| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Original Research |
1 Department of Radiology, Gifu University School of Medicine, 1-1 Yanagido,
Gifu 501-1194, Japan.
2 Department of Radiology Services, Gifu University Hospital, Gifu, Japan.
3 College of Computer Science and Information Engineering, Guangxi University,
Nanning City, Guangxi, P. R. China.
4 Department of Immunopathology, Gifu University Graduate School of Medicine,
Gifu, Japan.
5 Department of Information Science, Faculty of Engineering, Gifu University,
Gifu, Japan.
Received August 28, 2006;
accepted after revision February 22, 2007.
Supported in part by the Health and Labour Sciences Research Grants for
Third Term Comprehensive Control Research for Cancer from the National Cancer
Center Hospital, Tokyo, Japan.
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
MATERIALS AND METHODS. Liver parenchymal textures on the MR images of 52 patients who underwent partial hepatectomy were processed by the computer algorithm and reviewed by two radiologists. The texture features using the finite difference method were processed by an artificial neural network program containing a three-layer learning algorithm of the back propagation, composed of a seven-unit input layer, a six-unit hidden layer, and a one-unit output layer. The radiologists assigned confidence levels for the presence of hepatic fibrosis. Degrees of hepatic fibrosis were determined semiquantitatively by a pathologist. Algorithm outputs and radiologists' interpretations were correlated with degrees of fibrosis using Spearman's rank correlation analysis, and diagnostic performances were evaluated using receiver operating characteristic curve analysis.
RESULTS. By the computer algorithm, the Az (area under the curve) value was greater for gadolinium-enhanced equilibrium phase images (Az = 0.801) than for T1-weighted (Az = 0.597) or T2-weighted (Az = 0.525) images (p < 0.05), and the outputs of equilibrium phase images showed a moderate correlation (r = 0.502, p = 0.001) with the pathologic grades. By the radiologists' interpretations, the Az value for all images combined (Az = 0.715) was greater than that of portal venous phase images (Az = 0.503) (p < 0.05), and the confidence levels of all images combined were moderately correlated (r = 0.473, p = 0.002) with pathologic grades.
CONCLUSION. Computer algorithm analysis of equilibrium phase images was found to reflect the degree of fibrosis most accurately. MR image texture analysis performed using the computer algorithm was found to have a potential usefulness for the diagnosis of hepatic fibrosis.
Keywords: computer-aided diagnosis • fibrosis • liver • MRI
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
Regenerative nodules and surrounding fibrosis become prominent as cirrhosis progresses, and the computerized recognition of the degree of a reformed parenchymal texture in cirrhosis has the potential to assess the degree of cirrhosis. MRI has a great advantage for the depiction of hepatic parenchymal multinodularity in cirrhosis because of its high sensitivity for increased cellularity, extracellular free water, fibrosis, glycogen, fat, and metal deposition [6]. Meanwhile, the development of a computer-aided diagnostic system that reduces the burden on radiologists and maintains the quality of service has been attempted on a number of occasions [7-9]. Although expectations of computer-aided diagnostic systems are high, only a few articles on computer-aided diagnosis for the evaluation of hepatic fibrosis or cirrhosis by CT or sonography [10, 11] have been published, and few articles on MRI of these subjects are available. We performed a preliminary study of the diagnostic performance of a computer algorithm we developed for the MR-based diagnosis of hepatic fibrosis.
|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
MRI
All MR images were obtained in the axial plane with a section thickness of
8-10 mm, a 2- to 3-mm intersection gap, and a field of view of 22 x
29-26 x 35 cm, which covered the whole liver, using a superconducting
imager operating at 1.5 T (Signa Horizon, GE Healthcare) with a phased-array
body coil. The imaging protocol comprised breath-hold T1-weighted spoiled
gradient-recalled echo (GRE) (TR/TE, 150/1.6; matrix, 512 x 224 [number
of frequency x phase encoding]; flip angle, 90°; signal acquisition,
1; slices, 18 per 26-second acquisition time), respiratory-triggered
T2-weighted fast spin-echo (effective TR range/TE range, 3,333-8,571/77-80;
matrix, 512 x 256; echo-train length, 8-18; signals acquired, 3-4;
slices, 20 per 3.2- to 5.0-minute acquisition time), and gadolinium-enhanced
triple-phase T1-weighted spoiled GRE (150/1.6; matrix, 512 x 224; flip
angle, 90°; signal acquisition, 1; slices, 18 per 26-second acquisition
time) sequences.
