Content uploaded by Yung-Ming Jeng
Author content
All content in this area was uploaded by Yung-Ming Jeng
Content may be subject to copyright.
Liver Fatty Change Classification Using 25MHz High Frequency
Ultrasound
Wen-Chun Yeh, Yung-Ming Jeng
1
, Cheng-Han Li, Po-Huang Lee
2
and Pai-Chi Li
Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, R.O.C.
1
Department of Pathology and
2
Department of Surgery, National Taiwan University Hospital, Taipei,
Taiwan, R.O.C.
Abstract - Liver fatty change (steatosis), the
accumulation of fat within liver cells, is a
common histological finding in liver biopsies.
The histological changes of steatosis such as the
distribution of fatty droplets cannot be resolved
by conventional ultrasound. High frequency
ultrasound (≧20 MHz), on the other hand, has
the potential to reveal more detail of steatosis. In
this study, B-mode images of high frequency
ultrasound of 19 fresh human liver samples were
obtained to evaluate ultrasound’s ability in
determining the steatosis grade. B-mode images
were acquired by a mechanically controlled
25MHz single crystal probe. Image features
derived from gray level concurrence and
non-separable wavelet transform were extracted
to classify steatosis grade using a classifier
known as the support vector machine. Each liver
sample subsequently underwent histological
examination and liver steatosis was graded from
0 to 3 (0: no fatty change; 1: less than 33% of
hepatocytes affected; 2: 33% to 66% of
hepatocytes affected; and 3: more than 66% of
hepatocytes affected). The four grades were then
combined into 2 classes- grade 0 and grades 1-3
(non-steatosis and steatosis); 3 classes- grade 0,
grade 1 and grades 2-3 (non-steatosis, mild
steatosis, and prominent steatosis); and 4 classes
(all different grades). Classification results were
correlated with histology. It is shown that the
best classification accuracy of 2, 3 and 4 classes
were 90.5 %, 85.8 % and 82.6%, respectively.
Compared with the results acquired from
conventional ultrasound at 7 MHz (the best
classification accuracy of 2, 3 and 4 classes were
81.6 %, 75.8 % and 74.2 %, respectively), the
classification accuracy was clearly better at 25
MHz. Thus, liver steatosis can be more
accurately characterized using high frequency
B-mode ultrasound. Limitations and potential
solutions of liver steatosis using high frequency
ultrasound are also discussed.
Introduction
Liver fatty change (steatosis), the
accumulation of fat within liver cells, is a
common histological finding in liver biopsies.
Steatosis in nonalcoholic patients is widely
believed to be a benign condition with little or
no risk of disease progression. However, marked
steatosis may result in necroinflammation,
fibrosis and even cirrhosis [1]. B-mode
ultrasound is usually recommended as a first step
to evaluate the patient with mild elevated values
of serum liver enzymes and it has been used for
liver tissue classification. Steatosis change
increases the echogenicity of liver relative to that
of the kidneys and this characteristic has been
adopted in clinical diagnosis. Steatosis can also
be classified using image features. Mojsilovic et
al. proposed a non-separable wavelet transform
method to extract the B-mode image features for
classification of healthy liver, cirrhosis and
steatosis [2]. However, B-mode ultrasound at
lower frequencies (≦ 10 MHz) generally does
not have adequate resolution to reveal
morphological changes of steatosis.
High frequency ultrasound (≧ 20 MHz)
has superior resolution, thus allowing imaging
finer structures of tissue. Nevertheless, high
frequency ultrasound has not been widely
applied to liver tissue characterization primary
due to its limited penetration. The intra-vascular
approach (through inferior vena cava to hepatic
veins) provides a potential solution for the
application of high frequency ultrasound to
human liver. In this study, we evaluate the
feasibility of tissue characterization of human
liver using 25 MHz B-mode ultrasound in Vitro.
Results of the feasibility study can be a
foundation of intra-vascular high frequency
ultrasound for liver disease classification.
Image Feature Extraction and Classification
In this study, the gray level concurrence
(GLC) method and the non-separable wavelet
transform (NSW) method were adopted for
image features extraction. With these features,
steatosis grades were classified using the support
vector machine (SVM), a classification
technique proposed by Cortes and Vapnik [3].
In the GLC method, a spatial gray-level
dependence matrix (i.e. concurrence matrix) is
first calculated. The matrix,
,
represents the probability that the first pixel has
a gray-level of i and the second pixel has a
gray-level of j under the condition that the
distance between two pixels is d in the direction
of θ. For example, the following equation
indicates the concurrence matrix at θ=0°. The
concurrence matrix at other angles can be
similarly defined [4, 5].
()
θ
,;, djiCo
21690-7803-8412-1/04/$20.00 (c)2004 IEEE.
