Ultrasound in Med. & Biol., Vol. 44, No. 1, pp.

278–291, 2018
Copyright © 2018 World Federation for Ultrasound in Medicine & Biology. All rights reserved.
Printed in the USA. All rights reserved
0301-5629/$ - see front matter

https://doi.org/10.1016/j.ultrasmedbio.2017.09.001

● Original Contribution

SPATIAL COMPOUNDING OF 3-D FETAL BRAIN ULTRASOUND USING

PROBABILISTIC MAPS
Jorge Perez-Gonzalez,* Fernando Arámbula-Cosío, † Mario Guzmán,‡
Lisbeth Camargo,‡ Benjamin Gutierrez,§,¶ Diana Mateus,¶ Nassir Navab,¶ and
Verónica Medina-Bañuelos*
* Neuroimaging Laboratory, Electrical Engineering Department, Universidad Autónoma Metropolitana, Iztapalapa, Mexico;
†
Biomedical Imaging Laboratory, Centro de Ciencias Aplicadas y Desarrollo Tecnológico, Universidad Nacional Autónoma de
México, Mexico City, Mexico; ‡ Department of Fetal Medicine, Instituto Nacional de Perinatología, Mexico City, Mexico;
§
Department of Child and Adolescent Psychiatry, Psychosomatic and Psychotherapy, Ludwig–Maximilians–Universität, Munich,
Germany; and ¶ Chair for Computer Aided Medical Procedures, Technische Universität München, Munich, Germany
(Received 11 May 2017; revised 12 August 2017; in final form 1 September 2017)

Abstract—A new method to address the problem of shadowing in fetal brain ultrasound volumes is presented.
The proposed approach is based on the spatial composition of multiple 3-D fetal head projections using the
weighted Euclidean norm as an operator. A support vector machine, which is trained with optimal textural
features, was used to assign weighting according to the posterior probabilities of brain tissue and shadows. Both
phantom and real fetal head ultrasound volumes were compounded using previously reported operators and
compared with the proposed composition method to validate it. The quantitative evaluations revealed increases
in signal-to-noise ratio ≤35% and in contrast-to-noise ratio ≤135% using real data. Qualitative comparisons
made by obstetricians indicated that this novel method adequately recovers brain tissue and improves the visi-
bility of the main cerebral structures. This may prove useful both for fetal monitoring and in the diagnosis of
brain defects. Overall this new approach outperforms spatial composition methods previously reported.
(E-mail: vera@xanum.uam.mx) © 2018 World Federation for Ultrasound in Medicine & Biology. All rights reserved.

Key Words: Ultrasound shadows, Support vector machine, Image fusion, Speckle reduction, Fetal brain, Spatial
composition.

INTRODUCTION developmental changes during the second and third tri-

mesters of gestation; these changes include brain size, union
Ultrasound imaging has been widely applied to monitor-
of the cerebellar hemispheres, development of the corpus
ing and diagnosis of several diseases, pre-operative
callosum and increasing complexity of the cerebral cortex
modeling and computer-assisted surgery. In many cases,
(Monteagudo and Timor-Tritsch 2009). Currently, a com-
ultrasound (US) is preferred to other imaging techniques
bination of examinations by specialists and US fetal head
such as computed tomography (CT) and magnetic reso-
biometry is recommended in antenatal care to determine
nance imaging (MRI), because it gives real-time feedback;
the health of the CNS. However, accurate biometric mea-
is non–invasive, compact and portable; and is low in cost
surements are difficult to obtain, given that US images
(Contreras Ortiz et al. 2012). This work focuses on the ap-
present several limitations, such as badly defined edges,
plication of ultrasound in obstetrics, where it is used for
multiplicative speckle noise, intensity variations and acous-
estimating gestational age, fetal monitoring and diagno-
tic shadows.
sis of a number of conditions, for example, placental
One of the significant problems in the acquisition of
insufficiency and congenital brain defects.
fetal head US images is acoustic occlusion of the ultra-
In this context, the fetal central nervous system (CNS)
sound waves (Timor-Trisch et al. 2004). A suitable
is one of the structures that undergoes the most important
extraction of the fetal brain volumes depends on acous-
tic beam penetration and tissue impedance. Most of the
Address correspondence to: Verónica Medina-Bañuelos, brain tissue has similar acoustic properties and imped-
Neuroimaging Laboratory, Electrical Engineering Department, Universidad
Autónoma Metropolitana, Iztapalapa, Ciudad de Mexico C.P. 09340, Mexico. ance, but the fetal skull has higher impedance compared
E-mail: vera@xanum.uam.mx with soft tissue (Culjat et al. 2010). This phenomenon

