Forensic Science International 277 (2017) 2129

Contents lists available at ScienceDirect

Forensic Science International

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

Forensic age estimation by morphometric analysis of the manubrium

from 3D MR images
Naira P. Martnez Veraa,b,* , Johannes Hllera , Thomas Wideka , Bernhard Neumayera,b ,
Thomas Ehammera , Martin Urschlera,b,c
a
Ludwig Boltzmann Institute for Clinical Forensic Imaging, Universittsplatz 4, 8010 Graz, Austria
b
BioTechMed-Graz, Graz, Austria
c
Institute for Computer Graphics and Vision, Graz University of Technology, Inffeldgasse 16, 8010 Graz, Austria

A R T I C L E I N F O A B S T R A C T

Article history:
Received 9 December 2016 Forensic age estimation research based on skeletal structures focuses on patterns of growth and
Received in revised form 12 April 2017 development using different bones. In this work, our aim was to study growth-related evolution of the
Accepted 2 May 2017 manubrium in living adolescents and young adults using magnetic resonance imaging (MRI), which is an
Available online 12 May 2017 image acquisition modality that does not involve ionizing radiation. In a rst step, individual manubrium
and subject features were correlated with age, which conrmed a statistically signicant change of
Keywords:
manubrium volume (Mvol : p < 0.01, R2 0:50) and surface area (Msur : p < 0.01, R2 0:53) for the
Forensic age estimation
MRI studied age range. Additionally, shapes of the manubria were for the rst time investigated using
Manubrium principal component analysis. The decomposition of the data in principal components allowed to analyse
Morphometry the contribution of each component to total shape variation. With 13 principal components, 96% of
Principal component analysis shape variation could be described (Mshp : p < 0.01, R2 0:60). Multiple linear regression analysis
modelled the relationship between the statistically best correlated variables and age. Models including
manubrium shape, volume or surface area divided by the height of the subject (Y  MshpMsur/
Sh : p < 0.01, R2 0:71; Y  MshpMvol/Sh : p < 0.01, R2 0:72) presented a standard error of estimate of
two years. In order to estimate the accuracy of these two manubrium-based age estimation models, cross
validation experiments predicting age on held-out test sets were performed. Median absolute difference
of predicted and known chronological age was 1.18 years for the best performing model (Y  MshpMsur/
Sh : p < 0.01, R2p 0:67). In conclusion, despite limitations in determining legal majority age, manubrium
morphometry analysis presented statistically signicant results for skeletal age estimation, which
indicates that this bone structure may be considered as a new candidate in multi-factorial MRI-based age
estimation.
2017 Elsevier B.V. All rights reserved.

1. Introduction most frequently used biological features for estimating CA [7],

Forensic age estimation is an important tool in anthropology
for identication of human remains [1], however, recently it has
gained a lot of interest especially for age estimation of living
adolescents and young adults to verify chronological age (CA) in
legal proceedings [24] or sports applications [5,6]. For this
purpose, skeletal, sexual and dental maturity indicators are the
* Corresponding author at: Ludwig Boltzmann Institute for Clinical Forensic
Imaging, Universittsplatz 4, 8010 Graz, Austria.
E-mail addresses: naira.martinez-vera@c.lbg.ac.at (N.P. Martnez Vera),
martin.urschler@c.lbg.ac.at (M. Urschler).
with skeletal and dental development predominantly being
investigated by radiographic imaging methods [810]. Deriving
CA from biological development is challenging due to the
variation of human development [11,12] as well as uncertainties
introduced by both limited study populations [13] and used
radiological staging methods [8,9,14]. Therefore, there is an
increased interest in identication of novel indicators of
biological development to support accurate multi-factorial age
estimation approaches [15].
Skeletal age (SA) estimation is a core component in forensic age
estimation of the living from childhood into late adolescence. It is
based on skeletal maturation and may be followed until
development ceases to progress and maturity of the specic bone

http://dx.doi.org/10.1016/j.forsciint.2017.05.005
0379-0738/ 2017 Elsevier B.V. All rights reserved.
22 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129

is attained [16]. To increase the potential age range after the
completion of hand bone ossication, which occurs around the age
of 18 years in males [8,9,1719], research has been performed on
other patterns of growth and development using different bones.
The medial clavicular epiphysis (MCE) has been intensively studied
in plain chest radiography [20] and thin-slice computed tomogra-
phy (CT) of the thorax [21,22]. However, use of ionizing radiation
without medical indication for examining healthy children and
adolescents is problematic with these imaging modalities.
Therefore, magnetic resonance imaging (MRI) is nowadays seeing
increasing research interest for bone age estimation using hand
and wrist bones [5,23,24] and especially MCE [2528]. A very
obvious extension to MRI-based investigation of the clavicular
ossication is to additionally study the manubrium, i.e. the upper
part of the sternum, since it is well visible on MR images of the
upper thorax and therefore requires no extra image acquisition.
Age-dependent morphometric variability of the manubrium has
already been shown in the literature in relation to growth and
maturation [29]. These changes are due to ossication in the
cartilaginous sternum that begins during fetal development and is
completely nished around the age of 25 except for the
manubriosternal joint [30]. Previous studies have analysed
standard linear measurements such as manubrium length, area
and volume changes with age using radiographs or osteological
collections [31,32]. In [33], CT data of the sternum was analysed
and the results showed a volume increase in a population of 48
children between 4 months and 15 years which was modelled by
an exponential function. Selthofer et al. have further studied shape
changes across age [32]. They standardized the sternum according
to size and shape by applying morphometric analysis using indices
of manubrium breadth and length proportion, thickness, area and
volume. Based on the model for rib cage shape by Gayzik et al. [34],
a complete quantication of size and shape variation was
characterized for ages between 0 and 100 years from landmarks
of CT scans of the sternum by Weaver et al. [35]. In summary, all
these works suggest a signicant, reproducible change in
morphometry of the sternum during growth. Thus, the aim of
our study was to verify the hypothesis that growth-related
morphometric changes in the manubrium can be used to predict
chronological age in a forensically relevant data set of male
adolescents and young adults around majority age.
2. Material and methods
2.1. Study design
We retrospectively analysed the manubrium based on an
available dataset of upper thorax MR images that was acquired
for studying age estimation using the clavicle (see Fig. 1). The
accompanying study investigating the clavicle was performed in
accordance with the Declaration of Helsinki and was approved
by the ethical committee of the local medical university. All
eligible volunteering participants of the accompanying study
provided written informed consent and from underage subjects
written consent signed by the parents was additionally
obtained.

