Artificial Intelligence in MRI-Based Liver Fibrosis Diagnosis
In the rapidly evolving landscape of medical imaging, a groundbreaking integration of artificial intelligence (AI) and magnetic resonance imaging (MRI) is reshaping the diagnostic paradigm for liver fibrosis. This transformative approach, detailed in a recent comprehensive review published in the Chinese Journal of Magnetic Resonance Imaging, highlights how advanced computational techniques are overcoming the limitations of traditional diagnostic methods, offering a new era of non-invasive, precise, and personalized medicine. The study, led by Fengxian Fan, Yanli Jiang, Jun Wang, Wenjing Huang, Pengfei Zhang, and Jing Zhang from the Department of Magnetic Resonance at Lanzhou University Second Hospital and the Second Clinical Medicine College of Lanzhou University, provides a critical analysis of the current state and future potential of AI-driven MRI in the diagnosis and staging of liver fibrosis, a condition that affects millions worldwide and is a precursor to severe liver diseases such as cirrhosis and hepatocellular carcinoma.
Liver fibrosis, characterized by the excessive accumulation of extracellular matrix proteins and collagen, is a progressive condition that can lead to liver cirrhosis, liver failure, and liver cancer if left untreated. Early diagnosis and intervention are crucial for improving patient outcomes, as some stages of liver fibrosis are reversible with appropriate treatment. However, the current gold standard for diagnosing and staging liver fibrosis, liver biopsy, is an invasive procedure that carries risks such as bleeding and infection. Moreover, it is subject to sampling errors and inter-observer variability, making it less suitable for longitudinal monitoring of disease progression and treatment response. These limitations have driven the search for non-invasive, accurate, and reliable methods to diagnose and stage liver fibrosis.
The review by Fan et al. focuses on the application of AI, particularly radiomics and machine learning (ML), in conjunction with MRI to address these challenges. Radiomics involves the high-throughput extraction of quantitative features from medical images, transforming them into mineable data that can be used to support clinical decision-making. Machine learning, a subset of AI, enables the development of predictive models that can learn from large datasets and make accurate predictions without being explicitly programmed. Together, these technologies have the potential to unlock hidden information within MRI images that is not discernible to the human eye, thereby enhancing the diagnostic accuracy and reliability of liver fibrosis assessment.
One of the key areas of focus in the review is the use of radiomics in liver fibrosis staging. Radiomics can extract a wide range of features from MRI images, including morphological, first-order, second-order, and higher-order statistical features. Morphological features describe the size, volume, and shape of lesions, while first-order features, such as mean, median, maximum, minimum, entropy, skewness, and kurtosis, provide information about the distribution of pixel intensities within a region of interest (ROI). Second-order features, often referred to as texture features, capture the spatial relationships between pixels and are typically derived using methods such as gray-level co-occurrence matrices (GLCM) and gray-level run-length matrices (GLRLM). Higher-order features, obtained through techniques like fractal analysis, wavelet transforms, and Laplacian transforms, can identify and highlight more subtle details within the image.
The review highlights several studies that have demonstrated the potential of radiomics in liver fibrosis diagnosis. For instance, Yang et al. and Zheng et al. conducted histogram analysis on whole-liver apparent diffusion coefficient (ADC) maps, which are derived from diffusion-weighted imaging (DWI). They found that multiple parameters, such as mean, median, and entropy, were strongly correlated with the stage of liver fibrosis, validating the use of histogram analysis to reflect the heterogeneity of fibrotic tissue. However, the authors note that the low resolution of ADC maps can affect the accuracy of these correlations, suggesting the need for higher-resolution imaging techniques.
To address this limitation, Hu et al. applied histogram analysis to intravoxel incoherent motion (IVIM) imaging, a technique that uses a bi-exponential model to separate the effects of tissue diffusion and microvascular perfusion. By analyzing ADC maps, true diffusion coefficient (D) maps, pseudo-diffusion coefficient (D) maps, and perfusion fraction (F) maps, they found that D maps, in particular, showed excellent diagnostic performance for significant liver fibrosis (≥F2), with area under the curve (AUC) values ranging from 0.859 to 0.943. This indicates that IVIM imaging, combined with radiomics, can provide more detailed and accurate information about the microstructural changes associated with liver fibrosis.
Another promising technique discussed in the review is diffusion kurtosis imaging (DKI), which quantifies the non-Gaussian diffusion of water molecules, providing a more realistic representation of the complex microenvironment within the liver. Sheng et al. performed histogram analysis on DKI-derived corrected diffusion coefficient (D), diffusion kurtosis (K), and ADC maps in a rat model of liver fibrosis. They found that D maps had the strongest correlation with the stage of fibrosis, although the authors caution that the differences between rodent and human liver fibrosis may limit the direct applicability of these findings. Nevertheless, the results suggest that DKI, when combined with radiomics, has the potential to improve the accuracy of liver fibrosis staging.
The review also explores the use of texture analysis in liver fibrosis diagnosis. Texture analysis, a form of radiomics, quantifies the spatial patterns and heterogeneity of image intensities, which can be indicative of the underlying tissue structure. As liver fibrosis progresses, the liver becomes increasingly heterogeneous, with changes in texture, roughness, and contrast. Yu et al. conducted texture analysis on T1 and T2-weighted MRI images acquired at 11.7 T in mice, finding that the interquartile range of the histogram and the variance gradient of the gray-level gradient matrix in T1 images, as well as the kurtosis of the gray-level gradient matrix in T2 images, had excellent classification performance, with AUC values of 0.9 and 0.91, respectively. Barry et al. applied texture analysis to ADC maps in mice, reporting moderate to strong correlations between histogram, GLCM, and GLRLM features and the stage of liver fibrosis. These findings underscore the potential of texture analysis to detect subtle changes in liver tissue that are not visible to the naked eye.