2004 IEEE International Ultrasonics, Ferroelectrics,
and Frequency Control Joint 50th Anniversary Conference
(1)
where (k,l) and (m,n) are coordinates of two
pixels in an image; the image has a size of
xy
LL ×
, and I represents the intensity of a
particular pixel (0 to 255). The symbol #{‧}
denotes the total number of occurrences and T is
the total number of pixel pairs with the
pre-specified d and θ. Figures 1a and 1b show
the concurrence matrices (d=6, θ=0°) of 25MHz
ultrasound images of a steatosis liver and a
non-steatosis liver, respectively. Generally, the
probability pattern of the concurrence matrix of
a steatosis liver is more diversely distributed
than that of a non-steatosis liver. Image features
such as the angular second moment, contrast,
correlation, entropy and sum entropy can be
derived from the concurrence matrix [4, 5].
Fig 1. B-mode sub-images of 25 MHz ultrasound with its
corresponding concurrence matrix in a steatosis liver (grade
1) (a) and a non-steatosis liver (b).
In addition to GLC, NSW was also applied
to extract the image features. The quincunx
wavelet transform was applied [2, 5]. In this case,
the original image was filtered by a 2-D diamond
shaped high-pass filter and a low-pass filter of
the same shape before it was down-sampled to a
new image with half of the original area.
Filtering and down-sampling were performed
four times, thus producing high pass filtered
sub-images H1-H4 and low pass filtered
sub-images L1-L4. The variances of images
H1-H4 and L4 were taken as the image features
[2, 5]. Figure 2 compares the NSW output
images of a steatosis liver (a) to those of a
non-steatosis liver (b) (25 MHz ultrasound). In
this particular case, the variances of the output
Fig. 2 Non-separable wavelet transform of a steatosis liver
(grade 1) (a) and a non-steatosis liver (b). H1-4: images after
high-pass filtering and down-sampling, L4: image after final
low-pass filtering and down-sampling.
images (H1-H4 and L4) of a steatosis liver
exceed those of a non-steatosis liver.
The features extracted by the above two
methods were used by SVM for classification [5].
The SVM maps features of the training examples
to a higher dimensional space and finds a
hyper-plane to separate the two classes with the
decision boundary set by the support vectors.
The testing samples were then classified by the
trained SVM. The three-class SVM classifier can
be generated from a two-class classifier by
further dividing one of the original classes into
two classes. A classifier of higher classes can be
produced similarly. A multi-class SVM classifier,
developed by Junshui Ma (NIS-2, Los Alamos
National Lab.) and Yi Zhao (EE department,
Ohio State University), was used
(http://www.csie.ntu.edu.tw/~cjlin/libsvm).
Image Acquisition and Processing
Nineteen fresh human liver samples obtained
from surgical specimens were studied. These
specimens were cut to bar shape and immersed
in normal saline. After scanning by 25 MHz and
7 MHz ultrasound, these specimens received
histological examination for steatosis grading.
A. High frequency ultrasound (25 MHz)
A lithium-niobate single crystal focused
transducer (NIH Resource Center for Medical
Ultrasonic Transducer Technology, Penn State
University) was used. The transducer’s center
frequency is 44 MHz (–6 dB bandwidth: 34-59
MHz). A block diagram of the experimental
setup is shown in Fig. 3. To extend the depth of
focus of the fixed focus transducer, two images
were acquired and combined (also known as
depth scan) [6]. The two images corresponded to
two different transducer positions which were 2
mm apart in the axial direction. For 25 MHz
imaging, 3 cycles of a sine wave pulse with a
center frequency of 25 MHz was transmitted.
The spacing between two adjacent scan lines
was 20 µm. All images here are displayed with a
36-dB dynamic range.
Fig.3 A block diagram of the setup of 25MHz ultrasound.
B. Conventional ultrasound (7 MHz)
After the 25 MHz data was acquired, the
same specimen was also scanned using a 7 MHz
linear array transducer (10L, GE LOGIQ 500,
Chalfont St. Giles, UK) in a water tank. The
images were then acquired using a frame grabber
(UPMOST UPG401B, Taipei, Taiwan) with 256
gray levels and 320 x 240 pixels resolution.
()
()( )()
(
)
() ( )
()
°
==
=−=
×∈
=° 0,/
,,,
,
,,,
#0,;, dT
jnmIilkI
dnlmk
LLnmlk
djiCo
xy
21700-7803-8412-1/04/$20.00 (c)2004 IEEE.
2004 IEEE International Ultrasonics, Ferroelectrics,
and Frequency Control Joint 50th Anniversary Conference
C. Image processing
A sub-image of 64 x 64 pixels (11.0 mm x
11.0 mm and 2.8 mm x 2.8 mm equivalent to
51.3 λ x 51.3 λ and 46.7 λ x 46.7 λ for 7MHZ
and 25 MHz, respectively, where λ is the
ultrasound wave length) was selected from each
image. Five to ten sub-images in each frequency
of each specimen were selected; the cross-
correlation coefficients among the images were
all below 0.5. Image features were extracted
using GLC and NSW from these sub-images.
