Sunteți pe pagina 1din 15

ULTRASONIC IMAGING 20, 191-205 (1998)

Ultrasound Image Texture Analysis for Characterizing


Intramuscular Fat Content of Live Beef Cattle

NAM-DEUK KIM, VIREN AMIN, DOYLE WILSON, GENE ROUSE


(1) (2,3) (3) (3)
AND SATISH UDPA
(1)

(1)
Department of Electrical Engineering
(2)
Center for Nondestructive Evaluation (CNDE)
(3)
Department of Animal Science
Iowa State University, Ames, IA, 50011
nkim@iastate.edu

The primary factors in determining beef quality grades are the amount and distribution of intramuscu-
lar fat percentage (IMFAT). Texture analysis was applied to ultrasound B-mode images from ribeye
muscle of live beef cattle to predict its IMFAT. We used wavelet transform (WT) for multiresolutional
texture analysis and second-order statistics using a gray-level co-occurrence matrix (GLCM) technique.
Sets of WT- and GLCM-based texture features were calculated from ultrasonic images from 207 ani-
mals and linear regression methods were used for IMFAT prediction. WT-based features included en-
ergy ratios, central moments of wavelet-decomposed subimages and wavelet edge density. The
regression model using WT features provided a root mean square error (RMSE) of 1.44 for prediction of
IMFAT using validation images, while that of GLCM features provided an RMSE of 1.90. The predic-
tion models using the WT features showed potential for objective quality evaluation in the live animals.
KEY WORDS: Beef quality grading; co-occurrence matrix; tissue characterization; texture analysis; ultra-
sound; wavelets.

1. INTRODUCTION

Live animal evaluation for body composition has been an important research goal in ani-
mal agriculture. Recently, there has been a growing interest among researchers for the appli-
cation of ultrasound for food animal evaluation, particularly in the swine and beef industry.
Current ultrasound technology has the potential to accurately estimate carcass body compo-
sition (for yield and quality) in live animals. This should benefit the beef industry, enhancing
genetic selection programs1 and also allowing it to move to a value-based marketing system.
A functional value-based marketing system will provide a means of identifying the value of
individual animals or carcasses. The beef industry has set a high priority for developing an
instrument grading system and ultrasound has been identified as having the potential for
achieving the goal.2 Additionally, evaluation of body composition traits in live animals
could be of tremendous benefit for genetic studies of breeding stock and for sorting of feedlot
cattle. For example, serial ultrasound measurements on live animals could replace the need
for costly serial slaughter designs that are frequently used in growth and genetic studies.
Ultrasound technology was introduced early in the 1950’s as a means for estimating
compositional differences among livestock. However, with advances in real-time ultrasound
technology it has become possible to visualize and characterize quantitative and qualitative
differences in the composition of livestock animals. Houghton and Turlington3 present a
good review of application of ultrasound for feeding and finishing animals.
The ultrasound backscatter image (B-mode) has been successfully used for estimating
backfat thickness and cross-sectional area of the Longissimus dorsi (ribeye) muscle in beef
cattle.4,5 The backfat thickness and ribeye area are important carcass attributes evaluated by
the United States Department of Agriculture (USDA) to determine the yield grade or total

191 0161-7346/98 $18.00


Copyright 1998 by Dynamedia, Inc.
All rights of reproduction in any form reserved.
192 KIM ET AL

TABLE 1
6
Relationship between marbling and chemically-determined IMFAT.

Quality grade Degree of marbling Marbling score IMFAT (%) mean ± SD

Prime moderately abundant 800 − 899 10.42 ± 2.16


slightly abundant 700 −799 8.56 ± 1.60

moderate 600 − 699 7.34 ± 1.50


Choice modest 500 − 599 5.97 ± 1.15
small 400 − 499 4.99 ± 1.10

Select slight 300 − 399 3.43 ± 0.89

Standard traces 200 − 299 2.48 ± 0.59

product. Additionally, the beef carcasses are also evaluated for the quality of meat. This is
primarily determined by the marbling pattern or distribution of intramuscular fat (IMFAT)
on the cross-section of the ribeye muscle between the 12th and the 13th ribs. The marbling
score assigned by the USDA certified inspector primarily determines the quality grade of the
beef.
The relationship between subjective grading of the marbling and chemically-determined
fat content has been reported as linear with a correlation coefficient ranging from 0.7 to
6,7
0.88. Table 1 represents summary results of marbling grades and fat content obtained from
6
chemical analyses of ribeye muscle samples from 518 beef carcasses. The table shows the
carcass quality grade based upon the degree of marbling (as assigned by the USDA inspec-
tor), and the mean and standard deviation of the chemically-determined IMFAT. These re-
sults suggest a linear relationship between the marbling category and the mean of the
IMFAT. The standard deviation within each degree of marbling, however, suggests consid-
erable overlap across the various degrees of marbling. Also, the chemical IMFAT becomes
more variable as the marbling score increases. This relationship reflects the subjectiveness
of the marbling grades and suggests the need for an objective quality evaluation technique.
The potential use of A-mode and B-mode ultrasound for characterizing IMFAT in the
ribeye muscle of beef carcasses has been reported.7-11 Since the speckle patterns in ultra-
sound images are affected by scatter properties, such as size, density, as well as by interroga-
12
tion frequency, the analysis of speckle texture can be used to characterize the tissue. For
example, this approach has been successfully applied for differentiating diffuse liver dis-
13-16
eases, such as alcoholic liver disease or hepatitis from normal liver tissue.
This report presents the results of IMFAT evaluation in live beef animals from ultrasound
images and chemical fat analysis data collected over a period of four years. Image texture
analysis methods were applied to objectively estimate the content and distribution of
IMFAT for beef quality grading purpose. Image Gray Level Co-occurrence Matrix (GLCM)
and Wavelet Transform (WT) based texture analysis techniques and statistical methods were
applied for predicting the IMFAT from the images.