278
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al.
causes the images to contain shadows and artifacts,
especially in the second and third trimesters, as the cranial
bones thicken and calcify during the gestation process.
These changes, in turn, affect the appropriate identifica-
tion of several fetal structures and the correct measurement
of development indicators such as biparietal diameter, head
circumference and cerebellar volume (Jardim and
Figueiredo 2005).
To attenuate the ultrasound shadows and artifacts, this
study focuses on the improvement of fetal head US images
during the second and third trimesters of gestation. At this
age, cranial calcification leads to partial acoustic occlu-
sions, which make difficult the precise extraction of brain
indicators and, consequently, may result in an inaccurate
assessment of fetal health. A new spatial composition
method is proposed that extracts information from several
US fetal brain volumes. It allows characterization of the
acoustic parameters of brain tissue in areas where cranial
occlusion is present. This information may then be used
to mitigate the effect of occlusion on image quality. This
approach is based on the automatic classification and proba-
bilistic estimation of brain tissue and shadow areas from
multiple US projections.
Related work
Numerous physical or computational approaches have
been proposed to improve the quality of US acquisi-
tions. The focus of this study is on image compounding
methods, which combine information of different acqui-
sitions to improve image quality and reduce noise.
There are two types of compounding methods: fre-
quency and spatially based. Frequency compounding
combines multiple images of the same object acquired at
different frequencies, which has been reported to improve
the signal-to-noise ratio (SNR), but affects axial resolu-
tion (Erez et al. 2008; Perperidis et al. 2015). In contrast,
spatial compounding fuses several US images of the same
region taken from different angles or transducer posi-
tions with the purpose of building a new image with
enhanced contours, better SNR and attenuated artifacts,
without negative effects on resolution. According to
Contreras Ortiz et al. (2012), if N uncorrelated or partial-
ly correlated volumes are used in the composition, the
reduction in speckle is of the order of N using coher-
ent averaging.
Also, several authors have carried out spatial com-
pounding using the mean operator and have reported SNR
and contrast-to-noise ratio (CNR) improvements in US
volumes of the female breast (Krücker et al. 2000), gall
bladder of a healthy human (Rohling et al. 1997), athero-
sclerotic plaques (Jespersen et al. 2000; Kofoed et al. 2001)
and bovine muscle and epidural space with warping (Groves
and Rohling 2004; Tran et al. 2007). In another approach
279
reported by Vogt and Ermert (2008), a combination of high-
frequency ultrasound with a mechanical scan and multi-
angle compounding for skin image acquisition is used.
Wilhjelm et al. (2004) performed an extensive visual
and qualitative evaluation of composition operators based
on mean, geometric and logarithmic mean, median,
maximum values and root mean square; they reported that
the mean is the operator that provided better results re-
garding SNR and CNR improvement when the number of
projections used in the composition increases. Adam et al.
(2006) found that the application of non-linear filters, such
as adaptive anisotropic diffusion or Gaussian kernels, after
the composition process by coherent averaging, signifi-
cantly reduces the speckle noise; however, this can also
attenuate edges.
Grau and Noble (2005) proposed a spatial compo-
sition method using multiscale information and phase-
based image analysis for US heart compounding, achieving
speckle noise reduction without a decrease in contrast on
the main heart features. Rajpoot et al. (2011) presented a
cardiac volume compounding strategy based on a wavelet-
fusion approach that improves SNR and CNR. Gooding
et al. (2010) investigated the benefits of fusing several 4-D
heart US acquisitions, using different operators such as
the mean, median, maximum values, wavelet transform and
mean shift algorithms. They obtained noise reductions and
contrast increases up to 50% compared with a single scan.
Perperidis et al. (2016) employed an anthropomorphic left
ventricle phantom to examine the interslice angular dis-
placement and 3-D sector angular range on the elevational
spatial compounding using the mean operator in cardiac
ultrasound data.
Another compounding operator used is the weighted
mean. zu Berge et al. (2014) published results of an
orientation-aware compounding using confidence maps
(Karamalis et al. 2012), which can compensate uncertain-
ty in attenuated regions by estimating a per-voxel
confidence in the information depicted by US volumes fol-
lowing a random walk framework. These maps can be used
as weights to compound several US images. Preliminary
results have been reported on the weighted mean compo-
sition of multiple 2-D simulated brain images contaminated
by artificial multiplicative noise (Perez-Gonzalez et al.
2015).
These previous works focus mainly on enhancing
image quality from multiple projections; however, the spe-
cific goal of recovering or estimating missing information
in occluded areas is only partially solved. This work rep-
resents a new approach that has as its main aim dealing
with the problem of acoustic occlusions in US volumes.
The proposed method is based on the spatial composi-
tion of multiple projections accompanied by a suitable
estimation of the posterior probability of tissue and shadows
or artifacts. This method can be useful in building a new
280 Ultrasound in Medicine & Biology
US fetal brain volume reducing occlusion artifacts, which
is important in the assessment of fetal health.
METHODS
The problem of fusing multiple 3-D ultrasound scans
taken from different scanning positions into a single volume
is addressed. An overview of the proposed methodology
is illustrated in Figure 1. First, the ultrasound scans are
registered and re-sliced to a global coordinate system
(Fig. 1B). This step can be done by rigid registration of
landmark points. Once the scans are registered, the goal
is to compute for each voxel in the volume the intensity
that best represents the underlying anatomy. The main chal-
lenge here is inherent in ultrasound imaging because the
same anatomical location is represented by very differ-
ent intensity values when imaged from different views. This
is due mainly to the differing attenuation paths, but also
to shadow artifacts. The fusion is formulated in terms of
probabilistic spatial compounding (Fig. 1F), where the
effect of each input scan is modulated by a probabilistic
weight. In this context, the purpose is to learn the esti-
mation of the weights in a supervised manner from an
annotated data set. Although this is a regression problem,
it is difficult to give an a priori probability value to each
of the input values to be fused. Instead, a simpler binary
Fig. 1. Methodology overview. CNR = contrast-to-noise ratio; GLCM = gray-level co-occurrence matrix; PCA = principal com-
ponent analysis; SNR = signal-to-noise ratio; SVM = support vector machine.
Volume 44, Number 1, 2018
classification problem is targeted to discriminate infor-
mative versus uninformative (e.g., artifact) input voxels.
Effectively, the learning of the weight estimation is achieved
through feature extraction (Fig. 1C), feature selection
(Fig. 1D), and computation of the probabilities from the
results of the binary classifier (Fig. 1C). The details of the
steps are described in the following subsections.
Probabilistic spatial compounding
A set of aligned US volumes is designated as
Vn ( x, y, z ), with n = 1, 2, … , N being the number of
studies in three dimensions and with voxel coordinates (x,
y, z). The information from each acquisition can be com-
bined using different composition operators; in this work,
the Euclidean norm operator of the set of volumes Vn :  3
weighted by Ψ n ∈  3 → [0, 1] is used. It is defined as:
N
V ( x, y, z ) = ∑Ψ n ( x, y, z ) ⋅ Vn ( x, y, z ) 2 (1)
n =1
where V(x, y, z) is a new volume composed of N US
