This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging
Copyright (c) 2019 IEEE. Personal use of this material is permitted. However, permission to use this material for any other
purposes must be obtained from the IEEE by sending a request to [email protected].

The SVD Beamformer: Physical Principles and

Application to Ultrafast Adaptive Ultrasound
Hanna Bendjador, Thomas Deffieux, Mickaël Tanter


Abstract— A shift of paradigm is currently underway in It can be expressed as an inverse problem between the
biomedical ultrasound thanks to plane or diverging waves received ultrasonic echoes and the acoustic impedance
coherent compounding for faster imaging. One remaining distribution in the medium. This inverse problem is easily
challenge consists in handling phase and amplitude aberrations solved when assuming the local speed of sound and density to
induced during the ultrasonic propagation through complex be constant in the medium.
layers. Unlike conventional line-per-line imaging, ultrafast
This hypothesis is accepted in all conventional clinical
ultrasound provides backscattering information from the whole
imaged area for each transmission. Here, we take benefit from this devices and led to the standard Delay-And-Sum beamforming
feature and propose an efficient approach to perform fast process. In practice, it consists in computing and correcting the
aberration correction. Our method is based on the Singular Value time-of-flight for different travel paths between each pixel and
Decomposition (SVD) of an ultrafast compound matrix containing each transducer array element. In order to form high quality
backscattered data for several plane wave transmissions. First, we images at higher frame rates, ultrafast plane wave compound
explain the physical signification of SVD and associated singular imaging [1]-[7] was introduced. It gave rise to a wide range of
vectors within the ultrafast matrix formalism. We theoretically applications such as Shear Wave Elastography, Ultrafast
demonstrate that the separation of spatial and angular variables, Doppler, Functional Neuroimaging or Quantitative Ultrasound
rendered by SVD on ultrafast data, provides an elegant and
imaging [8]. However, similarly to conventional imaging,
straightforward way to optimize the angular coherence of
backscattered data. In heterogeneous media, we demonstrate that image quality can still suffer from sound speed heterogeneities
the first spatial and angular singular vectors retrieve respectively in the medium; leading sometimes to non-negligible phase
the non-aberrated image of a region of interest, and the phase and aberrations on propagating wavefronts [9]. Such aberrations
amplitude of its aberration law. Numerical, in vitro and in vivo affect the image itself but also the ensuing quantitative
results prove the efficiency of the image correction, but also the estimations in post-processing steps.
accuracy of the aberrator determination. Based on spatial and In the last thirty years, many different approaches have been
angular coherence, we introduce a complete methodology for developed to address this issue but this topic remains highly
adaptive beamforming of ultrafast data, performed on successive relevant today despite longstanding research efforts.
isoplanatism patches undergoing SVD beamforming. The
Measurement of the transmit echo phase [10] in Computed
simplicity of this method paves the way to real-time adaptive
ultrafast ultrasound imaging and provides a theoretical Ultrasound Tomography [11] informs on the arrival time of the
framework for future quantitative ultrasound applications. transmitted wave front and allows correction of small
aberrations compared to the PSF width. In conventional
Index Terms — Adaptive Beamforming, Ultrafast Imaging, ultrasound, we are limited to the use of backscattered signals
Singular Value Decomposition, Aberration Correction and the ultimate goal of aberration correction is actually to
recreate an equivalent point-like scatterer for each pixel of the
image. Then, the phase and amplitude distortion of its spherical
I. INTRODUCTION backscattered echo retrieve the so called Green’s function [12].
Although ultrasound imaging is a mature medical imaging When no bright reflector is available but rather a random
tool, new techniques are constantly explored to further improve distribution of Rayleigh scatterers, time-reversal of this speckle
image quality and extend its clinical usefulness to new noise [13],[14] can be used to virtually recreate an artificial
applications. The first objective of Ultrasound imaging is to ultrasonic star, whose echo retrieves the aberration laws.
build an accurate image of the medium from ultrasonic Though, it remains an iterative and quite complex non real-time
backscattered signals through a process called beamforming. process. Other approaches studying the spatial coherence of
backscattered signals can be exploited to optimize the
This work was supported by the Inserm ART (Technology Research summation of the different plane waves in beamforming [15].
Accelerator) "Biomedical Ultrasound", the European Research Council (ERC In speckle noise environment which is the vast majority of
Advanced Grant FUSIMAGINE) and the Fondation pour la Recherche configurations assessed in Biomedical Ultrasound, Van Cittert
Médicale (F.R.M.).
H. Bendjador, T. Deffieux and M. Tanter are with Physics for Medicine
Zernike theorem states indeed that this spatial coherence
Paris, Inserm, CNRS, ESPCI Paris, Paris Sciences et Lettres University, 17 rue measures the focusing quality [16], which is essential for an
Moreau, 75012 Paris, France. (e-mail: [email protected]) accurate ultrasound image. Founded on the use of this theorem,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

coherence-based imaging methods [17] were developed to illuminated by the ith transducer. Let ℎ𝑘,𝑖 be the temporal
improve image quality, especially Contrast-to-Noise and impulse response defining the propagation between a
Signal-to-Noise ratio, due to speckle reduction techniques. All transducer element i and a scatterer k. Thanks to spatial
these methods improving spatial coherence were shown to reciprocity, the temporal impulse response ℎ𝑗,𝑘 defining the
decrease aberration across an aperture. In particular, Dahl and propagation between the scatterer k and the transducer element
Trahey [18] proposed an imaging technique based on the pixel j is equal to hk,j. Let 𝑒𝑖 (t) be the emitted signal on the ith
mapping of this spatial coherence of ultrasound signals. This transducer, the received signal on an element j can be expressed
short-lag spatial coherence method (SLSC) consists in as:
estimating the spatial coherence between closely-spaced
elements to create images demonstrating superior SNR and 𝑁𝑒 𝐾
CNR compared to conventional ultrasound images. SLSC 𝑠𝑗 (𝑡) = ∑ ∑ 𝛽𝑖𝑘 ∗ ℎ𝑗,𝑘 (𝑡) ∗ ℎ𝑘,𝑖 (𝑡) ∗ 𝑒𝑖 (𝑡) (1)
permits to improve the spatial coherence of backscattered
𝑖=1 𝑘=1
signals by using only data demonstrating a strong angular
coherence for different compounded transmissions without where ∗ stands for the convolution operation [26]. As the spatial
trying to estimate the aberration law [19]. This efficient resolution of the ultrasound image is limited by the ultrasonic
approach requires the estimation of spatial coherence functions wavelength, it is convenient to introduce a discrete
on a pixel per pixel basis whose computational cost currently representation of the ultrasound-formed image, with a spatial
hampers a real-time implementation [20], [21], [22]. It was also pitch of the order of the ultrasonic wavelength. Each pixel 𝑝 =
recently expanded to angular coherence in the context of (𝑝𝑥 , 𝑝𝑧 ) contains a subset of {K 𝑝 } scatterers. For each subset,
compounded plane waves [23]. we notice that the travel time, for a given scatterer k, can be
written as: 𝜏𝑝,𝑘 = 𝜏𝑝 + ∆𝜏𝑘 , with ∆𝜏𝑘 ≪ 𝜏𝑝 . Equation (1)
Here, we propose a different approach to improve the
becomes:
ultrasonic image quality. Our method shares the common aim
to increase the angular coherence of signals coming from each 𝑁𝑒 𝑁𝑥𝑁𝑧
pixel for different transmits during plane wave compound
imaging. This new approach offers a fast and efficient 𝑠𝑗 (𝑡) = ∑ ∑ ∑ 𝛽𝑖𝑘 ∗ ℎ𝑗,𝑝 (𝑡 − 𝜏𝑝 ) ∗ 𝛿(𝑡 − ∆𝜏𝑘 )
correction process, both retrieving an optimized ultrasonic 𝑖=1 𝑝=1 𝑘 ∈ {𝐾𝑝 }
image and the local estimation of the amplitude and phase of ∗ ℎ𝑝,𝑖 (𝑡 − 𝜏𝑝 ) ∗ 𝛿(𝑡 − ∆𝜏𝑘 ) ∗ 𝑒𝑖 (𝑡)
the aberrations. Additionally, it also filters out low coherence
signals from the data. 𝑁𝑒 𝑁𝑥𝑁𝑧
Moreover, the technique keeps the data in the conventional = ∑ ∑ ℎ𝑗,𝑝 (𝑡) ∗ ℎ𝑝,𝑖 (𝑡) ∑ 𝛽𝑖𝑘 𝛿(𝑡 − 2∆𝜏𝑘 ) (2)
beamformed data space allowing to subsequently perform any 𝑖=1 𝑝=1 k∈{Kp }
conventional ultrasound estimation such as phase estimation,
power Doppler, or pulsed Doppler.
Importantly, we introduce the Ultrafast Compound Matrix 𝛾𝑖𝑝 (𝑡)
containing images, both beamformed in the receive mode and
phase delayed for transmit travel path compensation. We Where 𝛾𝑖𝑝 (𝑡) is a backscattering function describing the
explain theoretically the physical meaning of its Singular Value reflection induced by the scatterers distribution in pixel p when
Decomposition. In complement to former works applying SVD insonified by the transmit element i. Thus, the received signal
to subaperture or synthetic transmit data [15],[24][25], we becomes:
𝑁𝑒 𝑁𝑥 𝑁𝑧
hereby demonstrate a theoretical link between the SVD of
beamformed data from different transmissions and the local 𝑠𝑗 (𝑡) = ∑ ∑ 𝛾𝑖𝑝 (𝑡) ∗ ℎ𝑗,𝑝 (𝑡) ∗ ℎ𝑝,𝑖 (𝑡) ∗ 𝑒𝑖 (𝑡) (3)
ultrasonic aberrations. Finally, we propose a complete real-time 𝑖=1 𝑝=1
adaptive beamforming technique for ultrafast imaging based on
the application of SVD beamforming on subsets of the Ultrafast The difference between the equation (3) and the equation (1)
Compound Matrix corresponding to isoplanatic patches. This lies in the introduction of a backscattering function 𝛾𝑖𝑝 (𝑡) for
SVD beamformer merges Phase Aberration Correction (PAC) each pixel, instead of a simple scalar value 𝛽𝑖𝑘 for each
techniques and coherence-based imaging approaches in a scatterer. This function characterizes the spatial dependence of
unique matrix formalism implementation. Results of this novel the scattering coming from a given pixel of the image. This
and fast method in simulations, in vitro phantoms and in vivo formalism no longer requires a Green’s function per scatterer,
liver experiments are presented. but only one for each pixel. Since the number of pixels (𝑁 =
𝑁𝑥 𝑁𝑧 ) is far below the number of scatterers in the medium (K),
II. THEORY AND PHYSICAL INTERPRETATION OF ULTRAFAST this strongly simplifies the calculations.
ULTRASOUND BEAMFORMING

