Vol.3, No.4, 321-325 (2013) Open Journal of Animal Sciences
In vivo prediction of intramuscular fat in pigs using
computed tomography
Jørgen Kongsro*, Eli Gjerlaug-Enger
Norsvin, Furnesveien, Hamar, Norway; *Corresponding Author: jorgen.kongsro@norsvin.no
Received 14 August 2013; revised 21 September 2013; accepted 5 October 2013
Copyright © 2013 Jørgen Kongsro, Eli Gjerlaug-Enger. This is an open access article distributed under the Creative Commons Attri-
bution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
One hundred and four pure-bred Norwegian
Duroc boars were CT (computed tomography)
scanned to predict the in vivo intramuscular fat
percentage in the loin. The animals were slaugh-
tered and the loin was cut commercially. A mus-
cle sample of the m. Longissimus dorsi was
sampled and analyzed by the use of near-infra-
red spectroscopy. Data from CT images were
collected using an in-house MATLAB script.
Calibration models were made using PLS (p artial
least square) regression, containing independ-
ent data from CT images and dependent data
from near-infrared spectroscopy. The data set
used for calibration was a subset of 72 animals.
The calibration models were validated using a
subset of 32 animals. Scaling of independent
data and filtering using median filtering were
tested to improve predictions. The results showed
that CT is not a feasible method for in vivo pre-
diction of intramuscular content in swine.
Keywords: Intramuscular Fat; Computed
Tomography; PLS Regression; Calibration;
Intramuscular fat (IMF) in meat is an important trait
due to the impact on sensory quality and acceptance of
pork meat [1]. IMF is generally determined by extracting
IMF from muscle [2] or by spectroscopic methods [3].
However, these methods require meat samples collected
post mortem. In order to select for meat quality traits in
breeding, development of in vivo methods is of great
importance. Progeny and sib-testing programs have made
it possible to select for these traits, but these are costly in
terms of time and money [4]. Since its introduction to
animal sciences in the early eighties [5], computed to-
mography (CT) has been used to estimate and predict the
body composition of farmed animals. In recent years,
several studies have examined the use of CT to predict
intramuscular fat and fatty acid composition using CT
[1,6]. Both studies concluded that CT could be used to
predict the IMF content post mortem in carcasses or meat
samples. For meat, several studies have examined the use
of CT to predict the salt content in pork meat [7,8]. The
results showed that CT can also be used to predict salt
content of processed meat samples, especially when in-
cluding scans at several different energy levels [7]. For in
vivo studies, no results to the authors’ knowledge have
been published for breeding pigs using CT, but several
studies have been published using ultrasound. For beef,
Harada and Kumazaki [9] found in 1980 a good rela-
tionship between ultrasound estimates and marbling
scores of cross sectional areas of loin. Newcom et al. [4]
showed that estimation of in tramuscular fat percen tage in
the live pig using real-time ultrasound was feasible. In
vivo CT images may have a lot of noise due to b reathing
and muscle contractions etc. In order to remove noise,
some filtering techniques like median filtering [10] have
shown to reduce “salt and pepper” noise which may be
observed by in vivo CT images [11].
In this study, the objective was to construct an in vivo
calibration model to predict the IMF percentage in live
breeding pigs using CT at different energy levels and
using filtering to remove noise.
2.1. Animal and Carcass Samples
One hundred and four pure-bred Norwegian Duroc
boars (Figure 1) were sampled from the Norsvin nucleus
population in Norway in the period from October 2012 to
January 2013. The boars were located at the Norsvin
Delta testing station for boars described by Gerlaug-
Enger et al. [12] and Aasmundstad et al. [13]. The pigs
Copyright © 2013 SciRes. Openly accessible at http://www. scirp.org/journal/o jas/
J. Kongsro, E. Gjerlaug-Enger / Open Journal of Animal Sciences 3 (2013) 321-325
Figure 1. Norsvin duroc boar.
were CT scanned at 120 kg at the end of the testing pe-
riod. The pigs were slaughtered after the end of the test
period at a commercial abattoir in Norway. The carcasses
were transported and the loins were cut commercially at
the Norwegian Meat and Poultry Research Centre pilot
plant. The meat quality measurements were also per-
formed at the Norwegian Meat and Poultry Research
2.2. Meat Quality Measurements
Intramuscular fat was measured using the method
presented by Gjerlaug-Enger [3]. The FOSS FoodScan
Tm near-infrared spectrophotometer (FOSS, Denmark)
was used for the determination of intramuscular fat (Fig-
ure 2).