volumes, and Ψ n ( x, y, z ) are the weighting factors per
voxel, which aim to give more prevalence to regions with
fetal brain tissue and lower weight to areas with occlu-
sion acoustic artifacts (Fig. 1F). Therefore, the question
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al.
is how to assign weighting factors. In this context, this
method proposes learning the weights in a supervised
manner from an annotated data set of two classes: brain
tissue and shadowed regions. Feature extraction and se-
lection, together with the computation of the posterior
probabilities from the results of the classifier, were used
to estimate the weighting factors. This methodology is de-
scribed below.
Feature extraction. To classify brain tissue regions
and areas with shadows, a set of spatial histogram and
texture features were computed. Several authors have used
texture analysis in US images for the detection of symp-
tomatic carotid atherosclerosis (Afonso et al. 2015) and
breast tumor classification (Abdel-Nasser et al. 2017). In
this work, the intensity and 3-D spatial histogram fea-
tures (variance, rank and median gray values) were
calculated, which can provide a description of intensity
value variation on a US volume. To compute these filters,
a 9 × 9 × 9 neighborhood around the corresponding voxel
was used (Gonzalez and Woods 2002).
Additionally, to examine texture and spatial relation-
ships between voxels, the gray-level co-occurrence matrix
(GLCM) was computed. The GLCM characterizes the
texture of a volume by calculating how often pairs of voxels
with specific gray values (α, β) and a certain spatial re-
lationship occur in a volume (Haralick et al. 1973). It can
be defined as:
M N O samples; in this context, w is a vector of weights orthogo-
QD (α , β ) = ∑ ∑ ∑ C (α , β , x, y, z, D )
x =1 y =1 z =1
where
C (α , β , x, y, z, D )
⎧1, if F ( x, y, z ) = α and
⎪
=⎨ F ( x + Δx, y + Δy, z + Δz ) = β
⎪0,
⎩ otherwise
Thus, the co-occurrence probability is
PD = QD ∑ QD , where D = ( Δx, Δy, Δz ) is the displace-
ment and contains parameters of distance d and angles θ,
ϕ. By use of eqns (2) and (3) several texture functions F
(energy, entropy, homogeneity and contrast) were calcu-
lated (Wang 2015). A total of 260 texture maps were
computed for distances d = (1, 2, 3, 4, 5) , and 13 direc-
tions were defined by the azimuth and zenith angles
(θ , φ ) = (0°, 45°, 90°, 135°) in the vertical and horizontal
orientations, as described in Chen et al. (2007). These 260
texture maps together with intensity, variance, rank and
median gray values constitute the final 264-feature set
(Fig. 1C).
281
Finally, a principal component analysis (PCA) was
implemented to select the most significant features for the
classification process and to reduce the dimensionality of
the input space (264 initially) (Hotelling 1933; Pearson
1901). The number of features can be reduced by select-
ing the PCA coefficients with the higher variance. These
components were used to classify fetal brain tissue and
areas with shadows (Fig. 1D).
Learning the probabilistic weights. To select these
weighting factors we use support vector machines (SVMs),
which are employed in a great number of US applica-
tions because of their robustness and ability to learn from
experimental data, such as breast masses (Rodrigues et al.
2015) and lymph node segmentation (Chmielewski and
Dufort 2015), and to assess the malignancy risk of thyroid
nodules (Tsantis et al. 2005).
An SVM can separate a given set of binary labeled
( yi ∈ {+1, − 1} ) training data ( xi ∈ n , i = 1, 2, … , q ), by
means of a feature–space hyperplane that is maximally
distant from the training samples of two classes: voxels
containing brain tissue information (labeled with y = +1)
and voxels with acoustic shadows (labeled y = −1). The
SVM is fed by extracted features (described in previous
subsection) and provides as output the estimated poste-
rior probabilities of each class, which will be used as
weighting factors.
The task is to build a linear discriminant function
f ( x ) = w ⋅ F ( x ) + w0 that correctly classifies new
(2) nal to the decision plane, w0 is the threshold, “·” denotes
the inner product and Φ(x) is a non-linear transforma-
tion that maps the input data into a new feature space
(Vapnik 2000). Therefore, the decision function sign
{ f (x )} can be used to predict the label {+1, −1} of new
given data.
In this work, instead of label predictions, we pro-
(3) posed estimating the posterior probability that a new datum
x belongs to the class “brain tissue” P ( y = +1 x ), accord-
ing to Platt’s (1999) approximation by a sigmoid function
given by:
1
P ( y = +1 x ) ≈ PA,B ( f ) = (4)
1 + exp ( Af + B )
where each fi is an estimate of f(xi), and the best param-
eters A and B are computed in a discriminative form by
maximizing the likelihood. Hence, the weighting factors
“ 0 ≤ Ψ n ( x, y, z ) ≤ 1” take the posterior probabilities
values “ P ( y = +1 x )” corresponding to each new datum
(Fig. 1E).
For SVM, a radial basis function (RBF) kernel was
used, defined as K ( x i, x ) = e − x − xi 2σ , where σ controls
2 2
the function’s width (Chang and Lin 2011). The parameters
282 Ultrasound in Medicine & Biology
Fig. 2. Truncated cone phantom in axial view. (a) Ultrasound acquisition without occlusions and (b) ultrasound acquisition
obtained by positioning a plastic piece between the transducer and the phantom to generate acoustic shadows. (c) Phantom
designed in lateral and axial views.
σ and C can be optimized by gradually modifying their
values, filling a grid and then choosing the combination
that maximizes accuracy.
Depending on the structures of interest, the training
data set can be selected from a single US volume by choos-
ing representative regions of brain tissue (midline, head
contour, cerebellum, frontal lobe, etc.) with y = +1 and
shadows or artifacts with y = −1. The remaining acquisi-
tions can be used to classify and to estimate the probabilistic
maps (weighting factors). Thus, new volumes from dif-
ferent transducer positions or subjects do not require re-
training the SVM.
Data set and validation
To test the capability of the proposed algorithm to
recover brain information, an US phantom was used.
Several volumes taken from different transducer angles
were generated and processed. To validate this experi-
ment, the SNR, CNR and correlation metrics between a
composed volume and the baseline were computed. Af-
terward, a set of fetal brain US volumes (acquired from
axial and coronal projections) was processed; the aim was
to build a new fetal brain volume without occlusion arti-
facts. To analyze these results a quantitative validation
(using SNR and CNR) and a qualitative evaluation by en-
gineering and clinical experts were carried out. Additionally,
Volume 44, Number 1, 2018
to contrast probabilistic spatial compounding (PSC) results
using phantom and real data sets, some of the most fre-
quently used composition operators were tested. The details
of each data set and their validation are explained in the
following subsections.
Phantom data set. To validate the implemented al-
gorithm, a 3-D truncated conical phantom inside a box was
built (Fig. 2C). The phantom was prepared with 6 wt%
gelatin gel and 4 wt% agar, which is the standard mate-
rial used in US phantom applications. In contrast, the
surrounding tissue-mimicking gel was made with 4 wt%
gelatin and 1.5 wt% agar to ensure good distinction from
the conical shape in US (Dang et al. 2011). We used this
phantom to perform two series of acquisitions. The first
consisted on a set of six different volumes acquired from
several transducer positions (four orthogonal and two
oblique views); Figure 2A is an example image in axial
view. For the second series, six volumes were obtained
at the same previous positions, but this time an acoustic
occlusion was artificially created by placing a piece of
plastic between the transducer and the phantom; Figure 2B
is a representative image of an US with such an occlu-
sion. The occlusion covered 37.5% of the total phantom
volume. All acquisitions were performed in B-mode using
an Ultrasonix Sonix RP system with a motorized curvilinear
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al. 283