Fig. 1. T1-weighted coronal MRI slices showing clavicles (C) and manubrium (M) from a 13.02- (left) and a 23.71-year-old (right) male subject including semi-automatically
derived segmentations in red (top row). 3D point clouds visualizing the respective 3D shapes of the manubria at the different ages (bottom row). (For interpretation of the
references to colour in this gure legend, the reader is referred to the web version of this article.)
 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129
Fig. 2. Distribution per year among the study population of 130 subjects. Each
histogram bin represents the illustrated age 0.5 years.
2.2. Subjects
In this study morphometric changes [36] were analysed in
relation to CA on MRI data from 130 male Caucasian subjects with
ages ranging from 13.01 to 24.98 years (mean age 19.00 years,
standard deviation: 3.66 years). Subjects were roughly evenly
distributed within this age range as can be seen in the age
histogram in Fig. 2. Prior to scanning, height and weight of the
subjects were additionally collected by the expert on forensic age
estimation responsible for the study.
2.3. MR image acquisition
Subjects were examined on clinical 3 T MRI scanners (Tim Trio/
Skyra, Siemens Healthineers, Erlangen, Germany). Images of the
manubrium were acquired in supine position with a neck coil
(Siemens Healthineers, Erlangen, Germany). T1-weighted images
of the upper thorax region were obtained using a 3D VIBE sequence
with water excitation (FOV = 240 mm, TR/TE = 9.77/3.72 ms, voxel
size 0.9  0.9  0.9 mm3). The sequence was oriented on a coronal
oblique plane, parallel to the sternal bone. To avoid pulsation
artifacts from the heart and vessels, saturation bands were applied
in caudal and posterior position of the 3D region of interest. Fig. 1
shows two representative MR images of 13.02 and 23.71 year old
subjects, respectively. MR images of manubria included in this
study did not show a visible fusion at the manubriosternal joint
and those images which presented lateral ossications were
incorporated, however, these ossications were not included into
the analysis (Fig. 9, right).
2.4. Manubrium characterization
From the acquired T1-weighted images, we characterized
manubria by indices of width, height, thickness, volume, surface
area and shape variations. Therefore, a volumetric manubrium
segmentation was semi-automatically performed for each image
by using the active learning based segmentation method from
TurtleSeg software1 [37,38]. Due to inhomogeneities in the MR
images, some borders of the manubrium were difcult to get
detected automatically by the software, therefore some segmen-
tations were manually corrected using ITKSnap software2 [39].
From the semi-automatically segmented manubria (red mask in
top row of Fig. 1), 3D meshes were created with ITKSnap using the
1
http://www.turtleseg.org, last accessed 11.04.2017.
2 3
http://www.itksnap.org, last accessed 11.04.2017.
23
Marching Cubes algorithm [40]. The exported 3D triangle mesh
was then pre-processed with the MeshLab software.3 This pre-
processing consisted of Poisson reconstruction to achieve a
consistent, hole-free surface, followed by mesh simplication
according to a quadric surface error measure [41] that reduced the
number of vertices to 5000 (bottom row in Fig. 1). From these 3D
meshes, volume (Mvol) and surface area (Msur) of the manubria
were computed. After roughly aligning all meshes by shifting each
to its centre of gravity, a base mesh was randomly selected and all
other meshes were non-rigidly registered to the base mesh with
the method from [42]. As a result, correspondence of 3D points for
all 3D meshes of the manubria could be established. Dened in
analogy to [32], the determined correspondences allowed to
extract four conventional linear measurements (upper and lower
manubrium width Muw, Mlw, manubrium height Mh and
manubrium thickness Mt) by identifying eight homologous
landmarks (3D points). The input for principal component analysis
(PCA) based shape analysis was provided by the full set of
corresponding 3D points for all 3D meshes of the manubria. PCA is
a statistical procedure that transforms observations, i.e. the 3D
points for all 3D meshes, into a set of coefcients called principal
components [43]. The rst principal component encodes the
largest variance seen in the dataset, thus describing the principal
mode of maximal shape variation. Succeeding components are
orthogonal to preceding components but with decreasing variance,
thus representing less important variations in shape. From
corresponding landmarks we created the PCA-based shape model
Mshp by rst performing full Procrustes analysis [44] with respect
to translation, rotation and scale to achieve a model where aligned
shapes are centred around the same origin, have the same 3D
orientation and are normalized in size. This model solely contains
variations in shape due to anatomical differences in a population,
i.e. a manubrium shape space according to the denition of Kendall
[45]. After this Generalized Procrustes analysis step, PCA was
performed using Matlab R2014a, thus achieving a parametric
model of shape where each 3D mesh can be represented exactly
using a projection onto the principal components.
2.5. Linking manubrium development with age
We rst investigated the correlation of each manubrium feature
(conventional linear measurements Muw, Mlw, Mh, Mt in cm,
surface area Msur in cm2, volume Mvol in cm3 and the most
prominent shape change described by the rst principal compo-
nent of the PCA model M1shp ) and each subject feature (height Sh in
cm and weight Sw in kg) with known chronological age. Unity-
based normalization was used to scale these variables into the
range between 0 (minimum value) and 1 (maximum value). For
those two features which showed highest correlation with age
(Mvol, Msur), we additionally performed linear regression analysis
to describe how changes in these variables are related with age in
our target age range of subjects.
We further explored the relationship of individual PCA shape
components together with the highest correlated individual
manubrium and subject features and chronological age using
multiple linear regression (MLR) analysis. Therefore, in a rst step
we identied an appropriate number of principal shape compo-
nents to restrict further age regression analysis to the most
relevant shape variables. Besides the determined shape compo-
nents Mshp, additional candidate regressors for MLR were
manubrium features Mvol, Msur as well as subject-specic features
Sh and Sw . To rule out collinearities between multiple explanatory
http://www.meshlab.net, last accessed 11.04.2017.
24 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129

variables, Besley collinearity diagnostics [46] was performed.