2. IMAGE PROCESSING FOR TEXTURE

Texture is one of the important characteristics used in identifying and characterizing ob-
jects or regions of interest in an image. The commonly-used texture analysis approaches are
based on the probability distribution of gray levels and the texture pattern properties. The
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 193

features are measured from the first-order and second-order statistics. Examples of the
first-order statistical parameters that are employed include the mean, variance, skewness and
17
kurtosis of the one-dimensional gray level histogram. Layer et al observed a significant in-
crease in the ‘mean gray level’ (measure of image brightness) of the ultrasound B-scan im-
age of liver with a regional arrangements of large fat deposits. However, no difference in
image brightness was observed between normal liver tissue and liver steatosis with diffuse
homogeneous fatty infiltration. In general, the first-order statistical parameters are difficult
to apply in practice since they depend significantly on the gain or amplification settings of
the ultrasound equipment.
The second-order statistical methods include the gray-level co-occurrence matrices
(GLCM)18 and the gray-level run-length matrices.19 Haralick et al18 proposed a set of 14 fea-
tures calculated from a co-occurrence matrix whose elements represent estimates of the
probability of transitions from one gray level to another in a given direction at a given
interpixel distance. The features derived from GLCM include contrast, entropy, angular sec-
ond moment, sum average, sum variance and measures of correlation. This technique has
14,17,20 21
been shown to discriminate normal from abnormal liver, spleen pathologies, prostate
22 23 24 8 25
disease, myocardial ischemia, human placentae, and beef IMFAT. Parkkinen et al
showed that GLCM can be applied on different interpixel distances to reveal periodicity in
the texture. However, there is an inherent problem to choose the optimal interpixel distance
in a given situation. Also, the GLCM method, in general, is not efficient since a new
co-occurrence matrix needs to be calculated for every selected angle and inter-pixel distance.
The main difficulty of the above methods is due to the lack of an adequate tool that charac-
terizes different scales of textures effectively. Recent developments in spatial-scale (or fre-
quency) analysis such as Gabor transform, Wigner distribution,26 and wavelet transform27
have provided a new set of multiresolution analytical tools. The fundamental idea underly-
ing wavelets is the ability to analyze the signal at different scales or resolutions. The wavelet
analysis procedure uses a scalable wavelet prototype function, called the mother wavelet.
The fine frequency analysis is performed with a contracted, high-frequency version of the
wavelet function. The coarse frequency analysis is performed with a dilated, low-frequency
version of the same wavelet. This multiscale or multiresolution view of signal analysis is the
essence of the wavelet transforms.27
Many researches have studied wavelet-based multi-resolution for image texture classifi-
28-31 28
cation and segmentation. Chang et al propose a tree-structured wavelet transform and
applied it to texture classification. The new structure makes it possible to zoom into any de-
sired frequency channel for further decomposition. The authors claim the algorithm outper-
formed several other conventional methods. Wu et al31 used wavelet-decomposed images to
derive rotation and gray-scale transform invariant texture features.

3. MATERIALS AND METHODS

A. Data acquisition
A commercially-available, real-time ultrasound system (Aloka 500V by Corrometrics
Medical Systems, Inc., Wallingford, CT) with a 17 cm, 3.5 MHz linear array transducer was
used for scanning live beef cattle in the field. Ultrasonic B-mode images from 212 live beef
animals were acquired by the ISU Animal Science research team during 1991 to 1994.
Equipment settings for TGC (near gain, far gain, and overall gain) and the focal depth were
kept the same for all the scanning sessions. Vegetable oil was used as a coupling medium.
The images of the longissimus dorsi (ribeye) muscle across the 12th and the 13th ribs were
194 KIM ET AL

Frame
Grabber

Linear transducer Personal Computer


Real-Time
B-Mode Imaging (Data Acquisition,
Storage, and Display)
(Aloka 500V)

FIG. 1 Block diagram of the data acquisition system.

hide
128 x 128
ROI
back fat

12th rib

13th rib

FIG. 2 The typical ultrasound B-mode image of ribeye muscle identifying the ROI above and between the ribs
for texture analysis.

digitized using a PC equipped with a frame-grabber board (Cortex1, Imagenation Corp.,


Beaverton, OR). A block diagram of the data acquisition is shown in figure 1.
Within five days after scanning, animals were slaughtered at a commercial packing plant,
and the carcasses were evaluated by a qualified USDA grader to determine the marbling
scores following established USDA protocols. The ribeye slices between the 12th and the
13th ribs were removed and processed chemically at the ISU Meat Laboratory to determine
actual IMFAT using the n-hexane extraction method.

