- Research article
- Open Access
- Open Peer Review
Artificial neural network aided non-invasive grading evaluation of hepatic fibrosis by duplex ultrasonography
- Li Zhang†1,
- Qiao-ying LI†1,
- Yun-you Duan1Email author,
- Guo-zhen Yan1,
- Yi-lin Yang1 and
- Rui-jing Yang1
© Zhang et al.; licensee BioMed Central Ltd. 2012
- Received: 11 November 2011
- Accepted: 11 June 2012
- Published: 20 June 2012
Artificial neural networks (ANNs) are widely studied for evaluating diseases. This paper discusses the intelligence mode of an ANN in grading the diagnosis of liver fibrosis by duplex ultrasonogaphy.
239 patients who were confirmed as having liver fibrosis or cirrhosis by ultrasound guided liver biopsy were investigated in this study. We quantified ultrasonographic parameters as significant parameters using a data optimization procedure applied to an ANN. 179 patients were typed at random as the training group; 60 additional patients were consequently enrolled as the validating group. Performance of the ANN was evaluated according to accuracy, sensitivity, specificity, Youden’s index and receiver operating characteristic (ROC) analysis.
5 ultrasonographic parameters; i.e., the liver parenchyma, thickness of spleen, hepatic vein (HV) waveform, hepatic artery pulsatile index (HAPI) and HV damping index (HVDI), were enrolled as the input neurons in the ANN model. The sensitivity, specificity and accuracy of the ANN model for quantitative diagnosis of liver fibrosis were 95.0%, 85.0% and 88.3%, respectively. The Youden’s index (YI) was 0.80.
The established ANN model had good sensitivity and specificity in quantitative diagnosis of hepatic fibrosis or liver cirrhosis. Our study suggests that the ANN model based on duplex ultrasound may help non-invasive grading diagnosis of liver fibrosis in clinical practice.
- Artificial neural network
- Ultrasound diagnosis
- Hepatic fibrosis
Hepatic fibrosis is a common feature leading to liver cirrhosis as the result of chronic hepatitis or chronic liver injury. Regardless of the causes, liver fibrosis is characterized by increased extracellular matrix forming hepatic scars . Liver fibrosis is a reversible pathology process. If timely and effective treatment is adopted during the fibrosis stage, it will prevent the liver from developing hepatic cirrhosis. Liver biopsy is considered as the gold standard for final diagnosis of hepatic fibrosis, but the drawbacks, such as sampling errors, pain and complications of invasive procedures [2–4], limits its conventional use in clinical practice. Considering the need for repeated confirmation of the condition of the liver during treatment, a non-invasive modality for grading liver fibrosis or cirrhosis is urgently needed. Imaging technology has an advantage in being noninvasive and allowing repeated maneuverability. Among other methods, ultrasound scanning is the most frequently used with the superiority of inexpensive, real-time imaging and hemodynamic evaluation ability. Many researchers have been committed to establishing a system for liver fibrosis diagnosis or fibrotic stage evaluation by ultrasound [5, 6].
In recent years, artificial neural networks (ANNs) have appeared as tools for clinical decision-making  and are potentially more successful than traditional statistical models in predicting clinical outcomes [8, 9]. ANNs can learn experiential knowledge expressed through internal connections in a similar way to how natural neurons function in the brain and this knowledge can be made available for use . According to the lesion characteristics of liver fibrosis or cirrhosis, one single ultrasonographic index is incapable of reflecting the whole problem. An ANN is suitable for a “multi-parameter” diagnosing mode. In previous studies, liver fibrosis or cirrhosis was graded by an ANN based on laboratory results [11, 12]. The results showed that a three-layer ANN could effectively identify the risk for liver fibrosis in chronic hepatitis B (CHB) patients with positive HBsAg. An ANN consisting of an input layer, an output layer and one or more hidden layer could be adequate as a universal approximator of any nonlinear function [11, 13]. The input layer comprises the data available for analyzing and the output layer comprises the outcome such as diagnosis, prognosis and evaluation.
In this study, we established an ANN model based on the ultrasonographic changes in patients with liver fibrosis for the first time and aimed to predict the risk of early liver cirrhosis. We also assessed the early warning ability of the ANN from an ultrasonographic angle.
Ethical approval of the study protocol
All subjects included in the study provided written informed consent. The study protocol was approved by the ethics committee of the Fourth Military Medical University Tangdu Hospital (Xi’an, China).