Potential correlations were tested by singular values (s), condition
index (condIdx) and variance decomposition proportions
(VarDecomp). Finally, MLR models were chosen to be constituted
by variables which presented signicant correlation with age in the
linear correlation analysis and which showed independence of
each other in the collinearity tests. From these variables, different
combinations were used to t a number of multivariate models
that allow regression of chronological age.
In our nal experiments, 13-fold cross-validation analysis was
performed for all candidate MLR models to evaluate the ability to
predict age on images not contained in the dataset from which the
respective model is computed, thus indicating practical generali-
zation capabilities of the models. For this purpose, the data was
divided into ve age groups (1316 years, 1619 years, 1922 years,
2225 years) and 13 tests were conducted. Each test partitioned
the manubria set into two subsets, the test group with two
manubria per age group and the training group with the remaining
120 manubria. The test group was created by randomly drawing
from the whole dataset, such that each image was used once as a
test image during the 13 folds, but taking the distribution among
age groups into account. During each cross-validation fold, the
shape model Mshp was created from the training group and used to
predict the age of the subjects in the test group. This procedure was
repeated for all investigated MLR models. The absolute differences
between estimated age and chronological age for all subjects in the
test group were computed in each cross-validation round, and
summarized by calculating the median and the maximal absolute
difference. These values were then used in the statistical analysis.

2.6. Statistical analysis

Statistical analysis was performed using Matlab R2014a. To

measure the linear correlation between each manubrium and
subject variable and age, the Pearson's r coefcient was used.
Goodness of t of regression models was analysed using R2 values,
together with their statistical signicance. In MLR analysis, R2 is
inuenced by the number of variables. Adding any regressor
variable to an MLR model generates a bias towards a smaller error
and therefore a higher R2. To take this into account when
comparing models, in the MLR experiments goodness of t was
determined by the adjusted coefcient of determination R2 .
Additionally, the standard error of estimate (SEE) was used to
assess different MLR models. In our cross-validation experiment,
the prediction measure R2p was used to evaluate the ability of a
tted regression model derived from a training group to predict the
age of the associated test group. In this part of the study, all models
to be compared had the same number of variables, such that
unadjusted R2 could be used for comparison. Details of all used
statistical measures can be found in Appendix A. A paired t-test
was used for comparison between age prediction candidate
models. A p-value less than 0.05 was considered statistically
signicant, both in correlation/regression as well as t-test analysis
for comparing models.
3. Results
3.1. Analysis of individual manubrium variables
The scatter plots in Fig. 3 show paired variables formed from the
individual manubrium and subject features and CA. All correla-
tions (highlighted in red) are statistically signicant, with a p-value
<0.01. The histograms of each single variable appear next to each
scatter plot, showing the range of values for the normalized
variable to reveal the underlying distribution. Two of the four
Fig. 3. Strength of pairwise linear relationships indicated by Pearson's linear
correlation coefcients. Manubrium shape is represented only by the rst principal
component (M1shp ) of the PCA model. Data were normalized between 0
(minimum value) and 1 (maximum value). All correlations (highlighted in
red) were statistically signicant with p < 0.01. (For interpretation of the
references to colour in this gure legend, the reader is referred to the web
version of this article.)
conventional linear manubrium measurements as dened in [32]
showed correlations with a Pearson's r coefcient larger than 0.5,
i.e. upper and lower manubrium width Muw (r = 0.66), Mlw
(r = 0.66). Highest correlation with age was found for surface area
Msur (r = 0.73) and volume Mvol (r = 0.71) of the manubria, while
the rst principal component of the manubrium shape model
(M1shp ) showed only low correlation with age (r = 0.25).
Individual linear regression models of upper and lower
manubrium width, surface area and volume with age were
statistically signicant with respect to the linear tting
(Muw : R2 0:43, p < 0.01; Mlw : R2 0:43, p < 0.01;
2 2
Mvol : R 0:50, p < 0.01; Msur : R 0:53, p < 0.01). The two most
promising linear regression models (Mvol, Msur) are plotted in
detail in Fig. 4, together with their 95% prediction intervals. Both
models indicated similar variation with age (Msur(x) = 0.04  x
 0.25, Mvol(x) = 0.04  x  0.25). Values outside the prediction
intervals came from two subjects whose BMI (as derived from Sh
 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129
Fig. 4. Regression of manubrium surface area (Msur) and volume (Mvol) over age.
Dotted and dash-dotted lines indicate the 95% prediction intervals for surface area
and volume, respectively.
and Sw ) fell into obese (BMI = 34.32 kg/m2) and underweight
(BMI = 18.19 kg/m2) categories.
3.2. Shape variation and MLR analysis
To investigate the hypothesis that shape variations of the
manubrium are related to age, we analysed the PCA based shape
model with respect to the used number of principal components in
an MLR setup. Starting from a very low coefcient of determination
of the manubrium shape model when using just the rst principal
component (M1shp : p < 0:01, R2 0:05), the inclusion of more
2
principal components increased R until saturation was reached at
13 included components (Mshp : p < 0.01, R2 0:60). This behav-
iour is illustrated in Fig. 5. The magnitudes of the eigenvalues of
Mshp are shown in Fig. 6. With 13 principal components, 96% of
the shape variation could be explained, while the rest can be
assumed to come from noise as seen by its low eigenvalue
magnitudes, thus providing a suitable trade-off between model
exibility and compactness.
Fig. 5. Adjusted coefcients of multiple determination R2 obtained when the
shape models were correlated with age. These models were constituted from
one to 60 principal components. Dashed line points out the subsets with 13
principal components, where a large improvement in R2 was observed before
going into saturation.
25
Fig. 6. Eigenvalues of the rst 60 eigenvectors of the shape PCA-model. In red, the
rst 13 principal components which explain 96% of the shape variation, providing
a suitable trade-off between model exibility and compactness. (For interpretation
of the references to colour in this gure legend, the reader is referred to the web
version of this article.)
Besley collinearity diagnostics assessed the strength and
sources of collinearity among the manubrium (shape, volume
and surface area) and subject (height and weight) features in the
MLR model. Strong multicollinearity existed between the manu-
brium variables Msur and Mvol (see Fig. 7). From the test of
collinearity and the MLR analysis, the best candidate models to
build a predictive age model are summarized in Table 1 together
with their goodness of t and SEE when used on all available
manubria. Models which included shape with surface area or with
surface area divided by height yielded high R2 values (p < 0.01,
R2 0:71). Similar results were obtained using models constituted
by shape and volume (p < 0.01, R2 0:70), however an improve-
ment was found when the volume was divided by the height of the
subject (p < 0.01, R2 0:72). Lowest R2 values (R2  0:60) were
found in models where the subject's weight was included. The
Fig. 7. Strength and sources of collinearity among the manubrium and subject
variables in a multiple linear regression model. Manubrium shape is represented by
using only the rst principal component (M1shp ). Variance-decomposition
proportions with a condition index (condIdx) larger than the tolerance 30
are shown. The variance-decomposition proportions exceeding the default
tolerance proportion 0.5 (tolProp) are indicated by red markers. Results
indicated that surface area (Msur) and volume (Mvol) features presented
linear dependency (condIdx = 36.9). (For interpretation of the references to
colour in this gure legend, the reader is referred to the web version of this
article.)
26 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129