B. Image processing and feature extraction


All images were transferred to a DEC 5000 computer workstation (Digital Equipment
Cor., Maynard, MA) for further processing. PV-WAVE (Precision Visuals Inc., Boulder,
CO) and MATLAB (The Mathworks Inc., Natick, MA), which are commercial signal and
image processing software packages, were used for displaying the digitized images subjec-
tively assigning the image quality, and calculating the texture features. The images were 512
× 286 pixels in size with 256 gray scale levels. A typical acceptable quality image for
IMFAT prediction is shown in figure 2. The features that identify an image with acceptable
quality include: clearly visible hide and subcutaneous fat layer(s) without any sign of uneven
couplant or poor transducer contact; longissimus dorsi muscle area taken from across the
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 195

11th, 12th, and 13th ribs with clearly visible rib shadows; even speckle or texture patterns in
the muscle area. Five images that did not meet the criteria for acceptable quality were dis-
carded. Since each image contained hide, subcutaneous fat, rib bone, and other anatomical
parts apart from the ribeye muscle, a window or the region of interest (ROI) was selected
subjectively so that it contained a good quality image of the ribeye area between and above
the 12th and 13th ribs. In this study, the size of the ROI was 128 by 128 pixels (29 × 29 mm2),
which allowed use of fast discrete wavelet transform algorithms.

1. GLCM-based feature extraction


18
Haralick et al described the GLCM-based texture analysis technique. The following
GLCM features as defined in their study were calculated from ROI images for this study:
Angular second moment
Contrast
Correlation
Variance
Inverse difference moment
Sum average
Sum variance
Sum entropy
Entropy
Difference variance
Difference entropy
Information measure of correlation - I
Information measure of correlation - II
In an earlier study of application of GLCM method for beef quality,8 eight parameters were
found to be significantly correlated with the IMFAT. For this study, all GLCM parameters at
four angles were calculated in order to find the optimal subset for this application as well as
to compare with wavelet texture features.
Also, for each ROI, since muscle structure showed directional pattern in texture, GLCM
parameters were calculated at four different angles (0°, 45°, 90°, 135°). These features were
identified as Cxxx_nn, where xxx denotes the angle and nn denotes the texture feature listed
above. For example, C090_08 denotes the ‘Sum entropy for angle 90°’.

2. Wavelet-based feature extraction


Prior to the wavelet-based feature extraction, all ROIs were preprocessed using simple
gray level stretching (byte-scale transform) for image normalization. Consequently, pixel
values were scaled to fill the entire range of display brightness. The byte-scale transform
was performed by scaling the input ROI as follows:
Gi = T(fi − fmin)/(fmax − fmin) (1)
where fi is the original intensity in the range (fmax, fmin) and gi is the corresponding scaled inten-
sity to lie in (0, T). In this study, T is 255.
After preprocessing, WT-based features were computed using the procedure illustrated in
figure 3. In figure 4, the 2-D wavelet decomposition27 was illustrated with the scheme used
for the naming of subimages used in this paper. As shown in the figure, the decomposition
32
was taken up to three levels using the Haar wavelet as a basis function. Ten subimages were
generated as a result of decomposition. The following features were calculated in this study:
Central moments: The nth moment of x about its mean is defined by Eq. (2). The second
moment µ2(x) represents the variance that is a measure of the local activity in the amplitudes.
196 KIM ET AL

FIG. 3 The procedure employed for extracting WT features from ultrasound images of beef ribeye muscle.

3 3
S1 D2 2

3 3
D2
D3 D4 D2
1

2 2
D3 D4

1 1
D3 D4

FIG. 4 The three scale level 2-D wavelet transform of an image ROI (left) and the naming of the sub-band im-
ages (right) used in this study. Dmn contains vertical (m = 2), horizontal (m = 3) and diagonal (m = 4) information,
where n is the transformation level. S contains the lowest frequency information in both the x and y directions.

