Computer Methods and Programs in Biomedicine 221 (2022) 106893

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Reconstruction of the fetus face from three-dimensional ultrasound

using a newborn face statistical shape model
Antonia Alomar a,∗, Araceli Morales a, Kilian Vellvé b, Antonio R. Porras c,d, Fatima Crispi b,
Marius George Linguraru e,f, Gemma Piella a, Federico Sukno a
a
Department of Information and Communications Technologies, Universitat Pompeu Fabra, Barcelona, Spain
b
Fetal Medicine Research Center (BCNatal), Hospital Clinic and Hospital Sant Joan de Déu, Universitat de Barcelona, Barcelona, Spain
c
Department of Biostatistics and Informatics-Colorado School of Public Health, Department of Pediatrics-School of Medicine, University of Colorado Anschutz
Medical Campus Aurora, CO, U.S.A
d
Departments of Pediatric Plastic & Reconstructive Surgery and Neurosurgery, Children’s Hospital Colorado, Aurora, CO, U.S.A
e
Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Hospital, Washington, D.C., U.S.A
f
Departments of Radiology and Pediatrics, School of Medicine and Health Sciences, George Washington University, Washington, D.C., U.S.A

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: The fetal face is an essential source of information in the assessment of con-
Received 23 August 2021 genital malformations and neurological anomalies. Disturbance in early stages of development can lead
Accepted 13 May 2022
to a wide range of effects, from subtle changes in facial and neurological features to characteristic fa-
cial shapes observed in craniofacial syndromes. Three-dimensional ultrasound (3D US) can provide more
Keywords: detailed information about the facial morphology of the fetus than the conventional 2D US, but its use
Craniofacial defects for pre-natal diagnosis is challenging due to imaging noise, fetal movements, limited field-of-view, low
3D statistical shape model soft-tissue contrast, and occlusions.
Physical defects
3D segmentation Methods: In this paper, we propose the use of a novel statistical morphable model of newborn faces,
Fetal face the BabyFM, for fetal face reconstruction from 3D US images. We test the feasibility of using newborn
statistics to accurately reconstruct fetal faces by fitting the regularized morphable model to the noisy 3D
US images.
Results: The results indicate that the reconstructions are quite accurate in the central-face and less reli-
able in the lateral regions (mean point-to-surface error of 2.35 mm vs 4.86 mm). The algorithm is able to
reconstruct the whole facial morphology of babies from US scans while handle adverse conditions (e.g.
missing parts, noisy data).
Conclusions: The proposed algorithm has the potential to aid in-utero diagnosis for conditions that in-
volve facial dysmorphology.
© 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

1. Introduction craniofacial dysmorphology has been highlighted as an index of

Craniofacial anomalies are among the most common birth de-
fects [1,2]. They encompass a wide range of heterogeneous condi-
tions with many associated genetic syndromes. It is estimated that
up to 40% of genetic disorders produce alterations in the normal
morphology of the face and the head [3,4], and these congeni-
tal malformations can impact swallowing, breathing, hearing, vi-
sion, speech, and -more importantly- cognitive development. Thus,
∗
Corresponding author.
E-mail address: antonia.alomar@upf.edu (A. Alomar).
developmental disturbance at early stages of life. While these asso-
ciations have been established mainly using data from adult popu-
lations, there is a growing interest in early assessment by focusing
on children and newborns [5–7]. In this paper, we explore the fea-
sibility of assessing craniofacial morphology already before birth
by analyzing 3D US images of the fetal face.
The fetal face provides essential information in the evaluation
of congenital malformations and fetal brain function, as its devel-
opment is driven by genetic factors at the first stages of embryo-
genesis. In this sense, facial mass, brain, and skull are inseparable,
as they belong to the same neural compartment. In consequence,