Between July 2008 to June 2010, 308 patients infected by hepatitis B virus (HBV) or hepatitis C virus (HCV) and confirmed by laboratory tests at Tangdu Hospital of Fourth Military Medical University, were investigated in this study. The pathological changes in these patients to fibrosis or cirrhosis were evaluated by ultrasound-guided liver biopsy. Finally, 239 patients with liver pathological changes were enrolled in the study. Of these patients, 170 were affected by HBV and 69 by chronic HCV. Subjects with hematonosis and heart disease were excluded from the study. The patients were randomly divided into a training group (179 cases; 75%) and a validation group (60 cases; 25%).
Liver pathologic evaluation
Ultrasound-guided liver biopsy was performed after ultrasound examination. Three pathologists that had no clinical information about the patients evaluated the degree of hepatic fibrosis. Fibrosis was staged from F0 to F4 according to the METAVIR system: F0, no fibrosis; F1, portal fibrosis without septa; F2, few septa; F3, numerous septa without cirrhosis; and F4, cirrhosis . In the present study, fibrosis was defined as F1 to F3 stages; cirrhosis was considered as F4 stage.
Grading standard for the evaluation of ultrasonographic changes on liver fibrosis
obviously coarse or like wavy changes
no or < 500 ml
500 ml ~ 3000 ml
> 3000 ml
The Doppler gate was placed in the porta hepatis to measure the relevant parameters of the portal vein and hepatic artery. Velocity measurements were conducted at 30– 60°. The mean velocity of PVVel and HAPI were calculated automatically by the machine after the waveform trace for three cardiac cycles were obtained. Doppler hepatic vein (HV) waveforms were recorded for at least 5 s with end-expiration breath holding. The middle HV was measured because it has the most consistent triphasic flow in healthy subjects and the most favorable Doppler angle. The Doppler gate was placed in the vessel 2–3 cm away from the inferior vena cava (IVC) to measure the HV waveform. HV waveforms were classified as ‘triphasic’ (reversed flow in at least one phase), ‘biphasic’ (no reversed flow with or without decreased phasic oscillation), or ‘monophasic’ (flat with or without fluttering). Two examiners (Yilin Yang and Guozhen Yan) classified the recorded HV Doppler waveform tracings. The damping index (DI) was calculated by the minimum velocity/maximum velocity of downward HV flow.
Doppler examinations were undertaken by one author (Li Zhang) without prior knowledge of the clinical and biochemical status of the study population. The reproducibility of this method was evaluated with repeated ultrasound measurements of portal venous blood flow velocity in 10 healthy subjects over 5 consecutive days . In order to minimize inter-observer variation, at the beginning we unified the method of measuring each index, and all parameters were measured by the same observer who had no knowledge of the patients’ conditions on the same machine. Each index was calculated as the mean of three consistent measurements. The Doppler parameters we measured were consistent in all subjects.
Development of an ANN model
ANN models were constructed by using neural-network software (Statistic Neural Networks, version 4.0). The architecture of the ANN consisted of three layers; i.e., the input, hidden and output layers. Each layer contained 5 neurons, 11 neurons and 1 neuron, respectively. Neurons were tied together with weighted connections. The number of the network layers, hidden neurons and the stopping criteria were determined through a trial-and-error process. The input layer simply fed information, as well as related predictive factors, into the network, while nodes in the hidden and output layers processed information. The input data selected for the development of the neural network were ultrasongraphic parameters. The output layers contained one neuron (0, fibrosis; 1, cirrhosis).
The training rule that was used was back-propagation of error. During the training, the corresponding known outputs of the system were held in the output nodes to compare with the results produced by the network. The nodes in the hidden layer had no prescribed initial values and helped to allow complex relationships between the input and output nodes to evolve. Information was transported from the input layer to the output layer by calculating the sum at each node, which was derived from combining all the nodes in the previous layer. Training was terminated when the sum of square errors was at a minimum. At the end of each training session, the network was tested and the prediction accuracy was calculated. We then selected the best network in terms of accuracy.
Continuous variables were expressed as mean ± standard deviation (SD). Categorical variables were compared using χ2 analysis and continuous variables were compared by the Mann–Whitney U Test, or Kruskal-Wallis Test. Performance of the ANN prediction was tested using receiver operating characteristic (ROC) curve analysis. The ANN predictions for the diagnosis of liver fibrosis stage were expressed in terms of accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and Youden’s index (YI) for several considered cut-off values. A value of p<0.05 was considered significant in all the analyses.