After ultrasound scanning, the specimens
were fixed in formalin and embedded in paraffin.
Several sections with a thickness of 5 μm were
cut from each specimen and were stained with
Masson’s Trichrome stain for histological
examination. Slices of all specimens were
photographed to histological pictures (i.e.,
low-power magnifying) by a digital camera
(Sony S85, Tokyo, Japan). The steatosis grade
was classified by an experienced pathologist
according the following criteria [1]: 0: no fatty
change; 1: less than 33% of hepatocytes affected;
2: 33% to 66% of hepatocytes affected; and 3:
more than 66% of hepatocytes affected. A
semi-automatic method to calculate the
percentage of affected hepatocytes was also
adopted. Microscopic pictures (i.e. high-power
magnifying) of the liver slices were
photographed by a digital camera (Nikon
Coolpix 4500, Tokyo, Japan) mounted on a
microscope (Teng-Bo Roomkop XTL 300A,
Taipei, Taiwan). The digitized images were then
sent to a personal computer and were processed
using a commercial image processing software
(Adobe Photoshop, 4.0.1 TC, San Jose, CA ).
Fig. 4 (a) A microscopic picture of a steatosis liver (grade 2)
with Masson’s Trichrome stain, image size is 2.1 mm x 1.6
mm. (b) The steatosis affected hepatocytes (black arrow in
(a)) are marked in white, un-affected hepatocytes are in gray
and the other tissues (white arrow in (a)) are in black.
The hepatocytes affected by steatosis (the black
arrow in Fig. 4a) were marked by white color as
shown in Fig. 4b, the unaffected hepatocytes
were marked by gray color and the other tissues
(the white arrow in Fig. 4a) were marked by
black color in Fig. 4b. The percentage of
steatosis can be calculated as the sum of affected
hepatocytes (number of pixels of the white areas)
divided by the sum of total hepatocytes (number
of pixels of the white and the gray areas). Table
1 lists the mean steatosis percentage (counted
from 5 images in each specimen) and the grading
of the 10 steatosis specimens. It was further
verified that the grading results were the same as
Table 1: Mean percentage of steatosis and grading of 10
specimens of steatosis liver
those done by the pathologist. The all four
grades of steatosis were then combined into 2, 3
and 4 classes according to the following criteria:
2 classes- grade 0 and grades 1-3 (representing
non-steatosis and steatosis); 3 classes- grade 0,
grade 1 and grades 2-3 (representing
non-steatosis, mild and obvious steatosis); and 4
classes (all different grades). The classification
accuracy was tested using the support vector
machine by the leave-one-out method. After
testing, the classification results were recorded.
The process was repeated until all images had
been tested. The accuracy was then determined
as
(2)
where CI is the number of correct incidences in a
particular class and N is the total number of
samples in that class. Various combinations of
input features (angles and distances of GLC;
filter pairs of NSW) and the SVM parameters
(penalty term and kernel function parameters)
were tested to find the best performance
achievable.
Results
Among the nineteen specimens, nine
specimens were non-steatosis and ten specimens
was steatosis. The combination of all features of
the GLC (d=6, θ=0°) and all feature of the NSW
yielded the best results among all the different
combinations. For 25 MHz ultrasound images,
the best accuracy of 2, 3 and 4 classes were 90.5
%, 85.5 % and 82.6%, respectively. For
conventional ultrasound images at 7 MHz, the
best classification accuracy of 2, 3 and 4 classes
were 81.6 %, 75.8 % and 74.2 %, respectively.
Compared with the results acquired from
conventional ultrasound, the classification
accuracy was markedly better with high
frequency ultrasound. The comparison was
graphically shown in Fig. 5.
Figure 6 compares the images between 25
and 7 MHz. The hepatocytes affected by
steatosis in the microscopic picture (arrow in Fig.
6a) are shown as bright ovoid or strip shaped
structures in the histological picture (arrow in
Fig. 5 Comparison of the steatosis classification results
between the 25 MHz and 7 MHz.
N
CI
=Accuracy
21710-7803-8412-1/04/$20.00 (c)2004 IEEE.
2004 IEEE International Ultrasonics, Ferroelectrics,
and Frequency Control Joint 50th Anniversary Conference
Fig. 6 (a) A microscopic picture (i.e., high-power
magnifying) of a steatosis liver (grade 2) with Masson’s
Trichrome stain, image size is 2.1 mm x 1.6 mm. (b) A
histological picture (i.e., low-power magnifying) with
Masson’s Trichrome stain, image size is 19.6 mm x 16.0 mm.