Table 1
Multiple linear regression (MLR) models for age estimation. Results of best
performing MLR models in bold.

MLR model R2 SEE/years

Y  Mshp + Mvol 0.70** 2.12

Y  Mshp + Msur 0.71** 2.12
Y  Mshp + Mvol/Sh 0.72** 2.08
Y  Mshp + Msur/Sh 0.71** 2.07
Y  Mshp Mvol =Sw 0.63** 2.39
Y  Mshp Msur =Sw 0.60** 2.50

R2 = adjusted coefcient of determination; SEE = standard error of estimate;

Y = estimated age; Mshp = manubrium shape; Mvol = manubrium volume;
Msur = manubrium surface area; Sh = subject height; Sw = subject weight.
**
p-value <0.01.

standard error of estimate for these models was larger than two
years, with a small improvement when the height of the subject
was taken into account (Y  MshpMsur/Sh: SEE = 2.08 years; Fig. 8. Each box-whisker plot represents the difference between estimated SA and
Y  MshpMvol/Sh: SEE = 2.07 years). CA for subjects from the specied age group. Between 16 and 22 years of
chronological age, the median absolute difference of 1619 and 1922 years age
groups are 0.97 and 1.12 years, respectively.
3.3. Predictive models
years and 1.12 years, respectively, however with very different
Table 2 sums up the results of the cross-validation experi- distributions (75% percentile was 2.63 and 1.63 for the two groups,
ments with different predictive models. Models with the respectively). For the youngest age group, absolute differences of
inuence of subject's height produced higher values of the 75% of the age predictions were less than 3 years (median 1.75
predictive coefcient of determination (p < 0.01; R2p 0:67) years), while in the oldest group the median was highest with 2.31
compared to the models where only manubrium features were years. Outliers in the age group between 19 and 22 years came from
included (p < 0.01; R2p 0:65) and the models normalized by the overweight subjects (BMI = 34.32 kg/m2), subjects possessing a
manubrium with a rounder shape as expected at the same age
weight of the subject (p < 0.01; R2p  0:55). Regarding median
(Fig. 9, left), and subjects which presented lateral ossications
absolute difference of predicted and known chronological age, the
(Fig. 9, right).
model including both manubrium shape and surface area features
together with subject height predicted CA more accurately
4. Discussion
(1.18  0.41 years) than the same model but with the manubrium
volume feature (1.29  0.32 years). Nonetheless, the maximal
This retrospective study presented the manubrium bone as a
absolute age difference was smaller using the model which
candidate for age estimation in males during adolescence and early
included volume (4.76 years) than for the model where the surface
adulthood, especially focusing on the forensically important age
area of the manubrium was included (4.92 years). A statistically
range between 13 and 25 years, i.e. around majority age. To the best
signicant difference with p < 0.01 was found in median and
of our knowledge, we were the rst to propose the study of
maximal age difference when comparing those two models. Using
manubrium morphometrics using MRI data, which has the benet
either of the two methods with highest R2p , 17.57% of subjects of not involving ionizing radiation during acquisition. Age
younger than 18 years were estimated above the 18 years threshold prediction was studied in relation with shape variations of the
and 8.93% of the subjects older than 18 years were predicted as manubrium as derived from principal component analysis. This
younger than 18 years. Results of age estimation specic to the age extends upon previous morphometric studies [32,35] by using the
groups 13 years x< 16 years, 16 years x< 19 years, 19 years entire shape information instead of a limited user selected set of
x< 22 years and 22 years x< 25 years using the predictive model landmarks.
Y  MshpMsur/Sh can be observed in the box-whisker plots of The investigated 130 subjects cover an age range from 13.01 to
Fig. 8. Two age groups, between 16 and 19 years and between 19 24.98 years (approximately 10 subjects per year), where strong
and 22 years, showed a similar median absolute difference of 0.97 changes due to manubrium development may be expected. In
agreement with the literature, statistics of individual manubrium
features showed that volume and surface area of the manubria in
Table 2
our dataset correlated with age. According to Sandoz et al. [33], the
Predictive models for age estimation from cross-validation experiments. Results of
best performing predictive models in bold. volume of the sternum showed an increase between 0 and 15 year
old subjects, which they modelled using an exponential function.
MLR model R2p Std Median/year Std Max./year
After the age of 13 years, it is reasonable to consider that this
Y  Mshp + Mvol 0.65 **
0.16 1.39 0.28 5.25 exponential increase ceases, and the consequences of growth
Y  Mshp + Msur 0.65** 0.13 1.38 0.33 5.13 decline after puberty. So our hypothesis was to select a linear
Y  Mshp + Mvol/Sh 0.67** 0.14 1.29 0.32 4.76 regression model to describe the change of the different variables
Y  Mshp + Msur/Sh 0.67** 0.12 1.18 0.41 4.92
Y  Mshp Mvol =Sw 0.56** 0.12 1.66 0.87 6.35
along this age range (Fig. 4). Results afrmed this hypothesis with
Y  Mshp Msur =Sw 0.53** 0.15 1.79 0.93 6.83 statistically signicant changes over age both in volume and
surface area.
Y = estimated age; Mshp = manubrium shape; Mvol = manubrium volume; Msur =
The manubrium is a attened bone in an antero-posterior
manubrium surface area; Sh = subject height; Sw = subject weight; R2p = predictive
direction, with the posterior surface being smoother than the
coefcient of determination; Std = standard deviation; median and maximal
(Max.) absolute difference of predicted and known CA for the methods anterior one. The superior region is broader than the inferior,
selected with larger R2 values in the MLR analysis.
and the superior border is rounder and smoother than the
**
p < 0.01. roughened inferior border [47]. We found that the use of such
 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129
Fig. 9. Two examples of T1-weighted coronal MRI slices of individuals which presented outlier cases. At the left side, the manubrium of a subject which presented more
rounded shape compared to others with a similar age (20.57 years). At the right side, a 19.35 years old subject which showed lateral ossications (yellow arrows). Our
predictive model estimated these subjects 17.42 and 23.76 years old, respectively. Anatomical structures: M = manubrium, C = clavicle. (For interpretation of the references to
colour in this gure legend, the reader is referred to the web version of this article.)
conventional linear measurements (i.e. distances between
landmarks) describing width, height or thickness of the
manubrium are not enough to quantify the entire complex
morphology of this bone. While Weaver at al. already
characterized the sternal shape variation independent from
sternal size [35], in this work we extended their study by not
only evaluating distances between landmarks, but taking
advantage of the whole shape variation of the manubria in
our dataset. After including 13 principal components into our
shape model, shape variations of the manubrium from MR
images showed statistically signicant correlation with age
(see Fig. 5). Using PCA, the full dimensionality of the shape-
space is reduced to the more compact shape-space spanned by
13 principal components, these ones retaining the most
important shape variations of the manubrium within the entire