A. Basic Principles of Ultrasound Imaging inverse problem

Considering a transducer array made of 𝑁𝑒 elements, and a
medium containing a random distribution of K scatterers
(Rayleigh diffusion with K >> Ne), we define a dimensionless
scattering amplitude 𝛽𝑖𝑘 characterizing the kth scatterer

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

like scatterers can be seen as maximizing the spatial coherence

of backscattered signals coming from each focal spot. For this
reason, the ultimate goal of ultrasound imaging can be seen as
maximizing the spatial coherence function of backscattered
signals, coming from all pixels or focal spots. Thus, this enables
to retrieve each Green’s function associated to each pixel
(Fig.1).
B. Fourier Domain: Beamforming and Adjoint Operator
In the Fourier domain, convolution turns out to be a matrix
product and equation (3) can be rewritten as:

𝑆(𝜔) = 𝐻(𝜔)Γ(𝜔) 𝑡𝐻(𝜔)𝐸(𝜔) (4)

For a sake of clarity, we will consider in the following

demonstration that 𝑆 = 𝑆(𝜔0 ), where 𝜔0 is the central
Fig. 1: Schematic description of the general aim of ultrasound imaging: Most frequency of the transmitted pulse. We introduce the same
scattering in biomedical ultrasound comes from random distribution of notation simplification for all vectors and matrices in the
Rayleigh scatterers. i.e. the speckle noise. For each pixel, the inverse problem,
or beamforming, consists in retrieving from this speckle noise the signature of following sections of the paper. Let us introduce a model of the
an equivalent single point-like reflector. The scattering of such point-like propagation between each element of the ultrasonic array and
reflector gives access to the local Green’s function of the pixel. each pixel of the image. This propagation model can be defined
as a propagation operator H0 as similar as possible to the true
Depending on the strength and distribution of scatterers in propagation operator H. The beamforming operation in the
each pixel of the ultrasonic image, the scattering functions receive mode, consists in estimating
𝛾𝑖𝑝 (𝑡) can be very different. If pixel p contains only a strong
point-like reflector, the scattering function 𝛾𝑖𝑝 (𝑡) is 𝐼 = 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻 𝐸 (5)
independent of the transmit element i. In that case, 𝛾𝑖𝑝 (𝑡) =
𝛾𝑝 (𝑡) and all backscattered signals coming from pixel p are I is a [𝑁𝑥 𝑁𝑧 , 1]-vector, corresponding to the image for a
coherent. Their spatial coherence function is thus a square transmission vector E. The symbol * corresponds to the phase
function. If the pixel p contains a random distribution of conjugate operator.
Rayleigh scatterers, the function 𝛾𝑖𝑝 (𝑡) is different for each
∗
transmission coming from different transmit elements. The In the case where 𝐻0 is equal to 𝐻, 𝑡𝐻0 𝐻 ≈ 𝐼𝑑 where Id is
spatial coherence function of backscattered signals coming the Identity Matrix, and 𝐼 = Γ 𝑡𝐻 𝐸. Thus, the beamformed
from pixel p is then a triangle function according to the Van image I corresponds to the exact image of Γ, where each pixel
Cittert Zernike theorem [16]. has been multiplied by an amplitude and phase contained in the
The ultimate objective of ultrasound imaging is to solve the vector tHE, describing the travel time path differences during
inverse problem of ultrasound propagation, in order to retrieve forward propagation from the array to the pixels.
all 𝛾𝑖𝑝 (𝑡) functions for each pixel p. Solving this problem In experimental configurations due to diffraction limits, it is
∗
requires to retrieve all Green’s functions hpi(t) relating each known that 𝑡𝐻0 𝐻 is not perfectly equal to the identity matrix,
pixel p to each element i of the ultrasonic array. When pixel p but rather estimates the point spread functions. Indeed, in each
contains a point-like reflector, the Green’s function hpi(t) can be ∗
ith column of 𝑡𝐻0 𝐻, we find the point spread function of the
retrieved by applying the concept of iterative time reversal focusing in the ith pixel [28][29][30][31]. So, if 𝐻0 is almost
focusing [27] or using the decomposition of the time reversal equal to 𝐻, we can introduce the Point Spread function Matrix
operator [26]. When pixel p contains a random distribution of Ps :
Rayleigh scatterers, it becomes much more complex to estimate
hpi(t). Indeed, due to the inherent speckle noise nature of
backscattered signals, the scattering function 𝛾𝑖𝑝 (𝑡) is 𝑡
𝐻0∗ 𝐻 = = 𝑃𝑠 (6)
dependent on the transmit element i. This fundamental limit
makes the inverse problem of ultrasound imaging difficult to
solve in speckle noise environment. Unfortunately, speckle where all columns of Ps describe the point spread functions for
noise configuration corresponds to the vast majority of each individual pixel focal spot of the experimental
biomedical ultrasound imaging. configuration: the ith column vector of Ps corresponds to the ith
Indeed, in speckle noise, retrieving hpi (t) requires to recreate pixel’s PSF.
a single virtual point-like reflector in each pixel p of the image,
and to separate the contribution of each pixel in the
backscattered signals. Recreating such a virtual point-like C. Conventional vs plane wave imaging
reflector from speckle noise consists in increasing the spatial In conventional imaging, the focusing in the transmit mode is
coherence of its backscattered signals from a triangle function performed by successive transmissions of transmit vectors
to a square function. Thus, the problem of creating virtual point- 𝐸𝑖 = 𝑡𝐻0,𝑝
∗
, where 𝐻0,𝑝 is the pth column of 𝐻0 . 𝑡𝐻0,𝑝
∗
is the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

phase conjugate of the Green’s function 𝑡𝐻0,𝑝 and thus transmit focusing, the ultrafast compound matrix R can be
represents a transmit vector focusing on pixel p. For each defined as follows:
transmit, the final image Im is the diagonal of the matrix I
defined in Fig. 2: 𝑅 = ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻𝐴𝑒 𝑃) ◦ ( 𝑡𝐻0∗ 𝑃 ∗ ) (10)