Gadolinium-enhanced GRE images were obtained after an antecubital IV bolus injection of 0.1 mmol/kg of gadopentetate dimeglumine (Magnevist, Schering) followed by a 20-mL sterile saline flush. Scanning delays for the hepatic arterial dominant, portal venous, and equilibrium phases were 18 seconds, 60 seconds, and 3 minutes, respectively, after initiating the contrast material injection. The chemical shift selective fat-saturation technique was not used because the acquisition time and number of slices were traded.
Pathologic Evaluation
A pathologist who was blinded to patient history and radiology and surgery
reports retrospectively reviewed formalin-fixed slides of surgical specimens
to evaluate the degree of hepatic fibrosis in nontumorous liver parenchyma
using the 5-point Desmet scale: F0 indicates no fibrosis; F1, mild fibrosis;
F2, moderate fibrosis; F3, severe fibrosis; and F4, cirrhosis
[12].
Computer Algorithm Analysis
On a DICOM image of each MR sequence, a radiologist (unaware of patient
information and the pathologic diagnosis) placed 10 square regions of interest
(ROIs), each consisting of 32 x 32 pixels on the liver parenchyma, but
avoiding large blood vessels, focal hepatic lesions, and prominent hepatic
artifacts. A total of 10 ROIs were placed in the liver parenchyma devoid of
large blood vessels and artifacts: eight in the right lobe (typically, two
each in segments V, VI, VII, and VIII) and two in the left lobe (typically one
each in segments II and III). In each ROI, texture features using the finite
difference method were obtained as seven numeric values: mean gray-scale
value, SD, contrast, angular second moment, entropy, mean, and inverse
difference moment [13]
(Appendix 1).
The texture features were further processed by an artificial neural network program containing a three-layer learning algorithm of the back propagation, comprising a seven-unit input layer, a six-unit hidden layer, and a one-unit output layer [14] (Fig. 1). The seven numeric parameters by the finite difference method in each ROI were entered into the artificial neural network program, and a probability value for the presence of hepatic fibrosis in an ROI was outputted as a continuous number between 0 (absent) and 1 (present). The 10 probability values of the 10 ROIs were averaged, and the overall average value obtained was used as the total probability value for the presence of hepatic fibrosis. MR images in eight patients (four with no fibrosis [F0] and four with cirrhosis [F4]) were used for training the artificial neural network program. MR images in the remaining 44 patients (three with no fibrosis [F0], 10 with mild fibrosis [F1], 15 with moderate fibrosis [F2], 13 with severe fibrosis [F3], and three with cirrhosis [F4]) were then used to evaluate the computer algorithm.
Radiologists' Interpretations
A study coordinator, an author with 5 years of posttraining experience in
interpreting body MR images, prepared the MR images for review on a
commercially available DICOM viewer; and two independent gastrointestinal
radiologists with 15 and 7 years of posttraining experience at interpreting
body MR images, who were unaware of patient history and radiology and surgery
report details, reviewed the MR images in 44 patients for the presence of
hepatic fibrosis. When a disagreement occurred, consensus was reached by
discussion. All sequences—that is, T1-weighted, T2-weighted, hepatic
arterial dominant phase, portal venous phase, equilibrium phase, combination
unenhanced T1and T2-weighted (unenhanced images combined), and all combination
images (all images combined)—were reviewed in a randomized fashion.
MR images were magnified on the viewer frame to an appropriate degree so that only liver parenchymal textures were evaluated and liver contours were hidden to exclude the influence of liver configuration findings (Figs. 2A, 2B, 2C, 3A, 3B, and 3C). The radiologists on the DICOM viewer paged and reviewed only a few images at the middle level of the liver, taking care not to see the liver contours. After reviewing the MRI textures in the right and left lobes of the liver, a confidence level for the presence of fibrosis was assigned using a 4-point scale (1, definitely absent; 2, probably absent; 3, probably present; and 4, definitely present) for each sequence in a patient. The radiologists assigned confidence levels for the presence of hepatic fibrosis, paying attention to the degree of reticulonodular pattern of liver parenchymal textures on MR images that reflected pathologic characteristics of cirrhosis: fibrotic tissue surrounding regenerative nodules.
|
|
|
|
|
|
Statistical Analysis
Multiple regression analysis was used to examine correlations between the
seven numeric values of texture features by the finite difference method and
the pathologic degree of hepatic fibrosis. Receiver operating characteristic
(ROC) curve analysis, in LABMRMC 1.0B software (Metz CE), was used to
determine observer performances for the detection of hepatic fibrosis. For
this analysis, the total probability for the presence of fibrosis determined
by the computer algorithm, confidence levels for the presence of fibrosis
determined by the radiologists' interpretations, and pathologic grades (F0-F2
were assumed to be actually negative, and F3 and F4, actually positive) were
entered. Spearman's rank correlation coefficients were calculated to correlate
computer algorithm output values and the radiologists' confidence levels for
the presence of hepatic fibrosis with the pathologic fibrosis grades.
