Computer-implemented method for beamforming ultrasound channel data

Through the data consistency and prior steps of iterative processing, combined with deep learning technology, the problem of image quality and computational cost in ultrasonic beamforming methods is solved, and high-quality and low-cost beamforming is achieved, which is suitable for real-time ultrasonic imaging.

CN118974596APending Publication Date: 2024-11-15KONINKLIJKE PHILIPS NV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380031289.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-06-16
Filing Date
2023-04-06
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing ultrasonic beamforming methods are difficult to balance between image quality and calculation cost, especially in real-time imaging, the adaptive beamforming methods are complex and time-consuming to calculate, and mostly do not use prior information, which limits the image quality.

Method used

Using iterative processing data consistency steps and prior steps, the consistency and prior belief matching of beamforming images are improved through trained algorithms and neural networks, using maximum posterior probability estimation and proximal gradient descent method, combined with depth Learn techniques to generate high-quality beamforming images.

Benefits of technology

It is achieved to provide high image quality similar to adaptive beamforming with reduced computing costs, and the processing time is comparable to that of deep learning methods, supports high frame rate ultrasound imaging, significantly improving contrast and resolution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118974596A_ABST
    Figure CN118974596A_ABST
Patent Text Reader

Abstract

The invention relates to a computer-implemented method for beamforming ultrasound channel data (2) to obtain a beamformed image, the method comprising the steps of: receiving ultrasound channel data (2); determining an initial estimation result of the beamformed image as image data of intermediate beamformed; performing at least one iteration of a processing operation comprising a data coherence step (6), followed by a priori step (8), where the data coherence step (6) takes the channel data (2) and the intermediate beamformed image data as inputs; performing at least one processing step (14, 16, 50, 56) designed to improve the consistency of the intermediate beamformed image data (10) with the channel data (2); and outputting the updated intermediate beamformed image data (7); the priori step (8) takes as input the updated intermediate beamformed image data (7); and performing at least one processing step to improve the updated intermediate beamformed image data (7), the at least one processing step using a priori hypothesis with respect to the beamformed image data; and outputting the improved updated intermediate image data (12); and outputting the improved updated intermediate image data (12) as the beamformed image (20).
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for beamforming ultrasound channel data, a computer-implemented method for providing a trained algorithm, and related computer programs and systems. Background Art

[0002] Conventional ultrasound imaging methods typically use delayed-and-sum (DAS) beamforming because of its low complexity and fast reconstruction time, as described in KEThomenius, "Evolution of ultrasound beamformers" (Proceedings of the 1996 IEEE Ultrasound Symposium, Vol. 2, pp. 1615-1622, 1996). This method uses fixed, content-invariant apodization weights for the receive channels. Although its reconstruction speed allows real-time imaging, DAS beamforming does not provide optimal image contrast and resolution because it lacks content-adaptive array apodization. For example, backscatter from off-axis components is not adequately compensated.

[0003] Adaptive beamforming algorithms achieve this improvement by determining optimal content-adaptive apodization weights based on the acquired RF signals and applying them to the receive channels. In the context of beamforming, "apodization" can be described as introducing a weighting function ("apodization weights") when summing the channel data acquired by the transducer array. However, these content-adaptive methods are computationally more demanding and result in significantly longer reconstruction times. Therefore, they are generally not suitable for real-time ultrasound imaging.

[0004] One known adaptive beamforming algorithm is the minimum variance (MV) beamformer, in which the apodization weights are continuously optimized to minimize the variance of the received signal after apodization while maintaining unity gain in the desired direction. MV beamforming is described in A. Austeng and S. Holm, “Benefits of minimum-variance beamforming in medical ultrasound imaging” (IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, Vol. 56, No. 9, pp. 1868-1879, 2009). Although MV beamforming has shown significantly improved resolution and contrast compared to DAS, it is also notoriously slow, relying on the computation of the n×n spatial covariance matrix, which requires inversion, which has the n×n 3complexity, where n is the number of channels. Therefore, MV beamforming is not used in real-time imaging. Another known adaptive beamforming method is Wiener beamforming, as described in CCNilsen and S. Holm, "Wiener beamforming and the coherence factor in ultrasound imaging" (IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, Vol. 57, No. 6, pp. 1329-1346, 2010).

[0005] B. Luijten et al., "Adaptive Ultrasound Beamforming Using Deep Learning" (IEEE Transactions on Medical Imaging, Vol. 39, No. 12, pp. 3967-3978, 2020) describes a deep learning-based beamforming method, called ABLE, in which an artificial neural network calculates content-adaptive apodization weights based on input channel data. This is also described in WO2020 / 083918. This disclosure and the article by B. Luijten et al. are incorporated herein by reference.

[0006] Most known beamforming methods do not use a priori information at all in their signal processing, which limits the achievable image quality. Additionally, known content-adaptive beamforming methods are typically very slow and too complex for real-time imaging.

[0007] “Deep Algorithm Unrolling for Biomedical Imaging” by Yuelong Li et al. describes how model-based inversion played a dominant role in biomedical imaging before deep learning became widely popular and widely accepted.

[0008] Jinxi Xiang et al.’s “FISTA-Net: Learning A Fast Iterative ShrinkageThresholding Network for Inverse Problems in Imaging” describes that inverse problems are essential for imaging applications. A model-based deep learning network, FISTA-Net, is proposed. Summary of the invention

[0009] It is an object of the present invention to provide a beamforming method which is able to provide a high image quality while having a reduced computational cost compared to known content adaptive beamforming methods.

[0010] This object is achieved by a computer-implemented method according to claim 1 , a computer-implemented method for providing a trained algorithm according to claim 12 , a computer program according to claim 14 and a system according to claim 15 .

[0011] According to one aspect of the present invention, there is provided a computer-implemented method for beamforming ultrasound channel data to obtain a beamformed image. The method comprises the following steps:

[0012] a) receiving channel data collected by an ultrasonic transducer in response to ultrasonic transmission;

[0013] b) determining an initial estimation result of the beamformed image as intermediate beamformed image data;

[0014] c) performing at least one iteration of a processing operation comprising a data consistency step followed by a priori step, wherein

[0015] d) the data consistency step takes the channel data and the intermediate beamformed image data as input; performs at least one processing step, the at least one processing step being designed to improve the consistency of the intermediate beamformed image data with the channel data; and outputs updated intermediate beamformed image data; wherein the data consistency step comprises the following steps:

[0016] d.1) processing the channel data and the intermediate beamformed image data in a first processing step; and

[0017] d.2) adding the result of said first processing step to said intermediate beamformed image data to obtain said updated intermediate beamformed image data;

[0018] e) the a priori step takes as input the updated intermediate beamformed image data; and performs at least one processing step to improve the updated intermediate beamformed image data, the at least one processing step using an a priori assumption about the beamformed image data; and outputs improved updated intermediate image data; and

[0019] f) outputting said improved updated intermediate image data as said beamformed image.

