Muscle Study with Multispectral Image Analysis

Full text

MUSCLE STUDY WITH MULTISPECTRAL IMAGE ANALYSIS
S. Abouelkaram
1
, S. Chauvet
1
, M. El Jabri
2
, P. Strydom
1
, D. Bertrand
3
, J-L. Damez*
1
1. IRA, Centre de Clermont-Fd/Theix, 63122 Saint Genès Champanelle, France. 2. Laboratoire de
Mathématiques, Université Blaise Pascal. 63000 Clermont-Ferrand, France. 3. Unité de Sensométrie et de
Chimiométrie EITIAA- IRA, EITIAA, 44322 ATES CEDEX 3, France. Email: damez@clermont.inra.fr
Key Words: Multispectral imaging, muscle structure, meat quality.
Introduction
A feasibility study of multispectral image analysis (MIA) was undertaken in the aim to relate meat components
to sensorial and physical properties. The MIA technique was used during last decade for the assessments of food
products (Novales et al., 1996) especially to inspect poultry carcass in Visible/NIR range (Park et al., 1996).
Contaminations of apples were also studied using multispectral fluorescence imaging (Kim et al., 2005). In this
work we applied MIA to study meat component by targeting myofibres, collagen and lipids which play a major
role in meat sensorial quality. In a previous study we have established the specific spectral signals of muscle
components (Skjervold et al., 2003). This paper describes the multispectral imaging approach we have used in
order to characterise muscle slices. A specific bench was built and settled to optimise UV-Visible image of
muscle samples according to specific spectral tissue information. A method of exploiting the resulting multiway
image data was developed. It covers the different steps of data analysis i.e.: from the processing of the raw image
signal to the data statistical analysis. The image segmentation was performed using a multiway method based on
multispectral techniques (Novales et al., 1996). This method allowed to reach information of muscle
composition and structure and particularly the distribution and quantification of the connective tissue. The
analysis involved segmentation of the connective tissue making to directly identify collagen and lipids, while
myofibres information was obtained indirectly. Muscles of six animals were analysed in real conditions to
demonstrate the feasibility of the multispectral technique implemented for muscle characterisation.
Materials & Methods
Image analysis was performed on bovine muscles. Frozen biceps femoris muscles of six animals were cut into
samples of 20 mm thickness. The set of images was produced with an optical bench using a UV sensitive digital
camera (Sony XCD-SX900UV) (Figure1). Sample images of 5 cm X 4.5 cm were recorded. They were acquired
at different wavelengths in UV–Visible range. The light source was a 300W maximum power Xenon light used
for pure white light or ultra violet light filtered at 320 and 380 nm. Each wavelength corresponded to one
channel of the multispectral image.
In order to improve image quality both median and Gaussian
low-pass filters were applied. The image segmentation is a
chemometric technique based on pixel classification. It was
used here to classify the three components of muscle tissue:
myofibres, collagen and lipids. The chemometric technique
consists of a Matlab toolbox developed by the authors. Image
features were extracted from segmented images according to
(Abouelkaram et al., 2003). Parameters were selected in order
to find the most pertinent ones to use in the prediction models.
Data statistical studies, mainly based on multiple linear
regression, made it possible to build specific statistical models
which relate image features to meat properties.
Sensory parameters, including tenderness were assessed on
cooked meat by a sensory panel while mechanical tests were
performed on raw meat (Maunier-Sifre, 2005).
Results & Discussion
The multispectral imaging technique involves the following steps: image acquisition, image pre-processing,
multiway segmentation, image features extraction and statistical analysis (Figure 2). The main default on the raw
images was a non uniform lighting. We solved this problem by using a low pass Gaussian filtering (Figure 3). A
typical three grey level segmented image resulting from the multiway method is given on Figure 4. It shows the
discrimination between myofibrils (in black), collagen (in grey) and lipids (in white) allowing assessment of
their quantities and distributions.
Figure 1: Multispectral imaging bench.

On resulting segmented images of muscle of six animals, image features were extracted and related to
composition, mechanical and sensory measurements. Because of the low observations number, the predicting
models were built with only the most relevant image parameter.
Results in predicting composition and sensory parameters are presented on Figure 5. The R² coefficients of the
prediction models were satisfying. However, due to a small number of observations, these results must be
confirmed. The highest R² were obtained when predicting the proximate composition: collagen (R²=0.88) and
lipids (R²=0.87). Tenderness gave the lowest R² (0.75) while R² associated with juiciness was equal to 0.84. Note
that collagen and lipids could be assumed as primary parameters as they were visible on the image, whereas
sensory characteristics were supposed to be second rank parameters. This can explain the score of tenderness
which is also multifactorial.
Conclusions
This work was focused on the application of a multispectral imaging technique to muscle characterisation. It was
needed to develop specific equipment and method. The method covers the different steps of data analysis from
the image pre-processing to the data statistical analysis. To complete this study, preliminary experiments were
undertaken on muscles of six animals. The results obtained on this small sample population show the potential of
MIA in estimating composition and structural organisation and predicting sensory parameters. These promising
results have to be confirmed on a larger population.
References
1. Abouelkaram S., Berge P., Hocquette J. F., Culioli J. and Listrat A., 2003. Image study analysis of the
relationship between collagen content and distribution of the perimysial connective network in a bovine
muscle. Sciences des Aliments, 231 : 166-170.
2. Kim, MS, Lefcourt, AM, Chen, YR, Tao, 2005. T. Automated detection of fecal contamination of apples
based on multispectral fluorescence image fusion. Journal of Food Engineering, 71 (1): 85-91.
3. Maunier-Sifre L., 2005. Organisation spatiale du tissu conjonctif intramusculaire : relation avec la texture de
la viande bovine, Université d’Auvergne Université Blaise Pascal, N° d’ordre : 422, PhD Thesis.
4. Novales B., Bertrand D., Devaux M.F., Robert P., Sire A., 1996. Multispectral Fluorescence Imaging of
Food Products. Journal of the Science of Food and Agriculture, 71: 376-382.
5. Park, B. Chen, Y. R., Huffman R. W., 1996. Integration of Visible/NIR Spectroscopy and Multispectral
Imaging for Poultry Carcass Inspection. Journal of Food Engineering, 30:197-207.
6. Skjervold, P. O., Taylor, R. G., Wold, J. P., Berge, P., Abouelkaram, S., Culioli, J., and Dufour, E., 2003.
Development of intrinsic fluorescent multispectral imagery specific for fat, connective tissue and myofibers
in meat. Journal of Food Science, 68: 1161-1168.
Raw image Gaussian filtered image
Figure 3: Image pre-processing.
R²
Collagen
Lipid
Mecha.
Tender.
Juciness
0.5
0.6
0.7
0.8
0.9
1
Figure 5: Meat quality prediction results.
Image set Segmented image
Figure 4: Multiway segmentation.
Figure 2: MIA steps.
Image acquisition
Image pre-processing
Multiway segmentation
Image features extraction
Statistical processing
Results