https://doi.org/10.1016/j.cmpb.2022.106893
0169-2607/© 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
A. Alomar, A. Morales, K. Vellvé et al.
if a developmental gene is defective in a differentiation, migration
or in prominence formation, the fetus cannot undergo satisfactory
morphogenesis and can result in morphogenetic disturbances, such
as clefts, hypoplastia, fistulas and ectodermic cysts [8].
Early recognition and assessment of craniofacial conditions is
often crucial for the effective treatment of functional and devel-
opmental aspects [5]. For this reason, diagnosis is shifting from at
birth or childhood towards prenatal. This is important for parent
counseling and careful planning of delivery and post-natal treat-
ment. Moreover, in case of poor prognosis such as multiple anoma-
lies or associated aneuploidies, it may also aid in the management
of the pregnancy [9,10].
In clinical practice, the fetal face is assessed by visual inspec-
tion of routinary prenatal US scans [11,12]. Unfortunately, dysmor-
phology is hard to identify in this way, mainly because of the
wide range of morphological features involved and the challeng-
ing nature of US prenatal scans (low signal-to-noise ratio, fetal or
probe movements, fetal position, and limbs in front of the face).
Moreover, dysmorphology patterns tend to be subtle in most disor-
ders, and they can affect any of the spatial components of the face
(right-left, cranio-caudal, anterior-posterior). For these reasons, an
accurate 3D model of the fetal face could play a crucial role in pre-
natal diagnosis, allowing a detailed analysis of the facial geome-
try and providing a low-cost tool for genetic pre-screening, i.e., to
highlight suspicious cases for further study.
Such analysis must necessarily be addressed in relative terms
with respect to a reference population of normal morphology, for
which 3D Morphable Models (3DMM) are a valuable tool. 3DMMs
[13] encode the anatomical variability of the facial geometry in a
given population within a statistical model, which can be used to
fit and analyze new data. Nonetheless, a crucial aspect to consider
when using a 3DMM is that the statistics encoded in the model
must match those of the target population, e.g. in terms of eth-
nicity, gender and, especially important for our application, age.
The latter has been an important obstacle for the application of
pre-existing 3DMMs to fetal or neonatal data, since all available
3DMMs were built from adults and, although sometimes they also
included children, none of them included babies. However, in a
recent work [14], we published the Baby Face Model (BabyFM),
which constitutes the first 3DMM built exclusively from babies,
with an important proportion of newborns. Also, [15] has shown
how craniofacial deformations can be accurately computed from
3D surfaces using a normative statistical shape model build from
newborn subjects.
We hypothesize that an adequate adaptation of models built
from newborns could serve as statistical constraints to guide
the representation of fetal geometry, which is expected to pro-
duce not only improvements in accuracy (given the higher qual-
ity of the data used to construct the model), but also a way
to relate the fetal face geometry to the one expected at birth.
This would make it possible to use the available knowledge
about the craniofacial anomalies in babies to address an ear-
lier screening of craniofacial dysmorphologies already during fetal
life.
1.1. 3D Fetal face reconstruction
Little research has been done in 3D face reconstruction from fe-
tal images, mainly due to the limitations of prenatal imaging itself.
Nevertheless, advancements in 3D US have fueled a growing inter-
est in developing algorithms and techniques to obtain and analyze
3D fetal faces. One of the first works to automatically detect fetal
faces from 3D US is [16], who combined 3D facial surface detection
and 2D face profile detection to find the optimal acquisition plane
for face rendering, and proposed a carving algorithm to remove all
obstructions and enhance the 3D rendering. Their results showed
2
Computer Methods and Programs in Biomedicine 221 (2022) 106893
high accuracy and fast detection speed on a large fetal data set. In
[17], they also combined both 3D and 2D information to automat-
ically extract the facial surface from a stack of 2D US slices.
There are some works on generating physical fetal models (al-
though not face specific) from 3D US, magnetic resonance imag-
ing, and computer tomography, either separately or combined
[18,19]. However, they involve slice-by-slice manual segmentation
and post-processing with proprietary software. Despite the te-
dious processing, they concluded that physical prototypes are use-
ful when dealing with complex congenital malformations and may
improve antenatal surgical planning. In [20], they proposed a re-
construction method to build 3D printed models of fetal faces from
routine 3D US. Their method involved threshold-based segmenta-
tion, manual point cloud elaboration and merging, and 3D CAD
model construction. They also presented preliminary results of a
survey aiming to evaluate the possible benefits of the use of 3D
printed models for future parents. They concluded that 3D models
are a valuable tool for parents not only from an emotional point of
view, but also from a medical point of view, mainly in the presence
of diseases or malformations.
In [21], they evaluated the feasibility of quantitative assessment
of 3D US facial volumes using statistical shape models. They con-
structed their model from 20 surfaces meshes and found statisti-
cally significant differences between the facial shape of normal and
abnormal fetuses, although the number of cases was limited. They
also required detailed manual segmentation and 3D reconstruc-
tion from proprietary software to extract the facial surfaces. To ac-
celerate the process of data augmentation of the statistical shape
models created, the same investigation group proposed in [22] a
semi-automatic reconstruction algorithm based on multi-atlas
propagation segmentation. However, the results show that more
data need to be added to the model to obtain feasible reconstruc-
tions, as the model quality is limited to the US resolution and qual-
ity.
1.2. Contribution
In this paper, we investigate the feasibility of analyzing fetal 3D
US images with the help of a recently proposed statistical model
constructed from 3D scans of babies and newborns: the BabyFM.
Differently from previous works, we do not build our model di-
rectly from the noisy fetal images, but employ statistics from new-
borns to constrain the geometric reconstruction of the fetal face.
In this way, we circumvent the difficulties associated with build-
ing accurate models from the noisy 3D US images at the expense
of extrapolating the face statistics from newborns to fetuses. We
present a 3D US fitting algorithm to reconstruct the pre-natal fetal
face from 3D US using the BabyFM statistical model. As we do not
have the baby growth-truth we estimate the post-natal face from
multiple 2D images in order to perform the validation of the fetal
face obtained in the 3D US fitting. Tests on a set of 129 fetal scans
show promising results in both qualitative and quantitative terms,
even in adverse conditions (e.g. missing parts, noisy data).
2. Materials
2.1. 3D Baby face morphable model
3D morphable models (3DMMs) are tools for representing 3D
shapes and textures. They are parametric models that enable many
applications in domains of face recognition, entertainment, medi-
cal applications, forensics, cognitive science, neuroscience, and psy-
chology [23]. 3DMMs and statistical shape models (SSM) are also a
popular standard framework in medical imaging segmentation and
largely used as models of variations in anatomical structures.
A. Alomar, A. Morales, K. Vellvé et al.
Fig. 1. Illustration of the 23 anatomical landmarks considered in this work.
In the context of face analysis, the idea is learning a general
3D face model that is able to encode the statistics of facial shape
of a target population (optionally including appearance). The con-
struction of a 3DMM generally consists of two main parts: 1) es-
tablishing point-to-point correspondences between the 3D faces in
the training set, and 2) building a model with the registered data,
typically using principal component analysis (PCA) [13,24,25].
A 3DMM is described by the shape basis matrix  ∈
R3N×M , whose columns are the orthogonal eigenvectors -principal
components- obtained by PCA. Here, N corresponds to the number
of vertices and M to the number of principal components that are
retained, which define the dimensionality of the model. A 3DMM
allows us to create a deformable mesh whose shape is determined
by the parameters α ∈ RM as follows:
x ≈ α + x̄
where x ∈ R3Ncontains the coordinates of the N vertices of the 3D
face mesh and x̄ is the mean shape of the model.
Unfortunately, although there are several 3DMMs that are pub-
licly available for adults and children, the facial morphology of ba-
bies and fetuses differs much from that of adults (and even chil-
dren). For this reason, in [14] we addressed the construction of the
BabyFM, which is the first 3DMM to date built exclusively from in-
fants, including a large proportion of newborns. In this paper we
will used the proposed BabyFM to achieve the fetal face segmen-
tation from the 3D US scans.
The BabyFM was built with data obtained at the Children’s Na-
tional Hospital in Washington D.C. The dataset consisted of 45 3D
scans of baby faces (mean age 8.42 ± 6.45 months). Several ethnic-
ities were included: Caucasian (47%), African American (24%), His-
panic (20%), and Asian (9%). Also, the data were roughly gender-
balanced: 56% male and 44% female. The BabyFM comprises the
facial region delimited by the chin, the forehead and the ears, and
also provides the indexes of the vertices for 23 anatomical land-
marks (Fig. 1).
2.2. Database
2.2.1. Pre-natal data
To evaluate our methods, 129 3D US scans from 29 fetuses were
collected, i.e., there were multiple 3D US images for each fetus,
Computer Methods and Programs in Biomedicine 221 (2022) 106893
corresponding to different viewing angles (see Fig. 4). The mean
gestational age of the 3D US scans was 30.26 ± 10.22 weeks. These
fetuses are independent to any of the babies used for the construc-
tion of the BabyFM.
The 3D US scans were obtained at the Hospital Clinic de
Barcelona (BCNatal) according to its Ethical Research Committee
and the current legislation (Organic Law 15/1999). The ultrasono-
graphic evaluation of the fetuses was performed with static 3D
acquisitions of the volume on the midsagittal view of the fetal
face. Images were acquired using a General Electric Voluson E6
(General Electric, IL, USA) US machine with a low-frequency probe
(4–8 MHz).
From the volumetric images, 3D meshes were extracted by ap-
plying an automatic threshold segmentation with 3D-Slicer 4.10.21 .
The obtained meshes contained not only face but also other parts
(1)
of the fetus’ body, placenta, and noise.
Using the Markup module of 3D-Slicer, we positioned in each
fetal scan a subset of the 23 targeted landmarks, according to their
visibility. The landmarks used are anatomical features that can be
easily identified in a baby face, see Fig. 1. However, the identifica-
tion of these 23 anatomical landmarks in the fetal scan is challeng-
ing because of occlusions (e.g. the baby may be positioned with
the hand on the face) and the noisy nature of the data. Thus, a
variable fraction of the targeted landmarks could be positioned at
every fetus, depending on visibility, as indicated in Table 1. For
example, notice how the landmarks of the ears (tR, oiR, tL, and
oiL) where never positioned in the US scans as the viewing angles
taken did not allow acquisition of the ears due to occlusions and
noise, whereas landmarks in the nose (n, prn, acR, and sn) were
successfully identified in almost all of the US scans.
2.2.2. Post-natal data
Additionally to the US scans, 2D post-natal photographs were
also available for 8 out of the 29 scanned babies. These pictures
were taken by the parents from 3 different viewpoints (left/right
profiles and frontal) directly using their mobile phones; this sim-
ple setup was meant to avoid having to scan newborn babies with
special equipment. The post-natal images were used to obtain a
1
https://www.slicer.org.
3
A. Alomar, A. Morales, K. Vellvé et al. Computer Methods and Programs in Biomedicine 221 (2022) 106893