The third moment µ3(x) is a measure of the skewness of the histogram while the fourth mo-
ment µ4(x) is a measure of its kurtosis. Initially, four moments from each subimage of the
wavelet-decomposed image were computed to assess the variations in gray scale in the dif-
ferent subimages. Since the mean and the skewness didn’t show good correlation to the
IMFAT, they were not used in further analysis. The features were identified with VARnn
and KTSnn where nn denotes the subimage number shown in figure 4. For example, VAR32
3 1
denotes the ‘variance of subimage D2 ’ and KTS12 denotes the ‘kurtosis of subimage D2 ’,
k
µ n ( x ) = ∑ ( x i − m) n p( x i )
(2)
i =1

where
k
m = ∑ x i p( x i )
i =1
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 197

number of pixels with gray level x


p(xi) = i
total number of pixels

Energy ratios: The energy of the 2-D wavelet decomposed subimages was computed us-
3 3 2
ing Eq. (3). Three energy ratios between S1 and the feature extracted subimages (D4 , D4 ,
1
D4 ) in the lowest scale level were computed to characterize the lowest scale image. Three ra-
tios between vertical feature subimages (D 23 D 22 , D 23 D 21 , D 22 D 21 ), three from the horizon-
tal feature subimages ( D 33 D 32 , D 33 D 31 , D 32 D 31 )and three from diagonal feature extracted
subimages ( D 43 D 42 , D 43 D 41 , D 42 D 41 ) were computed to compare the amount of energy in
each directional subimage. The features are denoted as Enn/Exx where nn and xx index the
subimages described in figure 4.

[ ]
N 2
1 (3)
∑x
2
Energy = i = E xi
N i =1

Wavelet edge density: The coarseness/smoothness of texture can be measured from the
density of the edge pixels. After 2-D wavelet decomposition, the low frequency subimages
were replaced with zeros so that the high frequency components were emphasized. The 2-D
wavelet reconstruction technique33 was used to reconstruct the high frequency component
emphasized image, and the energy of reconstructed image was computed using Eq. (3) to es-
timate the density of edges. This is identified as Wedge in this study.
In summary, ten energy ratios, ten variances, ten kurtosis and the Wedge features were
calculated for this study.

C. Statistical analysis
Statistical correlation and multiple linear regression methods were applied using SAS
software (SAS Institute, Inc., Cary, NC). The stepwise selection of the features from all im-
ages was performed with the SAS PROC REG method.34 The actual (chemical) IMFAT was
used as the dependent variable and linear relationship between parameters and IMFAT was
tested. The best subset of variables was selected based on an optimum combination of the
coefficient of determination (R2), Mallows statistic (Cp), and the root mean square error
(RMSE). The Mallows statistic Cp is a measure of total squared error for a subset model con-
taining p independent variables and indicates when variable selection is deleting too many
variables.35
The selected subset of variables was used to develop a linear multiple regression model
that was analyzed further for validation testing. The model coefficients were developed us-
ing randomly-selected 80% of the data set and the testing was performed with the remaining
twenty percent of the data. This process was repeated five times, each time with different
randomly-selected training and testing sets. The RMSE and R2 were calculated for both the
training and testing sets. The correlation coefficients were calculated between the actual and
the predicted IMFAT for the testing data set. Also, further testing for prediction bias was
done using the regression analysis of the predicted versus actual IMFAT. The slope and in-
2
tercept of the regression, RMSE and R of validation testing were used for evaluating the ac-
curacy and robustness of the prediction models. The residual or the prediction error was
calculated as actual IMFAT minus predicted IMFAT and the distribution of the residuals
against actual IMFAT was plotted for further testing.
The GLCM and WT based features were analyzed separately for model development and
validation allowing comparison of usefulness for our application.
198 KIM ET AL

50
45
40
35

Frequency
30
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10 11
IM F AT (% )

FIG. 5 The distribution of the IMFAT.

4. RESULTS AND DISCUSSION

A. Distribution of IMFAT data


The chemically-determined IMFAT values for 207 ribeye samples ranged from 1.10% to
11.86% with a mean of 5.99 and standard deviation of 2.02. As shown in the figure 5, the
data in the range of 1 to 7% could be approximated as normal distribution. However, for the
higher IMFAT, the distribution did not follow the normal distribution and there were only
about 8% (17 of 207) of samples with more than 9% IMFAT. The differences in the texture
patterns in ROIs from ribeye muscles with different IMFAT are shown in figure 6. As shown
in figure 6, the lower IMFAT image has wider range of speckle patterns than the higher
IMFAT image. The higher IMFAT image appears to be more homogeneous than the lower
IMFAT image. These differences justified this study and the use of texture analysis methods
as pattern recognition tools. The continuously-varying texture patterns justify the use of the
regression analysis technique for IMFAT prediction as opposed to classification techniques.

