European Journal of Radiology 85 (2016) 511–517

Contents lists available at ScienceDirect

European Journal of Radiology

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

Using texture analyses of contrast enhanced CT to assess hepatic

fibrosis
Naznin Daginawala, Baojun Li, Karen Buch, HeiShun Yu, Brian Tischler,
Muhammad Mustafa Qureshi, Jorge A. Soto, Stephan Anderson ∗
Department of Radiology, Boston University Medical Center, United States

a r t i c l e i n f o a b s t r a c t

Article history: Purpose: To determine the ability of texture analyses of contrast-enhanced CT images for distinguishing
Received 1 September 2015 between varying degrees of hepatic fibrosis in patients with chronic liver disease using histopathology
Received in revised form 2 December 2015 as the reference standard.
Accepted 13 December 2015
Materials and methods: Following IRB approval, 83 patients who underwent contrast enhanced 64-MDCT
of the abdomen and pelvis in the portal venous phase between 12/2005 and 01/2013 and who had a
Keywords:
liver biopsy within 6 months of the CT were included. An in-house developed, MATLAB-based texture
Image post processing
analysis program was employed to extract 41 texture features from each of 5 axial segmented volumes
Texture analysis
Liver
of liver. Using the Ishak fibrosis staging scale, histopathologic grades of hepatic fibrosis were correlated
Cirrhosis with texture parameters after stratifying patients into three analysis groups, comparing Ishak scales 0–2
Computed tomography with 3–6, 0–3 with 4–6, and 0–4 with 5–6. To assess the utility of texture features, receiver operating
characteristic (ROC) curves were constructed and the area under the curve (AUC) was used to determine
the performance of each feature in distinguishing between normal/low and higher grades of hepatic
fibrosis.
Results: A total of 19 different texture features with 7 histogram features, one grey level co-occurrence
matrix, 6 gray level run length, 1 Laws feature, and 4 gray level gradient matrix demonstrated statistically
significant differences for discriminating between fibrosis groupings. The highest AUC values fell in the
range of fair performance for distinguishing between different fibrosis groupings.
Conclusion: These findings suggest that texture-based analyses of contrast-enhanced CT images offer a
potential avenue toward the non-invasive assessment of liver fibrosis.
© 2016 Published by Elsevier Ireland Ltd.

1. Introduction a heterogeneous disease distribution of fibrosis throughout the

The early detection and staging of fibrosis and cirrhosis is of
great clinical importance as this may lead to timely diagnosis and
the initiation of appropriate therapeutic regimens. The current
standard of reference for staging fibrosis is hepatic biopsy (usually
obtained via a percutaneous approach) and histopathology, but this
method has several limitations [1]. Firstly, percutaneous biopsy is
an invasive procedure and is with associated complications such
as patient discomfort, bleeding, and infection [2]. Furthermore, the
histopathologic analysis is well known to be subject to inter- and
intra-observer variability [3]. Finally, it has also been observed that
∗ Corresponding author at: Boston University Medical Center, Department of Radi-
ology, 820 Harrison Avenue, FGH Building, 3rd Floor, Boston, MA 02118, United
States. Fax: +1 617 638 6616.
E-mail address: stephan.anderson@bmc.org (S. Anderson).
liver leads to sampling error and variability as the biopsy needle
samples only a small fraction of the liver parenchyma [3,4].
Given the limitations of percutaneous biopsy, the development
of non-invasive methods for evaluating hepatic fibrosis with the
ability to distinguish between lower and higher grades of hep-
atic fibrosis is of significant clinical importance, as they yield the
potential for repeatable assessments with minimal risk and patient
discomfort. Therefore, a variety of noninvasive methods including
the use of serum markers, ultrasound-based transient elastogra-
phy, and MR-based imaging approaches have been evaluated and
shown promise in their ability to diagnose and stage fibrosis [5–8].
To date, only few reports have evaluated the ability of CT imaging to
determine the severity of hepatic fibrosis [9–12], especially when
compared to other imaging tests such as ultrasound-based tran-
sient elastography or MRI. However, CT offers unique advantages
including its nearly ubiquitous availability, whole organ imaging
capacity, and, when compared to MRI, relatively low cost and fewer

http://dx.doi.org/10.1016/j.ejrad.2015.12.009
0720-048X/© 2016 Published by Elsevier Ireland Ltd.

Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
512 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517

contraindications. Texture analysis yields a quantitative approach

to examine spatially organized patterns, sub-patterns, and distribu-
tions [13]. This allows for the quantitative analysis of a wide range
of morphologic and physiologic properties of living tissues, includ-
ing complex texture features that are not visible to the human
eye. Texture features of biomedical imaging data have been widely
employed in the analysis of a number of different diseases and
organ systems such as the liver, thyroid, breasts, kidneys, prostate,
heart, brain, and lungs [14–16]. In fact, several studies have been
reported which leverage texture analyses in the detection and anal-
ysis of chronic liver disease. For example, texture analysis has been
applied to diffusion-weighted and double contrast-enhanced MRI
of the liver with encouraging results [17,18]. Given the advantages
of CT imaging and the strengths of texture analyses in biomedical
imaging, this study aims to determine the ability of texture anal-
yses of contrast-enhanced CT images for distinguishing between
varying degrees of hepatic fibrosis in patients with chronic liver
disease.