(c) The corresponding 25MHz ultrasound image, image size
is 18.6 mm x 11.8 mm. (d) The corresponding 7 MHz
ultrasound image, image size is 20.0 mm x 15.0 mm.
Fig. 6b). Furthermore, the pattern of steatosis
distribution in Fig. 6b is more clearly
demonstrated in the 25 MHz ultrasound image
(arrow in Fig. 6c) than the 7 MHz ultrasound
image (arrow in Fig. 6d). The higher spatial
resolution at 25 MHz is believed to play a
critical role in the resultant higher accuracy.
Discussion and Conclusions
The 25 MHz images had better accuracy for
steatosis classification. Comparing the
histological pictures of steatosis livers with the
25 MHz ultrasound images, some similarities
were found. Figure 7a shows that the steatosis-
affected hepatocytes (grade 1) locate mainly in
the pericentral area of acinus and the clusters of
affected hepatocytes appear as an ovoid shape
(arrow). The 25 MHz ultrasound image shows a
similar pattern to the clusters of affected
hepatocytes shown in Fig. 7a (arrow in Fig. 7b).
Figure 7c shows that the steatosis-affected
hepatocytes (grade 1) in another liver locate
mainly in the periportal area (arrow). Again, the
25 MHz B-mode image in Fig. 7d presents a
similar pattern (arrow). Also, comparing the 25
MHz liver images of mild steatosis with severe
steatosis, the brightened lesions occupy larger
areas with higher grade of steatosis (Fig. 6c).
High frequency ultrasound has good spatial
Fig. 7 (a) and (c) show the histological pictures of 2
steatosis liver (grade 1) with Masson’s Trichrome stain,
image size is 21.0 mm x 16.0 mm; (b) and (d) The
corresponding 25MHz ultrasound images, image size is 18.6
mm x 11.8 mm.
resolution and reveals more details of steatosis
than conventional ultrasound at lower
frequencies, but the shorter penetration depth
and the smaller focal zone limit its clinical
applications. Intra-abdominal organs are hard to
be accessed by the high frequency ultrasound
because high frequency sound waves cannot
penetrate through the abdominal wall. The
intra-vascular approach may be a potential
solution. In addition to the limited penetration,
the small depth of focus of the single crystal
probe creates another limitation. For example,
fibrosis patterns may not be adequately
demonstrated using high frequency ultrasound.
The degree of fibrosis ranges from fibrous
expansion in the portal area to cirrhosis. The
regenerative nodules of a cirrhotic liver are
enclosed by fibrotic tissues and existence of
regenerative nodules is one of the key features
for pathological diagnosis of cirrhosis. The
diameter of regenerative nodules is generally
larger than 3 mm in patients with diseases such
as late phase of hepatitis B cirrhosis and
autoimmune cirrhosis, and primary biliary
cirrhosis. Although in our previous study, we had
successfully classified the fibrosis grade by
SVM using 7 MHz ultrasound [5]. However, the
increased spatial resolution of 25 MHz may not
improve the classification of fibrosis due to the
limited penetration and the small depth of focus.
In conclusion, the 25 MHz ultrasound
images had better accuracy for steatosis
classification. But the limitations of poor
penetration depth and small focal zone need to
be overcome for further clinical applications.
Acknowledgement
The authors would like to thank the National
Science Council of the Republic of China for
financially supporting this research under Grant
No. NSC 93-2213-E-002-119 and Mr.
Chao-Kang Liao for photographic assistance.
References
[1] Brunt EM, Janney CG, Di Bisceglie AM, et al.
Nonalcoholic steatohepatitis: a proposal for grading
and staging the histological lesions. Am J Gastroenterol
94: 2467-2474, 1999.
[2] Mojsilovic A, Popovic M, Markovic S, et al.
Characterization of visually diffuse diseases from
B-scan liver images using non-separable wavelet
transform. IEEE Trans Med Imag 17(4): 541-549,1998.
[3] Cortes C and Vapnik V. Support Vector Networks.
Mach Learn 20:273-297,1995.
[4] Wu C, Chen Y, Hsieh K. Texture features for
classification of ultrasonic liver images. IEEE Trans
Med. Imag 11(2):141-152, 1992.
[5] Yeh WC, Huang SW, Li PC. Liver Fibrosis Grade
Classification Using B-mode Ultrasound. Ultrasound
Med Biol 29(9):1229-1235, 2003.
[6] Passmann C, Ermert H. A 100-MHz Ultrasound
Imaging System for Dermatologic and Ophthalmologic
Diagnostics. IEEE Trans Ultrason Ferroelectr Freq
Control 43(4):545-552, 1996.
21720-7803-8412-1/04/$20.00 (c)2004 IEEE.
2004 IEEE International Ultrasonics, Ferroelectrics,
and Frequency Control Joint 50th Anniversary Conference