𝐼𝑚 = 𝑑𝑖𝑎𝑔(𝐼) = 𝑑𝑖𝑎𝑔( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻 𝑡𝐻0∗ ) (7) We introduce 𝐴𝜃 as the angular aberration matrix describing
these aberrations in the plane wave basis. It is related to 𝐴𝑒
In compounded plane wave imaging, the focusing is through the former matrix P of dimension (𝑁𝑒 , 𝑁𝜃 ) describing
performed by a set of 𝑁𝜃 > 1 plane wave transmissions. For our the projection from canonical space to plane wave space.
matrix formalism, this requires to introduce a Matrix P As long as the number of independent angles 𝑁𝜃 (independent
representing the change of referential basis between the meaning ∀𝑖, ∀𝑗 ⟨𝑃𝑖 |𝑃𝑗 ⟩ = 0) is sufficient to describe any kind of
canonical space (array elements) and the plane wave space. So ultrasonic wavefield initially defined in the canonical basis, the
P is a [𝑁, 𝑁𝜃 ] matrix. The final image obtained in plane wave matrix P is invertible [29]. In that case, we have 𝐴𝜃 = 𝑃 −1 𝐴𝑒 𝑃
compounding is thus quite similar to the conventional image and the angular aberration matrix 𝐴𝜃 is related to 𝐴𝑒 thanks to:
described in equation (7) with a potential filtering described by
𝑃 𝑡𝑃∗ when the number of plane waves is lower than the 𝑃𝐴𝜃 = 𝐴𝑒 𝑃 (11)
number of elements:
Combining (10) and (11), the ultrafast compound matrix R can
𝐼𝑚 = 𝑑𝑖𝑎𝑔(𝐼𝑃𝑊 ) = 𝑑𝑖𝑎𝑔( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻 𝑃 𝑡𝑃∗ 𝑡𝐻0∗ ) (8) be rewritten using the angular aberration matrix as:

𝑅𝜃 = ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻0 𝑃𝐴𝜃 ) ◦ ( 𝑡𝐻0∗ 𝑃∗ ) (12)

D. The Ultrafast Compound Matrix R
We now choose to introduce an important matrix in However, when 𝑁𝜃 ≪ 𝑁𝑒 , P is no longer invertible and 𝐴𝜃 is
compound plane wave imaging: the ultrafast compound data only an approximation of the true aberration projected on the
matrix R. It contains the 𝑁𝜃 images beamformed in the receive limited number of plane waves. In that case, 𝑅𝜃 ≠ 𝑅 and 𝑅𝜃 is
mode and individually time delayed pixel per pixel depending only an approximation of the true ultrafast compound matrix R.
on their transmit angle. R is thus a [𝑁𝑥 𝑁𝑧 , 𝑁𝜃 ] matrix. It
corresponds to the matrix of compounded plane wave data right Let us now consider a homogeneous medium with a phase
before the final summation on angles, which is the last aberrating screen in the canonical basis. Consequently, the
operation in ultrafast data beamforming to retrieve the coherent aberrator 𝐴𝑒 is a [𝑁𝑒 , 𝑁𝑒 ] diagonal matrix. When 𝑃 −1 is defined
compounded image. As done earlier, we note the number of and the phase aberrator does not contain very high spatial
pixels 𝑁 = 𝑁𝑥 𝑁𝑧 . R can be written as: frequency variations in the canonical space (which is typically
the case in medical ultrasound), 𝐴𝜃 can also be approximated
𝑅 = ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻 𝑃) ◦ ( 𝑡𝐻0∗ 𝑃 ∗ ) (9) by a [𝑁𝜃 , 𝑁𝜃 ] diagonal matrix. In the following equation and
under the previous hypothesis, we will note 𝐴 = 𝐴𝜃 , whose
Where ◦ stands for the matrix Hadamard product (Fig. 2). elements are 𝑎 = 𝑑𝑖𝑎𝑔(𝐴𝜃 ) = (𝑎𝑖 )𝑖=1..𝑁𝜃 .
The Hadamard product is a distributive operator, which is
also commutative for diagonal matrices (commutative ring of
ℳ(ℂ)). Indeed, if A and B are {m*n}-matrices, and D and E
diagonal matrices of respective size m and n, it can be
demonstrated that:
𝐷(𝐴 ◦ 𝐵)𝐸 = (𝐷𝐴𝐸) ◦ 𝐵 = (𝐷𝐴) ◦ (𝐵𝐸) = 𝐴 ◦ (𝐷𝐵𝐸)

Thanks to this property, Equation (12) becomes:

𝑅 = ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻0 𝑃𝐴) ◦ ( 𝑡𝐻0∗ 𝑃 ∗ )

= ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻0 𝑃) ◦ ( 𝑡𝐻0∗ 𝑃 ∗ )𝐴

This results in 𝑅 = 𝑀 . 𝐴 , where we define the matrix M as:

Fig. 2: Matrix description of beamforming in coherent compounding:
definition of the Ultrafast Compound Matrix R and the Plane wave imaging
matrix IPW. The Hadamard product describes the coherent compounding 𝑀 = ( 𝑡𝐻0∗ 𝐻 Γ 𝑡𝐻0 𝑃) ◦ ( 𝑡𝐻0∗ 𝑃∗ ) (13)
before the final summation on the transmit angles: term-to-term product with
𝑡 ∗ ∗
𝐻0 𝑃 compensates phases delays and amplitudes between transmitted plane Interestingly, the columns of the matrix M correspond to the
waves.
ideal image of the medium, without any aberration, for each
Let us define 𝐴𝑒 , the [𝑁𝑒 , 𝑁𝑒 ] matrix describing a physical plane wave transmit. Note that the travel path differences in the
realistic aberrator in the canonical basis (array elements space). transmit phase have been compensated in M. The ability to
Assuming that the aberrator does not dramatically affect describe the matrix R as a product between ideal images M and
a diagonal Matrix A will be important in the next sections.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

image. We just demonstrated that its angular covariance has to

be optimal. So, compensating for the aberrations consists in
E. The Local Angular Coherence Matrix Cθ
finding the correction vector X that maximizes the angular
In order to optimally correct ultrasound images, we intend to covariance of the image RX. This solution will verify 𝑋 =
maximize the coherence of backscattered echoes between plane 𝑑𝑖𝑎𝑔(𝐴∗ ). We are trying therefore to maximize the Rayleigh
wave transmits, as in [23]. For this reason, we introduce a Local quotient:
Angular Coherence Matrix 𝐶𝜃 . First, we choose a control region
of interest (ROI) in the image where we would like to control 𝑡
(𝑅𝑋)∗ . (𝑅𝑋) 𝑡 ∗ 𝑡 ∗
𝑋 ( 𝑅 𝑅) 𝑋
the angular coherence of backscattered signals. We extract the 𝐽(𝑋) = ∗
=
𝑋 .𝑋 ‖𝑋‖
local Ultrafast Compound matrix 𝑅̃ of this ROI. The angular
coherence of this region of interest is retrieved by calculating One notable property of the Rayleigh quotient is that it is
the covariance matrix:
maximized by the first eigen vector of the matrix ( 𝑡𝑅 ∗ 𝑅).
𝐶𝜃 = 𝑡 ̃∗
𝑅 . 𝑅̃ (14) ( 𝑡𝑅 ∗ 𝑅) being a Hermitian matrix, we also know that its eigen
vectors are the singular vectors of the matrix R. So, this
Fig. 3 shows the computation of 𝐶𝜃 using experimental data demonstrates that the first singular vector of R maximizes the
in the case of a speckle environment in the region of interest. angular covariance of the beamformed data, and thus
We will later discuss the influence of the ROI size. We notice corresponds to the aberration correction diag(𝐴∗ ). This
that the anti-diagonal terms exhibit the triangle-shaped demonstrates the striking importance of the Singular Value
coherence expected with Van Cittert Zernike theorem. When Decomposition of the Ultrafast Compound Matrix: 𝑅 =
the same medium is now imaged through an aberrating lens, the 𝑈 𝑆 𝑡𝑉 ∗ . The Rayleigh quotient optimization tells us that
coherence vanishes. This witnesses the strength of the speckle 𝑑𝑖𝑎𝑔(𝐴∗ ) = 𝑉1 , where 𝑉1 is the first singular vector, ie the first
coherence criterion in aberration correction. The complex column of V.
magnitude of this correlation matrix informs on the angular ∗
coherence between ultrasonic images acquired with different If we approximate 𝑡𝐻0 𝐻 ≈ 𝐼𝑑, we consider the PSF being
plane wave transmits summed over all pixels of the region of limited to the focal point and not affecting neighboring pixels,
interest. In speckle noise, as seen before, we expect it to the previous matrix M in equation (12) becomes:
decrease linearly as the angle difference increases.
𝑀 = ( Γ 𝑡𝐻0 𝑃) ◦ ( 𝑡𝐻0∗ 𝑃∗ ) = Γ . 𝑡𝐻0 𝑃 ◦ 𝑡𝐻0∗ 𝑃 ∗
Γ1 − Γ1
1 − 1
Γ2 ⋮ Γ2
= Γ . [| | |] = [ ] (15)
⋮ ⋮ ⋮
1 − 1 Γ𝑁 − Γ𝑁
All columns of M are identical, and we can write:

𝑚1
𝑅 = 𝑚 𝑡𝑎 = [ ⋮ ] . [𝑎1 ⋯ 𝑎𝑁𝜃 ] (16)
𝑚𝑁
Where m is the vector of the columns of M.