2. Materials and methods

The Investigations Review Board of our institution approved this

retrospective study. The study complied with the Health Insurance
Portability and Accountability Act; the requirement for informed
patient consent was waived.

2.1. Patient population

For this study, the electronic medical system was queried to

identify all patients with chronic liver disease who had non-
targeted ultrasound-guided liver biopsies performed within 6
months of a portal venous phase, contrast-enhanced CT exami-
nation of the abdomen between December 2005 and April 2013.
Patients were excluded from the study if the biopsy results showed
evidence of tumor or granulomatous involvement of the liver.

2.2. CT image acquisition and analysis

All CT examinations were performed with a 64-detector row

CT scanner (Light-Speed VCT; GE Medical Systems, Milwaukee,
Wisconsin). The following parameters were used: reconstruction
thickness, 1.25 mm; 120 kVp; noise index, 23 (automatic dose
modulation); pitch, 0.984:1; and gantry rotation time, 0.5 s. All
patients received a bolus of 100 mL of intravenous contrast material
(Optiray; Mallinckrodt Imaging, Hazelwood, Mo [350 mg iodine per
milliliter]) at a rate of 3–4 mL/s with use of a power injector via an
18- or 20-gage cannula in an antecubital vein. In addition, a 30-mL
saline chasing bolus was used immediately after administration of
the intravenous contrast material. Portal venous phase imaging of
the abdomen and pelvis employed a routine delay of 70 s following
the initiation of IV contrast administration, per our departmental
protocol.
Using a dedicated AW workstation (GE Healthcare, Cleveland,
OH), a single radiologist (blind) manually segmented five sequen-
tial 1.25 mm thick axial images (using a soft tissue reconstruction
kernel) at the level of the porta hepatis. The visualized vessels
within the segmented portions of the liver were identified by the
radiologist and manually omitted from the final segmented liver
image (Fig. 1). The inclusion of 5 slices at the level of the porta hep-
atis is felt to offer an adequate and practical sample size, whereas
segmenting and analyzing the entire liver may be prohibitively
Fig. 1. Contrast enhanced portal venous phase axial images of the liver with a fibro-
laborious and impractical to reproduce. Following segmentation, an sis score of 0 (a), 2 (b), 4 (c) and 6 (d) after segmentation on the AW workstation.
in-house developed, MATLAB-based texture analysis program was Subtle, subjective differences in image texture are appreciated when comparing
employed to extract 41 texture features from each segmented vol- these images, differences which were more rigorously analyzed using texture anal-
ume of liver. Twelve histogram features (mean, median, standard ysis in this work.