Table 1
Percentage of meshes where each anatomical landmark was positioned successfully.

Positioned Anatomical Landmarks

exR enR n enL exL acR aR prn aL acL

46.51% 73.64% 96.90% 65.12% 30.23% 87.60% 82.17% 92.25% 79.07% 78.29%

sn chR cphR ls cphL chL li sl pg tR

89.15% 57.36% 65.89% 67.44% 67.44% 48.84% 63.57% 59.69% 58.91% 0

oiR tL oiL

0% 0% 0%

Fig. 2. Overview of the pipeline proposed. The first and second rows illustrate the procedure for obtaining the pre-natal and post-natal reconstructions, respectively. Then,
both outputs are compared to obtain the surface error map between both reconstructions.

2D-3D reconstruction of the baby face to which we could quan-
titatively and qualitatively compare the facial geometry estimated
from the fetal imaging (see Section 4.2).
3. Methods
Our data processing pipeline consists of two main stages:
US fitting (pre-natal) and multiple image fitting (post-natal) (see
Fig. 2). First, the 3D reconstruction of the fetal face is obtained
from the 3D US images by fitting the BabyFM to it, i.e. finding
the shape parameters α in the 3DMM that best reproduce the face
observed in the US. At this stage, the BabyFM works as a statis-
tical shape constraint allowing a better robustness to noise and
other artifacts. Second, the 3D post-natal face is reconstructed from
the 2D images to compare the morphology estimated from fetal
imaging with respect to the actual outcome after birth, where the
BabyFM is used to estimate the facial 3D geometry.
3.1. 3D ultrasound fitting
In this section we describe our algorithm to obtain the 3D re-
construction of the fetal face from the 3D US scan and the facial
landmarks described in Section 2.2. The implemented algorithm
can be divided in the following steps: landmark-based initializa-
tion, automatic cropping, statistical model-based Non-Rigid Itera-
tive Closest Point fitting (NICP), and mesh refinement, as explained
next.
3.1.1. Landmark-based initialization
A first estimation of the 3D fetal face is obtained taking into
account only the landmarks positioned in the US mesh. For this, a
two-stage iterative procedure consisting of (1) landmark alignment
and (2) shape parameter calculation is performed. In the landmark
alignment stage, Procrustes analysis [26] is used to find a similarity
transformation to fit the current model estimate (initialized with
the mean of the BabyFM) to the US landmarks. Then, the shape
parameters (α ) that best match the fetal face in the US scan are
estimated by solving the morphable model equation described in
Eq. (2,
α = Tr (x − xˆr ), (2)
where r is the reduced shape basis matrix (i.e., the rows of 
that correspond to the landmarks), and xˆr is the current model
estimate of the landmark positions. To ensure that the shape pa-
rameters describe a plausible face, α is regularized as described in
Eq. (3).
The two-stage landmark-based fitting is iterated 20 times to en-
sure convergence. This first estimation will serve as a first align-
ment and shape adaptation of the model to the 3D US scans, the
better the initialization, the easier will be for the NICP to converge
to a solution that is faithful to the US shape observed.
3.1.2. Automatic cropping
Our input US scans cover a variable part of the fetus that is nor-
mally not restricted solely to the head and face. This leads to input
meshes that can easily reach 1 million vertices, many of which are
of no interest to quantify the craniofacial morphology and can un-
necessarily increase the computational cost. For this reason, we au-
tomatically crop the input US mesh before further processing. We
do so by taking advantage of the landmark-based initialization de-
scribed in the previous section, whose vertices are used as centers
to compute radial neighborhoods that determine what points from
the input US mesh are retained. In other words, any point from
the input US mesh whose distance is further than the specified ra-
dius from the landmark-based initialization is removed. Also, the

4
A. Alomar, A. Morales, K. Vellvé et al.
non-connected components are removed to reduce the presence of
scattered noise in the US mesh. In the experiments reported in this
paper, we have set the radial neighborhood to 10 mm. Notice that,
thanks to the landmark-based initialization, this simple step allows
an automatic cropping without the need for volume segmentation
on the 3D US scan.
3.1.3. Statistical model-based NICP fitting
The fetal face reconstruction obtained from the landmark-based
initialization is refined using a NICP algorithm. This is a widely
used algorithm for geometric alignment of 3D models when an ini-
tial estimate of the relative pose is known. It consists in five main
stages: i) point selection, ii) matching points (or correspondences);
iii) weights; iv) pairs rejection or constraints; v) minimization of
the error. There exist many versions of NICP varying from the se-
lection and matching to the minimization strategy [27,28].
In our implementation, every iteration consists of using the
NICP algorithm to fit the current face estimate to the 3D US
mesh and then recovering the model’s shape parameters α , as in
Eq. (2) but now using all rows of  rather than r , thus employ-
ing the whole surface. Point selection and matching are performed
by finding the closest sample of the model vertices in the US mesh.
Then, to reduce the presence of outliers, the point-matching is per-
formed under some distance, uniqueness, normal and neighbour-
hood consistency constraints (see Section 3.3). This process is re-
peated by alternating between the correspondence mapping and
the model’s parameter update followed by statistical regularization,
until the error difference between consecutive iterations is below
a predefined threshold.
The linear system α = (x − x̄ ) is solved using a least square
formulation. The estimated shape parameters obtained must be
regularized since not all linear combinations of the columns of
 (or r in the case of only using landmarks) represent a valid
face. The shape parameters are assumed to follow a multivariate
Gaussian distribution. Therefore, to obtain plausible faces, we need
to ensure that the shape parameter vectors lie within a hyper-
ellipsoid in the parameter space, the size of which is determined
by the variances (the eigenvalues) of the data. The regularised
shape vector αˆ is thus obtained by minimizing:
 αˆ 2
min ||αˆ − α||2 subject to i
≤ β 2. (3)
αˆ
i
λi
where λi are the model’s eigenvalues and β is a constant that de-
fines the hyper-ellipsoid size and is defined as the critical value
of a chi-square distribution with M degrees of freedom, i.e., β 2 =
χ 2 ( p, M ), where p is the probability threshold at which shapes are
considered plausible facial reconstructions and M is the number of
modes of variations in the model. In this paper, p=0.95 is used.
3.2. Mesh refinement
The use of the statistics encoded in the BabyFM implies a trade
off between accurate fitting to the image data and robustness to
noise and artifacts. As we will show in Section 4, the regulariza-
tion effect introduced by the BabyFM clearly improves the overall
quality of the reconstructions. Nonetheless, we have also detected
some regions, such as the nose, in which the model over-constrains
the solution, yielding a sub-optimal fitting to the input scan. This
seems to happen because the nose of fetuses is proportionally big-
ger than the nose of newborn.
To tackle this, we introduce a final refinement step following
[29], in which the facial mesh is allowed to deform solely con-
strained by the Laplacian-Beltrami operator (LBO). Using the LBO,
a mesh selected as template is refined towards the desired facial
parts. The idea behind this step is to transform those facial parts
5
Computer Methods and Programs in Biomedicine 221 (2022) 106893
of the template mesh to the positions in the target mesh, while
moving the rest of the vertices as rigid as possible.
Given a 3D mesh scan, Xt emplat e ∈ R3×N that refers to the mesh
that we want to refine and a 3D target mesh Xtarget ∈ R3×p that
contains the target facial parts that we want to copy to our tem-
plate, to balance the desired problem the following weighted linear
least squares system is solved:
   