Such angular and spatial variables separation in equation (16)

exactly corresponds to the definition of the SVD. Indeed, let’s
write the SVD of our ultrafast compound matrix R:
𝑅 = 𝑈 𝑆 𝑡 𝑉 ∗ = ∑𝑁 𝑡 ∗
𝑖=1 𝑠𝑖𝑖 𝑢𝑖 𝑣𝑖 where S is a diagonal matrix.
𝑡
As 𝑅 = 𝑚 𝑎, we have here:

𝑠11 = |𝑚
̅ |. |𝑎|, 𝑠𝑖𝑖 = 0 ∀𝑖 > 1 (17)
Fig. 3: Measurement of the complex magnitude of the angular coherence 𝑎∗ ̅
𝑚
matrix 𝐶𝜃 module in a control region of interest using the Ultrafast compound 𝑣1 = |𝑎|
; 𝑢1 = ̅|
|𝑚
Matrix 𝑅̃ extracted from R. 𝑅̃ is acquired in (a) non-aberrated and (b)
aberrated speckle noise using 100 transmit compound angles. On the right
side, diagonal terms are averaged and plotted, representing the angular 𝑣1 is the phase conjugate of the aberration vector 𝑎. 𝑢1 is the
coherence factor. normalized image in a non-aberrated medium: all pixels are
seen the same way for each plane wave transmission. In this
F. The Singular Value Decomposition of the Ultrafast case, it is trivial that the decomposition of 𝑅 is the SVD since
Compound Matrix all further singular vectors are zero, and thus orthogonal to the
first one.
If we no longer consider the PSF being a spatial Dirac, and
The estimation of the angular aberration matrix 𝐴 is essential ∗
add speckle noise in the media, 𝑡𝐻0 𝐻 ≠ 𝐼𝑑 and the columns
to perform proper adaptive imaging. From the definition of 𝐴,
assumed to be diagonal, 𝑅𝐴∗ corresponds to the corrected of M are no longer identical. There is thus an angular
dependence of the imaged medium with respect to the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

insonification angle. We can introduce the averaged PSF: ̅𝑚

̅̅̅𝑖 = 4). This means that all pixels of the singular image 𝑉1 are seen
1 𝑁𝜃 the same way for all plane wave illuminations angles due to the
∑𝑗=1 𝑚𝑖𝑗 . The elements of M can be written: 𝑚𝑖𝑗 =
𝑁𝜃
separation of variables achieved by the SVD. The vector 𝑈1
𝑚
̅(𝑖) + ∆𝑚𝑖 (𝜃𝑗 ), ∀𝑖 = 1. . 𝑁, ∀𝑗 = 1. . 𝑁𝜃 . contains the amplitude and phase delays between transmits, that
are required to achieve a constant coherence on the receive
If we consider the angular dependence of the point spread signal. In other words, it contains the aberration correction law
|∆𝑚 |
function to be reasonable, by assuming | ̅ 𝑖| ≪ 1, we can write: of the medium acting the same way on each pixel of the region
𝑚
of interest. For this reason, an important point is to understand
𝑅 = 𝑀 . 𝐴 = [𝑚
̅(𝑖) + ∆𝑚𝑖 (𝜃𝑗 )]𝑖𝑗 . 𝐴
the optimal size of the control region for SVD as it is closely
̅ (𝑖) + 𝑎𝑗 ∆𝑚𝑖 (𝜃𝑗 )] 𝑖𝑗
= [𝑎𝑗 𝑚 linked to the concept of isoplanatic patches [28][32].
=𝑚 ̅. 𝑡𝑎 + ∆𝑀. 𝐴 (15)
Strikingly, the Singular Value Decomposition of the
And we finally get: Ultrafast Compound matrix gives access, with its first singular
𝑚̅ 𝑎 ∆𝑀.𝐴 vector, to the knowledge of the aberrating screen. We notice
𝑅 = |𝑚
̅ |. |𝑎|. [ |𝑚̅| . |𝑎| + |𝑚̅ |.|𝑎|
] (16)
that selecting the first vector out of the SVD is equivalent to
̅
𝑚 𝑎 ∆𝑀.𝐴
with ||𝑚̅| . |𝑎|| = 1 and | |𝑚̅|.|𝑎|| ≪ 1. filtering the images and re-phasing them with the phase
conjugate of the aberrator 𝐴∗ . Indeed, SVD filtering creates
Again, we know that the first singular vector of this SVD, angular coherence for each point of an aberrated medium. This
with the highest singular value, represents the best linear fit for is the definition of creating a virtual reflector in each pixel
the rows of R. Each row of R contains the 𝑁𝜃 scattering complex location. This is actually the mathematical explanation for the
amplitudes of one pixel seen by each individual transmit. This bright reflector virtual creation in speckle noise described in
first singular vector maximizes the resemblance between the [13]. The interesting and unique aspect of our method stands in
rows of R and the fit. In other words, the coherence between the the physical understanding of the way to re-create coherence in
images seen from different angles is maximized. Physically, aberrated signal. The SVD-beamformer is a straightforward and
this is equivalent to a maximization of the energy of R rows non iterative solution for this complex problem addressed in
which corresponds to the highest energy of all singular vectors. [13].
As we know, due to the speckle noise features that the term One should note that the aberration was here assumed to have
an effect only in the transmit mode, for a sake of simplicity in
∆𝑀. 𝐴 is negligible compared to 𝑚 ̅ . 𝑡𝑎, we can again identify:
̅
𝑚 𝑎∗
the former mathematical developments. The aberration in the
𝑠11 = |𝑚
̅ |. |𝑎| ; 𝑢1 = |𝑚
̅|
; 𝑣1 = |𝑎|
. receive mode can be taken into account by re-writing Equation
(9) in:
The first singular vector of R is containing in 𝑣1 the phase
conjugate diag(A*) of the aberrator. Additionally, 𝑢1 contains 𝑅 = ( 𝑡𝐻0∗ 𝑃0∗ Γ𝐴 𝑡𝑃0 𝑡𝐻 𝑃𝐴) ◦ ( 𝑡𝐻0∗ 𝑃∗ )
the normalized image of the medium where the aberration has
been corrected. All pixels behave as individual point-like Where 𝑃0 is the orthonormal [𝑁𝑒 , 𝑁𝑒 ]-matrix, changing from
reflectors with no angular dependence in transmission. For each canonical to plane wave basis with 𝑟𝑎𝑛𝑘(𝑃0 ) = 𝑁𝑒 .
singular vector i, one can compute the coherence matrices In ultrafast imaging, 𝑟𝑎𝑛𝑘(𝑃0 ) is often much greater than
𝐶𝜃 (𝑉𝑖 𝑈𝑖∗ ), defined in part II.E. The mathematics of the SVD 𝑟𝑎𝑛𝑘(𝑃) as 𝑁𝑒 ≫ 𝑁𝜃 in most cases. Then, likewise, we can
operation imply that these matrices are constant and equal to 1 write :
as seen in Fig. 4. It means that, by definition, for each singular 𝑅 = 𝑀 .𝐴 = 𝑚 ̅. 𝑡𝑎 + ∆𝑀. 𝐴, and the SVD retrieves as well
vector, all transmitted angles “see” the same image. the aberration matrix.