B. Correlation of GLCM- and WT-based features to IMFAT


In plotting each image parameter against actual IMFAT, a linear relationship was ob-
served. A typical relationship is shown in figure 7 where parameters E22/E12 and Wedge are
plotted against actual IMFAT. Most parameters followed a general linear relationship; how-
ever, for some parameters, the higher IMFAT samples did not follow the linear relationship
(e.g., parameter E22/E12 in figure 7a).
The correlation table of the selected GLCM and WT features and IMFAT is shown in table
2. All the coefficient values in the table are significant (p < 0.05) unless otherwise specified.
The first row in the table indicates the correlation coefficients of the features with the actual
IMFAT while the remaining rows represent the cross-correlation between the features.
The correlation coefficients of the GLCM-based features to IMFAT range from 0.25 to
0.46 (Table 2). It is difficult to determine what these texture features represent. However,
one can make some intuitive guesses as to the properties represented by some of the features.
For example, the higher value of the Correlation feature might suggest higher linear depend-
encies of gray levels within the ROI, and one might expect the Entropy feature to take higher
value for more complex images.
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 199

(a) (b)

(c) (d)

FIG. 6 Ultrasound B-mode images from four ribeye muscles with different percentage of IMFAT. (a) 2.48%,
(b) 4.91 %, (c) 7.47 %, (d) 10.3 %.

140 0.35
120 0.3
100 0.25
E22/E12

Wedge

80 0.2
60 0.15
40 0.1
20 0.05
0 0
0 5 10 15 0 5 IMFAT (%) 10 15
IMFAT (%)
(a) (b)
FIG. 7 Typical plots of relationship between image texture parameters (E22/E21 (a), Wedge (b)) and the actual
IMFAT.

The correlation coefficients of the WT features to IMFAT were between 0.49 and 0.59.
The energy ratio features showed good correlation with the actual IMFAT and showed that
the differences in the energy amounts in each scale subimage were found to be very useful
for classifying the coarseness of the texture. The negative correlation value means that the
lower IMFAT images have more energy at the lower scale subimages than the higher
IMFAT images. This result can be explained by comparing the coarseness of original image
texture as shown in figure 6. Thus, 2-D wavelet transform acts like a bandpass filter and pro-
vides multiresolutional analysis of images.
The variance and kurtosis features have a negative correlation with IMFAT except those
of the lowest scale subimage (VAR31, KTS31). The negative correlation indicates that there
are many gray level changes in the images with lower IMFAT. The variance features of the
lower scale subimages (level two and three) showed better correlation than those of high
200 KIM ET AL

TABLE 2
Correlation table of selected WT and GCLM based features for further development.

E34/E14 E22/E12 E23/E13 Wedge VAR31 VAR34 KTS23 C90_8 C90_13 C135_3 C135_4 C135_6 C135_8 C135_13

IMFAT - 0.59 - 0.57 - 0.59 0.52 0.51 - 0.49 - 0.50 0.25 0.32 0.46 0.36 0.28 0.27 0.44
∗ ∗ ∗ ∗ ∗
E34/E14 1.00 0.74 0.67 - 0.22 - 0.30 0.69 0.49 0.08 0.02 - 0.25 0.02 0.05 0.07 - 0.13
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗
E22/E12 1.00 0.51 - 0.13 - 0.15 0.39 0.30 0.09 0.16 −0.09 0.09 0.07 0.10 0.02

E23/E13 1.00 - 0.34 - 0.35 0.37 0.60 - 0.20 -0.16 - 0.20 - 0.19 - 0.20 - 0.20 - 0.19

Wedge 1.00 0.59 - 0.38 - 0.59 0.68 0.73 0.64 0.75 0.81 0.71 0.77

VAR31 1.00 - 0.32 0.77 0.40 0.22 0.48 0.60 0.28 0.45 0.39
∗ ∗
VAR34 1.00 0.73 0.03* - 0.26 - 0.55 - 0.14 0.01 0.00 - 0.43

KTS23 1.00 - 0.28 - 0.42 - 0.59 - 0.35 - 0.29 - 0.30 - 0.55

C90_8 1.00 0.81 0.34 0.80 0.84 0.99 0.56

C90_13 1.00 0.74 0.76 0.61 0.83 0.83

C135_3 1.00 0.55 0.24 0.40 0.92

C135_4 1.00 0.84 0.83 0.77

C135_6 1.00 0.85 0.54

C135_8 1.00 0.62

*The values are not significant to the 5% level (p>0.05).

scale subimages (level one). This means that significant discriminatory information can be
extracted from the lower scale subimages.
The Wedge feature also showed good correlation with IMFAT. The negative correlation
indicates that there is higher density of edge pixels in the images with lower IMFAT. The
image with higher IMFAT has fewer edges and tends to be more homogeneous.