2.3. Computed Tomography and Image
The pigs were herded to individual pens and sedated
described by [13]. The CT scanner was a GE Healthcare
LightSpeed VCT 32 multi-slice. A meat quality scan was
performed on the loin area of the pig, using soft tissue
protocol for reconstruction of X-ray signals to images to
enhance the image quality of soft tissue. 5 mm slice
thickness was used and a dynamic current (mA), maxi-
mizing the mA to get as good image quality as possible.
The pigs were scanned twice using two different energy
levels; 80 kV and 140 kV (Figures 3 and 4). The pixel
spacing was 0.933 × 0.933 mm.
The images were analyzed using MATLAB [10] and
the Image Processing Toolbox [10]. The CT images were
downloaded as DICOM images and imported into MAT-
LAB using an in-house script. A landmark was set in the
coronal direction locating the last rib, and 10 images
were sampled from the last rib and backwards according
to the same meat quality sample described by Gjerlaug-
Enger [3]. The thickness of the sample was 5 mm times
10 images, which equals the 5 cm sample described by
Gjerlaug-Enger [3]. A squared region of interest was
chosen from the loin area (Figure 5) in each of the im-
Figure 2. FOSS FoodScan.
Figure 3. CT image of mid-section of pig. 80 kV energy level.
Figure 4. CT image of mid-section of pig. 140 kV energy level.
ages to cover as much of the loin area as possible. The
regions of interest were stored as image stacks of region
of interest × 10 slices for each animal. The images were
also filtered using the medfilt 2 function in MATLAB to
remove any salt- and pepper-type of noise. Both non-
filtered and filtered images were subject for the calibra-
tion study. The stacks were analyzed transforming them
into histograms (Figure 6) by counting pixels in the en-
tire stack. The histogram was further analyzed in the
calibration/validation study.
2.4. Calibration and Validation
The data set of 104 animals were separated into a cali-
Copyright © 2013 SciRes. Openly accessible at http://www. scirp.org/journal/o jas/
J. Kongsro, E. Gjerlaug-Enger / Open Journal of Animal Sciences 3 (2013) 321-325 323
Figure 5. Region of interest (ROI) sampled from m. longissi-
ums dorsi at last rib.
Figure 6. Histogram sampled at 140 kV. Contrast samples with
different IMF levels (low = red, blue = medium and green =
bration set and a validation set, where 72 animals was
used for calibration, and 32 was used for validation. The
sets were selected randomly using a random permutation
function randperm in MATLAB. The calibration of IMF
using CT images was performed using PLS regression
and the plsregress function in the MATLAB Statistical
Toolbox [10] and Several calibrations were made using
both scaled (mean centered and standardized) data using
the zscore function in MATLAB and non-scaled data.
The scaling was done to adjust for any variations in scale
in the data. Both non-filtered and median-filtered images
were subject for calibration. The independent data sets
(X) was split into 3 groups ; 80 kV energy level (X1), 140
kV energy level (X2), all with filtered and scaled d ata as
sub-groups, and all energy levels and filters as the last
group (X3). The X3 data set was made using unfolding
all the histograms from of the energy levels and filtering
methods, similar to the calibration made by Johansen et
al. [14]. The calibrations were validated using fu ll leave-
one-out cross validation. The number of PLS compo-
nents were selected based on the lowest cross-validated
calibration error; RMSECV. The models were validated
using the beta vectors from the cross-validated calibra-
tion models on the validation data set. The models were
evaluated based on its explained variance (R square) and
prediction error (RMSEP).