deviation, range, geometric mean, harmonic mean, 2D standard

Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517
deviation, 5-neighborhood std, 9-neighborhood std, 4th moment,
IQR, and entropy), five gray level co-occurrence matrix (GLCM) fea-
tures (entropy, contrast, correlation, energy, homogeneity), eleven
gray-level run-length (GLRL) features (short run emphasis (SRE),
long run emphasis (LRE), gray-level non-uniformity (GLN), run-
length non-uniformity (RLN), run percentage (RP), low gray-level
run emphasis (LGRE), high gray-level run emphasis (HGRE), short
run low gray-level emphasis (SRLGE), short run high gray-level
emphasis (SRHGE), long run low gray-level emphasis (LRLGE),
and long run high gray-level emphasis (LRHGE)), four gray level
gradient matrix (GLGM) features (mean, variance, skewness, and
kurtosis), and nine Laws features were computed and averaged over
the 5 axial images segmented for each patient, similar to our prior
work [19].
2.3. Reference standard
The pathology reports of the 83 patients meeting the inclusion
criteria were queried in our institutional electronic medical record
to identify the results of each individual’s ultrasound-guided percu-
taneous liver biopsy. The degree of fibrosis which had been reported
for each patient, graded on a scale from 0 to 6 and based on the
Ishak fibrosis scoring sytem (0 = no fibrosis, 1 = fibrous expansion
of some portal areas with or without short fibrous septa, 2 = fibrous
expansion of most portal areas with or without short fibrous septa,
3 = fibrous expansion of most portal areas with occasional por-
tal to portal bridging, 4 = fibrous expansion of most portal areas
with marked portal to portal bridging as well as portal to central
bridging, 5 = marked bridging with occasional nodules (incomplete
cirrhosis), 6 = cirrhosis, probable or definite) was recorded for each
patient.
2.4. Statistical analysis
Based on the patients’ histopathological results, the Ishak fibro-
sis staging scale was used to stratify the patient cohort into the
following three different categorizations of normal/low and higher
grade hepatic fibrosis: (1) scores 0–2 vs. 3–6, (2) scores 0–3 vs. 4–6,
and (3) scores 0–4 vs. 5–6. To assess the potential clinical utility
of texture features, receiver operating characteristic (ROC) curves
were constructed to determine the performance of each feature in
distinguishing between the three different categorizations of nor-
mal/low and higher degrees of hepatic fibrosis. The area under the
ROC curve (AUC) was used to assess the predicted validity of each
texture feature. The closer the AUC value is to 1.0, the more predic-
tive the features were with respect to high grade hepatic fibrosis.
The area under the ROC curve (AUC) results were considered excel-
lent for AUC values between 0.9–1, good for AUC values between
0.8–0.9, fair for AUC values between 0.7–0.8, poor for AUC val-
ues between 0.6–0.7 and failed for AUC values between 0.5–0.6
[20,21]. A two-sided hypothesis was used for all tests, and a proba-
bility value of less than 0.05 was considered statistically significant.
Statistical computations were performed using SAS 9.1.3 software
(SAS Institute, Cary, NC) and GraphPad prism software (version 3.0,
GraphPad Software).
3. Results
3.1. Patient population
A total of 83 patients (49 males, 34 females) aged 22–75 years
(mean age, 50 years) fulfilled the criteria for inclusion. The eti-
ologies of liver disease in the included patients were chronic
hepatitis C (n = 48), chronic hepatitis B (n = 5), cytomegalovirus
induced hepatitis (n = 1), Epstein–Barr virus induced hepatitis
(n = 1), nonalcoholic steatohepatitis (NASH) (n = 8), steatohepatitis
Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
513
due to alcohol and fatty liver disease (n = 2), autoimmune hepatitis
(n = 1), drug hypersensitivity (n = 2), primary biliary cirrhosis (n = 2),
hemochromatosis (n = 1), Budd–Chiari syndrome (n = 1), crypto-
genic cirrhosis (n = 2). The remaining nine patients underwent liver
biopsy to evaluate for abnormal liver function tests with no specific
etiology of liver disease identified.
3.2. Histopathology
The distribution of the Ishak hepatic fibrosis staging scores in
the 83 patients was the following: 16 patients (19%) with a score
of 0, 18 patients (22%) with a score of 1, 15 patients (18%) with a
score of 2, 11 patients (13%) with a score of 3, 2 patients (2%) with
a score of 4, and 5 patients (6%) with a score of 5, and 16 patients
(19%) with a score of 6.
3.3. Texture analysis: histogram features
ROC curve analysis results comparing histogram texture fea-
tures and histopathologic analysis of fibrosis are reported in
Table 1. For the comparison of fibrosis scores 0–2 with 3–6, statis-
tically significant AUC values were found for the mean (AUC = 0.68,
p = 0.005), median (AUC = 0.69, p = 0.004), STD (AUC = 0.64, p = 0.03),
and 4th moment (AUC = 0.66, p = 0.01) features. For the compar-
ison of fibrosis scores of 0–3 with 4–6, statistically significant
AUC values were found for the mean (AUC = 0.68, p = 0.01), median
(AUC = 0.69, p = 0.009), STD9 (AUC = 0.65, p = 0.03), 4th moment
(AUC = 0.64, p = 0.04), IQR (AUC = 0.66, p = 0.02), and entropy
(AUC = 0.67, p = 0.02) features. For the comparison of fibrosis scores
of 0–4 with 5–6, statistically significant AUC values were found
for the mean (AUC = 0.69, p = 0.009), median (AUC = 0.69, p = 0.009),
STD9 (AUC = 0.65, p = 0.04), IQR (AUC = 0.66, p = 0.03), and entropy
(AUC = 0.66, p = 0.03) features. The texture feature that was best
correlated with fibrosis score for all three stratifications was the
median noting that the aforementioned histogram texture features
with the highest AUC values remained within the range of poor ROC
curve results.
3.4. Texture analysis: gray-level co-occurrence matrix features
ROC curve analysis results comparing GLCM texture features
and histopathologic analysis of fibrosis are reported in Table 2. Only
the entropy texture feature was found to be statistically significant
for the comparison of fibrosis scores 0–3 with 4–6 with an AUC of
0.64 (p = 0.05), within the range of poor ROC curve results.
3.5. Texture analysis: gray-level run-length features
ROC curve analysis results comparing GLRL texture features
and histopathologic analysis of fibrosis is reported in Table 3. For
the comparison of fibrosis scores 0–2 with 3–6, statistically sig-
nificant AUC values were found for the RP (AUC = 0.64, p = 0.04),
LGRE (AUC = 0.63, p = 0.04), and SRLGE (AUC = 0.63, p = 0.04) tex-
ture features. For the comparison of fibrosis scores of 0–3 with
4–6, statistically significant AUC values were found for the
LGRE (AUC = 0.65, p = 0.04), SRLGE (AUC = 0.65, p = 0.04), SRHGE
(AUC = 0.75, p = 0.0004), LRLGE (AUC = 0.71, p = 0.004), and LRHGE
(AUC = 0.70, p = 0.006) texture features. For the comparison of fibro-
sis scores of 0–4 with 5–6, statistically significant AUC values were
found for the SRHGE (AUC = 0.76, p = 0.0004), LRLGE (AUC = 0.71,
p = 0.004), and LRHGE (AUC = 0.71, p = 0.005) texture features. Of
note, the texture features with the highest AUC values fall within
the range of fair ROC curve results. The texture feature which
demonstrated the highest AUC values for comparisons of both the
0–3 with the 4–6 groups as well as the 0–4 with the 5–6 groups was
514 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517

Table 1
Texture histogram features differentiating between hepatic fibrosis grades.