Interobserver variability was assessed using kappa statistics.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
|
Values for areas under the ROC curves (Az) are summarized in Table 2. By the computer algorithm, the Az value was greater for equilibrium phase images (Az = 0.801) than for T1-weighted (Az = 0.597) or T2-weighted (Az = 0.525) images (p < 0.05). By the radiologists' interpretations, the Az value for all images combined (Az = 0.715) was greater than that of the portal venous phase images (Az = 0.503) (p < 0.05). The Az value for equilibrium phase images in the computer algorithm (Az = 0.801) was greater than for equilibrium phase images of radiologists (Az = 0.546).
|
Correlation coefficients of the total probability determined by the computer algorithm and the confidence levels determined by radiologists with pathologic grades are summarized in Table 3. By the computer algorithm, the outputs of T1-weighted (r = 0.329, p = 0.031) and portal venous phase (r = 0.308, p = 0.044) images were weakly correlated and those of equilibrium phase images showed a moderate correlation (r = 0.502, p = 0.001) with the pathologic grades. By the radiologists' interpretations, the confidence levels of unenhanced images combined were weakly correlated (r = 0.357, p = 0.019) and those of all images combined were moderately correlated (r = 0.473, p = 0.002) with pathologic grades.
|
The kappa values for the two radiologists who independently rated images
ranged from 0.62 to 0.80 (mean, 
= 0.71), indicating substantial
agreement.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
According to the radiologists' interpretations, correlation coefficients and Az values were increased by interpreting both unenhanced T1- and T2-weighted images. This was presumably because the reticular patterns of hepatic fibrosis and regenerative nodules in cirrhosis were more efficiently recognized by radiologists after combination interpretation. Hepatic fibrosis, which often harbors vascular spaces, was observed as isointense or slightly hypointense on T1-weighted images and as slightly hyperintense on T2-weighted images because of increased levels of free water in fibrotic tissue. Moreover, regenerative nodules with possible hemosiderin deposition might show discrete hypointensity on T2-weighted images [15, 16].
Correlation coefficients and Az values were greatest with contrast-enhanced equilibrium phase images by the computer algorithm, whereas the values of equilibrium phase images by the radiologists' interpretations were not so high. We infer that the fibrotic tissue surrounding regenerative nodules showed reticular hyperintensity due to delayed enhancement on gadolinium-enhanced equilibrium phase images, and that the reticular enhancement pattern was better recognized by the computer algorithm than by the radiologists. Although radiologists can often detect the reticular enhancement of hepatic fibrosis, radiologists' performance was significantly lower than that of the computer algorithm.
We masked liver contours from the radiologists as well as information on hepatic configurations, such as enlargement or shrinking of liver segments and irregular liver surfaces. However, radiologists in the clinical setting interpret hepatic parenchymal textures and configurations simultaneously to determine the degree of cirrhosis. Thus, if the radiologists had been able to observe the whole liver on multiple-sequence images in our study, their performance at diagnosing hepatic fibrosis would probably have improved. Therefore, our finding that the diagnostic performance of the computer algorithm with equilibrium phase images was high may indicate that the computer algorithm could be used as a simple measurement tool for evaluating hepatic parenchymal textures and could aid in the diagnosis of hepatic fibrosis.
We included MR images of patients who underwent partial hepatectomy; the study cohort did not include cases of severe cirrhosis with highly impaired hepatic functional reserve. If our study had included a number of patients with severe fibrosis or cirrhosis, our results by ROC analysis and correlation might have been improved. Severe cirrhosis is often visualized as numerous regenerative nodules and as a reticular fibrotic pattern on MR images, and thus the diagnosis of severe hepatic fibrosis is not difficult. However, the diagnosis of hepatic fibrosis of an intermediate degree may not be easy because hepatic parenchymal textural alterations (due to regenerative nodules and fibrosis in such livers) are too subtle to convince radiologists of the presence of disease. Regarding the detection and grading of hepatic fibrosis of a moderate or lesser degree, the computer algorithm proved superior to the radiologists' interpretations in our study.