The distribution and clinical characteristics of 239 subjects
Stage of Liver Fibrosis
Statistical comparison of the ultrasonographic viriables between the fibrosis group and the cirrhosis group
0.645 ± 0.055
0.816 ± 0.129
0.639 ± 0.054
0.709 ± 0.112
Thickness of Spleen (cm)
3.279 ± 0.439
4.058 ± 0.672
0.413 ± 0.035
0.516 ± 0.082
HV waveform(0/ι/П) a
1.247 ± 0.155
1.147 ± 0.283
0.697 ± 0.050
0.711 ± 0.052
0.458 ± 0.131
0.574 ± 0.111
Ultrasound diagnosis of neural network used in the results of liver fibrosis
Predictive performance of ANN (artificial neural network)
Currently, with the development of laboratory and imaging means for staging the fibrotic evolution of chronic liver diseases, clinical validation has highlighted the fact that, overall, liver biopsy is probably an imperfect gold standard . Actually, even a 25 mm long liver biopsy has a 25% rate of discordance for fibrosis staging . Therefore, liver biopsy is prone to sampling errors and to intraobserver and interobserver variability [17, 18]. Also, when the specimen size is adequate, the level of experience of the pathologist may even be more important . Invasive procedures are not suitable for regular clinical monitoring of disease progression. Even though there is a high prevalence of chronic liver disease worldwide and it represents a significant public health problem, liver biopsy is obviously not appropriate for screening liver fibrosis and cirrhosis.
Liver fibrosis is a kind of diffuse lesion involved in multiple structures of the liver. Several factors are taken into account, either by imaging or laboratory tests, before the diagnosis of liver fibrosis or cirrhosis can be made. When multiple and diverse factors are likely to influence decision making, computer-based decision support systems, such as neural networks, are capable of handling large amounts of data and are helpful in arriving at or supplementing a correct decision by clinicians [20, 21]. ANNs have been used in medicine for various purposes, including prediction of mortality of patients with cirrhosis of the liver [22, 23]. An intelligent mode was also compared to MELD scoring, Child–Pugh’s scoring and other conventional logistic regression models and performance of an ANN was significantly better than those of the models.
Real time ultrasonography has become an integral part of the non-invasive evaluation of chronic liver disease in many clinical settings while the search for a non-invasive imaging marker for staging liver fibrosis or cirrhosis is inactive. The performance of ultrasonographic imaging as a non-invasive diagnostic or prognostic modality for liver fibrosis or cirrhosis, as well as for correlation with histological changes and functional disorders of the liver, remains controversial and is still debated. However, recent advances in ultrasound technology have improved the diagnostic accuracy of fibrosis in patients with chronic liver disease. Aube et al.  studied a high-resolution ultrasound probe of the liver parenchyma, liver surface smoothness, spleen size and portal vein blood flow rate, using 11 indicators in ultrasonic testing, and found an accuracy of 82 ~ 88% for surface nodular changes in the liver and spleen thickness in the diagnosis of cirrhosis of the liver.
In the present study, we constructed a multi-parameter dependent diagnostic model based on ultrasound in order to avoid the shortcomings of a single-parameter decision making model of ultrasonography. We took several ultrasonographic variables into consideration including grey-scale and Doppler indexes such as the liver parenchyma, liver edge, PVVel, and HAPI. Secondly, according to the different treatment principles for liver fibrosis and liver cirrhosis, we divided the patients into two groups: the fibrotic group (F1-F3 stage) and the cirrhotic group (F4 stage). Variables like the liver parenchyma, liver envelope, ascites and HV waveform were graded from 0 to 2 according to the imaging changes in different stages. The variables were quantitatively described, for example, as PVVel, HAPI, HARI and DI, and were compared by the actual values. DI was used to quantitatively assess the extent of the abnormal HV waveform. The relatively large number of intra- and extra-hepatic variables was considered in the study to work with the largest possible amount of information. In fact, data collection was performed by trying to include all variables that could have a connection with the problem. However, some of these variables may contain confusing information, or even completely irrelevant information. Selecting the significant variables after statistical analysis can increase diagnostic accuracy as well as sensitivity and specificity. Some experts would consider non-invasive serum tests of fibrosis with AUC-ROC values of 0.85 to 0.90 to be as good as a liver biopsy for staging fibrosis . In our study, the diagnostic performance achieved by the ultrasound-based ANN was considered as having AUC-ROC values around 0.92.