III. METHODS
A. Acquisitions
Ultrasound acquisitions were performed using a 256-channel
programmable research scanner (Verasonics Research
Systems). We developed customized sequences to drive a 192-
element linear probe, at a central frequency of 6.25 MHz and
pitch size of 0.2 mm (SL10-2 probe, Supersonic Imagine, Aix-
en-Provence, France). Plane waves steered at angles between -
18° and +18° were transmitted to insonify the media. In order
to guarantee the highest level of information, 100 angles at
Fig. 4: Singular Value Decomposition of the Ultrafast Compound Matrix.
Absolute value of angular Coherence matrix of R shows speckle decorrelation, Pulse Repetition Frequency (PRF) = 10 kHz were acquired
whereas singular vectors exhibit individually an optimized angular coherence. along with subsets of steering angles for higher frame rate
images. A phantom containing reflecting pins, hyperechoic and
Interestingly, if we compute the Singular Value anechoic cysts (CIRS 054GS – 1540 m/s) was used to assess
Decomposition of 𝑅̃ , and re-calculate 𝐶𝜃 considering only the our method, successively without and with a shaped-surface
first singular vector 𝑉1 , it gives a uniform matrix equal to 1 (Fig. silicone aberrating lens.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

Fig. 5: Fast aberration correction and filtering with SVD. Ultrafast compound matrix R retrieves IQ data beamformed for each transmit angle. Isoplanatic patches
are defined within the imaged area. SVD is performed on each patch of the image, and the first eigen vector gives directly the rephased image. Final corrected
image is computed by assembling corrected patches.

ultrafast compound in homogeneous medium was considered as

B. Correction method
the goal to reach for the aberrated images correction by the SVD
Classical Delay-And-Sum beamforming - at dx = λ, dz = λ/2 beamformer.
- was performed on plane wave data for each steered transmit. We call numerical aberration, the introduction at emission on
Resulting IQ signals were stored in a 3D-matrix R: the ultrafast acquired data, of a known angular delay law (Fig. 6) and we
compound matrix containing (x,z)- images for each angle θ evaluated the consistency between the phase aberration
(Fig. 5①). Summation along the third dimension, gives simply determined with our method, and the expected one. We also
the conventional plane wave compounded B-mode image. introduced physical aberrations: between the probe and the
phantom, we placed a silicone aberrating lens and applied the
Matrix R is reshaped into a 2D Casorati Matrix form with a SVD beamformer to compensate this near-field aberrator in the
[Nx Nz, Nθ] dimension. The Singular Value Decomposition is corrected image (Fig. 7). Finally, in vivo experiments were
then performed on delimited spatial ROIs called “control conducted in the human liver where aberrations are more
patches”, each corresponding to a smaller Ultrafast Compound spatially distributed.
matrix 𝑅̃ . We can adapt the dimension of the control ROIs with In all experiments, many patch sizes were tested and the final
respect to the isoplanatism limited extension induced by the result was obtained with the largest patch keeping the
aberrator. SVD of each 𝑅̃ matrix is performed and provides the consistency of the aberrator in both x and z dimensions.
separation of spatial and angular variables. For each singular Quantification of the image quality can be found in the
value, the SVD consists actually in the product of images V(x,z) angular coherence matrix, which mathematically becomes
and angular vectors U*(θ) weighted with singular values λ. The constant and maximal between all angles after correction.
decomposition is filtered, keeping only the image with the Though, we can also provide classical estimators such as lateral
highest singular value. For each angle, the corrected image of resolution – computed by the half width lateral resolution of a
𝜇
each ROI is thus corresponding to λ1.V1(x,z). Summing data on reflecting pin intensity, and contrast defined by: = 𝑖 , where
𝜇𝑜
the different angles retrieves this corrected image (Fig. 5②). 2 2
𝜇𝑖 = 𝑠𝑖 and 𝜇𝑜 = 𝑠𝑜 are the mean square intensities
The IQ signals have been rephased with the complex U1(θ) respectively inside (i) and outside (o) an anechoic cyst.
correction vector containing the phase and amplitude
corrections. As described in part II., the first eigen vector IV. RESULTS
guarantees a maximization of the angular coherence. This
coherence is indeed a constant square function equal to 1, since A. SVD Beamforming in speckle noise with numerical
all angles “see” the same outcome from the object. aberrators
If one computes the angular coherence in speckle for a casual
We studied four experimental cases for comparison: the B-mode image, limited to a control region, the obtained matrix
classical ultrafast compound imaging in homogeneous medium, exhibits a triangle profile: covariance is maximal along the
the SVD-beamforming using experimental data incorporating a diagonal – auto-coherence - and vanishes linearly down to zero
numerical aberration and finally in vitro and in vivo when the angular spacing increases (Fig. 3).
experimental data containing physical aberrations. The classical

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

Van Cittert Zernike theorem stipulates actually that the shape simulated aberration law is defined in the canonical space, we
of the spatial coherence is driven by the inverse Fourier used synthetic aperture acquisitions to reconstruct the realistic
transform of the focal spot intensity. The focal spot intensity aberration. Then, we built the ultrafast compound matrix and
varies as sinc², which is the Fourier Transform of a triangle extracted the phase of the first angular singular vector. This
function. Indeed in speckle, one obtains a linear decrease in gave us an angular aberration law, which was projected in the
spatial coherence as much as the lag between elements element space to be compared with the true numerical
increases. This theory is transposable when it comes to angular aberration (Fig. 6 (e)). Both laws are in very good agreement
coherence and transmit angular lag, as demonstrated in with a coefficient of determination r² = 98%. This validates the
reference [23]. This explains the evolution of angular ability of our method, on simulated examples, to retrieve
coherence of backscattered signals with respect to the lag in realistic aberrations initially defined either in canonical or in
transmitted angles. When it comes to an aberrated image, plane wave basis.
speckle is highly decorrelated and the angular coherence of Even if the numerical aberration is not fully distributed in the
backscattered signals sinks very quickly from an angle to the medium but rather a phase screen, we did perform the SVD-
next one. After SVD filtering of the image, this effect is totally beamforming on different geometrical patches. We selected in
recovered, and even further since the angular coherence of each the images different patches containing only speckle (no
block is constant equal to 1. Even if this result seems certain reflectors nor anechoic areas). Fig. 6 (d) and (e) show that the
mathematically, it illustrates that the SVD provides an image, phase laws extracted from this patch-SVD (corresponding to
for which all pixels are being seen the same way by all transmit V1*) are consistent with the expected numerical aberrations.
angles. This demonstrates that the efficiency of the aberration law
extraction does not rely on the presence of coherent targets
within the patch and only requires speckle.
We further tested 50 different numerical aberrators to assess
the ability of SVD beamforming to correct strong and highly
variable dephasing screens. We used, as numerical aberration
laws, some randomized linear combinations of linear, parabolic
and sinusoidal phase laws. The order of magnitude was chosen
to be consistent with typical in vivo aberrating layers, up to few
tens of microseconds. The correlation length of the tested phase
laws varied randomly from 5 to 20 angles.

SVD Speckle
Beamformer Brightness
Mean aberration law
0.9935 0.6727
correlation (r²)
Contrast on initial image -28.2 dB
Mean contrast after
-18.7 ± 4.1 dB
aberration ± std
Mean contrast after
-28.2 ± 0.3dB -21.1 ± 0.9dB
correction ± std
Mean contrast
9.5 dB 2.4 dB
improvement
Target width on initial image 0.54 mm
Target width on aberrated
0.70 ± 0.14 mm
image
Fig. 6: (a) Bmode image of a CIRS 054-GS phantom anechoic cyst. (b)
Example of image after introduction of an angular numerical aberration in Target width after correction 0.57 ± 0.005
0.64 ± 0.08 mm
transmit. (c) Same image as (b) after SVD-beamformer. (d) Example of ± std mm
simulated angular aberration: SVD-beamforming result is compared to the true Mean resolution
aberration introduced. (e) Example of simulated canonical laws: SVD- 22.8 % 9.4 %
improvement
beamforming result is projected in the element-space and then compared to
the true canonical aberration introduced. Table 1: Resume of the results of SVD beamformer on in vitro aberrated
Compared to the raw image, the numerically aberrated image phantom, with 50 aberrating phase screens tested. Speckle brightness aberration
correction was also implemented to compare the correction effects.
showed a strong degradation (Fig. 6 (a) and (b)), that is fully
recovered after SVD-beamforming correction (Fig. 6 (c)). The
lateral resolution, evaluated on reflecting pins, is effectively We added to the results a comparison with another existing
aberration correction: we indeed implemented Speckle
improved from 1.02 mm to 0.88 mm. Also, the contrast on the
Brightness approach as described in [19]. A first speckle
anechoic cyst in speckle is enhanced up to 11.7 dB, and reaches
brightness value was calculated on the aberrated image. For a
the contrast of the non-aberrated image. When the simulated
law is directly defined in the plane wave space, the SVD- given array element, the RF signal was delayed of 0.2λ at
beamformer result is shown to be very consistent with the emission and reception in beamforming. Speckle brightness
introduced phase change (Fig.6 (d), r² = 99%). When the was then re-calculated on the obtained image. If the speckle