[0020] The present invention provides a beamforming method, which processes channel data through at least one iteration of a data consistency step and a priori step. The data consistency step can be a step to improve the consistency of the solution (i.e., the image data of the intermediate beamformation) with the measurement data (i.e., the channel data). The data consistency step can be based on a gradient descent method. The data consistency (DC) step can be based on a data likelihood model. The DC step can include a trained algorithm, including trainable parameters, such as parameters of a trainable artificial neural network. The priori step can include a processing step designed to improve the consistency of the solution (in particular the image data of the intermediate beamformation) with the prior belief. Therefore, this can be a step to push the solution in the direction of the prior belief. The prior belief can be a known property of the expected beamformed image, for example, it includes sparse data and / or it includes a certain type of noise (e.g., a Gaussian noise distribution). The priori step can include a proximal operator. The method can include additional steps. For example, the method can include a DC step that is not followed by the priori step, for example as the last processing step.

[0021] The method for beamforming according to the invention can be considered as being derived from a maximum a posteriori probability (MAP) estimator and going to a linear measurement model. It can allow (in particular through a priori steps) to include a priori information about signal statistics. The method for beamforming according to the invention can use a model-based architecture that inherits from proximal gradient descent. In a preferred embodiment, the method can use a trained algorithm that allows learning the DC step and / or the priori step from measurement data. In particular, beamformed images obtained from known computationally expensive content adaptive methods (e.g., minimum variance (MV) or Wiener beamforming) can be used as output training data. Other possible output training data or training targets can be images generated with a full set of channel elements, or synthetic aperture acquisition results, where the goal is to achieve this quality with fewer channel elements. Other possible training data can be simulation results of ultrasound channel data as input training data, with corresponding real-world reflectivity maps as output training data.

[0022] The beamforming method of the present invention has been shown to outperform not only standard delayed-sum (DAS) beamforming, but also the prior art Adaptive Ultrasound Beamforming Using Deep Learning (ABLE) disclosed in the above-mentioned article by B. Luijten et al. At the same time, the processing time of the method is comparable to the ABLE method. Therefore, the beamforming method of the present invention provides excellent image quality comparable to adaptive beamforming techniques at high frame rates, while using conventional and / or GPU-accelerated hardware.

[0023] The invention can be used for all electronic forms of coherent beamforming. In particular, channel data can be acquired with an ultrasound probe having L channels. The probe can have several transducer elements, wherein the signal acquired by one or several transducer elements can contribute to one channel. Channel data can be acquired in a broadband pulse echo ultrasound imaging setting.

[0024] In particular, the channel data can be collected using plane wave sonification technology or line scan-based sonification, but other sonification schemes can also be used to collect channel data. In particular, other possible applications include intravascular ultrasound (IVUS), Doppler imaging, and two-dimensional (2D) and three-dimensional (3D) ultrasound. Expanding the scope of medical imaging, the method of the present invention can also be applied to channel data collected using sensor arrays of sensors other than ultrasound (e.g., acoustic arrays or radar arrays). The channel data collected by the ultrasound transducer in response to ultrasound transmissions can therefore also be channel data collected using other technologies such as acoustic transmissions or radar transmissions.

[0025] The channel data may be a radio frequency (RF) signal acquired during an ultrasound examination, in particular an RF signal acquired using an array of ultrasound (US) transducer elements. The ultrasound examination may be of a human or animal body, in particular in medical imaging applications, or alternatively it may be an ultrasound examination for other purposes such as material testing. The channel data subjected to the beamforming method of the present invention may be the acquired RF data, or may be derived from this data by demodulation, in particular by demodulation using the fundamental ultrasound frequency of the insonification. The channel data may be real-valued or complex-valued (IQ). For at least some parts of the beamforming method, the channel data may be time-of-flight corrected. This may be the case for a number of steps performed pixel by pixel, wherein the channel data may be time-aligned for each pixel before processing. The channel data may include as many channels as the ultrasound transducer has transducer elements, wherein the number of channels L may typically be between 16 and 1024, preferably between 32 and 512, and more preferably between 128 and 256. Thus, the channel data may be represented by a tensor Y of the input signal, preferably time-aligned. For pixel-by-pixel processing, the channel data for each pixel may be a vector y of time-aligned channel signals. The vector may have a length L, where L is the number of channels.

[0026] In order to save processing power, the initial estimate of the beamformed image data is advantageously obtained by a very simple processing operation. For example, it may be a data set containing only a constant value (e.g., the value 0). In particular, each pixel of the initial estimate of the beamformed image data may have the same value (e.g., the value 0 or the value 1). In an alternative embodiment, the initial estimate of the beamformed image data is determined by a delayed-sum beamforming method, which is a simple processing step that can be performed in real time. In examples, it has been found sufficient to use a data set in which each pixel has a value of 0 as the initial estimate.

[0027] The beamforming method of the present invention comprises at least one iteration of a processing operation comprising, preferably consisting of, a data consistency step and an a priori step.

[0028] The DC step takes as input the channel data (which may already be time-aligned channel data) and the intermediate beamformed image data and performs at least one processing step designed to improve the consistency of the intermediate beamformed image data with the channel data. This step may be performed pixel by pixel, wherein the channel data is input as a vector of time-aligned channel data for each pixel. The DC step may comprise processing steps performed by a trained algorithm, in particular an artificial neural network. The output of the DC step is referred to as updated intermediate beamformed image data.

[0029] This data is taken as input to an a priori step, wherein the a priori step (also referred to as the proximal step) performs at least one processing step designed to improve the updated intermediate beamformed image data. It may use an a priori assumption about the beamformed image data (e.g., the beamformed image is sparse in the image domain, or is sparse in the wavelet or Fourier domain). In the latter case, the a priori step may include steps of transforming the updated intermediate beamformed image data into the wavelet or Fourier domain and back. The a priori step may include filtering or thresholding operations. However, in cases where it is not possible to describe the a priori beliefs for the data analytically, the a priori step may be performed by a trained algorithm (in particular an artificial neural network, as described herein, which has been trained on the data). It outputs improved updated intermediate beamformed image data.