Fig. 3. Ultrasound fetal brain examples in axial transcerebellar plane. (a) An example with the following structures marked:
head contour (HC), midline (MID), cerebellum (CER), cavum septum pellucidum (CSP), peduncles (PC), cisterna magna (CM)
and nuchal fold (NF). (b) Fetal image with several shadows.

transducer m4DC7–3/40, with a frequency of 3.3 MHz and 1 N

isotropic resolution of 0.31 mm3. US phantom volumes
mean = ∑ Vn ( x, y, z )
N n=1
(5)
were co-registered based on a robotic system, as previ-
∑ τ ( x, y, z )Vn ( x, y, z )
ously described by Zettinig et al. (2016). N
n =1 n
weighted mean = (6)
∑ τ ( x, y, z )
N
Fetal brain US data. The data set consists of 10 3-D n =1 n
US fetal brain volumes taken between 20 and 24 gesta-
where Vn ( x, y, z ) is a set of volumes, N is the number of
tional weeks, in B-mode, using a motorized curvilinear
volumes used in the composition and τ n ( x, y, z ) are the
transducer with a frequency of 8–20 MHz and isotropic
weighting factors derived from the confidence map ap-
resolution of 0.2 to 0.5 mm3 depending on the patient. All
proach (zu Berge et al. 2014).
volumes were acquired by obstetric experts with General
On the other hand, the root mean square (RMS) op-
Electric Voluson E8 equipment. All subjects gave their
erator emphasizes high-intensity values by considering the
written informed consent, and all studies were approved
quadratic mean of several US volumes, and the maximum
by the Fetal Medicine Department at the National Insti-
values (MAX) preserves the highest-intensity values, thus
tute of Perinatal Medicine and agreed with the declaration
minimizing the shadowed regions; nevertheless, it does not
of Helsinki. Monteagudo and Timor-Tritsch (2009) re-
preserve edges and can alter the visibility of several struc-
ported that different orientations can provide
tures. These operators are defined as:
complementary information on several fetal brain struc-
tures. For this reason, images of each patient were acquired
1 N
from an axial and coronal plane, which were registered RMS = ∑ Vn ( x, y, z )
N n=1
(7)
manually by an obstetrician expert considering several
common fiduciary points in the neurofetal assessment
MAX = max {Vn ( x, y, z )} (8)
(Fig. 3A). An example of a fetal brain volume with shadows
is provided in Figure 3B. Alternatively, other operators that could be used are
the minimal values and the geometric mean; however, the
Compounding method performance comparison former takes intensity values that correspond mainly to oc-
The proposed PSC methodology was compared with cluded areas, whereas the latter can involve products
the mean, weighted mean, maximum values and root mean between voxels with missing data (shadows) and results
square operators, because Wilhjelm et al. (2004) re- in a composite volume with poor information and artifacts.
ported that these metrics can give acceptable results in the
composition process. Two of them, the mean and weighted Quantitative validation
mean operators, can improve considerably the SNR and To measure the quality of compounded volumes, SNR
CNR with respect to a regular US volume (Behar et al. and CNR were calculated (Fig. 1G). These indices are
2003; Li et al. 1994) and can be used to reconstruct a new defined as:
volume, containing the information of each of the pro-
jections included in the composition process. These metrics μVOI
SNR = (9)
are defined as: σ VOI
284 Ultrasound in Medicine & Biology
μVOI − μ B
CNR = (10)
σ VOI
2
+ σ B2
where µ is the mean, σ is the standard deviation, B is the
background and VOI is a volume of interest. These metrics
provide relevant information on multiplicative noise at-
tenuation capability on compounded volumes. In this work,
they are proposed as quality indices to measure PSC per-
formance in comparison to the mean (this method will be
subsequently referred as MEAN), Weighted Mean by Con-
fidence Map (WMCM), MAX and RMS methods.
The composition process was applied to a set of six
phantom volumes containing occlusion artifacts (Fig. 2B);
to measure SNR and CNR, a set of 3-D windows as VOIs,
covering 40% of the total volume of the conical phantom,
were selected.
For real fetal US data, 20 VOIs were selected with
15 × 15 × 15 windows containing relevant brain struc-
tures, such as the midline, peduncles, ventricles, frontal
lobe and head contour, which represent 65% of the total
fetal head volume.
On the other hand, to estimate the ability to recover
occluded information, the correlation coefficient (CC) and
normalized mutual information (NMI) were computed.
These measures can provide information on the relationship
between two variables and take values from 0 to 1, where
1 corresponds to identical volumes and 0 to totally unrelated
volumes. In this work, the baseline consisted of a volume
composed of the mean operators from six projections
without artifacts. CC and NMI were computed between
this reference and the volumes compounded by the tested
methods (PSC, MEAN, WMCM, MAX and RMS).
Qualitative validation
Considering that for real US volumes no baseline is
available, two qualitative tests were performed (Fig. 1G).
The first survey consisted of quality evaluation of the PSC
method versus the MEAN, WMCM, MAX, RMS and
single-view US acquisition (SA), carried out on image pairs
in the axial transventricular and transcerebellar planes. Ten
experts on fetal image acquisition and processing, with
3.5 ± 1.5 years of experience, were asked to perform 50
double-blind comparisons of three qualitative aspects:
visible brain tissue, head circumference and general quality
of the images; each pair of images was rated as very bad,
bad, similar, good or excellent.
Afterward, a second double-blind test was carried out
by three obstetricians with 22, 6 and 4 years of experi-
ence. They rated the composition quality on some of the
most relevant fetal brain structures—midline, head contour,
cavum septum pellucidum, peduncles, cerebellum, thala-
mus, cisterna magna and nuchal fold—as well as visible
brain tissue and general quality. The comparison between
pairs of compounded images obtained with PSC versus
Volume 44, Number 1, 2018
MEAN, WMCM, MAX, RMS and SA was also rated as
very bad, bad, similar, good or excellent for every struc-
ture evaluated. A total of 10 comparisons were performed
by each expert, in the axial transventricular and
transcerebellar planes, previously selected by a different
obstetrician expert.
RESULTS AND DISCUSSION
To determine the advantages of the proposed com-
pounding algorithm with probabilistic weights, a series of
experiments in which PSC was compared with other com-
pounding approaches described are presented. First, the
results of the composition using the phantom data set are
shown. In this experiment, the resulting volume of each
compounding method was contrasted against a baseline,
when the number of volumes consider in the composi-
tion was increased from 1 to 6. Quantitative and qualitative
evaluations illustrating that the PSC method is able to
reduce the effect of acoustic shadows in a compounded
volume are also presented.
Phantom US compounding
For the first series of experiments, the phantom data
previously described were evaluated with the purpose of
analyzing the behavior of the proposed algorithm in a con-
trolled setting where the ground truth is known. The aim
was to evaluate the ability of the classifier to estimate the
posterior probabilities of a voxel belonging to either tissue
or shadow. To do so, the ultrasound volumes obtained with
occlusions were used and a sixfold cross validation was
carried out. The area under the receiver operating char-
acteristic (ROC) curve (AUC) of posterior probabilities
was calculated (88.6 ± 2.4%); the results indicated that the
SVM classifier is able to accurately distinguish between
shadow and tissue regions, and therefore, its posterior prob-
abilities are reliable weights to be used by the composition
method.
As described in the validation section, the proposed
composition method was compared with other algo-
rithms. Results of the phantom data composition using
MEAN, WMCM, MAX, RMS and PSC operators are il-
lustrated in Figure 4 in an axial plane; for each case, six
volumes with shadows were used. Figure 4A illustrates
the baseline corresponding to the composition of six images
without occluded artifacts using the mean as operator. It
can be noted that the PSC approach (Fig. 4F) is visually
similar to the baseline and substantially better than the
MEAN, WMCM and RMS methods (a circular shape
should be observed). In contrast, the MAX approach
(Fig. 4D) recovers missing information, but affects the
edges of the circular shape on the phantom’s axial view.
A quantitative evaluation was also carried out and is
illustrated in Figure 5. It consisted of measuring the
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al. 285