Grade 0–2 vs. 3–6 Grade 0–3 vs. 4–6 Grade 0–4 vs. 5–6

AUC (95% CI), p-value

Mean 0.68 (0.57–0.78)* 0.005 0.68 (0.56–0.81)* 0.010 0.69 (0.56–0.82)* 0.009
Median 0.69 (0.57–0.80)* 0.004 0.69 (0.56–0.81)* 0.009 0.69 (0.56–0.82)* 0.009
STD 0.64 (0.52–0.76)* 0.033 0.61 (0.47–0.75) 0.110 0.58 (0.43–0.72) 0.295
Range 0.57 (0.43–0.71) 0.271 0.61 (0.46–0.75) 0.138 0.60 (0.45–0.75) 0.157
Geometric mean 0.50 (0.37–0.63) 0.971 0.51 (0.36–0.65) 0.911 0.55 (0.40–0.69) 0.537
Harmonic mean 0.61 (0.48–0.74) 0.090 0.59 (0.44–0.73) 0.230 0.56 (0.40–0.71) 0.438
2nd STD 0.56 (0.43–0.70) 0.331 0.60 (0.44–0.75) 0.179 0.59 (0.44–0.75) 0.201
STD5 0.59 (0.45–0.72) 0.179 0.64 (0.49–0.78) 0.058 0.63 (0.49–0.78) 0.067
STD9 0.60 (0.46–0.73) 0.136 0.65 (0.51–0.80)* 0.032 0.65 (0.51–0.80)* 0.037
4th moment 0.66 (0.54–0.78)* 0.012 0.64 (0.51–0.78)* 0.043 0.61 (0.47–0.76) 0.119
IQR 0.61 (0.48–0.75) 0.088 0.66 (0.52–0.81)* 0.021 0.66 (0.52–0.81)* 0.026
Entropy 0.62 (0.49–0.75) 0.063 0.67 (0.52–0.81)* 0.019 0.66 (0.51–0.81)* 0.032

Abbreviations: AUC—area under curve; CI—confidence interval; STD—standard deviation; STD5—5-neighborhood standard deviation; STD9—9-neighborhood standard devi-
ation; IQR—inter quartile range.
*
p-value < 0.05 indicates AUC is significantly different from 0.5 and there is evidence that the texture parameter has ability to distinguish between the two groups.

Table 2
Texture GLCM features differentiating between hepatic fibrosis grades.

Grade 0–2 vs. 3–6 Grade 0–3 vs. 4–6 Grade 0–4 vs. 5–6

AUC (95% CI), p-value

Entropy 0.60 (0.47–0.73) 0.131 0.64 (0.50–0.78)* 0.049 0.63 (0.48–0.78) 0.079
Contrast 0.60 (0.48–0.73) 0.111 0.58 (0.44–0.72) 0.246 0.54 (0.40–0.69) 0.579
Correlation 0.52 (0.39–0.65) 0.732 0.53 (0.38–0.68) 0.692 0.58 (0.43–0.73) 0.271
Energy 0.57 (0.43–0.70) 0.308 0.62 (0.47–0.76) 0.106 0.61 (0.46–0.76) 0.124
Homogeneity 0.57 (0.43–0.70) 0.304 0.61 (0.46–0.76) 0.132 0.61 (0.46–0.75) 0.151

Abbreviations: GLCM—gray level co-occurrence matrix; AUC—area under curve; CI—confidence interval.
*
p-value is less than 0.05 which indicates AUC is significantly different from 0.5 and there is evidence that the texture parameter has ability to distinguish between the
two groups.

Table 3
Texture GLRL features differentiating between hepatic fibrosis grades.

Grade 0–2 vs. 3–6 Grade 0–3 vs. 4–6 Grade 0–4 vs. 5–6

AUC (95% CI), p-value

SRE 0.56 (0.43–0.69) 0.364 0.52 (0.38–0.66) 0.776 0.50 (0.36–0.65) 0.983
LRE 0.56 (0.43–0.69) 0.355 0.53 (0.39–0.66) 0.714 0.51 (0.37–0.65) 0.917
GLN 0.56 (0.45–0.69) 0.322 0.53 (0.39–0.68) 0.633 0.52 (0.37–0.66) 0.834
RLN 0.56 (0.43–0.69) 0.359 0.53 (0.38–0.67) 0.707 0.51 (0.36–0.65) 0.933
RP 0.64 (0.51–0.76)* 0.036 0.64 (0.49–0.79) 0.052 0.60 (0.45–0.76) 0.154
LGRE 0.63 (0.50-0.76)* 0.041 0.65 (0.50-0.80)* 0.038 0.62 (0.46–0.78) 0.107
HGRE 0.62 (0.49–0.75) 0.057 0.63 (0.48–0.79) 0.058 0.60 (0.44–0.77) 0.164
SRLGE 0.63 (0.50-0.76)* 0.042 0.65 (0.49–0.80)* 0.041 0.61 (0.45–0.77) 0.129
SRHGE 0.61 (0.48–0.75) 0.080 0.75 (0.62–0.88)* 0.0004 0.76 (0.64–0.89)* 0.0004
LRLGE 0.61 (0.48–0.74) 0.090 0.71 (0.57–0.84)* 0.004 0.71 (0.58–0.84)* 0.004
LRHGE 0.62 (0.49–0.74) 0.075 0.70 (0.56–0.83)* 0.006 0.71 (0.57–0.84)* 0.005

Abbreviations: GLRL—gray level run length; AUC—area under curve; CI—confidence interval; SRE—short run emphasis; LRE—long run emphasis; GLN—gray-level non-
uniformity; RLN—run-length non-uniformity; RP—run percentage; LGRE—low gray-level run emphasis; HGRE—high gray-level run emphasis; SRLGE—short run low gray-level
emphasis; SRHGE—short run high gray-level emphasis; LRLGE—long run low gray-level emphasis; LRHGE—long run high gray-level emphasis.
*
p-value is less than 0.05 which indicates AUC is significantly different from 0.5 and there is evidence that the texture parameter has ability to distinguish between the
two groups.