[0030] If there are no further iterations of the processing operation, the improved updated intermediate image data is output as the beamformed image. If there are further iterations, the improved updated intermediate image data is used as the intermediate beamformed image data and further iterations of the DC step and the a priori step are performed.

[0031] The processing steps described herein may be performed pixel by pixel, where each pixel has a steering vector, which may be a unit vector if the channel data used as input to the step has been time aligned. Alternatively, tensors of channel data may be processed together.

[0032] According to an embodiment, the processing operation includes at least one processing step, which is performed by a trained algorithm, in particular a trained neural network. By using alternating data consistency steps and prior steps (enhanced with neural networks), challenges in the estimation of signal statistics can be overcome, thereby improving the quality of beamforming. As demonstrated by experiments conducted using an embodiment of the beamforming method of the present invention including a trained neural network called "Neural MAP", strong robustness and generalization capabilities beyond the training data have been demonstrated. The Neural MAP beamformer of the present invention is significantly superior to ABLE beamforming, providing stronger contrast and higher resolution. The parameters of the trained algorithm, in particular the neural network, can be trained / learned based on the training data by utilizing deep learning methods.

[0033] According to an embodiment of the present invention, the data consistency step includes at least one processing step, and the at least one processing step is performed by a trained neural network. In particular, the data consistency step can take a specific, model-based form (enhanced with at least one neural network) to overcome challenges in signal statistics estimation. This form is particularly suitable for beamforming ultrasound data. In a preferred embodiment, the beamforming method unfolds an iterative scheme in a feedforward neural network through a series of DC steps and prior steps. As described herein, the resulting model-based architecture allows the DC step and / or prior step to be trained based on input training data and output training data. The concept of algorithm unfolding is explained in "Algorithm Unrolling" by Monga et al. (IEEE Signal Processing Magazine, March 2021, pp. 18-44).

[0034] The processing architecture according to steps d.1) and d.2) has proven to be advantageous, since the first processing step is designed to give an estimate of the difference between the intermediate beamformed image data and the correct image and to add it to the intermediate beamformed image data. The first processing step, referred to as f(·), may comprise or consist of a trained neural network. For example, it may comprise a neural network comprising up to 6, preferably 2 to 4, convolutional layers and up to 3 fully connected layers. The neural network parameters may be trained by back propagation throughout the processing operation and in one or several iterations.

[0035] According to an embodiment of the present invention, the first processing step comprises the following steps: calculating a residual by multiplying the intermediate beamformed image data pixel by pixel by the steering vector and subtracting the result from the channel data; and processing the residual and the channel data by a second processing step. In this embodiment, the data consistency step has a specific form. In particular, the first processing step calculates the residual as input for the further second processing step. The residual can be obtained by multiplying the intermediate beamformed image data at the pixel pixel by the steering vector of the pixel and subtracting the result from the channel data, wherein the channel data in this case can be represented by a vector of length L (L=number of channels). The steering vector can be different for each pixel in the beamformed image. If the channel data has been time aligned, the steering vector can be a unit vector. The advantage of using the residual as input is that the consistency with the measurement data can be best evaluated thereby. Then, the residual and the channel data are used as input for the second processing step, and the result of the second processing step is added to the intermediate beamformed image data to obtain updated intermediate beamformed image data.

[0036] According to an embodiment of the invention, the method, in particular the second processing step, comprises calculating a set of apodization weights from the channel data, preferably by a trained neural network, and the second processing step comprises multiplying the residual by the apodization weights.

[0037] According to the embodiment, a further processor, referred to as h(·), acquires channel data and generates a set of apodization weights. The apodization weights may be content-adaptive apodization weights. The processor may include or consist of a trained artificial neural network disclosed in WO2020 / 083918A1, which is incorporated herein by reference. The same deep learning-based adaptive neural network (referred to as ABLE) is disclosed in the article by Ben Luijten et al. By using the ABLE network, a set of high-quality content-adaptive apodization weights can be obtained, in particular for each pixel, with a relatively small additional computational burden. However, in contrast to the article by Ben Luijten et al., the content-adaptive apodization weights are not used directly for beamforming. Instead, they are used in the framework of a sequence of steps for processing residuals, which are obtained by multiplying the intermediate beamformed image data by the steering vector and subtracting the result from the channel data. The residuals are correspondingly multiplied by the apodization weights, in particular pixel by pixel. According to an embodiment, the result of this multiplication is the result of the first processing step and is therefore added to the intermediate beamformed image data to obtain updated intermediate beamformed image data. The processor h(·) that generates a set of apodization weights may be a neural network comprising 0 to 6 (preferably 2 to 4) fully connected layers and 0 to 6 (preferably 2 to 4) activation layers. It may also contain up to 6 convolutional layers.

[0038] According to an embodiment of the invention, the first processing step comprises a processing step performed by a trained neural network, wherein preferably, the second processing step or part of the second processing step is performed by the trained neural network. According to another embodiment, the second processing step taking the residual as input can be performed by a trained neural network. Such a neural network can include 0 to 6 (preferably 2 to 4) fully connected layers and 0 to 6 (preferably 2 to 4) activation layers. It can also contain up to 6 convolutional layers. The neural network can also be trained by back propagation across the sequence of steps of the beamforming method of the present invention.

[0039] According to an embodiment of the invention, the a priori step is based on the a priori assumption that the beamformed image is sparse. It is shown that it is advantageous to use a priori information on the resulting beamformed image. Sparsity is often found in medical ultrasound imaging, so it is useful to make such a priori assumption. Other possible a priori assumptions may be that all pixels are independent of each other and / or that they are identically distributed (e.g., Laplace distribution).

[0040] According to an embodiment of the present invention, the a priori step includes a soft thresholding step. This is particularly advantageous when the a priori belief believes that the image is sparse. In soft thresholding, all pixels whose values ​​are lower than the threshold t minus half of the range r can be set to 0. If the pixel value S is within the range r around the threshold t, the result follows the function m(S), which can be a function of stable increase. If the pixel value S is higher than the threshold plus half of the range, the pixel value can remain unchanged. It can be mathematically represented as follows, where the original value of the pixel is S, and D is the output value of the soft thresholding filter.