age range under study. Generally, the shape tended mainly to
change from an oval shape in early childhood to an angular and
trapezoidal shape (Fig. 1) in young adults driven by the
ossication process that takes place during manubrium
development [29,35].
Different models composed of manubrium and subject features
were investigated using MLR analysis. Results from these models
indicated that changes in the manubrium shape described with 13
principal components together with changes in volume or surface
area of the manubrium normalized by the body height of the
subject were highly age-dependent (Table 1). However, models
including subject weight did not give any new information about
manubrium changes. The accuracy of predictions measured by SEE
showed that these models presented in the best case 2.07 years of
prediction error.
To test prediction accuracy and generalization capabilities of
the regression models in a more practically oriented setup, a 13-
fold cross-validation was carried out (Table 2). Models predicted
age with median absolute differences to the known ground-truth
age smaller than 1.80 years. These prediction results conrm that
strong changes in manubrium morphology occur in this age range,
which can be used for age estimation. However, unfortunately the
maximal age differences were also large. When examining those
manubria more closely, for which large differences occur, we see
that most outliers result from overweight subjects, from manubria
presenting lateral ossications or from signicant shape
27
differences compared to other subjects with similar age. This
indicates that the two models with highest R2p are highly inuenced
by such outlier cases. Although results obtained from all models
are similar, we have selected the model which includes shape,
surface area and subject height because of its lowest median age
estimation error (1.18 years). However, in cases in which human
remains have been mutilated and the stature of the subject is not
available, the two models constituted only by manubrium features,
i.e. without subject features, may also be practical methods for age
estimation.
When dividing the data into four groups according to age and
using the predictive model which showed the smallest median age
estimation error (see Fig. 8), manubria coming from subjects in the
youngest and oldest age groups showed the largest median
difference to chronological age. This indicates that only small
morphological changes occur in these age ranges and our model
lacks the specicity to detect them. However, manubria from
individuals between 16 and 22 years, resembling a very important
age range in practical forensic age estimation, were predicted with
a median error of 0.97 and 1.12 years, respectively, suggesting that
strong morphological changes occur and could be detected by our
model. There was a small but considerable percentage (17.57%) of
subjects younger than 18 years that were estimated older than 18,
indicating that these models currently have to be considered with
care for establishing age of majority, which is 18 years in the legal
systems of many countries.
4.1. Limitations
Due to its retrospective design and its connection to a forensic
age estimation study of adolescents and young adults, for this
study only volunteers between 13 and 25 years were investigated.
For this age range, strong shape changes due to manubrium
development can be expected. Extension to other age ranges would
have to be explored in follow-up studies, however, for examining
younger children the need for sedation or anesthesia to enable the
currently time-consuming acquisition of MRI data clearly is a
major limitation.
Despite our MR protocol being adapted for MCE imaging, the
manubrium bone was well depicted in our whole dataset of MR
images. However, some edges of the manubria in a number of MR
28 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129
images, mostly at the bottom of this structure, could not be
detected reliably by the segmentation software. This issue is a
result of the MRI scans not being optimized for the manubrium but
for the clavicle. So, in these cases manual image correction was
performed which suffers from observer variability. Consequently,
future work will focus on showing the reproducibility of the
manual image segmentation by letting another observer repeat the
ground truth segmentations and studying inter-observer variabili-
ty. This resembles a large effort and was not yet considered, since
our aim was rst to verify our hypothesis that shape variations are
in principle related with age.
Currently, clinical applicability of our method is limited by the
tedious and time-consuming semi-automatic manubrium seg-
mentation step. While all subsequent steps like 3D mesh
generation, non-rigid point registration and projection into PCA
shape space are already fully automatic, developing a software
pipeline for routine clinical use will require further research on
automatic segmentation of the manubrium from MRI data. Here,
statistical shape and appearance models [48] as well as recent
machine learning methods for anatomical landmark localization
[49,50] are promising approaches to follow. We see the work-ow
described in this manuscript as a rst step towards establishing
such an automated pipeline.
Subjects with a manubrium whose shape differed considerably
from the rest of our population at a comparable age, resulted in a
large age estimation error. Selthofer et al. suggested a standard
sternum shape (trapezoid shape) which was present in more than
two thirds of his analysed samples, with the remaining third being
divided into two groups, triangular and quadrangular shapes [32].
So future work will need to pre-classify these three different types
of shapes and dene them separately in our PCA model.
Additionally, subjects with high and low BMI values presented
volumes and shapes of manubria that were outside of the
prediction intervals in our models. Therefore, the use of our
suggested regression model should be limited to cases where the
BMI is neither in an under-, nor an overweight category. In order to
investigate BMI categorical variables in our predictive model more
systematically, we will need additional subject data with a more
uniform distribution of BMI values.
5. Conclusion
Due to the misclassication regarding legal majority age, the
age variation of the manubrium shape can currently not be
considered as an isolated criterion to determine age of majority.
However, the presented method of estimating SA from shape
variations of manubria can be included as a new candidate in MRI-
based age estimation, despite not being as accurate as other
methods for age estimation like the left hand or the clavicle, to
serve as an additional factor in multi-factorial age estimation. It
needs to be studied further, if combining the manubrium or the
whole sternum with MCE may improve the prediction of CA using
MRI data. A denite benet is, that the manubrium is readily
available on the clavicle MRI volumes, and therefore does not add
additional costs to a scan investigating traditional anatomical
structures for multi-factorial age estimation.
Acknowledgment
Bernhard Neumayer acknowledges funding from the City of
Graz, Kulturamt, GZ A 16 021628/2013. Martin Urschler
acknowledges funding from the Austrian Science Fund (FWF):
P28078-N33 (FAME).
Appendix A. Details of statistical analysis
The adjusted coefcient of determination (R2 ) that assesses
goodness of t in the MLR experiments was dened as:


n1  
R2 1  1  R2 ;
np
where the number of model coefcients (p) is used in computing
R2 and n is the number of subjects.
Following formula was used to calculate the standard error of
estimate (SEE) for all two degrees of freedom MLR models:
v
u
uX yi;CA  yi;SA 2
SEE t
i
n2
where yi,CA is the CA, yi,SA is the estimated skeletal age of subject i,
and n the number of subjects in our sample.
In our cross-validation experiment involving predictive models,
R2p for a model t was calculated by
Pn
yi;CA  yi;SA 2
R2p 1  Pi1 ;
n 2
i1 yi;CA  yCA
where n, yi,CA and yCA were the number of subjects, the CA of
subject i and the mean age in the random training subset,
respectively. Using the model derived from a training group, the
estimated SA of subject i in the test group was computed to give yi,
SA.
References
[1] N. Mrquez-Grant, An overview of age estimation in forensic anthropology:
perspectives and practical considerations, Ann. Hum. Biol. 42 (4) (2015) 308
322, doi:http://dx.doi.org/10.3109/03014460.2015.1048288.
[2] S. Ritz-Timme, C. Cattaneo, M.J. Collins, E.R. Waite, H.W. Schuetz, H.J. Kaatsch,
H.I. Borrman, Age estimation: the state of the art in relation to the specic
demands of forensic practise, Int. J. Legal Med. 113 (3) (2000) 129136, doi:
http://dx.doi.org/10.1007/s004140050283.
[3] A. Schmeling, A. Olze, W. Reisinger, G. Geserick, Forensic age diagnostics of
living people undergoing criminal proceedings, Forensic Sci. Int. 144 (2-3)
(2004) 243245, doi:http://dx.doi.org/10.1016/j.forsciint.2004.04.059.
[4] A. Aynsley-Green, T.J. Cole, H. Crawley, N. Lessof, L.R. Boag, R.M.M. Wallace,
Medical, statistical, ethical and human rights considerations in the assessment
of age in children and young people subject to immigration control, Br. Med.
Bull. 102 (1) (2012) 1742, doi:http://dx.doi.org/10.1093/bmb/lds014.
[5] J. Dvorak, J. George, A. Junge, J. Hodler, Age determination by magnetic
resonance imaging of the wrist in adolescent male football players, Br. J. Sports
Med. 41 (1) (2007) 4552, doi:http://dx.doi.org/10.1136/bjsm.2006.031021.
[6] M.F. Bergeron, M. Mountjoy, N. Armstrong, M. Chia, J. Ct, C.A. Emery, A.
Faigenbaum, G. Hall, S. Kriemler, M. Lglise, R.M. Malina, A.M. Pensgaard, A.
Sanchez, T. Soligard, J. Sundgot-Borgen, W. van Mechelen, J.R. Weissensteiner,
L. Engebretsen, International Olympic Committee consensus statement on
youth athletic development, Br. J. Sports Med. 49 (13) (2015) 843851, doi:
http://dx.doi.org/10.1136/bjsports-2015-094962.
[7] S. Black, A. Aggrawal, J. Payne-James, Age Estimation in the Living: The
Practitioner's Guide, John Wiley and Sons, Inc., 2010.
[8] W.W. Greulich, S.I. Pyle, Radiographic Atlas of Skeletal Development of the
Hand and Wrist, 2nd ed., Stanford University Press, Stanford, CA, 1959.
[9] J.M. Tanner, R.H. Whitehouse, C.N, W.A. Marshall, M.J.R. Healy, H. Goldstein,
Assessment of Skeletal Maturity and Prediction of Adult Height (TW2
Method), 2nd ed., Academic Press, 1983.
[10] A. Demirjian, H. Goldstein, T. JM, A new system of dental age assessment, Hum.
Biol. 45 (2) (1973) 211227.
[11] N. Cameron, Can maturity indicators be used to estimate chronological age in
children? Ann. Hum. Biol. 42 (4) (2015) 300305, doi:http://dx.doi.org/
10.3109/03014460.2015.1032349.
[12] D. Gibelli, D. De Angelis, C. Cattaneo, Radiological pitfalls of age estimation in
adopted children: a case report, Miner. Pediatr. 67 (2) (2015) 203208.
[13] E. Cunha, E. Baccino, E. Martrille, F. Ramsthaler, J. Prieto, Y. Schuliar, N.
Lynnerup, C. Cattaneo, The problem of aging human remains and living
individuals: a review, Forensic Sci. Int. 193 (1-3) (2009) 113, doi:http://dx.doi.
org/10.1016/j.forsciint.2009.09.008.
[14] R. Cameriere, S. De Luca, D. De Angelis, V. Merelli, A. Giuliodori, M. Cingolani, C.
Cattaneo, L. Ferrante, Reliability of Schmeling's stages of ossication of medial
clavicular epiphyses and its validity to assess 18 years of age in living subjects,
 N.P. Martnez Vera et al. / Forensic Science International 277 (2017) 2129
Int. J. Legal Med. 126 (6) (2012) 923932, doi:http://dx.doi.org/10.1007/
s00414-012-0769-4.