SRHGE; as an illustrative example, the ROC curve for SRHGE for the
0–4 vs. 5–6 categorization is graphically represented in Fig. 2.
3.6. Texture analysis: laws features
ROC curve analysis results comparing Laws texture features and
histopathologic analyses of fibrosis are reported in Table 4. Laws
feature 1 was found to have a statistically significant AUC when
comparing fibrosis scores of 0–3 with 4–6 (AUC = 0.67, p = 0.02) and
0–4 with 5–6 (AUC = 0.66, p = 0.03), both within the range of poor
ROC curve results.
3.7. Texture analysis: GLGM features
texture GLGM texture features were found to have statistically sig-
nificant AUC values. For comparison of fibrosis levels of 0–2 with
3–6, AUC values ranged from 0.64 to 0.67, with p-values ranging
from 0.008 to 0.027. For comparison of fibrosis levels of 0–3 with
4–6, AUC values ranged from 0.72 to 0.74, with p-values ranging
from 0.001 to 0.003. For comparison of fibrosis levels of 0–4 with
5–6, AUC values ranged from 0.71 to 0.74, with p-values ranging
from 0.001 to 0.005. Of note, the GLGM texture features with the
highest AUC values fall within the range of fair ROC curve results.
The texture feature which demonstrated the highest AUC values
for comparisons of both the 0–3 with the 4–6 groups as well as the
0–4 with the 5–6 group categorizations was MGR.

ROC curve analysis results comparing GLGM texture features

and histopathologic analysis of fibrosis are reported in Table 5. All

Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517 515

Table 4
Texture laws features differentiating between hepatic fibrosis grades.

Grade 0–2 vs. 3–6 Grade 0–3 vs. 4–6 Grade 0–4 vs. 5–6

AUC (95% CI), p-value

L1 0.62 (0.48–0.75) 0.072 0.67 (0.52–0.81)* 0.019 0.66 (0.51–0.81)* 0.030
L2 0.52 (0.39–0.66) 0.718 0.56 (0.41–0.70) 0.439 0.55 (0.41–0.70) 0.463
L3 0.51 (0.37–0.64) 0.897 0.55 (0.40–0.69) 0.508 0.55 (0.40–0.69) 0.523
L4 0.57 (0.43–0.70) 0.291 0.61 (0.46–0.75) 0.127 0.61 (0.46–0.75) 0.146
L5 0.53 (0.40–0.66) 0.657 0.50 (0.36–0.64) 0.976 0.50 (0.36–0.64) 0.983
L6 0.54 (0.41–0.67) 0.511 0.52 (0.38–0.66) 0.745 0.52 (0.39–0.66) 0.745
L7 0.52 (0.39–0.65) 0.753 0.51 (0.37–0.65) 0.895 0.51 (0.37–0.65) 0.908
L8 0.53 (0.40-0.67) 0.591 0.52 (0.37–0.67) 0.776 0.52 (0.38–0.67) 0.738
L9 0.55 (0.41–0.69) 0.442 0.60 (0.45–0.74) 0.179 0.60 (0.45–0.74) 0.187

Abbreviations: AUC—area under curve; CI—confidence interval; L—laws feature.

*
p-value is less than 0.05 which indicates AUC is significantly different from 0.5 and there is evidence that the texture parameter has ability to distinguish between the
two groups.

Table 5
Texture GLGM features differentiating between hepatic fibrosis grades.

Grade 0–2 vs. 3–6 Grade 0–3 vs. 4–6 Grade 0–4 vs. 5–6

AUC (95% CI), p-value

MGR 0.65 (0.53–0.77)* 0.021 0.74 (0.61–0.86)* 0.001 0.74 (0.61–0.87)* 0.001
VGR 0.67 (0.55–0.79)* 0.008 0.73 (0.60–0.85)* 0.001 0.71 (0.58–0.85)* 0.004
Skewness 0.64 (0.52–0.77)* 0.026 0.72 (0.59–0.85)* 0.002 0.71 (0.57–0.85)* 0.004
Kurtosis 0.64 (0.52–0.76)* 0.027 0.72 (0.58–0.85)* 0.003 0.70 (0.56–0.85)* 0.005

Abbreviations: GLGM—gray level gradient matrix; AUC—area under curve; CI—confidence interval; MGR—mean gradient; VGR—variance gradient.
*
p-value is less than 0.05 which indicates AUC is significantly different from 0.5 and there is evidence that the texture parameter has ability to distinguish between the
two groups.