[0041] If S<t-r / 2, then D=0

[0042] If t-r / 2≤S<t+r / 2, then D=m(S)

[0043] If S ≥ t + r / 2, then D = S

[0044] Wherein, the function m can be, for example, a sine function, a linear function or a polynomial function. Thus, a smooth transition between the original value and the deleted value is achieved. Values ​​slightly below the threshold t are not set to 0, but are simply attenuated. In other embodiments, the a priori step can use a different proximal operator (e.g., hard thresholding). It can also be a soft thresholding in a transform domain (e.g., a wavelet or Fourier domain). In this case, the a priori step can include transforming the updated intermediate beamformed image data into the domain (e.g., a wavelet or Fourier domain), and performing a soft thresholding operation here, and transforming back into the image domain. In other embodiments, the a priori step can include or consist of additional filtering operations (e.g., smoothing, denoising or image segmentation).

[0045] According to an embodiment of the present invention, the a priori step includes or consists of at least one processing step performed by a trained algorithm. The parameters of the soft thresholding step can also be trained according to the data, so the soft thresholding step can be such a trained algorithm. In other embodiments, the trained algorithm can be a neural network. This is very useful, especially when the a priori belief is difficult to express through analysis. In this case, the a priori step can still perform an operation that pushes the solution in the direction of the a priori belief, but the exact nature of the operation is trained according to the data as explained in this article. For example, the entire a priori step can be replaced by a neural network, which can be trained according to input training data and output training data. Such a neural network can include 0 to 6 (preferably 2 to 4) fully connected layers and 0 to 6 (preferably 2 to 4) activation layers. It can also contain up to 6 convolutional layers.

[0046] According to an embodiment of the present invention, the channel data is time-of-flight corrected for each image pixel before undergoing processing operations. The advantage of this is that no further operations are required (e.g., no multiplication with a steering vector is required). Time-of-flight correction can be performed for each pixel of the beamformed image. The channel data corrected by time-of-flight can be stored in a memory or buffer.

[0047] According to an embodiment of the present invention, the method is performed by an algorithm using parameters that have been trained based on training data. The parameters that can be trained based on input data and output data include weights and biases of one or more neural networks that can be included in the processing operation. It can also include soft threshold processing parameters, especially in embodiments where the prior step includes soft threshold processing, for example, thresholds, ranges, and parameters of functions applied within the range.

[0048] According to an embodiment of the present invention, a predetermined number of iterations of the processing operation is performed, the predetermined number of iterations is preferably 1 to 20 iterations, more preferably 2 to 10 iterations. It is advantageous to predetermine the number of iterations, because otherwise it is necessary to check the improved updated intermediate image data to determine when to stop the algorithm, which increases the total processing time. Surprisingly, it has been shown that very few iterations (e.g., 1 to 6 times, preferably 2 to 4 times) are sufficient. Therefore, it has been shown that simply performing a predetermined number of iterations and then stopping is advantageous. If the processing operation of the method of the present invention includes trainable parameters, it may be advantageous to train the algorithm together in all steps and iterations. Again, in this case, a predetermined number of iterations is advantageous. When performing the beamforming method of the present invention, it is also possible to train the algorithm using one iteration and then train the algorithm using one or several iterations. Other embodiments include a step of processing the improved updated intermediate image data to determine whether the beamformed image has a quality sufficient to stop the algorithm or whether to perform another iteration.

[0049] According to an embodiment of the present invention, the training data includes input training data and output training data, the input training data includes channel data collected by an ultrasonic transducer in response to ultrasonic emission, and the output training data includes beamformed image data obtained according to the input training data. The input training data may include channel data collected by any of the above-mentioned sound penetration schemes (e.g., plane wave imaging). The output training data is usually obtained by a content adaptive beamforming method (e.g., a minimum variance beamformer) or by Wiener beamforming. These methods provide very good results, but the amount of calculation is usually too large to run in real time. However, when training a beamforming algorithm, processing time is not a problem, so the beamforming method of the present invention can include trainable parameters trained according to minimum variance beamforming data. The training data can be ultrasound data collected from a human subject or an animal subject. It can also be simulated data (e.g., a simulated point scatterer from a single plane wave ultrasound acquisition).

[0050] According to a further aspect of the invention, there is provided a computer-implemented method for providing a trained algorithm, the method being adapted to perform steps c), d) and e) of one of the preceding claims and comprising trainable parameters (preferably comprising parameters of a trainable neural network), the method comprising:

[0051] (a) receiving input training data, wherein the input training data includes channel data collected by an ultrasonic transducer in response to ultrasonic transmission;

[0052] (b) receiving output training data, the output training data comprising beamformed image data preferably obtained from the input training data by a content adaptive beamforming algorithm, such as a minimum square error algorithm;

[0053] (c) training the algorithm by using the input training data and the output training data;

[0054] (d) providing the trained algorithm.

[0055] The trained algorithm may have the features described in relation to the beamforming method of the present invention, and vice versa. In particular, it may comprise at least one trainable neural network (NN), in particular at least one trainable neural network (NN) included in the data consistency step and / or the prior step. The training step may be performed using back propagation. Thus, input training data (in particular channel data) are propagated through an algorithm using predetermined initial values ​​of trainable parameters (in particular weights and biases of one or more NNs) and (where appropriate) parameters of soft thresholding. The output of the algorithm is compared with the output training data (i.e., the beamformed image), for example using an error function or a cost function (the output of the error function or cost function is propagated back through the algorithm), thereby calculating a gradient to find the trainable parameters that produce the minimum error. It has been found that the parameters of the trained algorithm according to the present invention converge quickly to a minimum value, enabling the algorithm to be trained on a limited amount of data relating to only one or a few ultrasound images. In some embodiments, one or more neural networks include the dropping of layers during training. Thus, certain parameters and dropped layers are randomly selected and their values ​​are set to 0. This has the advantage that the training converges to a useful minimum value. Therefore, using dropout layers improves the confidence of the algorithm.