Limitation of the study
This study has several limitations that must be taken into account. Firstly, the ultrasound variables in the present study did not fully cover all involved parameters, although some of these variables could have contributed to improvement of the ANN. The ANN model was constructed using 10 variables as the proposed input neurons. This Secondly, the number of patients was limiting. In an ANN model, each group should have 100 patients to avoid the risk of overfitting the data. This was not fully achieved for the validation group (60 of 239 patients). Finally, we could not evaluate the accuracy of pathological diagnosis caused by sampling error or variation in the experience of the pathologists.
The study highlighted the construction and assessment of an ANN for identifying the risk of liver cirrhosis by a non-invasive imaging modality. In the study, we provided evidence that this intelligent model can accurately predict liver cirrhosis by ultrasound. It could be used to improve clinical decisions for patients with chronic liver disease.
The study was supported generously by the Department of General Surgery in Tangdu Hospital and the Department of Pathology in the Fourth Military Medical University.
- Friedman SL: Seminars in medicine of the Beth Israel Hospital, Boston. The cellular basis of hepatic fibrosis. Mechanisms and treatment strategies. N Engl J Med. 1993, 328: 1828-1835. 10.1056/NEJM199306243282508.View ArticlePubMedGoogle Scholar
- Piccinino F, Sagnelli E, Pasquale G, Giusti G: Complications following percutaneous liver biopsy. A multicentre retrospective study on 68,276 biopsies. J Hepatol. 1986, 2: 165-173. 10.1016/S0168-8278(86)80075-7.View ArticlePubMedGoogle Scholar
- Bravo AA, Sheth SG, Chopra S: Liver biopsy. N Engl J Med. 2001, 344: 495-500. 10.1056/NEJM200102153440706.View ArticlePubMedGoogle Scholar
- Narasimhan G, Sargios TN, Kalakuntla R, Homel P, Clain DJ, Theise ND, Bodenheimer HC, Min AD: Treatment rates in patients with chronic hepatitis C after liver biopsy. J Viral Hepat. 2006, 13: 783-786. 10.1111/j.1365-2893.2006.00763.x.View ArticlePubMedGoogle Scholar
- Knodell RG, Ishak KG, Black WC, Chen TS, Craig R, Kaplowitz N, Kiernan TW, Wollman J: Formulation and application of a numerical scoring system for assessing histological activity asymptomatic chronic active hepatitis. Hepatology. 1981, 1: 431-435. 10.1002/hep.1840010511.View ArticlePubMedGoogle Scholar
- Chevallier M, Guerret S, Chossegros P, Gerard F, Grimaud JA: A histological semiquantitative scoring system for evaluation of hepatic fibrosis in needle liver biopsy specimens: comparism with morphometric studies. Hepatology. 1994, 20: 349-355. 10.1002/hep.1840200213.View ArticlePubMedGoogle Scholar
- Dayhoff JE, Deleo JM: Artificial neural network: opening the black box. Cancer. 2001, 91: 1615-1635. 10.1002/1097-0142(20010415)91:8+<1615::AID-CNCR1175>3.0.CO;2-L.View ArticlePubMedGoogle Scholar
- Green M, Bjork J, Forberg J, Ekelund U, Edenbrandt L, Ohlsson M: Comparison between neural networks and multiple logistic regression to predict acute coronary syndrome in the emergency room. Artif Intell Med. 2006, 38: 305-318. 10.1016/j.artmed.2006.07.006.View ArticlePubMedGoogle Scholar
- Abbod MF, Catto JW, Linkens DA, Hamdy FC: Application of artificial intelligence to the management of urological cancer. J Urol. 2007, 178: 1150-1156. 10.1016/j.juro.2007.05.122.View ArticlePubMedGoogle Scholar
- Cross SS, Harrison RF, Kennedy RL: Introduction to neural networks. Lancet. 1995, 346: 1075-1079. 10.1016/S0140-6736(95)91746-2.View ArticlePubMedGoogle Scholar