Specialized imaging sequences and contrast agents have also been explored for their utility in liver fibrosis texture analysis. For example, dual-contrast-enhanced MRI, which combines gadolinium-based and superparamagnetic iron oxide (SPIO) contrast agents, can better visualize the grid-like changes associated with fibrosis. Bahl et al. and Yokoo et al. demonstrated that texture analysis of dual-contrast-enhanced MRI images could accurately assess liver fibrosis, with good diagnostic performance even in early stages. Yu et al. further investigated the use of proton density (PD) images for texture analysis, finding that GLCM features such as correlation and contrast were moderately to strongly correlated with the stage of liver fibrosis. Li et al. proposed an enhanced version of Laws texture features, which improved feature extraction and overall quantification, leading to higher model accuracy in liver fibrosis staging.
Despite the promising results, the review identifies several challenges that must be addressed to fully realize the potential of radiomics in liver fibrosis diagnosis. One of the primary challenges is the lack of standardization in imaging and pre-processing parameters, which can significantly affect the reproducibility and comparability of radiomics studies. The choice of ROI, its size, and the method of segmentation can all influence the extracted features and, consequently, the performance of the diagnostic models. To address this issue, the authors recommend the adoption of radiomics quality scores (RQS) to ensure the rigor and transparency of radiomics research.
Another challenge is the high dimensionality of radiomics data, which can lead to overfitting, especially when the number of features exceeds the number of samples. Overfitting occurs when a model performs well on the training data but poorly on new, unseen data, reducing its generalizability. To mitigate this risk, the authors suggest the use of machine learning algorithms that incorporate regularization techniques, such as logistic regression, LASSO regression, elastic net regularization, support vector machines (SVM), decision trees, and random forests. Schawkat et al. demonstrated that a combination of texture analysis, SVM, and principal component analysis (PCA) could achieve similar accuracy to magnetic resonance elastography (MRE) in classifying liver fibrosis, highlighting the potential of these algorithms to enhance model robustness and accuracy.
Chronic liver disease often involves multiple pathological processes, including fibrosis, inflammation, iron deposition, and steatosis. These coexisting conditions can confound the diagnostic accuracy of liver fibrosis models. For example, inflammation can alter the texture features of liver tissue, making it difficult to distinguish between fibrosis and inflammatory changes. Wu et al. found that using linear discriminant analysis (LDA) combined with a 1-nearest neighbor (1-NN) classifier could effectively separate the effects of fibrosis and inflammation on texture features, improving the accuracy of fibrosis staging. Similarly, iron deposition can increase the contrast between different tissue types, while steatosis can reduce it. House et al. showed that incorporating age, liver fat content, and R2 values (a measure of iron overload) into a generalized linear model could significantly improve the diagnostic performance of liver fibrosis models. Cannella et al. found that in patients with non-alcoholic fatty liver disease (NAFLD), histological parameters such as steatosis and inflammation did not significantly correlate with texture parameters, suggesting that MRI-based texture analysis may be robust to these confounding factors.
The review also discusses the application of machine learning in liver fibrosis diagnosis, with a particular focus on deep learning, a subset of ML that uses neural networks with multiple layers to learn complex representations of data. Deep learning has shown remarkable success in image recognition tasks, and its application to medical imaging is rapidly growing. Convolutional neural networks (CNNs), a popular deep learning algorithm, have been used to analyze Gd-EOB-DTPA-enhanced hepatobiliary phase MRI images for liver fibrosis staging. Yasaka et al. trained a CNN model on 634 liver fibrosis patients and found that the machine learning-derived fibrosis score was significantly correlated with the stage of fibrosis, with AUC values of 0.84, 0.84, and 0.85 for F4, F3, and F2 stages, respectively. However, the authors note that CNNs have limited interpretability, meaning that it is difficult to understand how the model makes its predictions, and they often require large, labeled datasets for training.
In contrast, other machine learning approaches, such as statistical shape modeling, offer better interpretability. Soufi et al. used partial least squares regression (PLSR) to build a statistical shape model of the liver from 51 fibrosis patients’ MRI images. This model was able to detect subtle changes in liver morphology beyond conventional variations, such as enlargement of the caudate lobe and posterior segment of the right lobe, and shrinkage of the anterior segment of the right lobe. The model achieved an AUC of 0.90 in distinguishing between F0-1 and F2-4 stages, demonstrating the feasibility of using liver shape changes for fibrosis staging. The authors argue that this approach provides a more transparent and interpretable alternative to black-box models like CNNs.
The review concludes by emphasizing the need for further research to standardize methodologies, increase sample sizes, and validate models in independent cohorts. The authors also highlight the importance of sharing code and data to facilitate reproducibility and collaboration. They believe that with continued advancements in AI and MRI technology, the integration of radiomics and machine learning will play a pivotal role in the development of personalized and precision medicine, ultimately improving patient outcomes in the management of liver fibrosis.
Fengxian Fan, Yanli Jiang, Jun Wang, Wenjing Huang, Pengfei Zhang, Jing Zhang. Department of Magnetic Resonance, Lanzhou University Second Hospital; The Second Clinical Medicine College of Lanzhou University. Chinese Journal of Magnetic Resonance Imaging, 2021, 12(3): 105-108. DOI: 10.12015/issn.1674-8034.2021.03.026