Given the risks of ionizing radiation, a consideration in the

Fig. 2. Receiver operator characteristic (ROC) curve for selected texture feature of
short run high gray-level emphasis (SRHGE) for differentiating Ishak fibrosis scores
of 0–4 with 5–6.
4. Discussion
Texture analysis, which seeks to reduce image information by
extracting texture descriptors from a given image, offers a poten-
tial avenue to detect subtle texture features not perceived by
the human eye and offers a quantitative approach to analyzing
these differences. Imaging-based texture analysis offers a promis-
ing alternative to the current gold standard of percutaneous liver
biopsy. In this study, we utilized an in-house developed, MATLAB-
based texture analysis program to extract 41 texture features. A
total of 19 of these texture features with 7 histogram features, one
gray level co-occurrence matrix, 6 gray level run length, 1 Laws
feature, and 4 gray level gradient matrix demonstrated statistically
significant differences for discriminating between different group-
ings of patients with varying degrees of liver fibrosis; the highest
AUC values fell in the range of fair performance.
development of CT-based tools for discriminating between dif-
fering degrees of hepatic fibrosis is the role this may play in
clinical practice. Ultrasound-based elastography has demonstrated
promise in evaluating liver fibrosis, with particular utility in distin-
guishing cirrhotic livers from non-cirrhotic livers [5,22]. However,
several limitations need to be considered, including the fact that
ultrasound-based elastography is highly operator dependant, and
an adequate acoustic window can be difficult to identify depending
on patient body habitus and overlying bowel gas. In a prospective
study of 13,369 examinations, Castéra et al. found that nearly one in
five cases of liver stiffness measurement were uninterpretable due
to these limitations [23]. MRI also offers a powerful toolset for the
evaluation of liver fibrosis. Various MRI-based imaging approaches
including MR elastography, diffusion-weighted imaging, and tex-
ture analysis of MR images including T2-weighted and double
contrast, and diffusion-weighted images, among others, have been
developed with some success [7,8,17,18,24]. However, limitations
of MRI-based approaches include the relatively high cost, multi-
ple patient contraindications, and the need for specialized external
hardware in the case of MR elastography.
In the case of CT, a variety of techniques, dual-energy imag-
ing, quantitative morphologic analyses, as well as multiphasic,
contrast-enhanced imaging have been applied to the evaluation of
hepatic fibrosis [25–29]. In addition, there have been a few stud-
ies demonstrating the utility of texture analysis of CT images for
distinguishing degrees of fibrosis. Romero-Gómez et al. demon-
strated the utility of optical digital analysis of non-contrast CT in
detecting fibrosis in patients with chronic hepatitis C [9]. Kayaaltı
et al. performed texture analysis of contrast enhanced CT across a
wide range of texture features and used multivariate analysis to
classify liver fibrosis stages [10]. Zhang et al. compared the useful-
ness of gray level co-occurrence matrix (GLCM) texture features to
determine the stage of liver fibrosis in CT vs. MR [11]. In general,
our findings echo those reported in the aforementioned studies
[9–11] supporting to potential utility of CT in this space and the
need for further inquiry and refinement of these approaches. Com-
pared to other studies [9,11,12] evaluating texture analysis of CT

Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
516 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517
liver images, a strength of our study includes that we identified
a relatively large number of texture features (19 total) across dif-
ferent categories of features with statistical significance in their
capacity to discriminate between fibrosis stage groupings. No one
modality has yet shown an adequate sensitivity and specificity
for staging liver fibrosis and therefore non-invasive methods for
assessing fibrosis remain an active area of investigation.
CT may find integration into the clinical evaluation of patients
with chronic liver disease for a variety of reasons. A major con-
sideration is in the routine screening of patients undergoing
abdominopelvic CT imaging, for a variety of indications, in the
Emergency Department. In this case, a practical, accurate CT-based
tool may be employed to effectively screen and identify patients
with hepatic fibrosis, thereby triaging this subset of patients to
receive the appropriate care of their liver disease. In addition, as
CT is often employed in the imaging evaluation of patients with
suspected or known hepatocellular carcinoma or liver metastases,
the added benefit of a liver fibrosis assessment at this time offers
the potential for providing clinically relevant information prior to
direct therapeutic interventions (surgical resection, ablation, etc.)
[12]. Finally, in those patients in whom advanced MRI and ultra-
sound technology is unavailable, a robust, CT-based tool may be
employed in select cases for assessment of fibrosis, being mindful
of the ionizing radiation exposure.
In the case of CT, a variety of techniques, dual-energy imag-
ing, quantitative morphologic analyses, as well as multiphasic,
contrast-enhanced imaging have been applied to the evaluation of
hepatic fibrosis [25–29]. While fair performance of texture analysis
of portal venous phase, contrast-enhanced CT imaging is reported
herein, several facets of the study design must be considered when
judging the potential utility of this technique. Firstly, all patients
included in this study exhibited some degree of liver dysfunction,
thereby necessitating the need for a percutaneous liver biopsy. As
no true control group without any active liver disease was analyzed,
this introduces potentially confounding effects related to hepatic
parenchymal abnormalities such as inflammation or edema, even
in the absence of hepatic fibrosis (Ishak score 0). Furthermore, as
patients with a variety of etiologies of liver disease are included,
similar confounding effects may be introduced such as, for exam-
ple, differing degrees of steatosis between patients with NASH vs.
those with chronic viral hepatitis, among others. Finally, the acqui-
sition of portal venous phase images using a range of injection rates
(3–4 mL/s) and a fixed scan delay introduces potential confound-
ing effects of slight differences in hepatic parenchymal contrast
kinetics as a function of injection rate and physiology. Overall, the
fact that numerous texture parameters achieved fair performance
based on ROC analyses potentially speaks to the robust nature of
texture analyses in the assessment of liver fibrosis, regardless of
etiology or the presence of confounding histopathological features
(varying degrees of steatosis or tissue inflammation, for exam-
ple). The results reported herein strongly support future inquiry
in which such potential confounding effects are more strictly con-
trolled, such as analyzing a patients affected by a single etiology
of liver disease. With regards to CT acquisition techniques, strictly
fixed injection rates and automated bolus tracking approaches may
be employed to mitigate potential confounders related to the acqui-
sition of the portal venous phase datasets.
There are a few additional limitations to our study to consider.
Given that there were up to 6-month intervals between the imaging
studies and the percutaneous liver biopsies, it is possible that the
level of liver fibrosis differed in the intervening time between the
biopsy and CT study. In addition, it has been previously reported
that differences in noise indices affect texture analysis [30]. Even if
automated, attenuation-based dose modulation is employed, as in
this study (fixed noise indices), noise will nevertheless be expected
to vary, especially at the extremes of body habitus. However, as
Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.
above, given that our results demonstrate fair performance in spite
of this potential confounding limitation, this further supports the
robust nature of these analyses. Another limitation to consider is
the fact that, given that there were only two patients in our dataset
with an Ishak fibrosis score of 4, our ability to categorize patients
with this level of fibrosis may be limited. This fact may, in part,
explain why the AUC values reported above for the various texture
features are similar when comparing fibrosis scores of 0–3 with 4–6
and 0–4 with 5–6.
In conclusion, the promising results reported herein strongly
support further inquiry into and refinements of these tech-
niques, possibly through the development of multi-parametric
approaches leveraging numerous individual texture features. The
reported findings related to texture analysis of contrast-enhanced
CT datasets offer a potential avenue toward the noninvasive eval-
uation of degrees of liver fibrosis.
Conflicts of interest
None.
Grant support
None.
References
[1] G. Menghini, One-second needle biopsy of the liver, Gastroenterology 35
(1958) 190–199.
[2] P. Thampanitchawong, T. Piratvisuth, Liver biopsy: complications and risk
factors, World J. Gastroenterol. 5 (1999) 301–304.
[3] A. Regev, M. Berho, L.J. Jeffers, C. Milikowski, E.G. Molina, N.T. Pyrsopoulos,
et al., Sampling error and intraobserver variation in liver biopsy in patients
with chronic HCV infection, Am. J. Gastroenterol. 97 (2002) 2614–2618.
[4] V. Ratziu, F. Charlotte, A. Heurtier, S. Gombert, P. Giral, E. Bruckert, et al.,
Sampling variability of liver biopsy in nonalcoholic fatty liver disease,
Gastroenterology 128 (2005) 1898–1906.
[5] M. Friedrich-Rust, K. Wunder, S. Kriener, F. Sotoudeh, S. Richter, J. Bojunga,
et al., Liver fibrosis in viral hepatitis: noninvasive assessment with acoustic
radiation force impulse imaging versus transient elastography, Radiology 252
(2) (2009) 595–604.
[6] S.H. Park, C.H. Kim, D.J. Kim, K.T. Suk, J.Y. Cheong, S.W. Cho, et al., Usefulness
of multiple biomarkers for the prediction of significant fibrosis in chronic
hepatitis B, J. Clin. Gastroenterol. 45 (4) (2011) 361–365.
[7] S.W. Anderson, H. Jara, A. Ozonoff, M. O’Brien, J.A. Hamilton, J.A. Soto, The
effect of disease progression on liver ADC and T2 values in a murine model of
hepatic fibrosis at 11.7T MRI, J. Magn. Reson. Imaging 35 (2012) 140–146.
[8] L. Huwart, C. Sempoux, E. Vicaut, N. Salameh, L. Annet, E. Danse, et al.,
Magnetic resonance elastography for the non invasive staging of liver fibrosis,
Gastroenterology 135 (1) (2015) 32–40.
[9] M. Romero-Gómez, E. Gómez-González, A. Madrazo, M. Vera-Valencia, L.
Rodrigo, R. Pérez-Alvarez, et al., Optical analysis of computed tomography
images of the liver predicts fibrosis stage and distribution in chronic hepatitis
C, Hepatology 47 (3) (2008) 810–816.
[10] O. Kayaaltı, B.H. Aksebzeci, I.O. Karahan, K. Deniz, M. Öztürk, B. Yılmaz, S.
Kara, et al., Liver fibrosis staging using CT image texture analysis and soft
computing, Appl. Soft Comput. 25 (2014) 399–413.
[11] X. Zhang, X. Gao, B.J. Liu, K. Ma, W. Yan, L. Liling, et al., Effective staging of
fibrosis by the selected texture features of liver: which one is better, CT or MR
imaging? Comput. Med. Imaging Graph. 46 (2015) 227–236.
[12] A.L. Simpson, L.B. Adams, P.J. Allen, M.I. D’Angelica, R.P. DeMatteo, Y. Fong,
et al., Texture analysis of preoperative CT images for prediction of
postoperative hepatic insufficiency: a preliminary study, J. Am. Coll. Surg. 220
(3) (2015) 339–346.
[13] G. Castellano, L. Bonilha, L.M. Li, F. Cendes, Texture analysis of medical images,
Clin. Radiol. 59 (2004) 1061–1069.
[14] O. Yu, Y. Mauss, G. Zollner, I.J. Namer, J. Chambron, Distinct patterns of active
and non-active plaques using texture analysis on brain NMR images in
multiple sclerosis patients: preliminary results, Magn. Reson. Imaging 17
(1999) 1261–1267.
[15] F. Risse, J. Pesic, S. Young, L.E. Olsson, A texture analysis approach to quantify
ventilation changes in hyperpolarised 3He MRI of the rat lung in an asthma
model, NMR Biomed. 25 (1) (2012) 131–141.
[16] A. Skoch, D. Jirák, P. Vyhnanovská, M. Dezortová, P. Fendrych, E. Rolencová,
et al., Classification of calf muscle MR images by texture analysis, Magn.
Reson. Mater. Phys. 16 (6) (2004) 259–267.
[17] G. Bahl, I. Cruite, T. Wolfson, A.C. Gamst, J.M. Collins, A.D. Chavez, et al.,
Noninvasive classification of hepatic fibrosis based on texture parameters
 N. Daginawala et al. / European Journal of Radiology 85 (2016) 511–517 517