[15] A. Schmeling, M. Garamendi, J.L. Prieto, M.I. Landa, Forensic age estimation in
unaccompanied minors and young living adults, in: D.N. Vieira (Ed.), Forensic
Medicine - From Old Problems to New Challenges, InTech, 2011, pp. 77120
(Chapter 5).
[16] L. Scheuer, S. Black, Developmental Juvenile Osteology, 1st ed., Elsevier
Academic Press, 2000, doi:http://dx.doi.org/10.1016/B978-012624000-9/
50015-0.
[17] S.M. Garn, Radiographic atlas of skeletal development of the hand and wrist,
Am. J. Hum. Genet. 11 (3) (1959) 282283, doi:http://dx.doi.org/10.1097/
00007611-196011000-00033.
[18] D. Freitas, J. Maia, G. Beunen, J. Lefevre, A. Claessens, A. Marques, A. Rodrigues,
C. Silva, M. Crespo, M. Thomis, A. Sousa, R. Malina, Skeletal maturity and socio-
economic status in Portuguese children and youths: the Madeira growth
study, Ann. Hum. Biol. 31 (4) (2004) 408420, doi:http://dx.doi.org/10.1080/
03014460410001713050.
[19] S. Schmidt, B. Koch, R. Schulz, W. Reisinger, A. Schmeling, Comparative analysis
of the applicability of the skeletal age determination methods of Greulich-Pyle
and Thiemann-Nitz for forensic age estimation in living subjects, Int. J. Legal
Med. 121 (4) (2007) 293296, doi:http://dx.doi.org/10.1007/s00414-007-
0165-7.
[20] A. Schmeling, R. Schulz, W. Reisinger, M. Muehler, K.D. Wernecke, G. Geserick,
Studies on the time frame for ossication of the medial clavicular epiphyseal
cartilage in conventional radiography, Int. J. Legal Med. 118 (1) (2004) 58, doi:
http://dx.doi.org/10.1007/s00414-003-0404-5.
[21] M. Kellinghaus, R. Schulz, V. Vieth, S. Schmidt, H.P.A. Schmeling, Enhanced
possibilities to make statements on the ossication status of the medial
clavicular epiphysis using an amplied staging scheme in evaluating thin-slice
CT scans, Int. J. Legal Med. 124 (2010) 321325, doi:http://dx.doi.org/10.1007/
s00414-010-0448-2.
[22] M. Kellinghaus, R. Schulz, V. Vieth, S. Schmidt, A. Schmeling, Forensic age
estimation in living subjects based on the ossication status of the medial
clavicular epiphysis as revealed by thin-slice multidetector computed
tomography, Int. J. Legal Med. 124 (2) (2010) 149154, doi:http://dx.doi.org/
10.1007/s00414-009-0398-8.
[23] Y. Terada, S. Kono, D. Tamada, T. Uchiumi, K. Kose, R. Miyagi, E. Yamabe, H.
Yoshioka, Skeletal age assessment in children using an open compact MRI
system, Magn. Reson. Med. 69 (6) (2013) 16971702, doi:http://dx.doi.org/
10.1002/mrm.24439.
[24] E. Tomei, A. Sartori, D. Nissman, N. Al Ansari, S. Battisti, A. Rubini, A. Stagnitti,
M. Martino, M. Marini, E. Barbato, R.C. Semelka, Value of MRI of the hand and
the wrist in evaluation of bone age: preliminary results, J. Magn. Reson.
Imaging 39 (5) (2014) 11981205, doi:http://dx.doi.org/10.1002/jmri.24286.
[25] S. Schmidt, M. Muehler, A. Schmeling, W. Reisinger, R. Schulz, Magnetic
resonance imaging of the clavicular ossication, Int. J. Legal Med. 121 (4)
(2007) 321324, doi:http://dx.doi.org/10.1007/s00414-007-0160-z.
[26] E. Hillewig, J. Degroote, T. Van Der Paelt, A. Visscher, P. Vandemaele, B. Lutin, L.
DHooghe, V. Vandriessche, M. Piette, K. Verstraete, Magnetic resonance
imaging of the sternal extremity of the clavicle in forensic age estimation:
towards more sound age estimates, Int. J. Legal Med. 127 (3) (2013) 677689,
doi:http://dx.doi.org/10.1007/s00414-012-0798-z.
[27] S. Tangmose, K.E. Jensen, C. Villa, N. Lynnerup, Forensic age estimation from
the clavicle using 1.0 T MRI preliminary results, Forensic Sci. Int. 234 (2014)
712, doi:http://dx.doi.org/10.1016/j.forsciint.2013.10.027.
[28] V. Vieth, R. Schulz, P. Brinkmeier, J. Dvorak, A. Schmeling, Age estimation in U-
20 football players using 3.0 Tesla MRI of the clavicle, Forensic Sci. Int. 241
(2014) 118122, doi:http://dx.doi.org/10.1016/j.forsciint.2014.05.008.
[29] G.T. Ashley, The relationship between the pattern of ossication and the
denitive shape of the mesosternum in man, J. Anat. 90 (1956) 87105.
[30] P. Stark, D. Jaramillo, CT of the sternum, Am. J. Roentgenol. 147 (1) (1986) 72
77.
[31] I.C.F. Riach, Ossication in the sternum as a means of assessing skeletal age, J.
Clin. Pathol. 20 (4) (1967) 589590.
