Artificial neural network aided non-invasive grading evaluation of hepatic fibrosis by duplex ultrasonography

Li Zhang, Qiaoying Li, Yunyou Duan, Guozhen Yan, Yilin Yang, Ruijing YangBMC Medical Informatics Decis. Mak 2012 · DOI

TL;DR

This paper presents an artificial neural network (ANN) model that uses duplex ultrasonography to non-invasively grade hepatic fibrosis. By analyzing five key ultrasound parameters, the ANN achieved high accuracy in distinguishing between fibrosis and cirrhosis, offering a promising alternative to invasive liver biopsies for clinical decision-making.

What Problem Does This Paper Address?

Liver fibrosis, a precursor to cirrhosis, is a significant health issue. The current gold standard for diagnosing and staging liver fibrosis is liver biopsy. However, liver biopsy has several drawbacks: it's invasive, painful, carries risks of complications, and is prone to sampling errors, leading to potential inaccuracies. This makes it unsuitable for routine monitoring of disease progression. Consequently, there's a critical need for non-invasive methods to accurately assess and grade liver fibrosis. While ultrasound is a widely used, non-invasive imaging technique, its effectiveness in reliably grading fibrosis has been debated. This paper aims to bridge this gap by developing an ANN model that leverages multiple ultrasound parameters to improve the accuracy of fibrosis grading.

Methods

The core of this research is the development of an Artificial Neural Network (ANN) model. The ANN architecture employed a three-layer design: an input layer, a hidden layer, and an output layer. Specifically, the input layer had 5 neurons, the hidden layer had 11 neurons, and the output layer had 1 neuron. The network was trained using the back-propagation of error algorithm.

The selection of input parameters was crucial. Initially, several ultrasonographic parameters were considered. Through a data optimization procedure, five parameters were identified as significant and were used as input neurons for the ANN:

  • Liver parenchyma: A grey-scale parameter reflecting the texture of the liver.
  • Thickness of the spleen: Another grey-scale parameter.
  • Hepatic vein (HV) waveform: A Doppler parameter describing blood flow patterns in the hepatic vein.
  • Hepatic artery pulsatile index (HAPI): A Doppler parameter measuring blood flow pulsatility in the hepatic artery.
  • HV damping index (DI): A Doppler parameter quantifying the damping of the hepatic vein waveform.

The output layer was designed to classify patients into two categories: 0 for fibrosis (stages F1-F3) and 1 for cirrhosis (stage F4). The ANN's performance was evaluated using standard metrics like accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Youden's index (YI), along with Receiver Operating Characteristic (ROC) curve analysis.

Datasets

The study utilized data from 239 patients who had confirmed liver fibrosis or cirrhosis via ultrasound-guided liver biopsy. These patients were infected with Hepatitis B virus (HBV) or Hepatitis C virus (HCV). The patients were randomly divided into two groups:

  • Training group: 179 patients (75%) used to train the ANN model.
  • Validating group: 60 patients (25%) used to evaluate the trained ANN's performance.

The paper does not introduce a new benchmark dataset but rather uses a clinical dataset collected from patients at Tangdu Hospital.

Key Insights and Ideas

  • Multi-parameter approach is key: The study highlights that a single ultrasound parameter is insufficient for accurately grading liver fibrosis. Combining multiple parameters within an ANN allows for a more comprehensive assessment, mimicking how complex biological systems are evaluated.
  • ANNs excel at complex pattern recognition: The ANN's ability to learn intricate relationships between various ultrasound parameters and the actual fibrosis stage is demonstrated. This makes them powerful tools for decision support in complex medical diagnoses.
  • Non-invasive grading is feasible and valuable: The success of the ANN model suggests that accurate, non-invasive grading of liver fibrosis is achievable using duplex ultrasonography. This has significant implications for patient care, reducing the need for invasive procedures.
  • Specific ultrasound parameters are predictive: The identification of liver parenchyma, spleen thickness, HV waveform, HAPI, and DI as significant predictors underscores the diagnostic value of these specific ultrasound measurements.

Key Results

The ANN model demonstrated strong performance in diagnosing liver fibrosis and cirrhosis:

  • Accuracy: 88.3%
  • Sensitivity: 95.0% (correctly identifying those with fibrosis/cirrhosis)
  • Specificity: 85.0% (correctly identifying those without the condition being tested for, i.e., correctly classifying healthy or less severe cases)
  • Youden's Index (YI): 0.80 (a measure of diagnostic test accuracy)
  • Area Under the ROC Curve (AUC): 0.922 (indicating excellent discriminatory ability)

The study also identified statistically significant differences in liver parenchyma, spleen thickness, HV waveform, HAPI, and DI between the fibrosis and cirrhosis groups, reinforcing their importance as input features.

Limitations and Future Work

The authors acknowledge several limitations:

  • Incomplete parameter coverage: The ultrasound variables included in the study did not encompass all potentially relevant parameters.
  • Sample size for validation: The validation group (60 patients) was considered relatively small, potentially increasing the risk of overfitting the data. The authors suggest that ideally, ANN models require around 100 patients per group.
  • Pathological diagnosis variability: The study could not fully evaluate the accuracy of the pathological diagnosis itself, which can be affected by sampling errors and inter-observer variability among pathologists.

Future work could involve expanding the dataset to include more patients, incorporating additional ultrasound parameters, and potentially integrating other non-invasive markers (e.g., serum markers) into the ANN model for even greater diagnostic precision. Further validation in larger, diverse patient populations would also be beneficial.

Summary updated

BibTeX
@article{zhang2012artificial,
  author = {Li Zhang and Qiaoying Li and Yunyou Duan and Guozhen Yan and Yilin Yang and Ruijing Yang},
  title = {{Artificial neural network aided non-invasive grading evaluation of hepatic fibrosis by duplex ultrasonography}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2012},
  doi = {10.1186/1472-6947-12-55},
}
Figure