrated into the beamforming device, and its value can be calculated by equation (3-16):

Where b _i (p) can be obtained by the formula (3-17):

The other calculations are consistent with the ESBMV algorithm, except that its final output is obtained by equation (2-18):

v _{ESBMV_SCF} (p)=c _SCF (p)·v _ESBMV (p). (3-18)

This method combines the robustness of the coherence coefficient with respect to the phase error with the high resolution of the ESBMV.

(2) GPU implementation

The computational process of MV algorithm with feature space and symbol coherence coefficient is mainly based on the MV algorithm based on feature space processing and then multiplied by a symbol coherence coefficient. Therefore, its GPU implementation is consistent with the MV algorithm based on feature space processing. In the end, only one step of the symbol coherence coefficient is calculated according to the formula as described above. This process is the same as the MV algorithm which combines the generalized coherence factor, and the unnecessary calculation is saved by thread parallelism, replacement calculation, etc. to accelerate the imaging. speed.

The system uses a Field II simulator to simulate ultrasound channel data samples to perform a series of related experiments to obtain the imaging performance of the system. The following simulation simulates a vesicle virtual image in a speckle as shown in Fig. 2(a) as an imaging scene. The field II simulator was used to simulate the echo data samples of the ultrasonic channel. The simulation parameters are set in Table 1.

Table 1 Setting of Field II simulation parameters in the experiment

Each of the aforementioned MV algorithms is analyzed experimentally, each of which is programmed as a beamformer and parallelizes the pixel calculation process to save program run time. In order to obtain a comparatively significant result, the parameters of each algorithm are set to their usual values, for example, the number of subaperture channels is set to 32. The results of the test are shown in Figures 2(b) to (h), and all images show a dynamic range of 60 dB. Comparing these figures, it can be seen that the imaging sharpness of various beamformers is sufficient to distinguish the target, which embodies the ability of these MV algorithms for high-definition imaging, but the contrast of different algorithms still has some differences in this imaging scene. In clinical use, different algorithms need to be selected according to actual needs for imaging. The beamforming algorithm is a key part of medical ultrasound imaging technology. The MV adaptive beamforming algorithm utilizes the real-time statistical characteristics of the input echo data, so it has higher output imaging than the traditional DAS beamforming algorithm. quality. The system realizes rapid imaging of a variety of high-definition MV algorithms through the powerful high-performance computing power of the GPU, which can provide a fast and effective imaging basis for clinical diagnosis.

Claims (6)

1. GPU-based MV high-definition algorithm rapid medical ultrasound imaging system, which is characterized by comprising an initial startup module, an analog simulation module, an algorithm imaging module and a result display module; wherein, after the initial startup module completes the system initialization setting, the algorithm imaging module is called to perform imaging. The simulation module is used as an auxiliary module to generate analog data for use by other modules. The algorithm imaging module is the core module of the whole system, and receives the parameter settings from the initial startup module and the data of the module simulation module or the data acquisition device for imaging calculation, and the result The display module performs final processing and display on the results of other modules, and each module works independently and cooperates with each other, and finally realizes rapid imaging functions of various MV algorithms.
2. The GPU-based MV high-definition algorithm rapid medical ultrasound imaging system according to claim 1, wherein the initial startup module mainly performs initialization work required during system startup, including initializing parameters, setting an imaging scenario, setting an enabled algorithm, Evaluation indicators and image display parameters, most of which provide the same options as traditional ultrasound imaging systems, and for imaging algorithm settings, the system provides a variety of high-definition imaging algorithms, including conventional MV algorithms, iterative MV Algorithm, FFT algorithm with diagonal loading, forward and backward smooth MV algorithm, MV algorithm with generalized coherence factor, MV algorithm based on feature space processing and MV algorithm with feature space and symbol coherence coefficient fusion.
3. The GPU-based MV high-definition fast medical ultrasound imaging system according to claim 1, wherein the simulation module uses the Field II analog ultrasound imaging process to obtain simulation data, and first simulates physical data required for the actual ultrasound imaging device setting. Create a transmit and receive array element. In addition, you need to create a virtual object according to the actual application scenario. Then, calculate and collect the echo data one by one according to the scan line.
4. The GPU-based MV high-definition fast medical ultrasound imaging system according to claim 1, wherein the algorithm imaging module comprises a conventional MV algorithm, an iterative MV algorithm, a fused diagonally loaded MV algorithm, a forward and backward smooth MV algorithm, MV algorithm combining generalized coherence factor, MV algorithm based on feature space processing and MV algorithm combining feature space and symbol coherence coefficient.

   The algorithm imaging module is started by the initial module, and the original signal data required for imaging is obtained by the data acquisition device or the simulation module, and then the pixel amplitude value is calculated according to the MV algorithm, the imaging is obtained, and the imaging result is given to the result display module. The final ultrasound image is displayed.
5. The GPU-based MV high-definition fast medical ultrasound imaging system according to claim 4, wherein the main calculation process of the algorithm imaging module is divided into calculating signal delay, calculating a covariance matrix, calculating an adaptive weight, and calculating an optimization calculation. The sub- and calculated pixel amplitude values are 5 parts, and these calculations are done by multi-threaded cooperation in the GPU thread block.

   The algorithm imaging module can be called after the initial startup module is set according to the actual diagnosis scenario requirement, and the MV high-definition imaging algorithm is encapsulated by the strategy mode, and the pluggable call is realized, and the addition, deletion or upgrade operation can be conveniently performed.
6. The GPU-based MV high-definition fast medical ultrasound imaging system according to claim 4, wherein the result display module further processes the calculation result of the algorithm imaging module, and outputs the imaging result to the display for display; After beamforming, the imaging usually performs envelope detection, logarithmic compression, and grayscale range correction. The result display module performs Hilbert transform, logarithmic compression, and grayscale range correction by calling the corresponding function of Maltab. After the operation, the image related data is output to the corresponding horizontal and vertical axis, and finally the ultrasonic imaging result is displayed on the screen; at the same time, the result display module also collects relevant information of the actual diagnosis scene and displays it on the display suitable In the area, the user is provided with comprehensive and detailed information for diagnostic reference.

Images