[0056] According to an embodiment, the trained algorithm comprises the following steps: performing at least one iteration of a processing operation, the processing operation comprising a data consistency step followed by an a priori step, wherein the data consistency step takes as input the channel data and the intermediate beamformed image data; performing at least one processing step, the at least one processing step being designed to improve the consistency of the intermediate beamformed image data with the channel data; and outputting updated intermediate beamformed image data; wherein the data consistency step comprises the following steps: processing the channel data and the intermediate beamformed image data in a first processing step; and adding the result of the first processing step to the intermediate beamformed image data to obtain the updated intermediate beamformed image data; and the a priori step takes as input the updated intermediate beamformed image data; and performing at least one processing step to improve the updated intermediate beamformed image data, the at least one processing step using a priori assumptions about the beamformed image data; and outputting improved updated intermediate image data, and wherein at least one of steps c), d) and e) comprises parameters, preferably parameters of a trainable neural network, which are trained by using the input training data and the output training data.

[0057] According to a further aspect, the present invention relates to a computer program comprising instructions and / or program codes, which, when run by a computing unit, cause the computing unit to perform a method according to an embodiment of the present invention. In particular, the computer program may perform a method for beamforming according to the present invention. It may also perform a method for providing a trained algorithm. The computer program may be provided in the form of a computer program product. The computer program may be provided on a non-transitory digital storage medium. Such a storage medium may be any optical or magnetic digital storage medium, for example a hard disk, a floppy disk, a DVD, a CD-ROM, a USB stick, an SD card, an SSD card, etc. The storage medium may be part of a computer, a server or a cloud computer. The computing unit may be any digital processing unit (in particular a CPU or a GPU of a PC, a server or a cloud computer).

[0058] The invention also relates to a non-tangible digital storage medium having stored thereon a computer program comprising instructions and / or program codes which, when executed by a computing unit, cause a single execution of a method according to an aspect or embodiment of the invention.

[0059] The beamforming method of the present invention is preferably computationally very cheap, so that it can be performed at the point of care during an ultrasound examination of a patient. In an embodiment, it can be performed in real time by a computing unit (e.g., a PC, server, or GPU or CPU that is part of an ultrasound scanner that acquires ultrasound channel data) present at the point of care. The channel data can also be transmitted to a central computing unit in the cloud or on a server at a hospital, and the computer-implemented method for beamforming can be executed on the central server, thereby transmitting the beamformed image back to the point of care. In a useful embodiment, the method for beamforming can be performed in real time, so that the beamformed image can be obtained directly after the channel data is acquired.

[0060] According to a further aspect, the present invention relates to a system for beamforming ultrasound channel data to obtain a beamformed image, the system comprising: a first interface configured to receive channel data acquired by an ultrasound transducer in response to ultrasound transmission; a computing unit configured to: determine an initial estimate of the beamformed image as intermediate beamformed image data; and perform at least one iteration of a processing operation, the processing operation comprising a data consistency step followed by an a priori step, wherein the data consistency step takes the channel data and the intermediate beamformed image data as input; performs at least one processing step, the at least one processing step being designed to improve the consistency of the intermediate beamformed image data with the channel data; and outputs Updated intermediate beamformed image data; wherein the data consistency step comprises the following steps: processing the channel data and the intermediate beamformed image data in a first processing step; and adding the result of the first processing step to the intermediate beamformed image data to obtain the updated intermediate beamformed image data; the a priori step takes the updated intermediate beamformed image data as input; and performs at least one processing step to improve the updated intermediate beamformed image data, the at least one processing step using a priori assumptions about the beamformed image data; and outputs improved updated intermediate image data; and a second interface configured to output the improved updated intermediate image data as the beamformed image.

[0061] The system may be a digital processing unit, in particular a GPU or a CPU of a computer (eg a PC, a cloud computer, a server), or it may be a processing unit of an ultrasound scanner.

[0062] According to a further aspect, the present invention relates to an ultrasound scanner incorporating a system for beamforming ultrasound channel data according to the present invention. The ultrasound scanner may also include an ultrasound probe having L channels, the ultrasound probe being suitable for ultrasonic insonification of a target tissue of a patient. The probe may include a two-dimensional or three-dimensional array of ultrasound transducer elements. Each element may correspond to an ultrasound channel. In some embodiments, several ultrasound transducer elements contribute to a single channel of data.

[0063] All features described with respect to the computer-implemented method for beamforming also apply to the computer program, the system for beamforming, the training method and the ultrasound scanner, and vice versa. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Embodiments of the present invention will now be described with reference to the accompanying drawings, in which:

[0065] Figure 1is a schematic diagram of an embodiment of the method of the present invention;

[0066] Figure 2 is a schematic flow chart of an embodiment of a data consistency step;

[0067] Figure 3 is a schematic flow chart of an embodiment of a data consistency step;

[0068] Figure 4 is a schematic diagram of a processor designed to generate apodization weights;

[0069] Figure 5 is a neural network according to an embodiment;

[0070] Figure 6 is an example of a soft thresholding filter;

[0071] Figure 7 is an image of a simulated point scatterer that has been beamformed according to (a) a standard DAS beamformer, (b) prior art ABLE beamforming, and (c) an embodiment of the beamforming algorithm of the present invention (Neural MAP);

[0072] Figure 8 is a schematic diagram of a system according to an embodiment of the present invention.

[0073] Reference numerals list

[0074] 2-channel data

[0075] 4 Initial estimation results of beamformed image

[0076] 5 Residual

[0077] 6 Data consistency steps

[0078] 7 Updated intermediate beamforming image data

[0079] 8 Prior steps

[0080] 10 Intermediate beamforming image data

[0081] 12 Improved and updated intermediate beamforming image data

[0082] 14 First processing step

[0083] 16 Second processing step, neural network

[0084] 42 Flight time correction

[0085] 43 channel data corrected for time of flight

[0086] 45 Rearranged TOF-corrected channel data

[0087] 50 Multiplication

[0088] 51 Beamforming Residual Image

[0089] 52 Sum

[0090] 54 pixels

[0091] 55 Beamforming residual

[0092] 56 Addition

[0093] 57 Anti-rectifier

[0094] 58 Dimensionality reduction

[0095] 60 Fully connected layers

[0096] 100 Ultrasound System

[0097] 102 Hardware Unit

[0098] 104 CPU

[0099] 106 GPU

[0100] 108 Digital storage media

[0101] 110 CD-ROM

[0102] 112 Internet

[0103] 114 User Interface

[0104] 116 Keyboard

[0105] 118 Touchpad

[0106] 120 Ultrasound Probe

[0107] 122 Ultrasonic transducer array

[0108] 124 images

[0109] 126 Screen

[0110] 128 Computer DETAILED DESCRIPTION

[0111] Many classical beamforming methods can be derived from the maximum likelihood (ML) estimation result p(y|x) of a narrowband linear measurement model with additive noise, which is given by:

[0112] y=a θ x+n

[0113] Where y is the vector of channel data, x is the tissue reflectivity, i.e., the pixel value of the target image, a is the steering vector for that pixel, and n is the additive noise. The above model assumes independent sources x (i.e., each pixel is independent), which is a typical assumption in a beamforming setting. In a broadband pulse-echo ultrasound imaging setting, preferably, a TOF correction / time alignment is first performed so that the steering vector simplifies to a θ = 1. In beamforming, our goal is to find the candidate that is most likely to explain our measurement y Because of this, our goal is to maximize the probability that:

[0114] p(x|y)∝p(y|x)p(x)

[0115] This follows from Bayes' theorem. The resulting solution requires knowledge of the signal statistics. Additionally, since this problem has no closed-form solution, this can be solved by iterative methods (e.g., proximal gradient descent). Without any assumptions about the prior distribution of x (p(x) = 1), the above estimator becomes The standard delay-and-sum beamforming yields the maximum likelihood (ML) estimator of under the assumption of Gaussian noise. ML estimation results, where n~N(0,σ 2 I), and yields a closed-form solution:

[0116]

[0117] This corresponds to summing the individual ultrasound channel signals. Similarly, we can derive the colored noise profile n~N(0, C -1 ), where C is the noise covariance matrix. It can be solved by maximizing the log-likelihood or equivalently minimizing the negative log-likelihood.

[0118] argmax x log p(y|x)=argmin x (y-ax) T (C -1 )(y-ax)

[0119] Setting the gradient at x to zero produces:

[0120]

[0121] This corresponds to minimum variance (MV) beamforming, where w MV Including MV apodization weights. When C = σ 2I, this is equivalent to DAS. In general, the noise covariance is unknown and is estimated from the data. This process is prone to errors and can seriously affect image quality. Similarly, one can derive a maximum a posteriori probability (MAP) beamformer, i.e., when p(x) ≠ 1, which assumes prior knowledge about the distribution of x. Now one can maximize the log-posteriori probability, which is given by:

[0122] argmax x log p(x|y)=argmin x (y-1x) T (C -1 )(y-1x)-log p(x)

[0123] For the Gaussian likelihood model, the general solution is given by:

[0124]

[0125] This is called Wiener beamforming. This general case is equivalent to the post filter (scaling) of the MV beamformer.

[0126] For many relevant prior distributions, the MAP estimator has no closed form solution. According to an embodiment of the present invention, the MAP estimator can be obtained by an iterative method (e.g., proximal gradient descent). The proposed proximal gradient descent method alternates between a data consistency (DC) step based on a data likelihood model and a priori (proximal) steps. In the context of beamforming, each iteration k can be written as:

[0127]

[0128] where Prox(·) denotes the proximal operator operating in the image domain of the beamformer, C is the noise covariance matrix, and μ is the gradient step size of the DC update of the multivariate Gaussian channel noise process. As with MV beamforming and Wiener beamforming, a straightforward implementation of the above scheme requires estimating the signal statistics for each pixel, i.e., the inverse covariance matrix C -1 According to embodiments of the invention, this slow and often unstable process can be circumvented by learning portions of the DC step with a neural network. For simple independent and identically distributed prior distributions (e.g., the Laplacian), the proximal step can be a soft thresholding operation. To model proximal operators for more complex / rich prior distributions, a neural network can also be used according to embodiments.

[0129] Figure 1An embodiment of the beamformer of the present invention for ultrasound data is shown, which processes channel signals through a series of data consistency ("DC") steps and "a priori" steps. The beamformer has the following form: X = f (Y), where Y is a tensor of (time-aligned) input channel signals (real valued or IQ), and X is the processed (beamformed) output image. Moreover, a steering vector a is represented for each pixel θ , where, for the time-aligned signal, a θ =1. In addition, = represents the initial estimation result. Therefore, the initial estimation result of the channel data 2 and the beamformed image 4 is the input for the DC step 6. Initial estimation result Can be an image of zero, i.e., In an alternative embodiment, is a beamformed image obtained by any other prior art beamforming method (e.g., delay-sum). For each pixel, the channel data may in particular be time aligned. The output of the DC step 6 is input to the a priori step 8, which outputs improved updated intermediate image data 12 (denoted as ). These steps may be performed pixel by pixel, wherein one, several or all pixels may be processed simultaneously. The updated intermediate image data 12 may be used as input 10 for the next iteration of the DC step 6 and the a priori step 8. Each iteration converts the intermediate beamformed image data As input, and given the improved updated intermediate image data As output. According to an embodiment, a fixed number k (e.g., 2 to 5 times, more preferably, 3 to 4 times) of iterations are performed. The processing methods and processing parameters of the DC step 6 and the a priori step 8 may be the same in each iteration. In some embodiments, in each iteration, the processing method is the same, but the parameters are different. In an embodiment, only the DC step includes a trained artificial neural network, which may be the same in each iteration. In an embodiment, the a priori step may consist of or include a soft threshold processing filter, which may also be the same in each iteration. The improved updated intermediate image data output from the last iteration of the a priori step 8 is the final beamformed image 20.

[0130] Figure 2 An embodiment of the DC step 6 is shown, followed by a priori step g(·), wherein only one iteration is shown. Each iteration utilizes the measurement channel data obtained by an ultrasound probe having L channels and is represented as The intermediate solution 10 is taken as input. In a first processing step 14, denoted as processor f(·), these inputs are processed. The output of processor f(·) is added to the input To obtain So we have:

[0131]

[0132] is the updated intermediate beamformed image data 7, the output of the DC step 6. The second processor g(·) corresponding to the prior step 8 further processes To generate improved updated intermediate image data So that we have Processors f(·) and g(·) can be (but are not limited to) neural networks. Neural network parameters can be trained by back propagation across a sequence of steps. The neural network parameters for each step can be the same or different. Processor g(·) can also alternatively be a soft thresholding function. The latter is useful for imposing sparse priors.