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

brightness increased, the RF delay was incremented of +0.2λ (- pin. We found again a very good agreement (r² = 95%) between
0.2λ if it decreased). The process was iterated until a maximal those two angular aberration laws. In the case of a thick
value of speckle brightness was reached. This was repeated for aberrating lens, smaller isoplanatic patches were used to
each element of the transducer. Finally, beamforming with perform the correction, in order to keep the aberration
optimal delays gave the corrected image. We performed both correction as optimal as possible on the patch spatial extension.
methods on the exact same set of aberrated data and retrieved An interesting issue remains in the optimization of this patch
the results in Table 1. First we observe, with the r² value, that size in order to perform the most efficient possible correction.
the SVD beamformer is very efficient to recall the aberration In order to investigate this point, we designed patches around a
law. Also, both contrast (-9.5 dB) and resolution (-22.8%) central point of five different sizes, and performed correction
improve significantly with SVD beamforming. These on each of them with SVD beamforming (Fig. 7).
improvements clearly outperform speckle brightness method. C. Influence of the patch size on aberration correction
The low standard deviations demonstrate that the SVD
beamforming approach gives confident results regardless the In order to validate our method, and identify the patch sizes
numerical aberration. providing a realistic result, we considered a speckle phantom
image containing a single reflecting pin at its center (Fig. 8. (a)).
First, we used the backscattered signal coming from the
B. SVD beamforming in speckle noise with physical reflecting pin at the center of our image. From these RF data,
aberrators we extracted the aberration delay law in the element space. We
These improvements of the SVD beamformer are visible as used the matrix P defined in part II.C. to project the canonical
well on physically aberrated data (Fig. 7). As the non- law into the plane wave basis. This gave us an estimation of the
negligible thickness of the aberrator tends to decrease the physical aberration encountered in the region surrounding the
isoplanatic angle, the SVD processing is now performed on reflector, and is displayed on Fig. 8.(f).
many different isoplanatic patches. After SVD beamforming,
reflecting pins resolution improved from 2.1mm to 1.3mm, and
contrast increased by 8.5 dB+/- 1.9 dB.

Fig. 8: (a) B-mode image of a phantom containing speckle and one reflecting
pin. (b) Small patch definition: [25λ x 30λ] region excluding the central
reflecting pin (c) Median patch definition: [35λ x 45λ] region excluding the
central reflecting pin (d) Large patch definition: [50λ x 100λ] region excluding
the central reflecting pin (e) Normalized eigenvalues distribution for each patch
size (zoomed for low eigen indices). (f) Angular aberration laws extracted from
SVD beamformer for a small patch compared to the experimental canonical
Fig. 7: (a) Bmode images of a CIRS 054-GS phantom reflecting pins and aberration obtained from the central pin and then projected in plane wave space.
anechoic cyst. (b) Images after propagation through an aberrating lens. (c)
Images after SVD-beamformer. (d) Aberrating laws: extracted from SVD-
beamforming on a patch, and from the closest pin echo.
Then, we defined a small patch (25λ x 30λ, Fig. 8 (b)), a
median patch (35λ x 45λ, Fig. 8 (c)), and a large patch (50λ x
We were also able to compare the aberration law extracted 100λ, Fig. 8 (d)) surrounding the reflector. For the SVD-
from the SVD patch correction and its true estimation - obtained beamforming operation, we excluded the reflecting pin area at
by using the backscattered wavefront coming from the closest the center of the patch in order to rely only on speckle noise.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

For the small, and median patches, the phase of the first angular configuration will essentially depend on the clinical
singular vector retrieves the same result in terms of aberration application.
and image correction. That means that for these patch sizes, the
patch is seeing the same aberrator and the SVD-beamformer
provides an optimal correction and estimation of the aberration.
When increasing the patch size, we observe a change in the
phase. The spatial extent becomes then too large to ensure a
correct solution for the SVD beamformer. Indeed, the
aberration can no longer be considered similar for all pixels of
the patch, and it is logical that the SVD performance degrades.
In the opposite case, for a too small patch, the number of pixels
would not be sufficient for the SVD to extract the correct
aberration law. One can see that for the smaller patch sizes, the
physical aberration estimation is in good agreement with the
SVD-Beamformer results.
For this case, we are thus able to identify a typical patch size
for aberration correction. It is actually the highest size for which
the phase laws remain consistent. This optimization of the patch
size can be done by looking at the singular value distribution of
the SVD for different sizes. Fig. 8 (e) shows that the singular
value distribution drops rapidly to zero for a correct patch size,
whereas in the case of too large patches, the slope at the origin
is less steep. The ratio between the first and second singular
value of the Ultrafast Compound Matrix can be considered as
an interesting parameter to choose the size of isoplanatic patch.
In the case of the time reversal operation HtH*, such ratio was
shown of particular importance as it affects the convergence of
iterative time reversal processing [26], [33]. Fig. 9: Impact of lowering the amount of plane wave transmits on SVD
beamforming for an aberrated pin phantom (1 = result of the compound, 2 =
result of SVD beamformer). (a) 100 plane wave images, (b) 10 plane waves
D. Influence of the number of plane wave transmissions images, (c) 5 plane wave images, (d) Evolution of the resolution and contrast
The previous results showed that, for an appropriate patch after correction with SVD beamforming on different amount of plane waves.
Contrast was estimated on an anechoic inclusion acquired with the same
size, the SVD beamformer allows us to retrieve both the aberrator.
aberration law and the image correction. Though, this last effect
should decrease when less plane waves are transmitted. The
SVD operation is indeed more efficient when the E. In vivo SVD beamforming in the human liver
decomposition basis is larger. Therefore, we tested different Finally, we tested our SVD beamforming on in vivo data.
acquisitions with various amount of plane wave transmits on The ultrasound emission sequences were calibrated with the
aberrated phantoms. Fewer angles gives less angular frequency acoustic measurement system from Acertara Acoustic
components to describe the phase law, but we still see an Laboratories. Both the spatial peak time-average (ISPTA) and
improvement in the first singular vector image. Fig. 9 shows in the Mechanical Index (MI) were shown to be, in all considered
(a), (b), (c) the comparison of the aberrated image, and the cases, below FDA Track 3 Recommendations (ISPTA =
SVD-beamformed image for different amount of plane wave 720mW/cm² and MI = 1.9). Indeed, maximum values obtained
transmits. We estimated, with the same aberrator in both cases, with our sequences were respectively ISPTA = 254.2 mW/cm²,
the lateral resolution on the pin phantom, and the contrast on an and MI = 0.819.
anechoic inclusion. This allowed us to perform our acquisitions on liver of a
healthy volunteer. Fig. 10 shows Bmode comparison of
We compared, on the graph (d) of Fig. 9, the correction classical compounded image, and SVD-beamformed image.
efficiency as a function of transmitted plane waves. For all The overall quality of the image improves with a contrast
configurations with different number of plane wave enhancement of 4.78 dB. Interestingly, some structures that
transmissions, the SVD beamformer improves the image were difficult to image in classical compound, tend to be more
quality both for contrast and resolution. Contrast improvement visible in the corrected image. These preliminary in vivo results
decreases progressively, whereas there is a clear drop of show the reliability of the proposed method to correct more
resolution improvement around 6 transmits. Still, one can note spatially distributed aberrators. This is therefore a promising
that even for as low as 5 plane waves, the resolution keeps first step towards clinical use of SVD-beamformer for adaptive
getting better with the SVD correction even if it is less imaging of patients.
remarkable on contrast. It is therefore a matter of trade-off
between the frame rate expected (amount of transmitted plane
waves) and the efficiency of the correction. The chosen

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

We believe that the computation time could be even further

reduced by optimizing the parallelization of the algorithms.
This allows us, already, to consider novel applications such as
real-time adaptive functional imaging, motion-correction on
cardiac imaging, or transcranial Doppler imaging of the brain.
Also, we present here computation times for small number of
angles. Though, too few transmit angles will inexorably lead to
a weaker performance of Singular Value Decomposition, Fig.
9 shows the performance of the correction depending on the
number of transmitted plane waves. Interestingly, the SVD
beamformer improves the image quality – both in terms of
resolution and contrast - up to a quite small number of angles
(typically 10 to 20 angles). Even in the extreme case of N=5
plane waves, the SVD beamformer leads to an improvement of
resolution nevertheless without any contrast gain anymore. In
this case, it is due to the lower number of available orthonormal
basis vectors insufficient to fully describe the aberrator.