λL0 λL0 Xt emplat e
Xre f ined = (4)
SXt emplat e SXtarget Xtarget
where Xre f ined is the adapted template that we are solving for,
L0 ∈ R p×p is the cotangent approximation of the LBO, SXt emplat e ∈
[0, 1]m×p and SXtarget ∈ [0, 1]m×N matrices are the selection matrices
of the m vertices that we want to refine, and λ is the parameter
that determines the stiffness of the template mesh to the trans-
formation. The selection matrices are binary patterns that contain
1 in the positions that belong to the facial parts in each corre-
sponding scan and 0 in the rest. The first row of Eq. (4) aims at
maintaining the template mesh structure while the second row en-
forces the matching between the vertices defining the facial parts.
The parameter λ weighs the relative contribution of the mesh re-
finement (second row) and the regulation constraint (first row). If
λ−→ 0 the facial parts of the template are moved exactly to cor-
responding positions in the target mesh, but the template shape
is not retained. Otherwise, if λ − → ∞ the refined mesh retains the
template shape but do not adapt well to the facial parts. A good λ
is chosen to give a good trade off between the refinement of the
facial parts and the retaining of the template shape, without sharp
transitions or distortions.
In this paper to refine our mesh after the NICP, we consider
the resulting NICP estimated mesh with 99.99999% of variance as
the target mesh and the NICP estimated mesh regularized to only
retain 95% of variance as the template mesh, and set λ = 0.1. The
facial parts that we want to refine correspond to the nose region
vertices. Applying this refinement step we obtain a smoother mesh
while maintaining the nose proportion presented in the US scan.
3.3. Constraints
Different types of pair rejections were implemented to reduce
the presence of outliers:
• Distance Constraint: rejection of correspondences that are fur-
ther than a given threshold (in this paper d=10 mm is used).
It discards pairs from the US scan that are far from the current
model estimate; usually the rejected pairs correspond to noise
or non-face regions.
• Uniqueness Constraint: rejection of points with more than one
correspondence (soft pruning). It removes pairs that introduce
a conflict; for example, moving multiple vertices to the same
location.
• Normal Constraint: rejection of correspondences with incom-
patible surface normals ( > 45). It eliminates pairs of corre-
spondences based on local estimates of the normal vectors of
each vertex.
• Neighbourhood Constraint: rejection of pairs that are not con-
sistent with neighboring pairs. The mean difference distance
between the k nearest neighbours to a given pair of corre-
spondences is rejected if it exceeds a specified threshold (we
use E (q ) > 2). This enforces that the neighbours of the cor-
respondences should be locally consistent between them, im-
plying similar local structure. This mean difference distance is
computed as:
1
K
E ( q ) = |( pus − q(k,us)  −  pmodel − q(k,model)  )| (5)
K
k=1
A. Alomar, A. Morales, K. Vellvé et al.
where pus and pmodel are the correspondence being evaluated,
while q(k,us ) and q(k,model ) are the k nearest pairs of correspon-
dences.
3.4. Multiple image fitting
In order to quantitatively validate the fetal shapes estimated
from the 3D US scans, we reconstruct the newborn 3D face from
a set of three 2D images (frontal, left, and right pose) taken after
birth.
A 3DMM allows estimating the 3D facial geometry from one or
more uncalibrated 2D pictures by fitting the model to them. We
address the 3D-from-2D reconstruction problem using sparse geo-
metric features -edges and landmarks- rather than fitting all facial
pixels as in the analysis-by-synthesis approach, which is computa-
tionally expensive and sensitive to initialization. Our approach is
based on the algorithm proposed in [30], but modifying the al-
gorithm in order to consider multiple images rather than only a
single image. The algorithm starts with an initialization of the 3D
face, using only the landmarks, solving the linear system defined
by mapping the vertices of the model to the 2D image space, un-
der the assumption of a scaled orthographic projection2 (SOP).
The landmark fitting is then refined in an iterative closest point
manner adding the closest edge pixel correspondence information.
These correspondences are filtered using two heuristic rules: 1) re-
moval of 5% of the matches with largest distance to the closest
image edge pixel; 2) removal of matches for which the image dis-
tance divided by the scale factor s exceeds a threshold (fixed em-
pirically to 10).
Finally, a non-linear optimization of the pose (rotation R, scal-
ing s, and translation t) and shape parameters (α ) is performed
by minimizing an hybrid objective function containing landmark,
edge, and prior terms:
E (α , R, t, s ) = w1 Elmk (α , R, t, s ) +w2 EEdge (α , R, t, s ) + w3 EPrior (α ).
The prior term acts as a regularizer of the shape parameters
based on the statistics encoded in the 3DMM:
 and the point-to-surface error. The point-to-point Multiple Model
M
α
EPrior (α ) = (  i )2 ,
i=1 λi
where λi are the model’s eigenvalues (variance of the modes of
variation). The landmark term penalizes differences between the
actual landmark position in the image and the one obtained by
projecting the 3D model landmark:
1
L
Elmk (α , R, t, s ) = ||li − SOP (i α + x̄i , R, t, s )||2
L
i=1
where li is the i-th image landmark, i ∈ R3×M refers to the sub-
matrix of the shape basis corresponding to the ith vertex, and
x̄i ∈ R3 is the vertex of the mean face shape such that i α + x̄i
are the coordinates of the ith vertex. The last term compares the
edges detected on the image with those induced by the model due
to occluding boundaries:
1 
EEdge (α , R, t, s ) = minl
|β (α , R, t, s )| i∈β (α,R,t,s) l∈ ence (growth of the baby).
−SOP(i α + xi , R, t, s )2
where is the set of edges, and β (α , R, t, s ) is the set of occluding
boundary vertices.
2
The SOP model assumes that the variation in depth over the object is small
relative to the mean distance from the camera to the object. Under this assumption,
the projected 2D position of a 3D point does not depend on the distance of the
point to the camera.
Computer Methods and Programs in Biomedicine 221 (2022) 106893
3.5. Quantitative and qualitative analysis
Evaluating the performance of the presented algorithm is diffi-
cult as we have really challenging data and also because we use
a postnatal reconstruction as the estimate of the ground truth.
The US scan is a noisy mesh that contains both face and non-face
regions, distortions due to fetal movement, occlusions, and view,
pose, age or expression changes with respect to other US scans of
the same baby or to the postnatal reconstruction. As so many ex-
ternal factors influence the reconstructions, it is difficult to find a
