Utility of Convolutional Neural Network-Based Algorithm in Medical Images for Liver Fibrosis Assessment

Utility of Convolutional Neural Network-Based Algorithm in Medical Images for Liver Fibrosis Assessment

Liver fibrosis is a critical stage that can lead to hepatic dysfunction and is significant in the progression to portal hypertension, biliary cirrhosis, and hepatocellular carcinoma. Accurate assessment of liver fibrosis remains a clinical concern for physicians. Historically, liver biopsy has been considered the gold standard for diagnosing and assessing liver fibrosis. However, due to its invasive nature, sampling variability, and inter- and intra-assessor variability in pathological interpretations, liver biopsy is not always well-received by patients and clinicians. Consequently, non-invasive methods for hepatic fibrosis assessment have been developed as alternatives, offering convenient operation and acceptable diagnostic accuracy through medical imaging. Early detection of hepatic fibrosis via clinical imaging can reduce liver failure and prevent disease progression. Traditionally, the interpretation and analysis of medical imaging information are performed by clinical experts. However, with the rapid development of computer-aided diagnosis, particularly advancements in deep learning algorithms within artificial intelligence, physicians can now extract more accurate assessments for clinical decision-making.

Convolutional Neural Networks (CNNs) have emerged as a preferred method for medical imaging processing in hepatic fibrosis assessment, including tasks such as medical image segmentation, clinical classification, and prediction. CNNs have demonstrated robustness against data heterogeneity and have been effective in assessing different stages of liver fibrosis. They have shown high diagnostic accuracy in image classification and have gained popularity in various medical fields.

The primary objective of this research is to investigate the basic techniques of CNNs and recent CNN algorithms for hepatic fibrosis assessment in medical imaging to improve diagnostic accuracy. This includes exploring imaging feature extraction, convolution operations, normalization, and rectified linear units (ReLU) operations. Additionally, the study reviews current research utilizing CNN algorithms for evaluating hepatic fibrosis stages and discusses potential deep learning algorithms in medical imaging for hepatic fibrosis assessment.

The research involved a comprehensive search of several databases, including MEDLINE, EMBASE, Chinese Biomedical Literature Database, WANFANG, and CNKI, covering publications from January 1, 1966, to January 1, 2020. The search strategy included terms related to liver fibrosis, convolutional neural networks, and diagnostic accuracy. The abstracts and full texts of the identified studies were assessed by the researchers, with a third investigator resolving any conflicts. Inclusion criteria specified that patients should be aged between 18 and 65 years, that CNN algorithms were used for liver fibrosis stage assessment, and that the diagnostic model’s accuracy was assessed using indicators such as the area under the receiver operating characteristic curve (AUROC), sensitivity (SEN), specificity (SPE), false-positive rate (FPR), or false-negative rate (FNR). Exclusion criteria included patients with severe cardiovascular and cerebrovascular diseases, psychological disorders, or malignant digestive tumors.

Eight studies utilizing CNN-based algorithms for assessing various liver fibrosis stages were identified. One study developed an automated CNN classifier framework for assessing liver fibrosis induced by non-alcoholic fatty liver disease (NAFLD), while the other seven focused on liver fibrosis induced by hepatitis B virus (HBV).

Liu et al. proposed a computer-aided cirrhosis diagnosis system based on ultrasound images, using a deep convolutional neural network (DCNN) model to extract image features. Their CNN model outperformed other methods, achieving the highest accuracy of 0.968. Brattain et al. developed an automated framework to check shear wave elastography (SWE) image quality and select a region of interest (ROI). They constructed several classifiers, including random forests (RF), support vector machine (SVM), and CNN, with the CNN approach achieving the best AUROC of 0.890. Byra et al. constructed an inception-ResNet-v2 DCNN pre-trained on the ImageNet dataset for evaluating liver steatosis in ultrasound images, achieving an AUROC of 0.977. Wang et al. evaluated the performance of deep learning radiomics of elastography (DLRE) for assessing liver fibrosis stages, with AUROCs of 0.970 for cirrhosis (F4), 0.980 for advanced fibrosis (≥F3), and 0.850 for significant fibrosis (≥F2). Yu et al. validated a deep learning-based algorithm using pre-trained AlexNet-CNN, which automatically calculated liver fibrosis stages with superior AUROC compared to conventional CNN, non-liver multinomial logistic regression (MLR), SVM, and RF. Yasaka et al. investigated DCNN based on dynamic contrast-enhanced computed tomography (CT) images for evaluating hepatic fibrosis stages, showing significant correlation with fibrosis stages. Treacher et al. designed a randomized search of 100 CNN architectures for parameter optimization, demonstrating the accuracy of experts classifying high versus low fibrosis. Gatos et al. detected and isolated areas of low and high stiffness temporal stability in SWE images validated by DCNN for different liver fibrosis stage patients, showing improved accuracy in SWE images compared to unmasked ones.

The convolution layer is the foundational component of CNN construction, implementing image feature extraction through a combination of linear and non-linear functions. The convolution layer’s function is to represent features by increasing the semantic level of features with the depth of layers. In the convolution layer, a kernel consisting of a small array of pixel numbers is applied across the input, which can be learned with backpropagation algorithms. Multiple image maps are connected with a plurality of neurons in local regions, calculated from upper-layer image features through the convolution kernel and weight matrix. Each local weight matrix is activated to a non-linear function to solve the output value and propagate to subsequent convolutional layers.

In a fully connected layer, the output feature information of the final convolution layer is transformed into a flattened one-dimensional array of output numbers calculated through previous convolution and pooling operations, then connected to one or more fully connected layers. Once the imaging information abstracted by the kernel and convolution operation is created, it can be mapped into the subset of a fully connected layer toward the final output, indicating the probable classification for each input imaging feature.

The activation function utilized in the last fully connected layer directly bridges the clinical output, significantly influencing the diagnostic accuracy for liver fibrosis assessment. Depending on the final classification of fibrosis, an appropriate activation function is applied to normalize the target fibrosis probabilities ranging from 0 to 1. The ReLU function ensures the output value is more than zero even when the input value is less than zero. In the sigmoid function, input variables are fitted by linear regression before the activation function is applied, transforming the variables into an “s” curve with values between 0 and 1. The hyperbolic tangent (tanh) function ranges from -1 to 1.

CNNs have demonstrated outstanding performance in computer-aided diagnoses of liver fibrosis patients, automatically extracting imaging features and calculating the weights between each neuron through their contribution, making the network construction more cost-effective and less tedious. One of the potential advantages of the CNN model is its capability of extracting image features by the convolutional process and reducing data dimension by the pooling process, especially for deciphering features and pattern recognition based on its architecture structure inspired by the visual cortex. One of the challenges for clinicians is the interpretation of vast neurons and non-linear activation functions in the CNN algorithm. To address this, researchers have proposed several techniques to provide insight into image features, helping clinicians comprehend the CNN model.

doi.org/10.1097/CM9.0000000000001536

Was this helpful?

YesNo

0 / 0