Fig. 4. Truncated cone phantom compounding results in axial view (a circular shape should be observed). (a) Baseline. (b–f)
Compounding using the (b) mean, (c) weighted mean by confidence map, (d) maximum values, (e) root mean square and (f)
probabilistic spatial compounding.

operators’ performance when the number of acquisitions
with shadows is increased from 1 to 6 in the composi-
tion process. In the top row of Figure 5A and B, the
behavior of SNR and CNR is illustrated. It can be seen
that as the number of volumes used for the composition
increases, SNR and CNR improve considerably, which is
consistent with previous work (Jespersen et al. 2000). An
exception is when the fourth volume is added, because it
has a higher content of occlusion artifacts, thus provok-
ing a decrease in SNR and/or CNR; this behavior is
observed only for the MEAN, WMCM, MAX and RMS
methods. For PSC, these parameters remain stable, pos-
sibly because of the method’s ability to select and weight
information in the composition process.
On the other hand, an increase in SNR and CNR can
be observed (Fig. 5) when six volumes are considered for
the composition. the PSC method presents considerable
improvements compared with the MEAN (49.6% and
145.9%), WMCM (26.1% and 113.4%), MAX (9.1% and
489.7%) and RMS (21% and 272.1%) methods. The MAX
and RMS methods exhibit good performance on the SNR,
but affect image contrast, and the MEAN and WMCM
methods present a better balance between SNR and
CNR; however, the phantom shape is altered by occlu-
sion artifacts.
When these indices were computed for the baseline
image, an SNR of 6.32 and a CNR of 4.54 were ob-
tained, which represent increments of 16.7% and 92.8%
286 Ultrasound in Medicine & Biology Volume 44, Number 1, 2018

Fig. 5. Quantitative assessment of phantom compounding using the MEAN, weighted mean by confidence map (WMCM),
maximum values (MAX), root mean square (RMS) and probabilistic spatial compounding (PSC) approaches. The horizontal
axis indicates the number of volumes used to generate the compounded volume. (a) Signal-to-noise ratio (SNR). (b) Contrast-
to-noise ratio (CNR). (c) Cross correlation (CC). (d) Normalized mutual information (NMI).

respectively, compared with the values obtained for the
proposed PSC method; it must be recalled that the base-
line image was composed of non-occluded projections.
Also, the behavior of CC and NMI indices was ob-
served for MEAN, WMCM, MAX, RMS and PSC
composition results, compared with the baseline image
(Fig. 5C, D). It can be observed that all five methods pro-
vided substantially better results when the number of
volumes considered for the composition was increased.
However, the best performance was obtained using PSC
with CC = 0.63 and NMI = 0.42. These values represent
enhancements of 20% and 28.3%, respectively, with respect
to the MEAN method, and enhancements of 15.8% and
17.9% compared with the WMCM method. In contrast,
the MAX and RMS approaches obtained the worst results
as revealed by CC and NMI metrics. These two param-
eters indicate the similarity of information contained in
two different volumes, and therefore, they provide an es-
timation of the capacity of each method to recover
information preserving the phantom shape from differ-
ent projections, compared with the baseline image. In this
sense, the results indicate that the proposed PSC method
perform better than the MEAN, WMCM, MAX and RMS
methods. However, it must be remarked that all methods
are dependent on the quantity of information provided by
each volume used in the composition process. If they
contain minimal information or are completely oc-
cluded, the resulting composed image will be very poor.
Fetal brain US compounding
The second series of experiments consisted of quan-
titative and qualitative evaluations on the fetal brain US
volumes composed using the PSC algorithm and other
methods. The first test consisted of evaluating the
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al. 287

Fig. 6. Ultrasound fetal brain results in axial transcerebellar plane. (a) Single ultrasound acquisition. (b–f) Compounding using
the (b) mean, (c) weighted mean by confidence map, (d) maximum values, (e) root mean square and (f) probabilistic spatial
compounding.