C. Linear regression models for predicting IMFAT


Based on the statistical criteria described earlier, several subsets of features were selected
for each of GLCM and WT methods for developing and analyzing the regression models.
Figure 8 shows a decrease in RMSE with increasing number of variables. However, the
RMSE does not decrease significantly when more than seven variables are employed. The
figure also shows that the WT features provided less RMSE than the GLCM features. Fur-
ther analysis was done on regression models using seven parameters from each of GLCM
and WT methods.
The selected seven GLCM features for IMFAT prediction included five of 135° features
and two of 90° features. The 135° features were Correlation (C135_03), Variance
(C135_04), Sum Average (C135_06), Sum Entropy (C135_08), and Information Measure of
Correlation-II (C135_13). The 90° parameters were Sum Entropy (C090_08) and Informa-
tion Measure of Correlation-II (C090_13). Combining the GLCM features from different
angles provided more information than the features from one angle.
The selected seven WT features for IMFAT prediction included three energy ratios
(E34/E14, E22/E12, and E23/E13), two variances (VAR31 and VAR34), one kurtosis (KTS23), and the
Wedge feature.
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 201

2.1

GLCM
1.9

1.8
RMSE

1.7

1.6

1.5
WT

1.4

1.3
0 2 4 6 8 10 12 14
Number of parameters
FIG. 8 The variation of RMSE with the number of GLCM and WT variables used in regression models.

For each method, the validation testing of the regression models was done using five ran-
domly selected training and testing data sets as described earlier. Table 3 shows the results
2
from the validation testing. The first two columns represent RMSE and R of the prediction
output using the training data set, and the last two columns show those of the testing data set.
The average RMSE of 1.40 for the WT model indicates that about 68% of the new samples
can be predicted within 1.40 error in IMFAT (assuming normal distribution). The R2 indi-
2
cates the variation in the data that is explained by the model. For example, average R of 0.62
for the WT model suggests that about 62% of the variation in the IMFAT of testing data set is
explained by the prediction model. As the table shows, the RMSE of GLCM models are
2
higher than that of WT models and R of GLCM models are smaller than that of WT models
in both training and testing models.
Table 4 shows the slopes and intercepts of the models and the correlation coefficients of
the predicted IMFAT with the actual IMFAT. The slopes of WT models are around one,
which are considered as a good fit. The intercept values in both GLCM and WT models were
not significant (p >> 0.05). The correlation coefficients of the predicted IMFAT from the
WT models were higher than those of the GLCM-based models. This means again that the
WT feature set provides much more distinguishable information about the ultrasonic texture
than the GLCM feature set. The correlation coefficient values of the WT models are consid-
ered good for this application where biological variability is inherent in the problem.
Figure 9 shows the distribution of the residuals (calculated as actual IMFAT minus pre-
dicted IMFAT). The plot of the residuals against the actual IMFAT (Fig. 9a) shows even dis-
tribution of the residuals around zero for the IMFAT less than about 9%. The residuals were
high and positive for the IMFAT greater than about 9%, indicating underprediction of the
very high IMFAT samples. Figure 9b shows the plot of cumulative frequency of absolute re-
siduals. As seen from the plot, about 87% of samples were predicted to within 1.5% IMFAT
error for WT model.
202 KIM ET AL

TABLE 3
Results of linear regression models using GLCM and WT based features.

Training set Testing set


2 2
Data set RMSE R RMSE R

GLCM set 1 1.79 0.42 2.02 0.25

GLCM set 2 1.89 0.37 1.75 0.42

GLCM set 3 1.89 0.37 1.76 0.39

GLCM set 4 1.81 0.39 2.07 0.29

GLCM set 5 1.85 0.38 1.91 0.38

GLCM average 1.85 0.38 1.90 0.34

WT set 1 1.44 0.63 1.27 0.69

WT set 2 1.35 0.67 1.64 0.51

WT set 3 1.47 0.62 1.14 0.75

WT set 4 1.37 0.65 1.57 0.59

WT set 5 1.37 0.66 1.59 0.56

WT average 1.40 0.65 1.44 0.62

TABLE 4
The statistical test results of regression models using testing data images.

Data set Slope Intercept* Correlation coefficient

GLCM set 1 0.63 1.76 0.50


GLCM set 2 1.12 − 0.26 0.64
GLCM set 3 0.96 0.14 0.62
GLCM set 4 0.96 0.03 0.54
GLCM set 5 1.13 − 0.64 0.61
Mean ± SD 0.96 ± 0.20 0.21 ± 0.92 0.58 ± 0.06
WT set 1 0.96 0.37 0.83
WT set 2 0.90 0.29 0.71
WT set 3 1.05 − 0.27 0.86
WT set 4 1.06 − 0.35 0.77
WT set 5 0.89 0.50 0.75
Mean ± SD 0.97 ± 0.08 0.11 ± 0.39 0.78 ± 0.06

*None of the intercepts were significant (p>0.05).

The average of absolute residual values of each model was calculated and is presented in
figure 10. The IMFAT range was divided into four groups to show how the samples in each
range affect the overall prediction accuracy. For the lower IMFAT group samples, there are
slightly higher errors than the middle group samples. In the case of higher group samples,
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 203

3
100%
2 80%

RESIDUAL (%)
1 60%
0 40%
0 2 4 6 8 10 12
-1 20%
-2 0%
-3 <0.5 <1.0 <1.5 <2.0 <2.5 <3.0
Actual IMFAT (% ) Prediction residual (%)

(a) (b)

FIG. 9 Residual (error of prediction) analysis for model validation: (a) Distribution of residuals with actual
IMFAT, (b) Cumulative frequency versus residuals.