[0133] In an embodiment, the first processing step 14, here represented as processor f(·), takes as explicit input the residual 5. This residual 5 can be obtained by using a steering vector a of length L calculated for each pixel. θ , according to each pixel, the intermediate solution at that pixel Multiply by the pixel's steering vector a θ , and subtract it from the channel data Y. In the case of time-aligned input channel data, a θ = 1. For each pixel, this can be written as:

[0134]

[0135] exist Figure 3 In the specific case of the previous embodiment illustrated, f(·) may further include a second processing step 16 (here represented as processor h(·)), which acquires the channel data and generates a set of weights. This set of weights can be calculated only once, stored in a buffer, and used in several iterations. Processor h(·) can be a neural network. It can be a neural network disclosed in WO2020 / 083918A1. In an embodiment, the sequence includes 4 steps, and the neural network includes 3 convolutional layers.

[0136] The second processing step also comprises multiplying 50 the residual 5 and the set of weights calculated by the processor 16 element-wise with each other. In step 56, the result is added to the intermediate beamformed image data To give updated intermediate beamformed image data In per-pixel implementation this can be written as:

[0137]

[0138] Figure 4 A schematic diagram of a DC step 6 according to an embodiment is shown. Channel data, which may be RF data, is illustrated as input data 2. In step 42, the channel data is time aligned, with different planes 43 representing data for different channels. The time-of-flight corrected RF signal 43 may be pre-calculated and stored in a buffer, or calculated each time the algorithm requires it, thereby reducing memory overhead. In this embodiment, all data 43 from various channels associated with a pixel are rearranged into a new format 45 so that the data 43 for each pixel can be processed as a single batch in the neural network 16. The NN 16 may be as described below, or may be different. It outputs a set of apodization weights, which are multiplied by the residual 5 in step 50. The residual 5 may be calculated as described above. The result of the multiplication 50 and the weighted sum 52 is a beamforming residual 55 associated with a pixel 54. This can be used to reconstruct a beamformed residual image 51 which is added in step 56 to the same intermediate beamformed image data 4 used to calculate the residual 5. The output 7 is the updated intermediate beamformed image data.

[0139] Figure 5 An embodiment of an artificial neural network (NN) 16 that can be used to calculate adaptive apodization weights is shown in more detail. It includes at least one fully connected layer and at least one activation layer including an activation function that propagates positive and negative input values ​​and unbounded output values. "Unbounded" means that the output value of the activation function is not limited to any specific value (e.g., +1 or -1). Preferably, in principle any value can be obtained, thereby maintaining the dynamic range of the channel data. An example of such a function may be an anti-rectifier function or a cascaded rectifier linear unit (CReLU), as described in the article "Understanding and Improving Convolutional Neural Networks via ConcatenatedRectifier Linear Unit" by Shang et al. (Proceedings of the 33rd International Conference on Machine Learning, New York, USA, 2016). Above each layer, its output size (for 128 active channels) is indicated. Fully connected layers are represented by dark shading, anti-rectifier layers are represented by white, and dropout layers (which only appear during network training) are represented by light shading. The NN 16 includes 4 fully connected layers, including 128 nodes for the input layer and the output layer and 32 nodes for the inner layer. After the first anti-rectifier, the dimension is reduced by 58 to 1 / 8 (2 3), or a dimensionality reduction to 1 / 4 relative to the input layer, which forces the network to find a more compact representation of the data. Each fully connected layer (except the last layer) is followed by an anti-rectifier. The final fully connected layer 60 is either the output layer or directly connected to the output layer (not shown). During training, dropout can be applied between each pair of fully connected layers, for example with a probability of 0.2. In other words, during training, a fixed percentage of nodes in the dropout layer are dropped. Therefore, the dropout layer only exists during training the network. Dropout helps reduce overfitting of the neural network to the training data. The result of the NN is a set of apodization weights W.

[0140] The NN can be implemented in Python using the Keras API and the TensorFlow (Google, CA, USA) backend. For training, the Adam optimizer can be used with a learning rate of 0.001, randomly optimizing over a batch of pixels belonging to a single image. Figure 5 The neural network shown can be trained on in vivo ultrasound image data as input training data. As output training data, apodization weights calculated by known adaptive beamforming techniques using conventional algorithms can be used. The input training data are the corresponding time-aligned RF signal / channel data. Optionally, according to an embodiment of the present invention, the neural network of the second processing step 16 can be trained together with the complete sequence of steps of the processing operation.

[0141] Figure 6 A soft thresholding filter is illustrated, where for values ​​greater than 0, the filtered signal is given for the original signal. Thus, all signal values ​​below a threshold value t-r / 2 are set to 0, where r is the range. In the interval between t-r / 2 and t+r / 2, the filtered signal follows a sine function. Above t+r / 2, the filtered signal is equal to the original signal. It can also be slightly smaller than the original signal by a predetermined value. By using this function as a priori step g, the image can be pushed towards the prior belief that the ultrasound image is sparse.

[0142] Figure 7 A comparison of the image quality of a simulated point scatterer from a single plane wave acquisition is shown. The three images show Figure 7 Standard delay-and-sum beamforming in (a), Figure 7 Prior art ABLE beamforming in (b), and Figure 7 (c) Neural MAP beamforming of the present invention. Figure 7 (c) The neural MAP beamformer uses Figure 3An embodiment of where the prior step is a soft thresholding function, h is a neural network with 4 convolutional layers, and 4 iterations are used. The deep learning based beamformers ABLE and Neural MAP are both trained towards high quality in vivo targets generated using a minimum variance beamformer with 11 plane wave acquisitions. Figure 7 Inference results on simulated data (which differs significantly from the training set) are shown to demonstrate strong generalization and robustness outside the training domain. Furthermore, the Neural MAP beamformer significantly outperforms ABLE, providing stronger contrast and higher resolution.

[0143] Figure 8 1 is a schematic diagram of an ultrasound system 100 according to an embodiment of the present invention, and the ultrasound system 100 is configured to perform the method of the present invention. The ultrasound system 100 includes a conventional ultrasound hardware unit 102, which includes a CPU 104, a GPU 106, and a digital storage medium 108 (e.g., a hard disk or a solid state disk). The computer program can be loaded into the hardware unit from a CD-ROM 110 or via the Internet 112. The hardware unit 102 is connected to a user interface 114, which includes a keyboard 116 and an optional touch pad 118. The touch pad 118 can also be used as a display device for displaying imaging parameters. The hardware unit 102 is connected to an ultrasound probe 120, which includes an array 122 of ultrasound transducers, which allows live ultrasound images to be collected from an object or patient (not shown). A live image 124 acquired by the ultrasound probe 120 and beamformed according to the method of the present invention executed by the CPU 104 and / or GPU is displayed on a screen 126, which may be any commercially available display unit (e.g., a screen, a television, a flat-panel screen, a projector, etc.).