metric that is not influenced by the above challenges. Thus, dif-
ferent metrics to evaluate the performance of the method were
computed, each one evaluating a different aspect of the algorithm.
However, the quantitative analysis is not sufficient to asses the al-
gorithm because it does not ensure that plausible fetal faces are
obtained. For this reason, the quantitative analysis was comple-
mented with visual evaluation to ensure that the obtained fetal
faces were similar to the parts of the face visible in the US or to
the post-natal reconstruction. In the following subsection, the dif-
ferent metrics and comparisons performed to asses the algorithm
will be explained.
3.5.1. Model to US error
The Model to US (M2U) error is the Euclidean distance between
the US scan vertices that were assigned as correspondences to
model vertices of the reconstructed fetal face. This distance is the
error that the NICP is trying to minimize. Because of the distance
constraint applied, this error is limited to 10 mm and is subject
to the uncertainty of being computed between points that may or
may not be true correspondences. It provides information about
how well the algorithm adjusts the model to the points of the 3D
scan, but not about the quality of the reconstructions.
3.5.2. Multiple model reconstructions error
Given the availability of several US scans per subject, we can
(6)
evaluate the performance of the reconstruction algorithm by com-
paring the reconstruction from different views. For each pair of re-
constructions, two errors are quantified: the point-to-point error
(7) Reconstructions (MMR) error corresponds to the Euclidean dis-
tance between the vertices of the two meshes. In contrast, the
point-to-surface MMR error corresponds to the Euclidean distance
between the vertices of one of the meshes to the nearest surface
point in the other mesh. Low MMR errors indicate that the algo-
rithm is behaving similarly even if the acquisition view of the US
scan changed. More distant views with less points in common are
(8) expected to have larger errors.
3.5.3. Pre-natal vs post-natal
We compare the pre-natal reconstruction with the post-natal
reconstruction (obtained from 2D pictures, see Section 3.4) to vali-
date the pre-natal estimates of the facial geometry. The same met-
rics explained above, point-to-point and point-to surface errors
will be computed, and we refer to them as Fetus to Newborn (F2N)
errors. This comparison is challenging because of errors due to the
pose change, noise, occlusions, facial expression and time differ-
(9) 4. Results
In this section, we first show the fetal face reconstructions ob-
tained from different US scans of the same baby. Then, we show
the fetal face reconstructions error obtained using different param-
eters of correspondence constrains. Finally, we compare the pre-
natal and post-natal reconstructions of the same individual to eval-
uate the performance of the reconstruction algorithm.
6
A. Alomar, A. Morales, K. Vellvé et al.
Fig. 3. Mean Model to US Error Map.
Fig. 4. Reconstructions from different view angles and mean error map of the reconstructed meshes. The first row correspond to the US meshes of the same baby from
three different views and the second row are the corresponding reconstructions.
4.1. US fitting
4.1.1. Model to US error
Fig. 3 shows the mean error map between the US scans and
the reconstructed faces, the model to US (M2U) error. At every
vertex of the mean face, this error is obtained by averaging the
Euclidean distances between its position in the reconstructed face
and the correspondence found in the US scan (see Section 3.5.1).
Note that the errors at each vertex are averaged across those ba-
bies for which the correspondence (model to US) was found. Ver-
tices in gray correspond to those that were never assigned as
correspondences in the US scans. These not-assigned vertices are
mainly located at the ears. On the other hand, the largest errors
were found on the lateral regions (where the US scans are often
noisy) and on the vertices corresponding to the landmarks (since
they are manually set in the US). On average, the error per ver-
tex was 0.39 mm, which reflects the high accuracy of the fitting.
Note, however, that this does not necessarily imply a faithful re-
construction since the M2U error does not evaluate whether cor-
respondences are anatomically correct.
7
Computer Methods and Programs in Biomedicine 221 (2022) 106893
4.1.2. Multiple model reconstructions error
The Multiple Model Reconstructions (MMR) error aims to com-
pensate the shortage of M2U in evaluating the anatomical validity
of the correspondences by analyzing how stable is the geometry
of the estimates given different input scans from the same fetus.
Before discussing the quantitative results, let us analyze qualita-
tively one of the cases to better illustrate the reconstruction chal-
lenges. Fig. 4 (top) shows three US meshes corresponding to dif-
ferent viewing angles of the same fetal face. These different US
meshes not only show different poses of the baby, but also dif-
ferent expressions. Moreover, the smoothness of the surfaces and
the occlusions/noise vary from one to another.
The reconstructions obtained for each view are shown at the
bottom of Fig. 4. We can observe how the reconstructions adapt to
the US information while being guided by the facial geometry en-
coded in the BabyFM. Indeed, for the face regions visible in the US
meshes, the reconstruction clearly resembles the US meshes while
the model provides an informed guess (statistical estimate) for re-
gions with occlusions and large missing areas. For example, even
if the ears are not present in any of the US meshes, the recon-
A. Alomar, A. Morales, K. Vellvé et al.
Fig. 5. Mean Point Surface Error Map. It reports the average error per vertex of the M2U error.
Fig. 6. Correspondence incidence. It indicates the percentage of meshes in which the vertex was selected as a correspondence in the US mesh.
structions obtained have plausible ears. Also, the reconstructions
are quite smooth even if the US meshes are rough (especially for
View 1 and 2), and the nose and mouth are well captured (e.g.
Reconstruction 2 reproduces the open mouth whereas the other
two are maintained closed). Some artifacts can be observed in the
eyes; whereas the US meshes present the eyes closed, they are
reconstructed as if they were opened or with strange wrinkles.
The reason is that the BabyFM has been built from newborn faces
with (mostly) open eyes, thus the reconstructions tend to open the
eyes.
We quantitatively compared the reconstructions as described in
3.5.2. Fig. 4 (g) shows the (point-to-point) error map for the ex-
ample discussed above. Overlapping regions in the three US views
present the smallest errors. For example, the nose and the upper
lip region have an error between 0 and 1.5 mm, which means that