4
3.5
ABS(RESID)

3
2.5 GLCM
WT
2
1.5
1
0.5
0

0-3% 3-6% 6-9% 9+ %


IMFAT range (%)

FIG. 10 The average of absolute residual value of the prediction models using the WT and the GLCM features.

the errors in both GLCM- and WT-based models are high. The high level of errors may be
due to the fact that we had only about 8% (17 of 207) of images with more than 9% IMFAT.
In addition, the visible differences between the images in the range are very small. In each
range, the residuals of GLCM models were higher than those of WT models.
These statistical tests indicate that the WT features provide better texture analysis tool and
IMFAT prediction accuracy than GLCM features.

5. CONCLUSION

In this study, the texture of ultrasonic B-mode images were analyzed using co-occurrence
matrices and wavelet-based methods to predict IMFAT in live beef ribeye muscle. The cor-
relation coefficients and prediction performance of both features were also compared.
The GLCM-based method provided spatial texture information at different angles. Com-
bining the GLCM features from different angles provided more information than the fea-
tures from one angle. However, using only GLCM features provided an average RMSE of
prediction of 1.90% IMFAT and a correlation coefficient of 0.58 on the test data set.
WT-based features provided significant information about the textural differences, such
as coarseness, orientation and variations using spatial-scale localized information of tex-
tures. Since the amount of energy in the wavelet-decomposed subimages were different for
204 KIM ET AL

coarse and smooth textures, the energy ratio features could distinguish homogeneous high
IMFAT images from heterogeneous low IMFAT images very well. The variance and
kurtosis features showed that variations in pixel values at all subimages of low IMFAT tex-
tures were different from those of high IMFAT textures. The regression model using seven
WT features provided an average RMSE of 1.44 and the correlation coefficient of 0.78 on the
test data set, which is comparable to the correlation of the visual marbling score on the car-
cass to the actual IMFAT. These prediction models showed potential for objective quality
grading in the live animals. Also, based on the validation testing, WT-based features outper-
formed GLCM-based features in differentiating textures.
Several pattern classification techniques, such as tree-based classification,36 artificial neu-
ral networks and fuzzy-C means algorithm are also being applied to preclassify samples into
two groups (less than 9%, and higher than 9% in IMFAT). It is expected that by combining
the preclassifying techniques and linear regression prediction methods, the overall predic-
tion accuracy can be further improved, particularly for the very low and very high IMFAT
images.

REFERENCES

1. Wilson, D. E., Application of ultrasound for genetic improvement, J. Animal Sci. 70, 973-983 (1992).
2. Cross, H. R. and Whittaker, A. D., The role of instrument grading in a beef value-based marketing system, J.
Animal Sci. 70, 984-989 (1992).
3. Houghton, P. L. and Turlington, L. M., Application of ultrasound for feeding and finishing animals: a review,
J.Animal Sci. 70, 930-941 (1992).
4. Perkins, T. L., Green, R. D. and Hamlin, K. E., Evaluation of ultrasonic estimates of carcass fat thickness and
longissimus muscle area in beef cattle, J. Animal Sci. 70, 1002-1010 (1992).
5. Smith, M. T., Oltjen, J. W., Dolezal, H. G., Gill, D. R. and Behrens, B. D., Evaluation of ultrasound for predic-
tion of carcass fat thickness and longissimus muscle area in feedlot steers, J. Animal Sci. 70, 29-37 (1992).
6. Savell, J. W., Cross, H. R. and Smith, G. C., Percentage ether extractable fat and moisture content of beef
longissimus muscle as related to USDA marbling score, J. Food Sci. 51, 838 (1986).
7. Park, B., Whittaker, A. D., Miller, R. K. and Hale, D. S., Predicting intramuscular fat in beef longissimus mus-
cle from speed of sound, J.Animal Sci. 72, 109-116 (1994).
8. Amin, V., Wilson, D., Roberts, R. and Rouse, G., Tissue characterization for beef grading using texture analy-
sis of ultrasonic images, in Proc. 1993 IEEE Ultrason. Symp., pp. 969-972 (1993).
9. Amin, V., Izquirdo, M., Wilson, D., Rouse, G. and Roberts, R., Ultrasound evaluation of quality attributes in
live beef animals using B-mode image processing and pattern recognition, Ann. Biomed. Engin. 23, Suppl. 1, S-67.
10. Patel, A., Amin, V., Roberts, R., Wilson, D. and Rouse, G., Application of A-mode ultrasound to character-
ize intramuscular fat content, Rev. Prog. Quant. Nondestructive Eval. 14, 1781-1788 (1995).
11. Whittaker, A. D., Park, B., Thane, B. R,. Miller, R. K. and Savell, J. W., Principles of ultrasound and mea-
surement of intramuscular fat, J. Animal Sci. 70, 942-952 (1992).
12. Feleppa, E. J. and Yaremko, M.M., Ultrasonic tissue characterization for diagnosis and monitoring, IEEE
EMB Magazine, 18-26 (1987).
13. Bleck, J. S., Ranft, U., Gebel, M., Hecker, H., et al, Random field models in the texture analysis of ultrasonic
images of the liver, IEEE Trans. Med. Imag. 15, 796-801 (1996).
14. Wu, C., Chen, Y. and Hsieh, K., Texture features for classifications of ultrasonic liver images, IEEE Trans.
Med. Imag. 11, 141-152 (1992).
15. Thijssen, J. M., Oosterveld, B. J., Hartman, P. C. and Rosenbusch, G. J. E., Correlations between acoustic
and texture parameters from RF and B-mode liver echgrams, Ultrasound Med. Biol. 19, 13-20 (1993).
16. Oosterveld, B. J., Thijssen, J. M., Hartman, P. C., Romijn, R. L. and Rosenbusch, G. J. E., Ultrasound attenu-
ation and texture analysis of diffuse liver disease: methods and preliminary results, Phys. Med. Biol. 36, 1039-1064
(1991).
TEXTURE ANALYSIS OF FAT CONTENT OF BEEF CATTLE 205