compounding volume quality measured by the SNR and method (CNR = 12.13). In the second column of Table 1
CNR indices. Afterward, two qualitative tests were carried it can be seen that the proposed methodology yields the
out: general evaluation by acquisition and processing US best result (with increases of 3.5% to 135%), withstatistical
image experts, and fetal brain structure evaluation by ob- differences versus SA and all the other methods tested.
stetrician experts. These results reveal that the best balance of SNR and CNR
In Figure 6 is an example of compounding results with corresponds to PSC followed by the WMCM and MEAN
two projections using the MEAN, WMCM, MAX, RMS, methods. This suggests that the proposed approach con-
PSC and SA. It can be appreciated that in all cases the siderably improves the quality of US volumes by enhancing
composition process helps to recover occluded areas com- information, reducing multiplicative noise and mitigat-
pared with SA. Such is the case for the MAX method ing shadowed regions.
(Fig. 6D), where more brain tissue can be observed, but To evaluate the overall quality of a new compounded
with a granulated aspect and multiplicative noise, espe- volume a qualitative assessment was carried out by experts
cially in the cerebellum and peduncle regions. On the other in US fetal image acquisition and processing. Table 2 out-
hand, in Figure 6F, the results obtained using PSC illus- lines the results of three evaluated parameters: head contour
trate that this method achieves a reduction of shadowed definition, visible brain tissue and general quality. For each
regions and improves image quality without altering the item a comparison between the PSC method and all other
morphology of the fetal brain structures.
In Table 1, a SNR and CNR evaluation of single ac-
quisition and compounding approaches is outlined. It can Table 1. Quantitative evaluation (µ ± σ) of
be seen that all methodologies exhibit considerably better compounding methods in fetal volumes*
results compared with SA. This reveals that the compo- Method SNR CNR
sition of several acquisitions can improve US image quality,
SA 7.83 ± 2.8†
5.15 ± 1.9‡
which agrees with previously reported research (Wilhjelm MEAN 10.27 ± 1.2† 11.44 ± 2.7‡
et al. 2004). On the other hand, to measure statistical dif- WMCM 10.34 ± 1.9 11.71 ± 1.8‡
ferences, a Friedman test, followed by a Wilcoxon test with MAX 10.55 ± 3.6 7.34 ± 3.9‡
RMS 10.33 ± 2.6 10.41 ± 2.5‡
Bonferroni post hoc correction, was carried out. In this PSC 10.53 ± 2.7 12.13 ± 2.3
analysis only the significant differences between PSC and
all other results are presented. With respect to the SNR, CNR = contrast-to-noise ratio; MAX = maximum values; PSC = proba-
bilistic spatial compounding; RMS = root mean square; SA = single-view
PSC exhibits statistical differences in comparison to the acquisition; SNR = signal-to-noise ratio; WMCM = weighted mean by con-
SA and MEAN methods and a slight increase compared fidence map.
with the WMCM and RMS approaches. However, MAX * Comparisons using the Friedman test and Wilcoxon test with Bonferroni
correction statistical tests at significance level (p < 0.05).
has the best performance (SNR = 10.55), although it has †
Indicates significant difference in SNR compared with PSC.
a poorer result for CNR (7.34) compared with the PSC ‡
Indicates significant difference in CNR compared with PSC.
288 Ultrasound in Medicine & Biology Volume 44, Number 1, 2018

Table 2. General qualitative assessment of ultrasound Finally, a thorough validation was carried out for
fetal brain data* the assessment of the most important brain structures
Comparison Very bad Bad Similar Good Excellent considered in the clinical diagnosis of several patholo-
gies. For this purpose, pairs of images were compared
Head contour quality between the proposed method (PSC) and SA, MEAN,
PSC vs. SA 3 10 13 33 41
PSC vs. MEAN 1 17 19 38 25 WMCM, MAX and RMS results, assigning a qualitative
PSC vs. WMCM 4 10 22 37 27 score: very bad, bad, similar, good or excellent. Previ-
PSC vs. MAX 2 8 21 40 29 ously, an expert obstetrician (who did not participate in
PSC vs. RMS 3 16 19 36 26
Visible brain tissue the evaluation) selected the images presented for com-
PSC vs. SA 4 7 20 33 36 parison, in an axial view of transcerebellar and
PSC vs. MEAN 2 12 12 44 30 transventricular planes. Figure 7 summarizes the evalua-
PSC vs. WMCM 0 11 23 36 30
PSC vs. MAX 15 20 24 26 15 tions carried out by three obstetricians on brain structures
PSC vs. RMS 1 12 27 35 25 and common aspects for the images presented on a given
General quality of the volume pair. In Figure 7A, the overall comparison of US volumes’
PSC vs. SA 1 17 12 32 38
PSC vs. MEAN 8 17 12 37 26 general quality can be observed, where the good catego-
PSC vs. WMCM 6 19 13 39 23 ry votes vary between 50% and 83%. These results concur
PSC vs. MAX 0 5 12 38 45 with those obtained for the SNR and CNR evaluations
PSC vs. RMS 10 14 12 33 31
and with the other qualitative evaluation made by experts
CNR = contrast-to-noise ratio; MAX = maximum values; PSC = proba- on US fetal image acquisition and processing discussed
bilistic spatial compounding; RMS = root mean square; SA = single-view
acquisition; SNR = signal-to-noise ratio; WMCM = weighted mean by con-
in the previous section.
fidence map. Head contour is one of the most important anatom-
Values in bold represent the highest number of votes for each comparison. ical structures in clinical practice, because it serves as the
* Fetal image processing experts’ comparisons of three different aspects;
all comparisons were carried out between the PSC and the SA, MEAN,
basis for estimation of several other biomarkers, such as
WMCM, MAX or RMS approach. Values correspond to the number of votes biparietal and fronto-occipital diameters, cranial circum-
assigned to one of the five categories (columns 2–6). N = 100. ference and gestational age (Monteagudo and Timor-Tritsch
2009). The percentages of votes obtained when compar-
approaches, as well as an SA, was performed. It can be ing head contours in all images pairs are illustrated in
seen that in comparison to SA, PSC received most of the Figure 7B. Note that PSC is clearly superior to SA, with
votes in the excellent category with 41% for head contour most of the votes in the good and excellent categories.
quality, 36% for visible brain tissue and 38% for general Because the head is one of the structures that is less af-
quality; this is evidence that the composition using proba- fected by acoustic occlusion (in fact, shadows are generated
bilistic maps can improve image aspects such as quality mainly by the skull itself), several composition scores are
and capability to recover missing information, con- similar, which is reflected in evaluations comparing PSC
trasted with a single US volume. Regarding head contour with all other methods. Specifically, some enhancement
quality, PSC is mainly in the good category (between 36% was observed in images composed using WMCM; this can
and 40%). In the case of visible brain tissue, it can be ap- be explained by the intensity homogenization of regions
preciated that the majority of the votes were assigned to around the fetal skull, caused by the use of confidence
the good category when comparing PSC with all other com- maps, inherent to this method.
pounding methods. Nevertheless, if we focus only on the Other important structures used in fetal brain assess-
comparison between PCS and MAX, similar numbers of ment are the midline and CSP; these serve as elementary
votes were obtained in all categories; this is due to the fact references for plane realignment or detection of brain dis-
that MAX always provides composed volumes with orders associated with midline deviation, among others
maximal information, but, as mentioned before, it alters (Monteagudo and Timor-Tritsch 2009). In Figure 7C and
edges, has poor contrast and can modify the morphology D are the results obtained when comparing these struc-
of several fetal brain structures. Finally, for general quality tures; it can be observed that PSC was evaluated as better
evaluation, the PSC comparisons present positive results than the other methods, with 50%–83% of votes in the good
with scores between 33% and 39% in the good category, category, for both midline and CSP.
when contrasting PSC with MEAN, WMCM and RMS; An important aspect of assessment in this work was
in the case of PSC versus MAX, a total of 45% votes in the overall quality of visible brain tissue after the PSC pro-
the excellent category were obtained. These results reaf- cedure was applied for image compounding. The results
firm that the proposed method has a better capability to obtained are illustrated in Figure 7E, where it can be seen
improve quality, and at the same time it minimizes oc- that PSC performs better than SA, MEAN and RMS, with
cluded areas by recovering information from the volumes most of the votes in the good and excellent categories; the
used for the composition. comparison of PSC with WMCM yields 66% good votes.
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al. 289