from double contrast-enhanced magnetic resonance images, J Magn. Reson.
Imaging 36 (November (5)) (2012) 1154–1161.
[18] B. Barry, K. Buch, J.A. Soto, H. Jara, A. Nakhmani, S.W. Anderson, Quantifying
liver fibrosis through the application of texture analysis to diffusion weighted
imaging, Magn. Reson. Imaging 32 (1) (2014) 84–90.
[19] H. Yu, K. Buch, B. Li, M. O’Brien, J. Soto, H. Jara, S.W. Anderson, Utility of
texture analysis for quantifying hepatic fibrosis on proton density MRI, J.
Magn. Reson. Imaging 42 (5) (2015) 1259–1265.
[20] N.A. Obuchowski, Receiver operating characteristic curves and their use in
radiology, Radiology 229 (2003) 3–8.
[21] C.E. Metz, Basic principles of ROC analysis, Semin. Nucl. Med. 8 (1978)
283–298.
[22] M. Friedrich-Rust, M.F. Ong, S. Martens, C. Sarrazin, J. Bojunga, S. Zeuzem,
et al., Performance of transient elastography for the staging of liver fibrosis: a
meta-analysis, Gastroenterology 134 (2008) 960.
[23] L. Castéra, J. Foucher, P.H. Bernard, F. Carvalho, D. Allaix, W. Merrouche, et al.,
Pitfalls of liver stiffness measurement: a 5-year prospective study of 13,369
examinations, Hepatology 51 (2010) 828.
[24] M.H. Hsiao, P. Chen, J. Jao, Y. Huang, C. Lee, S. Chao, et al., Quantifying liver
cirrhosis by extracting significant features from MRI T2 image, Sci. World J.
2012 (2012) 343847.
[25] J.H. Yoon, J.M. Lee, E. Klotz, J.H. Jeon, K.B. Lee, J.K. Han, et al., Estimation of
hepatic extracellular volume fraction using multiphasic liver computed
tomography for hepatic fibrosis grading, Invest. Radiol. 50 (4) (2015) 290–296.
[26] P. Lamb, D.V. Sahani, J.M. Fuentes-Orrego, M. Patino, A. Ghosh, P.R. Mendonça,
Stratification of patients with liver fibrosis using dual-energy CT, IEEE Trans.
Med. Imaging 34 (3) (2015) 807–815.
[27] Y.W. Chen, J. Luo, C. Dong, X. Han, T. Tateyama, A. Furukawa, et al.,
Computer-aided diagnosis and quantification of cirrhotic livers based on
morphological analysis and machine learning, Comput. Math. Methods Med.
2013 (2013), 264809.
[28] M. Hori, T. Okada, K. Higashiura, Y. Sato, Y.W. Chen, T. Kim, et al., Quantitative
imaging: quantification of liver shape on CT using the statistical shape model
to evaluate hepatic fibrosis, Acad. Radiol. 22 (3) (2015) 303–309.
[29] S. Guan, W.D. Zhao, K.R. Zhou, W.J. Peng, J. Mao, F. Tang, CT perfusion at early
stage of hepatic diffuse disease, World J. Gastroenterol. 11 (22) (2005)
3465–3467.
[30] O.S. Al-Kadi, Assessment of texture measures susceptibility to noise in
conventional and contrast enhanced computed tomography lung tumour
images, Comput. Med. Imaging Graph. 34 (6) (2010) 494–503.

Descargado para Anonymous User (n/a) en Hosp Italiano Buenos Aires - JCon de ClinicalKey.es por Elsevier en enero 11, 2019.
Para uso personal exclusivamente. No se permiten otros usos sin autorización. Copyright ©2019. Elsevier Inc. Todos los derechos reservados.