The sample mean IMF was 1.78 percent and the stan-
dard deviation was 0.44. The lowest value was 0.44 per-
cent and the highest value was 3.23. An IMF level of 1.8
in boars, equals 2.4 in females and 3.0 in castrates (Nors-
vin meat quality database). This difference in level of
IMF between sexes are related to average IMF level in
the population, and higher levels gives lager differences
between sexes. The sample mean reflect the level low
level of IMF in the pig population, and is a bit higher
than the average expected value in the mixed-breed
slaughter pig. Selection based on leanness in the last
decades has led to a low level of IMF in pork meat for
many modern big breeds, which is a concern for the in-
dustry. There is a high unfavorable genetic correlation
between leanness and IMF [15], and a single focus on
leanness will eventually lead to a low level of IMF in
pork meat. Norsvin Duroc is selected for slaughter pig
efficiency and leanness in an broad breeding goal with
balanced selection for meat quality in 17 years. There-
fore, the IMF level is almost at the original level, only
0.5 lower, while the leanness has been improved with 7
percentage points in this period. The relatively low level
of IMF and small variation in boars, compared to ordi-
nary slaughter pigs, may also lead to difficulties making
good calibration models on samples selected from the
The calibration errors showed that pred iction using the
cross-validated calibration error (RMSECV) was feasible
(Ta bl es 1 - 3 ). The smallest RMSECV achieved for the
unfolded data set X3 from the non-scaled data. This is
due to the amount of information that is present in all the
histograms and the degree of variation covered by the
PLS method. The results show that scaling does not seem
to have a significant effect on the calibration. This is due
to that the histograms have the same level and variation,
and scaling may have a negative effect on the calibration
by inflating errors hidden in smaller values in the histo-
gram. The histograms are also a size of the number of
pixels present at each CT value, and by scaling the inter-
pretation of these CT values might be diff icult. The effect
of scaling vs. non-scaling on CT values was also shown
Copyright © 2013 SciRes. Openly accessible at http://www. scirp.org/journal/o jas/
J. Kongsro, E. Gjerlaug-Enger / Open Journal of Animal Sciences 3 (2013) 321-325
Tab le 1. Results from multivariate calibration of IMF content
in pig loin using CT; 80 kV energy level. Non-filtered (1) and
median filtered images (2); non-scaled (a) and scaled (b).
Model 1
a 0.37 0.25 0.51 0.01 1
b 0.37 0.28 0.49 0.00 1
Model 2
a 0.36 0.31 0.52 0.00 1
b 0.36 0.30 0.48 0.00 1
Table 2. Results from multivariate calibration of IMF content
in pig loin using CT; 140 kV energy level.
Model 1
a 0.36 0.32 0.45 0.08 2
a 0.31 0.48 0.46 0.08 2
Model 2
a 0.35 0.33 0.53 0.00 2
a 0.34 0.39 0.48 0.02 2
Table 3. Results from multivariate calibration of IMF content
in pig loin using CT; all energy levels and filters.
a 0.30 0.53 0.50 0.02 1
b 0.33 0.41 0.48 0.18 1
by Johansen et al. [14] in calibration of soft tissues in
lamb carcasses, where similar results were found. The
number of components used in the PLS model was also
very low, as the RMSECV values increased with in-
creasing number of PLS components. By adding more
components, better calibration results could be achieved,
but the risk of overfitting a prediction model was highly
With respect to prediction, none of the models were
considered to be feasible. The best prediction model was
achieved by scaling with an RMSEP of 0.48 and a R
squared of 0.18. The prediction errors had the same level
as found by Font-i-Furno ls et al. [1] in prediction of IMF
in post mortem pork loins, but the explained variance (R
squared) was significantly lower. This might be due to
the variation in the sample, our sample set had a lower
variance (std. dev. of 0.44 vs 1.15). However, prediction
errors similar or close to the standard deviation of the
sample set is not acceptable for the prediction models,
and the models in this study should be rejected based on
these results. In vivo images seem to be more “noisy”
and have a lower image quality compared post mortem
images. Blood flow, dynamic water content in muscle,
respiratory motion and muscle contractions may all lead
to more noise and artifacts in the images. Improving im-
age quality by using anesthesia rather than sedation dur-
ing CT scanning may reduce motion artifacts. However,
in a breeding perspective, anesthesia cannot be used due
high costs related to withhold ing of non-selected animals
that is to be sent to abattoir. Median filtering did not
seem to improve the quality of images, thus the salt and
pepper noise does not seem to be the main issue for in
vivo CT images of pigs using the protocols described in
this study. The soft tissue protocol of the reconstruction
of X-ray signals to CT images in the CT scanner hard-
ware may have an effect on the results, however previ-
ously unpublished trials have shown that standard proto-
cols does not seem to improve the results of predictions
of IMF. Therefore the results show that the models used
in this study does not accurately predict the in vivo IMF
percentage of loin muscles using CT at different energy
levels and by median filtering to remove noise.