Fig. 7. Evaluation of the main fetal brain structures and quality aspects in ultrasound volumes. Three obstetrician experts com-
pared five different parameters (a–e) of the results obtained with probabilistic spatial compounding (PSC) versus the single-
view acquisition (SA), MEAN, weighted mean by confidence map (WMCM), maximal values (MAX) and root mean square
(RMS) approaches. In each panel are 100% stacked bar charts of the votes obtained for each comparison performed.

On the contrary, results obtained between PSC and MAX
are more balanced, with 50% of the votes falling in the
similar category. This is in agreement with results re-
ported in the previous sections, where it was remarked that
composition using MAX always provides volumes with
higher-intensity values, thus mitigating shadowed regions,
but also altering borders and structure shape and reduc-
ing CNR.
Other structures analyzed were the thalamus in the
transventricular axial plane and the peduncles, cerebel-
lum, nuchal fold and cisterna magna in the transcerebellar
axial plane. Results are similar to those obtained for other
structures, with 66% good votes when comparing PSC with
SA, MEAN, WMCM, MAX and RMS. The cerebellum
was one of the best qualified structures with all votes falling
in the good and excellent categories, whereas the tha-
lamic region received 46% of similar and bad scores and
only 54% of good votes.
These comparisons reveal that PSC composition
allows a general enhancement of volume quality and, more
specifically, of those brain structures most frequently used
for fetal brain evaluation; this can lead to better quanti-
fication and segmentation processes and, therefore, to
adequate clinical diagnostics.
290 Ultrasound in Medicine & Biology
CONCLUSIONS
A new method for spatial compounding using pos-
terior probability map estimation was described. The main
goal was to compose a US volume of the fetal brain from
several projections affected by acoustic occlusions gen-
erated by fetal skull calcification. These projections are
contaminated by shadows that can make adequate mea-
surement and clinical evaluation of important CNS
structures difficult. The final composed US volume con-
tains as much useful information as possible, thus mitigating
the shadowed regions.
Comparisons with other reported methods indicate
that PSC exhibits the best combination of SNR and CNR,
followed by the WMCM and MEAN approaches. Unlike
these methods, PSC focuses specifically on mitigating
acoustic occlusion in US volumes and was found to perform
better on phantom data (improving SNR up to 49.6% and
CNR up to 489.7%) as well as in fetal brain volumes
(increments of SNR and CNR of 35% and 135%, respec-
tively). Qualitative validations were also favorable to the
PSC method.
It must, however, be considered that for all methods,
the composition quality and capability of shadow sup-
pression are highly dependent on the quantity of information
provided by those volumes considered in the composi-
tion process and also on the performance of volume
registration methods. Furthermore, the PSC method re-
quires the computation of texture maps, making it more
time consuming.
The proposed method can be very useful in other clin-
ical diagnostic applications where shadows can appear as
the result of structure occlusion, such as the gestational
sac, fetal heart, breast cancer and knee bone reconstruc-
tion. Depending on the application, shape, texture,
frequency features or local patterns can be added to maxi-
mize useful information and to eliminate artifacts and
shadowed regions. In summary, PSC is a robust method,
based on machine learning approaches for the probability-
weighted composition of several fetal brain projections,
proposed to mitigate acoustic shadows.
Acknowledgments—The financial support of the Science Technology
and Innovation Secretariat of Mexico City (SECITI), under Collabora-
tion Agreement 110/2015, is gratefuly acknowledged. Jorge Perez-
Gonzalez is supported by CONACYT Scholarship No. 35879.
REFERENCES
Abdel-Nasser M, Melendez J, Moreno A, Omer OA, Puig D. Breast tumor
classification in ultrasound images using texture analysis and super-
resolution methods. Eng Appl Artif Intell 2017;59:84–92.
Adam D, Beilin-Nissan S, Friedman Z, Behar V. The combined
effect of spatial compounding and nonlinear filtering on the speckle
reduction in ultrasound images. Ultrasonics 2006;44:166–
181.
Volume 44, Number 1, 2018
Afonso D, Seabra J, Pedro LM, Fernandes JF, Sanches JM. An
ultrasonographic risk score for detecting symptomatic carotid ath-
erosclerotic plaques. IEEE J Biomed Health Inform 2015;19:1505–
1513.
Behar V, Adam D, Friedman Z. A new method of spatial compound-
ing imaging. Ultrasonics 2003;41:377–384.
Chang CC, Lin CJ. LIBSVM: A library for support vector machines. ACM
Trans Intell Syst Technol 2011;2:1–27.
Chen W, Giger ML, Li H, Bick U, Newstead GM. Volumetric texture
analysis of breast lesions on contrast-enhanced magnetic resonance
images. Magn Reson Med 2007;58:562–571.
Chmielewski A, Dufort P, Scaranelo AM. A computerized system to assess
axillary lymph node malignancy from sonographic images. Ultra-
sound Med Biol 2015;41:2690–2699.
Contreras Ortiz SH, Chiu T, Fox MD. Ultrasound image enhancement:
A review. Biomed Signal Process Control 2012;7:419–428.
Culjat MO, Goldenberg D, Tewari P, Singh RS. A review of tissue sub-
stitutes for ultrasound imaging. Ultrasound Med Biol 2010;36:861–
873.
Dang J, Frisch B, Lasaygues P, Zhang D, Tavernier S, Felix N, Lecoq
P, Auffray E, Varela J, Mensah S, Wan M. Development of an an-
thropomorphic breast phantom for combined pet, B-mode ultrasound