our method is able to obtain similar reconstructions from differ-
ent US scans from the same baby. In contrast, frequently missing
regions in the US views present the largest error.
We repeated this quantitative analysis for all the 29 fetuses
and obtained an average point-to-surface MMR error of 3.08 ±
1.33 mm; see Fig. 5. When restricting to the central-face region
(see Appendix Fig. A.12), the error reduces to 2.34 ± 0.71 mm. In
Fig. 5 we see the average error map of all the reconstruction com-
parisons performed (339 pairs in total). On the other hand, rep-
resenting the vertices that were finally considered as correspon-
dences in the US to estimate the fetal face gives us insight about
the behaviour of the algorithm. Fig. 6 represents the incidence
in percentage of each vertex being found as a correspondence in
8
Computer Methods and Programs in Biomedicine 221 (2022) 106893
the US when considering all the 129 US scans. This representa-
tion allows us to make some statements about the data and the
behaviour of the algorithm. First, we can see that the region that
is most present in the US scans is the nose region (between 80%-
95%) and the less present ones are the lateral regions (between
0%-15%). Second, the constraints applied (see Section 3.3) are able
to remove most of the correspondences that would be assigned to
the ears. These are wrong correspondences since the ears are not
visible or present in any of the US scans used. Third, we observe
that right side of the face of the babies was more often present in
the US scans than the left side.
As shown in Fig. 5, the inner part of the face presents the least
error. This together with the fact that this is the region most fre-
quently present in the US meshes (see Fig. 6) indicates that the
algorithm is adapting to the US information in a consistent man-
ner. In contrast, we see how the lateral regions corresponding to
the ears have the highest error since they are not present in the
US meshes. Also, in these regions the model shape is probably be-
ing adapted to some outliers introduced by the presence of noise,
movement or occlusions.
4.1.3. Correspondence constraints experiments
We evaluated the impact of different correspondence con-
straints settings (normal constraint and neighborhood constraint in
Section 3.3) on the MMR error. To this end, we first modified the
angle acceptance threshold between surface normals from 0 to 180
degrees, for a fixed distance threshold of 2 mm. Then, we evalu-
ated different distance thresholds from 1 to 10 mm for a fixed an-
A. Alomar, A. Morales, K. Vellvé et al.
Fig. 7. Constraints Parameters.
gle threshold of 35 degrees. To ensure that the parameters found
can be established as global and not specific, for each US scan a
leave-one-out cross-validation was performed, obtaining the same
behaviour and minima.
Fig. 7 shows the mean point-to-surface error obtained after
comparing the reconstructions of the same individual from the
different US views when varying the acceptance angle and the
neighbours distance difference. It is observed that as larger angle
or distance difference is allowed between the pairs of correspon-
dences, the error between the reconstructions of the same subject
increases, suggesting that we are adding outlier pairs as we are be-
ing less restrictive. In contrast, if we are too restrictive with the an-
gles or distance difference, the error also increase, suggesting that
correspondences that add information of the face surface are be-
ing rejected. Thus, for the reconstructions shown in this paper, we
chose a surface normal of 35 degrees and a neighbors distance of
2 mm, which correspond to the minima found for each constraint
evaluated.
4.1.4. Model-data mismatch in the nose size
With respect to other facial features, the size of the nose in
the US meshes appears relatively larger than expected for the face
of a baby. This is not well captured by the BabyFM, built from
newborns. We therefore applied the refinement step explained in
Section 3.2 to better reconstruct the nasal protuberance and obtain
smoother reconstructions.
Fig. 8 shows the reconstructed faces with (right) and with-
out (middle) the Laplace Beltrami mesh refinement. In the latter
case, the average point-to-surface error is reduced from 3.08 mm
to 3.01 mm, when considering the entire mesh, and from 2.34 ±
1.39 mm to 2.28 ± 0.73 mm, when restricting it to the central-face.
9
Computer Methods and Programs in Biomedicine 221 (2022) 106893
Fig. 8. NICP vs NICP + Mesh Refinement.
4.2. Pre-natal vs post-natal reconstruction
Fig. 9 shows three examples of pre-natal and post-natal recon-
structions, the latter being obtained as described in Section 3.4. As
in previous examples, we observe that the pre-natal reconstruc-
tions resemble the US meshes in those regions where the face
is visible, while providing plausible facial features in the missing
parts. As for the post-natal reconstructions, we see that they cap-
ture well the facial features (nose, chin and chicks) and expression
(mouth and eyes position).
Comparing pre-natal and post-natal reconstructions is difficult
since there is an age gap and a potential change of expression and
pose that can influence the reconstruction. Fig. 10 shows the point-
to-point error between the pre-natal and post-natal reconstruc-
tions of the three previous examples. The error map in Fig. 10(a)
shows that the errors in regions where there is US face information
are typically below 3 mm, whereas in the regions estimated by the
model the errors increase. The error maps in Fig. 10(b) and (c) il-
lustrate how the expression affects the resemblance between pre-
natal and post-natal reconstruction. For example, in Fig. 9 second
case the mouth appears closed in the pre-natal data whereas in the
post-natal frontal image is open; this makes the error map (10(b))
to have large values around the chin and mouth, since they are dis-
placed. In contrast, the baby in the third row of Fig. 9 appears with
a similar expression in the pre-natal scan and the post-natal im-
age, which favours having low error (see Fig. 10(c)) in most of the
central-face region (error below 2mm). The larger errors around
the eyes, forehead and mouth are due to the lack of information
from the US scan.
The mean point-to-surface error obtained from the comparisons
between pre-natal and post-natal of the 8 babies for which we
have both reconstructions was 2.50 ± 1.15 mm (and 2.00 ±1.33
mm in the central-face). Fig. 11 shows that the average error per
vertex between the pre-natal and post-natal reconstructions.
5. Discussion
In this study, our aim was to obtain accurate fetal face recon-
structions from 3D US images to allow assessing craniofacial mor-
phology as early as possible. The proposed algorithm recovers ac-
A. Alomar, A. Morales, K. Vellvé et al. Computer Methods and Programs in Biomedicine 221 (2022) 106893