We did not find strong correlations between pathologic fibrosis grades and outputs by computer algorithm or by radiologists, and we inferred some possible limitations in our study methods: We adopted only one pathologic parameter of fibrosis among a variety of factors in the pathologic diagnosis of chronic hepatitis or cirrhosis, such as periportal necrosis, intralobular necrosis, and portal inflammation [17]. Although the radiologists assessed hepatic fibrosis, paying attention to the reticulonodular pattern indicating fibrotic tissue surrounding regenerative nodules, sizes of regenerative nodules and thickness of fibrosis vary in individual patients, depending on the kind of underlying diseases. We did not include patients with severe cirrhosis because such patients could not undergo surgical treatment for liver tumors. The pathologist evaluated liver parenchyma near hepatic tumors that were surgically resected, whereas the two radiologists reviewed a few MR images at the middle level of the liver: Liver parenchyma evaluated by the pathologist and the radiologists did not necessarily match, and such a difference might have yielded some correlation errors because hepatic fibrosis does not occur evenly in the liver.
Our study has several other limitations. The study population was small because the study was conducted in a single institution. A multiinstitutional study of a program with a more user-friendly interface may be warranted to determine the true clinical impact. Although the radiologists gave diagnoses referring to multiple sequences, the computer algorithm processed images by a single sequence. Thus, we need to develop computer algorithms that process multiple sequences in combination. We used a non-fat-suppressed 2D GRE sequence for the gadolinium-enhanced study because that was the only sequence available as a result of the limited performance of our MR imager at that time. Recently, a variety of 3D GRE sequences allowing acquisitions of thin slices have been available for gadolinium-enhanced studies, and we have been using such sequences. Nevertheless, we believe that the MR images in our study exhibited fine textures of liver parenchyma because we used high imaging matrices (512 x 224).
In conclusion, this texture analysis of MR images of the liver using the finite difference method and an artificial neural network reflected the degree of hepatic fibrosis, and equilibrium phase images were found to be most suitable for analytic purposes. Our study indicates that MR image texture analysis using suitable computer algorithms offers the potential for predicting the degree of hepatic fibrosis.
APPENDIX 1: Texture Features for the Finite Difference Method
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
Among the five features by the finite difference method, "contrast" is the difference moment of the matrix and is a measure of local variations present in the image. Contrast also measures the spread of matrix values. A low contrast value reflects uniformly gray images; conversely, high-contrast images show much local variation. "Angular second moment" is a measure of the textural uniformity of an image. When entries in the co-occurrence matrix are equal, this has a low value. Moreover, this measure is higher when some entries are high and others are low. If an image has a uniform gray level, there will be only one entry in the co-occurrence matrix, and this will produce the highest value for the angular second moment. "Entropy" is a measure of the randomness of the image texture. When the spatial co-occurrence matrix is equal, entropy is the highest. Entropy is at a minimum when the elements in a matrix are very unequal. Hence, higher values for entropy indicate greater randomness in the image. "Mean" is the average value in the probabilities function, and "inverse difference moment" measures image homogeneity.
Regarding the two features by histogram analysis, "mean gray-scale value" is defined as the average pixel value in a region of interest (ROI) and indicates the intensity or brightness of a region. For example, a value of 255 is very bright and a value of 0, very dark. "Standard deviation" indicates variation from the mean gray-scale value. If an ROI is homogeneous, the SD is small; if the region is heterogeneous, the SD is large. The mean and the inverse difference moment by the finite difference method are similar to the mean gray-scale value and SD by histogram analysis. The mean and inverse difference moment are calculated using second-order statistics, whereas the mean gray-scale value and SD are calculated using first-order statistics.
Acknowledgments
The contribution of Noriyuki Moriyama, Research Center for Cancer
Prevention and Screening, National Cancer Center Hospital, Tokyo, Japan, to
our research is greatly appreciated. We thank Wenguang Li, Tetsuji Tajima, and
Takeshi Hara, Department of Information Science, Faculty of Engineering, Gifu
University, Gifu, Japan, for providing helpful technical support and for
performing many of the tasks associated with this project.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionAPPENDIX 1: Texture Features...References |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