Processing Time for processing (s)

Number of patch(es) 1 9 25 100

Fig. 10: In vivo B-mode results on liver. Acquisitions of 41 transmitted plane Size of patch (pixels) 592*192 196*63 117*37 58*18
waves : (a) Classical compound beamforming showing red-squared region. (b)
Renormalized Bmode of the anechoic vessel area. (c) Result of SVD- Beamforming
0.000036
beamforming on the whole image. (d) Result of SVD-beamforming in the 100 angles
region of interest. SVD-correction
0.50 0.41 0.34 0.50
100 angles
V. DISCUSSION SVD-correction
0.025 0.018 0.023 0.045
The results presented above demonstrate that the SVD 10 angles
beamformer is a reliable and efficient technique to correct SVD-correction
0.012 0.010 0.014 0.029
images and retrieve simultaneously the phase aberration law 5 angles
from ultrafast data acquired using plane wave compounding. In
vitro experiments in tissue mimicking phantoms have shown Table 2: Calculation times for an image - beamformed at λ/2 - acquired at a
both a 9.5 dB+/- 0.9 dB rise in contrast for simulated aberrations central frequency of 6.25 MHz and a sampling frequency of 25 MHz, at 5cm
and a 8.5 dB+/- 1.9 dB contrast rise for physical aberrations. depth with a 192-element probe. Calculations were computed on a PC with an
IntelCorei7-5820K CPU (3.30GHz), 32 Go RAM, on a Graphics Processing
The lateral resolution was also improved in both cases Unit: GeForce GTX 1080 TI.
respectively from 0.7mm to 0.57mm in the simulated
aberrations case, and from 2.1mm to 1.3mm with the physical For transcranial imaging, it could seem limitative that our
aberration. Another simple way to assess the accuracy of our method only corrects in emission. Though, due to time-reversal
method in retrieving the aberration pattern was to beamform symmetry of the wave equation, the aberration is also
raw data while correcting the phase and amplitude aberrations symmetric. It is therefore possible to correct in emission, as
determined by SVD, and to re-perform the SVD beamforming well as in reception by re-beamforming the data while taking
process on this new set of corrected data. As expected, this the aberration delay into account in both propagation ways. To
resulted in a flatten phase distribution, as if no aberration had consider the aberration phase in reception, the delay law would
distorted the corrected backscattered signals. This showed also need to be expressed in the canonical basis. This can be
that a single round of correction was required whereas existing obtained by a change-of-basis thanks to the matrix P introduced
methods often imply iterative processes. Of particular interest in II.C. We performed tests on simulated delay laws, and
is the simplicity and straightforward implementation of this observed that the law remains highly consistent down to a
SVD beamforming. Using parallel computation devices, and dozen transmit angles. For lower number of angles, the
high-end processors, the implementation time of our method is projection of the estimated angular aberration in the canonical
significantly low enough to perform real-time aberration basis fails to reach the true aberration. This is a limitation, for
correction. It depends on the number of transmit angles in the now, to the ultrafast character of the adaptive beamforming
data, but also on the size of the isoplanatic patches for Singular when both transmit and receive correction are desired.
Value Decomposition. We showed in Fig. 9, that for a Another limitation of our technique relies on the assumption
reasonable amount of plane waves, and for a patch size of tens of near field phase screen aberrator. Despite longstanding
of λ in axial and lateral directions, the computation speed is yet research in the field of aberration corrections, this assumption
high enough to correct in real-time the beamformed data which has been shown insufficient to correct phase aberration in a way
is a major advantage of this SVD beamformer. that is significant meaningful for diagnostic image quality in
vivo. The use of a near screen phase screen layer fails to model

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

spatially distributed aberrations which is the vast majority of retrieves only the most coherent data from different plane wave
cases in biomedical applications. Nevertheless, in the case of transmits.
reasonable medium heterogeneities such as soft tissues, the
description of a complex distributed aberrator as a large set of
phase screen aberrators corresponding to spatially distributed
isoplanatic patches permits to model aberrations more complex
than a near-field aberrator. The optimization of the number of
isoplanatic patches and their overlapping for the final image
reconstruction was not discussed in the manuscript but it is also
an interesting problem with room of further improvements. The Fig. 11: (a) Raw aberrated image of reflecting pins in speckle. (b) Corrected
size of the isoplanatic region depends on geometric parameters image after SVD-beamforming. (c) Image only rephased with the aberration
(such as depth, aperture) and aberration characteristics (such as extracted from SVD-beamforming (no filtering). (d) Difference between
images (b) and (c), normalized by image (c) intensity.
correlation length). Indeed, the more distributed the aberrator
is, the smaller the isoplanatism angle is and consequently the
smaller the isoplanatic patch should be. For smaller isoplanatic In our study, we considered a simple scattering regime for
patches, the standard deviation in the aberration estimation the backscattered signals. This hypothesis is sufficient in most
increases as less pixels are available to stabilize the SVD configurations of biomedical ultrasound. Nevertheless, the
processing. Though, we observed on Fig. 8 that there is a trade- discrimination between signals resulting from simple and
off to find between the stability of the angular aberration multiple scattering could be performed at a non-negligible
correction and the size of the patch chosen for SVD to perform supplemental computational cost before our SVD beamforming
an efficient correction. To date, we decided to choose a typical method by studying the propagation matrix singular vectors
20λ x 40λ patch size for the SVD beamforming of data with our [35].
In the case of a non-aberrated medium, for a correct image,
silicone aberrating lens and its geometry. For the in vivo liver
the first angular eigen vector out of SVD exhibit obviously a
experiment, we chose a 10λ x 70λ patch size. The determination
constant phase at 0. Though, if there is a mismatch between the
of the ideal patch size highly depends on experimental
beamforming sound speed and the medium sound speed, the
conditions and could be refined in further works. A 50%
phase immediately shows a hyperbolic trend – either convex or
overlapping between patches was chosen and the final image
concave depending on the sign of the error. So, the phase of the
was reconstructed by choosing for each pixel the mean value of eigen value straightforwardly informs on the sound speed
the same pixel in the several overlapping patches. This enables accuracy in the patch of interest. The framework of this SVD
to avoid spatial discontinuities at the frontier between patches.
beamformer processing is thus particularly adapted for
Other weighting methods before summation over the different
quantitative imaging, and particularly acoustic sound speed
patches could be studied in further works in order to optimize
mapping.
the computation time.
Since it is the common configuration in clinical ultrasound
VI. CONCLUSION
applications, we used in this work a 𝜆-pitch linear array.
However, it has been shown that using 𝜆/2-pitch probes can The SVD Beamformer provides a fast and adaptive method
decrease the number of required emissions in plane wave for beamforming whilst correcting phase and amplitude
imaging, and improve the image quality by reducing grating aberrations. On a theoretical point of view, applying the SVD
lobes in the near-field of the ultrasonic probe [34]. As it would on a particularly suited ultrafast compound Matrix R gives a
also reduce the field of view for a similar number of elements, straightforward solution to the aberration correction problem.
the probe choice depends mostly on the clinical application. This technique for adaptive ultrasound imaging reunites Phase
Further work will thus be needed to fully investigate the Aberration Correction and Coherence-based imaging in a single
capability of the SVD-beamformer for different probe correction operation. Selecting the first image singular vector is
geometries such as curved or phased arrays. sufficient to maximize the angular coherence between
One important point to notice is that the SVD beamformer transmitted insonifications and enhances the image quality,
also goes beyond a phase and amplitude aberration correction. without compromising very high frame rate imaging. We
Indeed, for each patch, the final SVD image corresponding to propose for the first time a physical meaning to the
the first singular vector U1 does not correspond to the simple mathematical SVD operation in the context of ultrasound image
phase delaying and summation of the initial data. In other formation. Future work will focus on the application of this
words, the SVD processing performs a more complex technique for quantitative ultrasound imaging.
aberration correction than just a phase screen aberrator
correction. Fig. 11 illustrates this important point. It compares REFERENCES
the corrected image provided by SVD with the corrected image [1] C. Bruneel, R. Torguet, K. M. Rouvaen, E. Bridoux,
provided when applying only a phase correction delay during and B. Nongaillard, “Ultrafast echotomographic
the conventional coherent compounding process. The system using optical processing of ultrasonic signals,”
difference between both corrected images in terms of Appl. Phys. Lett., vol. 30, no. 8, pp. 371–373, 1977.
normalized amplitude (Fig. 11 (c)) is non negligible. It shows [2] S. W. Smith, H. G. Pavy, and O. T. von Ramm,
that SVD beamforming improves also the speckle grain. There “High-Speed Ultrasound Volumetric Imaging
is therefore a filtering effect of the SVD beamformer that System—Part I: Transducer Design and Beam