and elastographic imaging. IEEE Trans Nucl Sci 2011;58:660–
667.
Erez Y, Schechner YY, Adam D. Space variant ultrasound frequency
compounding based on noise characteristics. Ultrasound Med Biol
2008;34:981–1000.
Gonzalez RC, Woods RE. Digital image processing. Upper Saddle River,
NJ: Prentice-Hall; 2002.
Gooding MJ, Rajpoot K, Mitchell S, Chamberlain P, Kennedy SH, Noble
JA. Investigation into the fusion of multiple 4-D fetal echocardiography
images to improve image quality. Ultrasound Med Biol 2010;36:
957–966.
Grau V, Noble JA. Adaptive multiscale ultrasound compounding using
phase information. In: Proceedings, Medical Image Computing and
Computer-Assisted Intervention 8th International Conference—
MICCAI 2005, Palm Springs, CA, USA, October 26–29, 2005, Pro-
ceedings, Part I. Berlin/Heidelberg: Springer, 2005:589–596.
Groves AR, Rohling RN. Two-dimensional spatial compounding with
warping. Ultrasound Med Biol 2004;30:929–942.
Haralick RM, Shanmugam K, Dinstein I. Textural features for image clas-
sification. IEEE Trans Syst Man Cybern 1973;SMC-3:610–621.
Hotelling H. Analysis of a complex of statistical variables into princi-
pal components. J Educ Psychol 1933;24:417441.
Jardim SM, Figueiredo MA. Segmentation of fetal ultrasound images.
Ultrasound Med Biol 2005;31:243–250.
Jespersen S, Wilhjelm J, Sillesen H. In vitro spatial compound scan-
ning for improved visualization of atherosclerosis. Ultrasound Med
Biol 2000;26:1357–1362.
Karamalis A, Wein W, Klein T, Navab N. Ultrasound confidence maps
using random walks. Med Image Anal 2012;16:1101–1112.
Kofoed SC, Grønholdt MLM, Wilhjelm JE, Bismuth J, Sillesen H. Real-
time spatial compound imaging improves reproducibility in the
evaluation of atherosclerotic carotid plaques. Ultrasound Med Biol
2001;27:1311–1317.
Krücker J, Meyer C, Carpentier GL, Fowlkes J, Carson P. 3D spatial
compounding of ultrasound images using image-based nonrigid reg-
istration. Ultrasound Med Biol 2000;26:1475–1488.
Li W, Gussenhoven EJ, Zhong Y, The SH, Pieterman H, van Urk H,
Bom K. Temporal averaging for quantification of lumen dimen-
sions in intravascular ultrasound images. Ultrasound Med Biol 1994;
20:117–122.
Monteagudo A, Timor-Tritsch IE. Normal sonographic development of
the central nervous system from the second trimester onwards using
2D, 3D and transvaginal sonography. Prenatal Diagn 2009;29:326–
339.
Pearson K. On lines and planes of closest fit to systems of points in space.
Philos Mag Ser 6 1901;2:559–572.
Perez-Gonzalez JL, Arámbula Cosío F, Medina-Bañuelos V. Spatial com-
position of US images using probabilistic weighted means. In:
Proceedings, 11th International Symposium on Medical Information
 Spatial compounding of 3-D fetal brain US ● J. Perez-Gonzalez et al.
Processing and Analysis (SIPAM 2015). Proc SPIE 9681; 2015. http://
dx.doi.org/10.1117/12.2207958.
Perperidis A, Cusack D, White A, McDicken N, MacGillivray T, Anderson
T. Temporal compounding: A novel implementation and its impact
on quality and diagnostic value in echocardiography. Ultrasound Med
Biol 2015;41:1749–1765.
Perperidis A, McDicken N, MacGillivray T, Anderson T. Elevational
spatial compounding for enhancing image quality in echocardiography.
Ultrasound 2016;24:74–85.
Platt J. Probabilistic outputs for support vector machines and compari-
sons to regularized likelihood methods: Advances in large margin
classifiers, vol. 10. Cambridge: MIT Press; 1999. p. 61–74.
Rajpoot K, Grau V, Noble JA, Szmigielski C, Becher H. Multiview fusion
3-D echocardiography: Improving the information and quality of real-
time 3-D echocardiography. Ultrasound Med Biol 2011;37:1056–
1072.
Rodrigues R, Braz R, Pereira M, Moutinho J, Pinheiro AM. A two-
step segmentation method for breast ultrasound masses based on multi-
resolution analysis. Ultrasound Med Biol 2015;41:1737–1748.
Rohling R, Gee A, Berman L. Three-dimensional spatial compounding
of ultrasound images. Med Image Anal 1997;1:177–193.
Timor-Trisch I, Monteagudo A, Cohen H. Neuroecografia Prenatal y Neo-
natal, vol. 1. Madrid: Marban Libros; 2004.
Tran D, Kamani A, Lessoway V, Rohling RN. Adaptive spatial com-
pounding for improving ultrasound images of the epidural space. In:
Emelianov SY, McAleavey SA, eds. Proceedings, Medical Imaging
291
2007: Ultrasonic Imaging and Signal Processing. Proc SPIE; 2007;
6513:65130W.
Tsantis S, Cavouras D, Kalatzis I, Piliouras N, Dimitropoulos N,
Nikiforidis G. Development of a support vector machine-based image
analysis system for assessing the thyroid nodule malignancy risk on
ultrasound. Ultrasound Med Biol 2005;31:1451–1459.
Vapnik VN. The nature of statistical learning theory, 2nd ed. New York:
Springer; 2000.
Vogt M, Ermert H. Limited-angle spatial compound imaging of skin with
high-frequency ultrasound (20 MHz). IEEE Trans Ultrason Ferroelect
Freq Control 2008;55:1975–1983.
Wang S. Texture feature extraction of hyper-spectral image with three-
dimensional gray-level co-occurrence. J Inform Comput Sci 2015;
12:1439–1448.
Wilhjelm J, Jensen M, Jespersen S, Sahl B, Falk E. Visual and quanti-
tative evaluation of selected image combination schemes in ultrasound
spatial compound scanning. IEEE Trans Med Imaging 2004;23:181–
190.
zu Berge CS, Kapoor A, Navab N. Orientation-driven ultrasound com-
pounding using uncertainty information. Lect Notes Comput Sci 2014;
236–245.
Zettinig O, Fuerst B, Kojcev R, Esposito M, Salehi M, Wein W,
Rackerseder J, Frisch B, Navab N. Toward real–time 3D ultra-
sound registration-based visual servoing for interventional navigation.
In: Proceedings, IEEE International Conference on Robotics and Au-
tomation, Stockholm, Sweden. New York: IEEE, 2016.