[0144] Additionally, there may be a connection (eg via the Internet 112) to a remote computer or server 128. The method according to the invention may be performed by the CPU 104 or GPU 106 of the hardware unit 102, but may also be performed by a processor of a remote server 128.

Claims

1. A computer-implemented method for beamforming ultrasound channel data (2) to obtain a beamformed image (20), the method comprising the following steps: a) receiving channel data (2) collected by an ultrasonic transducer (122) in response to ultrasonic transmission; b) determining an initial estimation result (4) of the beamformed image as intermediate beamformed image data (10); c) performing at least one iteration of a processing operation comprising a data consistency step (6) followed by a priori step (8), wherein: d) the data consistency step (6) takes the channel data (2) and the intermediate beamformed image data (10) as input; performs at least one processing step (14, 16, 50, 56), the at least one processing step being designed to improve the consistency of the intermediate beamformed image data with the channel data (2); and outputs updated intermediate beamformed image data (7); wherein the data consistency step (6) comprises the following steps: d.1) processing the channel data (2) and the intermediate beamformed image data (10) in a first processing step (14); and d.2) adding (56) the result of said first processing step (14) to said intermediate beamformed image data (10) to obtain said updated intermediate beamformed image data (7); e) the a priori step (8) takes the updated intermediate beamformed image data (7) as input; and performs at least one processing step to improve the updated intermediate beamformed image data (7), the at least one processing step using a priori assumptions about the beamformed image data; and outputs improved updated intermediate image data (12); f) outputting the improved updated intermediate image data (12) as the beamformed image (20).

2. The method according to claim 1, wherein: The data consistency step (6) comprises at least one processing step (16) which is performed by a trained neural network.

3. The method according to one of the preceding claims, wherein: The first processing step (14) comprises the following steps: Calculating a residual (5) by multiplying the intermediate beamformed image data (10) pixel by pixel by a steering vector and subtracting the result from the channel data (2); The residual (5) and the channel data (2) are processed by a second processing step (16, 50).

4. The method according to claim 3, wherein: The method comprises calculating a set of apodization weights from the channel data (2), preferably by a trained neural network (16), and Therein, the second processing step comprises multiplying (50) the residual by the apodization weight.

5. The method according to claim 1, 3 or 4, wherein: The first processing step (14) comprises processing steps performed by a trained neural network, wherein preferably the second processing step is performed by the trained neural network (16).

6. The method according to one of the preceding claims, wherein: The a priori step (8) is based on the a priori assumption that the beamformed image (20) is sparse.

7. The method according to one of the preceding claims, wherein: The a priori step (8) comprises at least one processing step performed by a trained neural network.

8. The method according to one of the preceding claims, wherein: The a priori step (8) comprises a soft threshold processing step.

9. The method according to one of the preceding claims, wherein: A predetermined number of iterations of the processing operation is performed, the predetermined number of iterations preferably being 1 to 20 iterations, more preferably being 2 to 10 iterations.

10. The method according to one of the preceding claims, wherein: The method is performed by an algorithm using parameters that have been trained based on training data.

11. A computer-implemented method for providing a trained algorithm, the method being adapted to perform steps c), d) and e) of one of the preceding claims and comprising trainable parameters, preferably parameters of a trainable neural network, the method comprising: (a) receiving input training data, the input training data comprising channel data (2) acquired by an ultrasonic transducer (122) in response to ultrasonic transmission; (b) receiving output training data, the output training data comprising beamformed image data preferably obtained from the input training data by a content adaptive beamforming algorithm, such as a minimum square error algorithm; (c) training the algorithm by using the input training data and the output training data; (d) providing the trained algorithm.

12. The computer-implemented method of claim 11, wherein: The trained algorithm comprises the following steps: c) performing at least one iteration of a processing operation comprising a data consistency step (6) followed by a priori step (8), wherein: d) the data consistency step (6) takes the channel data (2) and the intermediate beamformed image data (10) as input; performs at least one processing step, the at least one processing step being designed to improve the consistency of the intermediate beamformed image data with the channel data (2); and outputs updated intermediate beamformed image data (7); wherein the data consistency step (6) comprises the following steps: d.1) processing the channel data (2) and the intermediate beamformed image data (10) in a first processing step (14); and d.2) adding (56) the result of said first processing step (14) to said intermediate beamformed image data (10) to obtain said updated intermediate beamformed image data (7); e) the a priori step (8) takes the updated intermediate beamformed image data (7) as input; and performs at least one processing step to improve the updated intermediate beamformed image data (7), the at least one processing step using a priori assumptions about the beamformed image data; and outputs improved updated intermediate image data (12); And wherein at least one of steps c), d) and e) comprises parameters, preferably parameters of a trainable neural network, which are trained by using the input training data and the output training data.

13. A computer program comprising instructions which, when said program is run by a computing unit, cause said computing unit to perform the method according to any one of claims 1 to 12.

14. A system for beamforming ultrasound channel data (2) to obtain a beamformed image, the system comprising: A first interface configured to receive channel data (2) collected by an ultrasonic transducer (122) in response to ultrasonic transmission; A computing unit configured to: Determining an initial estimation result (4) of the beamformed image as intermediate beamformed image data; and At least one iteration of a processing operation is performed, the processing operation comprising a data consistency step (6) followed by a priori step (8), wherein: The data consistency step (6) takes the channel data (2) and the intermediate beamformed image data (10) as input; performs at least one processing step (14, 16, 50, 56), the at least one processing step being designed to improve the consistency of the intermediate beamformed image data (10) with the channel data (2); and outputs updated intermediate beamformed image data (7); wherein the data consistency step (6) comprises the following steps: processing the channel data (2) and the intermediate beamformed image data (10) in a first processing step (14); and adding (56) the result of said first processing step (14) to said intermediate beamformed image data (10) to obtain said updated intermediate beamformed image data (7); The a priori step (8) takes the updated intermediate beamformed image data (7) as input; and performs at least one processing step to improve the updated intermediate beamformed image data (7), the at least one processing step using an a priori assumption about the beamformed image data; and outputs improved updated intermediate image data (12); A second interface is configured to output the improved updated intermediate image data (12) as the beamformed image (20).