29
[32] R. Selthofer, V. Nikoli c, T. Mrcela, R. Radic, I. Leksan, I. Rudez, K. Selthofer,
Morphometric analysis of the sternum, Coll. Antropol. 30 (1) (2006) 4347.
[33] B. Sandoz, A. Badina, S. Laporte, K. Lambot, D. Mitton, W. Skalli, Quantitative
geometric analysis of rib, costal cartilage and sternum from childhood to
teenagehood, Med. Biol. Eng. Comp. 51 (9) (2013) 971979, doi:http://dx.doi.
org/10.1007/s11517-013-1070-5.
[34] F.S. Gayzik, M.M. Yu, K.A. Danelson, D.E. Slice, J.D. Stitzel, Quantication of age-
related shape change of the human rib cage through geometric morpho-
metrics, J. Biomech. 41 (7) (2008) 15451554, doi:http://dx.doi.org/10.1016/j.
jbiomech.2008.02.006.
[35] A.A. Weaver, S.L. Schoell, C.M. Nguyen, S.K. Lynch, J.D. Stitzel, Morphometric
analysis of variation in the sternum with sex and age, J. Morphol. 275 (11)
(2014) 12841299, doi:http://dx.doi.org/10.1002/jmor.20302.
[36] F.L. Bookstein, Morphometric Tools for Landmark Data: Geometry and Biology,
Cambridge University Press, New York, 1991.
[37] A. Top, G. Hamarneh, R. Abugharbieh, Spotlight: Automated condence-based
user guidance for increasing efciency in interactive 3D image segmentation,
Medical Image Computing and Computer-Assisted Intervention Workshop on
Medical Computer Vision (MICCAI MCV), vol. LNCS 6533, (2010), pp. 204213,
doi:http://dx.doi.org/10.1007/978-3-642-18421-5_20.
[38] A. Top, G. Hamarneh, R. Abugharbieh, Active learning for interactive 3D image
segmentation, Medical Image Computing and Computer-Assisted Intervention
(MICCAI), vol. LNCS 6893, (2011), pp. 603610, doi:http://dx.doi.org/10.1007/
978-3-642-23626-6_74.
[39] P.A. Yushkevich, J. Piven, H.C. Hazlett, R.G. Smith, S. Ho, J.C. Gee, G. Gerig, User-
guided 3D active contour segmentation of anatomical structures: Signicantly
improved efciency and reliability, NeuroImage 31 (3) (2006) 11161128, doi:
http://dx.doi.org/10.1016/j.neuroimage.2006.01.015.
[40] W.E. Lorensen, H.E. Cline, Marching cubes: A high resolution 3d surface
construction algorithm, Proceedings of the 14th Annual Conference on
Computer Graphics and Interactive Techniques, SIGGRAPH87, ACM, New York,
NY, USA, 1987, pp. 163169, doi:http://dx.doi.org/10.1145/37401.37422.
[41] M. Garland, P.S. Heckbert, Surface simplication using quadric error metrics,
Proceedings of the 24th Annual Conference on Computer Graphics and
Interactive Techniques, SIGGRAPH97, ACM Press/Addison-Wesley Publishing
Co., New York, NY, USA, 1997, pp. 209216, doi:http://dx.doi.org/10.1145/
258734.58849.
[42] B. Amberg, S. Romdhani, T. Vetter, Optimal step nonrigid ICP algorithms for
surface registration, Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recogn. (2007) 18, doi:http://dx.doi.org/10.1109/CVPR.2007.383165.
[43] I.T. Jolliffe, Principal Component Analysis, Springer, New York, 2002, doi:
http://dx.doi.org/10.1007/b98835.
[44] I.L. Dryden, K.V. Mardia, Statistical Shape Analysis, Wiley, Chichester, 1998.
[45] D.G. Kendall, Shape manifolds, procrustean metrics, and complex projective
spaces, Bull. Lond. Math. Soc. 16 (2) (1984) 81121.
[46] D.A. Belsley, E. Kuh, R.E. Welsh, Regression Diagnostics: Identifying Inuential
Data and Sources of Collinearity, John Wiley and Sons, Inc., 1980.
[47] M. Schaefer, S. Black, L. Scheuer, Juvenile Osteology: A laboratory and Field
Manual, Elsevier, Inc., 2009.
[48] T.F. Cootes, G.J. Edwards, C.J. Taylor, Active appearance models, IEEE Trans.
Pattern Anal. Mach. Intell. 23 (6) (2001) 681685, doi:http://dx.doi.org/
10.1109/34.927467.
[49] T. Ebner, D. tern, R. Donner, M. Urschler, Towards automatic bone age
estimation from MRI: localization of 3D anatomical landmarks, Medical Image
Computing and Computer-Assisted Intervention (MICCAI), vol. LNCS 8674,
Springer, 2014, pp. 421428, doi:http://dx.doi.org/10.1007/978-3-319-10470-
6_53.
[50] D. tern, T. Ebner, M. Urschler, From local to global random regression forests:
exploring anatomical landmark localization, Medical Image Computing and
Computer-Assisted Intervention (MICCAI), vol. LNCS 9901, Springer, 2016, pp.
221229, doi:http://dx.doi.org/10.1007/978-3-319-46723-8_26.