Fig. 9. Pre-natal vs Post-natal reconstructions. Each row correspond to a different case, the two first columns are the pre-natal US and the reconstruction obtained, whereas
the remaining columns are the post-natal frontal image and the reconstruction from multiple image fitting.

Fig. 10. Pre-natal vs Post-natal color error maps. Each error map is the error between the pre-natal and the post-natal reconstruction of the three cases presented in Fig. 9.

curate 3D reconstructions and is robust to the viewing angle of the
US scan. Both qualitative and quantitative assessment of the results
are satisfactory. For a limited number of cases, we had post-natal
photographs and we could further validate our pre-natal estimates
of their facial geometry by comparing with the post-natal recon-
struction. The reliability and accuracy of the reconstructions de-
pend on the quality of the US as well as on the fetal pose and
expression. It has been shown that it is reliable in the visible re-
gions in the US scan and less accurate or less reliable in the re-
gions were no information is available, as the algorithm is esti-
mating the shape of these regions based on the statistics of the
newborn faces encoded in the BabyFM. The more noise, or less vis-
ible facial surface, is present in the US scans, the less accurate our
reconstruction is.
Our results suggest that an adequate adaptation of models built
from newborns (such as BabyFM) can serve as statistical con-
straints to guide the representation of fetal geometry, which is ex-
pected to produce not only improvements in accuracy (given the

10
A. Alomar, A. Morales, K. Vellvé et al. Computer Methods and Programs in Biomedicine 221 (2022) 106893

Fig. 11. Pre-natal vs Post-natal mean Error Map.

higher quality of the data used to construct the model), but also a the whole facial morphology of the babies under different condi-
way to relate the fetal face geometry to the one expected at birth. tions (large amounts of noise, missing parts or occasions), obtain-
This would make it possible to use the available knowledge about ing promising results. The facial shape reconstructed by the model
the craniofacial anomalies in babies to address an earlier screening would constitute the nearest newborn shape for the given fetal
of craniofacial dysmorphologies already during fetal life. data. This paper is a step forward to a technique that has the po-
Compared to other fetal face reconstructions approaches such tential to aid the pre-natal assessment and in-utero diagnosis of
as [21] and [17] our method can handle noisy data, occlusions and facial dysmorphology.
missing parts, adapting well to the standard quality of clinical rou-
tine imaging. Acknowledgements
However, our study has some limitations. First, the number of
cases was modest and only a few had post-natal images. Sec- This work is partly supported by the eSCANFace project
ond, there is a mistmatch between the age of the model (babies (PID2020-114083GB-I00) funded by the Spanish Ministry of Sci-
and newborns) and the data (fetuses). This makes the fetal re- ence and Innovation, and the Eunice Kennedy Shriver Na-
constructed faces to look older. Also, some distortions might ap- tional Institute of Child Health and Human Development grant
pear in the eyes (which the model tends to open) and nose (which R42HD081712. A. Alomar was supported by AGAUR under the FI
the model tends to make smaller). For the former, we included a scholarship and G. Piella was supported by ICREA under the ICREA
mesh refinement step. For the latter, a possible future work will be Academia programme.
adding more landmarks to the eyes and also to enlarge the train-
ing data with more babies scanned with the eyes closed in order
Appendix A. Appendix
to capture this deformation mode as a shape vector of the BabyFM.
Nevertheless, under the assumption that this mistmatch is pri-
marily due to development of the face, the facial shape recon-
structed by the model would constitute the nearest newborn shape
for the given fetal data. This geometry difference becomes more
evident when we try to measure the difference between the pre-
natal and post-natal reconstructions. The time difference between
the US and the photos used to get the 3D reconstructions was
about 20 weeks, and might have a clear effect in the facial mor-
phology of the baby. For this reason, applying age correction to
the mesh extracted from the 3D fetal faces or to the BabyFM could
help to more faithfully describe the fetal face.
To compare pre-natal and post-natal reconstructions we need
to consider also the difference in pose and expression between the
fetal and the baby face capture. As we are not under a controlled
environment, the same pose and expression cannot be ensured.
While the use of a 3D model is expected to deal with pose dif-
ferences, a fair comparison between pre-natal and post-natal data
would require identity information to be decoupled from expres-
sions variations. Possible lines of future work in this direction are
building a bilinear model of identity and expression of babies or
performing expression transfer between meshes.

6. Conclusions

In this paper, we proposed an automatic fetal face reconstruc-

tion algorithm from 3D US images using a newborns statistical Fig. A1. Central-face region. In green it can be seen the vertices considered as in-
shape model. It was demonstrated that our algorithm reconstructs face region and in blue the vertices left out.