17. Layer, G., Zuna, I., Lorentz, A., et al., Computerized ultrasound B-scan texture analysis of experimental
fatty liver disease: influence of total lipid content and fat deposit distribution, Ultrasonic Imaging 12, 171-188
(1990).
18. Haralick, R. M., Shanmugam, K. and Dinstein, I., Textural features for image classification, IEEE Trans.
Syst. Man Cyber. 3, 610-621 (1973).
19. Dasarathy, B. and Holder, E., Image characterization based on joint gray level-run length distributions, Patt.
Recog. Lett. 12, 497-502 (1991).
20. Haberkorn, U., Zuna, I., Lorentz, A., et al., Echographic tissue characterization in diffuse parenchymal liver
disease: correlation of image structure with histology, Ultrasonic Imaging 12, 155-170 (1990).
21. Nicholas, D., Nassiri, D., Garbutt, P. and Hill, C., Tissue characterization from ultrasound B-scan data, Ul-
trasound Med. Biol. 12, 135-143 (1986).
22. Basset, O., Sun, Z., Mestas, J. L. and Gimenez, G., Texture analysis of ultrasonic images of the prostate by
means of co-occurrence matrices, Ultrasonic Imaging 15, 218-237 (1993).
23. Mc Pherson, D., Aylward, P., Knosp, B., et al., Ultrasound characterization of acute myocardial ischemia by
quantitative texture analysis, Ultrasonic Imaging 8, 227-240 (1986).
24. Morris, D. T., An evaluation of the use of texture measurements for the tissue characterization of ultrasonic
images of in vivo human placentae, Ultrasound Med. Biol. 14, 387-395 (1988).
25. Parkkinen, J. and Oja, E., Detecting texture periodicity from the cooccurrence matrix, Patt. Recog. Letts. 11,
43-50 (1990).
26. Reed, T. R. and Wechsler, H., Segmentation of textured images and gestalt organization using spatial/spa-
tial-frequency representations, IEEE Trans. PAMI 12, 1-12 (1990).
27. Mallat S., A theory for multiresolution signal decomposition: The wavelet representation, IEEE Trans.
PAMI 11, 674-693 (1989).
28. Chang T. and Kuo, C. J., Texture analysis and classification with tree-structures wavelet transform, IEEE
Trans. Image Proc. 2, 429-441 (1993).
29. Cater, P. H., Texture discrimination using wavelets, applications of digital image processing, Proc. SPIE
1567, 432-438 (1991).
30. Laine, A. and Fan, J., Texture classification by wavelet packet signatures, IEEE Trans. PAMI 15, 1186-1191
(1993).
31. Wu, W. and Wei, S., Rotation and gray scale transform invariant texture identification using wavelet decom-
position and Hidden Markov Model, IEEE Trans. Image Proc. 5, 1423-1434 (1996).
32. Vetterli, M., Wavelets and filter banks: theory and design, IEEE Trans. Sig. Proc. 40, 2207-2232 (1992).
33. Mallat, S., Multifrequency channel decompositions of images and wavelet models, IEEE Trans. ASSP,
2091-2110 (1989).
34. SAS/STAT User’s guide, Release 6.03 (SAS Institute Inc., Cary, NC, 1988).
35. Freund, R. J. and Littell, R. C, Sas System for Regression, 2nd Ed. (SAS Institute Inc., Cary, NC, 1991).
36. Kim, N-D., Amin, V., Wilson, D., Rouse, G. and Udpa, S., Texture analysis using multiresolution analysis
for ultrasound tissue characterization, Rev. Pro. Quant. Nondestructive Eval. 16B, 1351-1358 (1997).

S-ar putea să vă placă și