Computed tomography has proven to have good po-
tential to measure the intramuscular fat in post mortem
meat samples or cuts. However, in vivo results in this
paper show that the image quality is poorer and the pre-
diction results using PLS regression show ed that the pre-
diction errors were close to the standard deviation, mak-
ing the models less feasible for practical use. The use of
anesthesia instead of sedation during CT scanning may
improve image quality, however, from a breeding point a
view, the cost of withholding non-selected animals be-
fore sending them to abattoir due to regulations using
anesthesia on farmed animals is too high to be sustain-
IMF, intramuscular fat; CT, computed tomography;
PLS, partial least square regression; ROI, region of in-
The authors acknowledges support from the Norwegian Research
Council, project n o 210637/O10 and the Nortura meat cooperative.
[1] Font-i Furnols, M., Brun, A., Tous, N. and Gispert, M.
(2013) Use of linear regression and partial least square
regression to predict intramuscular fat of pig loin com-
puted tomography images. Chemometrics and Intelligent
Laboratory Systems, 122, 58-64.
[2] Van Deckel, M.J., Casteels, M., Warnants, N., Van
Copyright © 2013 SciRes. Openly accessible at http://www. scirp.org/journal/o jas/
J. Kongsro, E. Gjerlaug-Enger / Open Journal of Animal Sciences 3 (2013) 321-325
Copyright © 2013 SciRes. http://www.scirp.o rg/journal/ojas/Openly accessible at
Damme, L. and Boucqué, C.V. (1996) Omega-3 fatty ac-
ids in pig nutrition: Implications for the intrinsic and
sensory quality of the meat. Meat science, 44, 55-63.
[3] Gjerlaug-Enger, E., Aass, L., Odegård, J. and Vangen, O.
(2010) Genetic parameters of meat quality traits in two
pig breeds measured by rapid methods. Animal: An in-
ternational journal of animal bioscience, 4, 1832-1843.
[4] Newcom, D.W., Baas, T.J. and Lampe, J.F. (2002) Predic-
tion of intramuscular fat percentage in live swine using
real-time ultrasound. Journal of Animal Science, 80,
[5] Skjervold, H., Gronseth, K., Vangen, O. and Evensen, A.
(1981) In vivo estimation of body composition by com-
puterized tomography. Zeitschrift für Tierzüchtung und
Züchtungsbiologie, 98, 77-79.
[6] Prieto, N., Navaj as, E., Richardson, R., Ross, D., Hy slop,
J., Simm, G., Roehe, R., Glasbey, C. and Fisher, A. (2010)
Predicting beef cuts composition, fatty acids and meat
quality characteristics by spiral computed tomography.
Meat Scie nce, 86, 770-779.
[7] Håseth, T.T., Høy, M., Kongsro, J., Kohler, A., Sørheim,
O. and Egelandsdal, B. (2008) Determination of sodium
chloride in pork meat by computed tomography at differ-
ent voltages. Journal of food science, 73, E333-E339.
[8] Vestergaard, C., Risum, J.R. and Adler-Nissen, J. (2004)
Quantication of salt concentrations in cured pork by
computed tomography. Meat Science, 68, 107-11 3.
[9] Harada, H. and Kumazaki, K. (1980) Relationship be-
tween fat thickness, cross sectional area of M. Longis-
simusthoracis and marbling score and their ultrasonic es-
timates. Nihon Chikusan Gakkaiho, 51, 261-266.
[10] MATLAB (2012) The MathWorks, Inc., Natick.
[11] Ning, C.-Y., Liu, S.-F. and Qu, M. (2009) Research on
removing noise in medical image based on median lter
method. IEEE International Symposium on IT in Medi-
cine & Education, Jinan, 14-16 August 2009, 384-388
[12] Gjerlaug-Enger, E., Kongsro, J., Odegård, J., Aass, L. and
Vangen, O. (2012) Genetic parameters between slaughter
pig efficiency and growth rate of different body tissues
estimated by computed tomography in live boars of Land-
race and Duroc. Animal, 6, 9-18.
[13] Aasmundstad, T., Kongsro, J., Wetten, M., Dolvik, N.I.
and Vangen, O. (2013) Osteochondrosis in pigs diagnosed
with computed tomography: Heritabilities and genetic
correlations to weight gain in specic age intervals. Ani-
mal: An International Journal of Animal Bioscience, 1-7.
[14] Johansen, J., Egelandsdal, B., Røe, M., Kvaal, K. and
Aastveit, A. (2007) Calibration models for lamb carcass
composition analysis using Computerized Tomography
(CT) imaging. Chemometrics and Intelligent Laboratory
Systems, 87, 303-311.
[15] Knapp, P., Willam, A. and Sölkner, J. (1997) Genetic
parameters for lean meat content and meat quality traits
in different pig breeds. Livestock Production Science, 52,