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

Steering,” IEEE Trans. Ultrason. Ferroelectr. Freq. Dahl, “Short-lag spatial coherence of backscattered
Control, vol. 38, no. 2, pp. 100–108, 1991. echoes: Imaging characteristics,” IEEE Trans.
[3] J. Y. Lu, “2D and 3D High Frame Rate Imaging with Ultrason. Ferroelectr. Freq. Control, vol. 58, no. 7,
Limited Diffraction Beams,” IEEE Trans. Ultrason. pp. 1377–1388, 2011.
Ferroelectr. Freq. Control, vol. 44, no. 4, pp. 839– [18] J. Dahl, “Coherence beamforming and its applications
856, 1997. to the difficult-to-image patient,” IEEE Int. Ultrason.
[4] J. Y. Lu, “Experimental study of high frame rate Symp. IUS, 2017.
imaging with limited diffraction beams,” IEEE Trans. [19] L. Nock, G. E. Trahey, and S. W. Smith, “Phase
Ultrason. Ferroelectr. Freq. Control, vol. 45, no. 1, aberration correction in medical ultrasound using
pp. 84–97, 1998. speckle brightness as a quality factor.,” J. Acoust. Soc.
[5] M. Tanter, J. Bercoff, L. Sandrin, and M. Fink, Am., vol. 85, no. 5, pp. 1819–33, 1989.
“Ultrafast Compound Imaging for 2-D Motion Vector [20] N. Riaz, S. L. Wolden, D. Y. Gelblum, and J. Eric,
Estimation: Application to Transient Elastography,” “Coherent Flow Power Doppler (CFPD): Flow
vol. 49, no. 10, pp. 1363–1374, 2002. Detection using Spatial Coherence Beamforming,”
[6] S. I. Nikolov and J. A. Jensen, “In-vivo synthetic IEEE Trans Ultrason Ferroelectr Freq Control., vol.
aperture flow imaging in medical ultrasound,” IEEE 118, no. 24, pp. 6072–6078, 2016.
Trans. Ultrason. Ferroelectr. Freq. Control, vol. 50, [21] J.-L. Gennisson et al., “Robust sound speed estimation
no. 7, pp. 848–856, 2003. for ultrasound-based hepatic steatosis assessment,”
[7] G. Montaldo, M. Tanter, J. Bercoff, N. Benech, and Phys. Med. Biol., vol. 62, no. 9, pp. 3582–3598, 2017.
M. Fink, “Coherent Plane Wave Compounding for [22] A. Aubry and A. Derode, “Detection and imaging in a
Very High Frame Rate Ultrasonography and Transient random medium: A matrix method to overcome
Elastography,” IEEE Trans. Ultrason. Ferroelectr. multiple scattering and aberration,” J. Appl. Phys., vol.
Freq. Control, vol. 56, no. 3, pp. 489–506, 2009. 106, no. 4, 2009.
[8] M. Tanter and M. Fink, “Ultrafast Imaging in [23] Y. L. Li and J. J. Dahl, “Angular coherence in
Biomedical Ultrasound,” vol. 61, no. 1, p. 10, 2014. ultrasound imaging: Theory and applications,” J.
[9] G. E. Trahey and S. W. Smith, “Properties of Acoust. Soc. Am., vol. 141, no. 3, pp. 1582–1594,
acoustical speckle in the presence of phase aberration 2017.
part I: first order statistics,” Ultrason. Imaging, vol. [24] F. Vignon, J. S. Shin, S. W. Huang, and J. L. Robert,
10, no. 1, pp. 12–28, 1988. “Adaptive ultrasound clutter rejection through spatial
[10] M. Jaeger, E. Robinson, H. G. Akarçay, and M. Frenz, eigenvector filtering,” IEEE Int. Ultrason. Symp. IUS,
“Full correction for spatially distributed speed-of- no. Table 1, pp. 31–34, 2017.
sound in echo ultrasound based on measuring [25] J.-L. Robert and M. Fink, “Green’s function
aberration delays via transmit beam steering,” Phys. estimation in speckle using the decomposition of the
Med. Biol., vol. 60, no. 11, pp. 4497–4515, 2015. time reversal operator: Application to aberration
[11] M. Jaeger, G. Held, S. Peeters, S. Preisser, M. Grünig, correction in medical imaging,” J. Acoust. Soc. Am.,
and M. Frenz, “Computed Ultrasound Tomography in vol. 123, no. 2, pp. 866–877, 2008.
Echo Mode for Imaging Speed of Sound Using Pulse- [26] C. Prada and M. Fink, “Eigenmodes of the time
Echo Sonography: Proof of Principle,” Ultrasound reversal operator: A solution to selective focusing in
Med. Biol., vol. 41, no. 1, pp. 235–250, 2015. multiple-target media,” Wave Motion, vol. 20, no. 2,
[12] M. O’Donnell and S. W. Flax, “Phase-Aberration pp. 151–163, 1994.
Correction Using Signals From Point Reflectors and [27] G. Montaldo, M. Tanter, and M. Fink, “Real time
Diffuse Scatterers: Measurements,” IEEE Trans. inverse filter focusing by iterative time reversal,” J.
Ultrason. Ferroelectr. Freq. Control, vol. 35, no. 6, Acoust. Soc. Am., vol. 112, no. 5, pp. 2446–2446,
pp. 768–774, 1988. 2013.
[13] G. Montaldo, M. Tanter, and M. Fink, “Time reversal [28] M. Tanter, J.-L. Thomas, and M. Fink, “Focusing and
of speckle noise,” Phys. Rev. Lett., vol. 106, no. 5, pp. steering through absorbing and aberrating layers:
1–4, 2011. Application to ultrasonic propagation through the
[14] B. F. Osmanski, G. Montaldo, M. Tanter, and M. Fink, skull,” J. Acoust. Soc. Am., vol. 103, no. 5, pp. 2403–
“Aberration correction by time reversal of moving 2410, 1998.
speckle noise,” IEEE Trans. Ultrason. Ferroelectr. [29] M. Tanter, J.-L. Thomas, and M. Fink, “Time reversal
Freq. Control, vol. 59, no. 7, pp. 1575–1583, 2012. and the inverse filter,” J. Acoust. Soc. Am., vol. 108,
[15] N. Q. Nguyen and R. W. Prager, “A Spatial Coherence no. 1, pp. 223–234, 2002.
Approach to Minimum Variance Beamforming for [30] M. Tanter, J.-F. Aubry, J. Gerber, J.-L. Thomas, and
Plane-Wave Compounding,” IEEE Trans. Ultrason. M. Fink, “Optimal focusing by spatio-temporal
Ferroelectr. Freq. Control, vol. 65, no. 4, pp. 522– inverse filter. I. Basic principles,” J. Acoust. Soc. Am.,
534, 2018. vol. 110, no. 1, pp. 37–47, 2002.
[16] R. Mallart and M. Fink, “The van Cittert–Zernike [31] J.-F. Aubry, M. Tanter, J. Gerber, J.-L. Thomas, and
theorem in pulse echo measurements,” J. Acoust. Soc. M. Fink, “Optimal focusing by spatio-temporal
Am., vol. 90, no. 5, pp. 2718–2727, 1991. inverse filter. II. Experiments. Application to focusing
[17] M. A. Lediju, G. E. Trahey, B. C. Byram, and J. J. through absorbing and reverberating media,” J.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMI.2020.2986830, IEEE
Transactions on Medical Imaging

Acoust. Soc. Am., vol. 110, no. 1, pp. 48–58, 2002.

[32] F. Chassat, “Theoretical evaluation of the isoplanatic
patch of an adaptive optics system working through
the atmospheric turbulence,” J. Opt., vol. 20, no. 1, pp.
13–23, 1989.
[33] C. Prada, J. Thomas, and M. Fink, “The iterative time
reversal process: Analysis of the convergence,” J.
Acoust. Soc. Am., vol. 97, no. 1, pp. 62–71, 2005.
[34] J. Jensen, M. B. Stuart, and J. A. Jensen, “Optimized
Plane Wave Imaging for Fast and High-Quality
Ultrasound Imaging,” IEEE Trans. Ultrason.
Ferroelectr. Freq. Control, vol. 63, no. 11, pp. 1922–
1934, 2016.
[35] A. Aubry and A. Derode, “Singular value distribution
of the propagation matrix in random scattering
media,” Waves in Random and Complex Media, vol.
20, no. 3, pp. 333–363, 2010.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.