11
A. Alomar, A. Morales, K. Vellvé et al.
References
[1] S. Liu, R. Narumi, N. Ikeda, O. Morita, J. Tasaki, Chemicalinduced craniofacial
anomalies caused by disruption of neural crest cell development in a zebrafish
model, Dev Dyn 249 (2020), doi:10.1002/dvdy.179.
[2] A. Yoon, B. Pham, K. Dipple, Genetic screening in patients with craniofacial
malformations, J Pediatr Genet 05 (2016), doi:10.1055/s- 0036- 1592423.
[3] K.T.D.C.S.M.L.M. Porras, A.R. Rosenbaum, Machine learning for genetic syn-
drome screening and risk stratification at the point-of-care in press (2021).
[4] B. Hallgrmsson, J.D. Aponte, D.C. Katz, J.J. Bannister, S.L. Riccardi, N. Ma-
hasuwan, B.L. McInnes, T.M. Ferrara, D.M. Lipman, A.B. Neves, J.A.J. Spitz-
macher, J.R. Larson, G.A. Bellus, A.M. Pham, E. Aboujaoude, T.A. Benke,
K.C. Chatfield, S.M. Davis, E.R. Elias, R.W. Enzenauer, B.M. French, L.L. Pickler,
J.T.C. Shieh, A. Slavotinek, A.R. Harrop, A.M. Innes, S.E. McCandless, E.A. Mc-
Court, N.J.L. Meeks, N.R. Tartaglia, A.C.-H. Tsai, J.P.H. Wyse, J.A. Bernstein,
P.A. Sanchez-Lara, N.D. Forkert, F.P. Bernier, R.A. Spritz, O.D. Klein, Auto-
mated syndrome diagnosis by three-dimensional facial imaging, Genet Med 22
(2020), doi:10.1038/s41436- 020- 0845- y.
[5] E. Learned-Miller, Q. Lu, A. Paisley, P. Trainer, V. Blanz, K. Dedden, R. Miller,
Detecting acromegaly: screening for disease with a morphable model 4191
LNCS - II (2006), doi:10.1007/11866763_61.
[6] L. Tu, A.R. Porras, A. Boyle, M.G. Linguraru, Analysis of 3D facial dysmorphology
in genetic syndromes from unconstrained 2D photographs 11070 LNCS (2018),
doi:10.1007/978- 3- 030- 00928- 1_40.
[7] L. Tu, A.R. Porras, A. Morales, D.A. Perez, G. Piella, F. Sukno, M.G. Linguraru,
Three-dimensional face reconstruction from uncalibrated photographs: Appli-
cation to early detection of genetic syndromes, in: H. Greenspan, R. Tanno,
M. Erdt, T. Arbel, C. Baumgartner, A. Dalca, C.H. Sudre, W.M. Wells, K. Drech-
sler, M.G. Linguraru, C. Oyarzun Laura, R. Shekhar, S. Wesarg, M.Á. González
Ballester (Eds.), Uncertainty for Safe Utilization of Machine Learning in Medi-
cal Imaging and Clinical Image-Based Procedures, Springer International Pub-
lishing, Cham, 2019, pp. 182–189.
[8] J.-M. Levaillant, J.-P. Bault, B. Benoit, G. Couly, Normal and abnormal fetal face
atlas, 2017, doi:10.1007/978- 3- 319- 43769- 9.
[9] A.R. Davis, S.K. Horvath, P.M. Castaño, Trends in gestational age at time of sur-
gical abortion for fetal aneuploidy and structural abnormalities 216 (3) (2017),
doi:10.1016/j.ajog.2016.10.031.
[10] A.S.L. Mak, K.Y. Leung, Prenatal ultrasonography of craniofacial abnormalities,
Ultrasonography 38 (1) (2019), doi:10.14366/usg.18031.
[11] D. Rotten, J.M. Levaillant, Two- and three-dimensional sonographic assessment
of the fetal face. 1. a systematic analysis of the normal face, Ultrasound Obstetr
Gynecol 23 (2004), doi:10.1002/uog.984.
[12] E. Merz, J. Abramovicz, K. Baba, H.G. Blaas, J. Deng, L. Gindes, W. Lee, L. Platt,
D. Pretorius, R. Schild, P. Sladkevicius, I. Timor-Tritsch, 3D Imaging of the fetal
face - Recommendationsfrom the international 3D focus group, Ultraschall in
der Medizin 33 (2) (2012), doi:10.1055/s- 0031- 1299378.
[13] V. Blanz, T. Vetter, A morphable model for the synthesis of, 3D Faces (1999),
doi:10.1145/311535.311556.
[14] A. Morales, A.R. Porras, L. Tu, M.G. Linguraru, G. Piella, F.M. Sukno, Spectral
correspondence framework for building a 3D baby face, Model (2020), doi:10.
1109/fg47880.2020.0 0 079.
[15] A.R. Porras, L. Tu, D. Tsering, E. Mantilla, A. Oh, A. Enquobahrie, R. Keat-
ing, G.F. Rogers, M.G. Linguraru, Quantification of head shape from
12
Computer Methods and Programs in Biomedicine 221 (2022) 106893
three-dimensional photography for presurgical and postsurgical evalua-
tion of craniosynostosis, Plast Reconstr Surg 144 (2019), doi:10.1097/PRS.
0 0 0 0 0 0 0 0 0 0 0 06260.
[16] S. Feng, S.K. Zhou, S. Good, D. Comaniciu, Automatic fetal face detection from
ultrasound volumes via learning 3D and 2D information, In IEEE Comput Soc
Conf Comput Vision Pattern Recognit (2009).
[17] L. Bonacina, A. Froio, D. Conti, F. Marcolin, E. Vezzetti, Automatic 3D foetal face
model extraction from ultrasonography through histogram processing, J Med
Ultrasound 24 (4) (2016), doi:10.1016/j.jmu.2016.08.003.
[18] H. Werner, J. Lopes, G. Tonni, E.A. Junior, Physicalmodel from 3D ultrasound
and magnetic resonance imaging scan data reconstruction of lumbosacral
myelomeningocelein a fetus with chiari ii malformation, Child’s Nervous Syst
31 (4) (2015) 511–513.
[19] G.A. Menezes, E.A. Junior, J. Lopes, G.T. S. Belmonte, H. Werner, Prenatal di-
agnosis and physicalmodel reconstruction of agnathia”otocephaly with limb
deformities (absent ulna, ibula and digits) following maternalexposure to
oxymetazoline in the irst trimester, J Obstetr Gynaecol Res 42 (8) (2016)
1016–1020.
[20] D. Speranza, D. Citro, F. Padula, B. Motyl, F. Marcolin, M. Calì, M. Martorelli, Ad-
ditive manufacturing techniques for the reconstruction of 3D fetal faces, Appl
Bionics Biomech 2017 (2017), doi:10.1155/2017/9701762.
[21] A. Dall’Asta, S. Schievano, J.L. Bruse, G. Paramasivam, C.T. Kaihura, D. Dunaway,
C.C. Lees, Quantitative analysis of fetal facial morphology using 3Dultrasound
and statistical shape modeling: afeasibility study, Am. J. Obstet. Gynecol. 217
(1) (2017), doi:10.1016/j.ajog.2017.02.007.
[22] A.E. Clark, B. Biffi, R. Sivera, A. Dall’asta, L. Fessey, T.L. Wong, G. Paramasivam,
D. Dunaway, S. Schievano, C.C. Lees, Developing and testing an algorithm for
automatic segmentation of the fetal face from three-dimensional ultrasound
images: automatic segmentation of the fetal face, R Soc Open Sci 7 (11) (2020),
doi:10.1098/rsos.201342.
[23] B. Egger, W.A. Smith, A. Tewari, S. Wuhrer, M. Zollhoefer, T. Beeler, F. Bernard,
T. Bolkart, A. Kortylewski, S. Romdhani, C. Theobalt, V. Blanz, T. Vetter, 3D Mor-
phable face modelsa-Past, present, and future, ACM Trans Graph 39 (5) (2020),
doi:10.1145/3395208.
[24] P. Paysan, M. Lüthi, T. Albrecht, A. Lerch, B. Amberg, F. Santini, T. Vetter, Face
reconstruction from skull shapes and physical attributes 5748 LNCS (2009),
doi:10.1007/978- 3- 642- 03798- 6_24.
[25] A. Morales, G. Piella, F.M. Sukno, Survey on 3d face reconstruction from uncal-
ibrated images, Comput Sci Rev 40 (2021), doi:10.1016/j.cosrev.2021.10 040 0.
[26] J.C. Gower, Generalized procrustes analysis, Psychometrika 40 (1) (1975)
33–51.
[27] H. Zhu, B. Guo, K. Zou, Y. Li, K.-V. Yuen, L. Mihaylova, H. Leung, A review
of point set registration: from pairwise registration to groupwise registration,
Sensors 19 (2019), doi:10.3390/s19051191.
[28] S. Rusinkiewicz, M. Levoy, Efficient variants of the ICP, Algorithm (2001) 145–
152, doi:10.1109/IM.2001.924423.
[29] H. Dai, N. Pears, W. Smith, C. Duncan, Statistical modeling of craniofacial
shape and texture, Int J Comput Vis 128 (2) (2020) 547–571, doi:10.1007/
s11263- 019- 01260- 7.
[30] A. Bas, W.A. Smith, T. Bolkart, S. Wuhrer, Fitting a 3D morphable model
to edges: a comparison between hard and soft correspondences 10117 LNCS
(2017), doi:10.1007/978- 3- 319- 54427- 4_28.