- Raoufy MR, Vahdani P, Alavian SM, Fekri S, Eftekhari P, Gharibzadeh S: A novel method for diagnosing cirrhosis in patients with chronic hepatitis B: artificial neural network approach. J Med Syst. 2011, 35: 121-126. 10.1007/s10916-009-9348-8.View ArticlePubMedGoogle Scholar
- Wang D, Wang Q, Shan F, Liu B, Lu C: Identification of the risk for liver fibrosis on CHB patients using an artificial neural network based on routine and serum markers. BMC Infect Dis. 2010, 10: 251-10.1186/1471-2334-10-251.View ArticlePubMedPubMed CentralGoogle Scholar
- Dariani S, Keshavarz M, Parviz M, Raoufy MR, Gharibzadeh S: Modeling force-velocity relation in skeletal muscle isotonic contraction using an artificial neural network. Biosystems. 2007, 90: 529-534. 10.1016/j.biosystems.2006.12.004.View ArticlePubMedGoogle Scholar
- Dietrich CF, Lee JH, Gottschalk R, Herrmann G, Sarrazin C, Caspary WF, Zeuzem S: Hepatic and portal vein flow pattern in correlation with intrahepatic fat deposition and liver histology in patients with chronic hepatitis C. AJR Am J Roentgenol. 1998, 171: 437-443.View ArticlePubMedGoogle Scholar
- Castera L, Pinzani M: Biopsy and non-invasive methods for the diagnosis of liver fibrosis: does it take two to tango?. Gut. 2010, 59: 861-866. 10.1136/gut.2010.214650.View ArticlePubMedGoogle Scholar
- Bedossa P, Darge`re D, Paradis V: Sampling variability of liver fibrosis in chronic hepatitis C. Hepatology. 2003, 38: 1449e57-View ArticleGoogle Scholar
- Maharaj B, Maharaj RJ, Leary WP, Cooppan RM, Naran AD, Pirie D, Pudifin DJ: Sampling variability and its influence on the diagnostic yield of percutaneous needle biopsy of the liver. Lancet. 1986, 1: 523e5-Google Scholar
- Regev A, Berho M, Jeffers LJ, Milikowski C, Molina EG, Pyrsopoulos NT, Feng ZZ, Reddy KR, Schiff ER: Sampling error and intraobserver variation in liver biopsy in patients with chronic HCV infection. Am J Gastroenterol. 2002, 97: 2614e18-View ArticleGoogle Scholar
- Rousselet MC, Michalak S, Dupré F, Croué A, Bedossa P, Saint-André JP, Calès P: Sources of variability in histological scoring of chronic viral hepatitis. Hepatology. 2005, 41: 257e64-View ArticleGoogle Scholar
- Hunt D, Haynes R, Hanna S, Smith K: Effects of computer-based decision support systems on physician performance and patient outcomes. JAMA. 1998, 280: 1339-1346. 10.1001/jama.280.15.1339.View ArticlePubMedGoogle Scholar
- Blumenthal D: The future of quality measurement and management in a transforming health care system. JAMA. 1997, 278: 1622-1625. 10.1001/jama.1997.03550190090053.View ArticlePubMedGoogle Scholar
- Cucchetti A, Vivarelli M, Heaton ND, Phillips S, Piscaglia F, Bolondi L, La Barba G, Foxton MR, Rela M, O'Grady J, Pinna AD: Artificial neural network is superior to MELD in predicting mortality of patients with end-stage liver disease. Gut. 2007, 56: 253-258. 10.1136/gut.2005.084434.View ArticlePubMedGoogle Scholar
- Banerjee R, Das A, Ghoshal UC, Sinha M: Predicting mortality in patients with cirrhosis of liver with application of neural network technology. J Gastroenterol Hepatol. 2003, 18: 1054-1060. 10.1046/j.1440-1746.2003.03123.x.View ArticlePubMedGoogle Scholar
- Aubé C, Winkfield B, Oberti F, Vuillemin E, Rousselet MC, Caron C, Calès P: New Doppler ultrasound signs improve the non-invasive diagnosis of cirrhosis or severe liver fibrosis. Eur J Gastroenterol Hepatol. 2004, 16: 743-751. 10.1097/01.meg.0000108357.41221.e5.View ArticlePubMedGoogle Scholar
- Afdhal NH, Nunes D: Evaluation of liver fibrosis: a concise review. Am J Gastroenterol. 2004, 99: 1160e74-View ArticleGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1472-6947/12/55/prepub
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.