2.1 Nonneoplastic diffuse liver diseases

Hepatic steatosis is caused by excessive accumulation of triglycerides in liver cells in people with, for example, NAFLD and ALD. Obesity is associated with hepatic steatosis, but the deleterious health implications of obesity extend beyond hepatic steatosis to encompass more severe conditions, namely nonalcoholic steatohepatitis (NASH), NASH-related cirrhosis, and HCC [8]. Therefore, early diagnosis and quantification of hepatic steatosis are of great importance.

NAFLD is characterized by fatty accumulation in the liver accompanied by underlying metabolic disorders, also known as metabolic-associated fatty liver disease [9]. People with advanced stages of NAFLD, such as NASH or advanced fibrosis, and people who have both NAFLD and type 2 diabetes, are at higher risk of developing cardiovascular disease. NAFLD is also an independent risk factor for heart conditions like myocardial infarction, coronary heart disease, and atrial fibrillation [10-12].

Hepatitis is a common type of chronic liver disease. Its causes include NAFLD, ALD, virus infection, chemotherapy, and metabolic or toxic impairment, all of which lead to liver cell damage and liver function impairment. Viral hepatitis, caused by hepatitis A, B, C, D, and E viruses, is a major global health concern that affects millions of individuals worldwide. The disease can present with clinical manifestations that range from asymptomatic to acute with rapid onset or to chronic disease. Unfortunately, a significant proportion of infected individuals remain asymptomatic and may develop liver cirrhosis or liver cancer, which can be fatal. Although all types of hepatitis are associated with morbidity, hepatitis B and C cause most of the deaths associated with this disease. The development of novel diagnostic tools and highly effective direct-acting antivirals has opened up new opportunities for global eradication of chronic hepatitis C. Recognizing the epidemiological and clinical characteristics of hepatitis A, B, C, D, and E is essential to facilitate the accurate diagnosis of these diseases [13, 14]. Therefore, it is crucial to explore how to use imaging and bioinformatic technologies to study these diseases. Further research and investigation are required in this field, and radiologists will be required to play an increasingly critical role in these studies.

Liver fibrosis is a dynamic process in which various pathogenic factors of chronic liver disease activate hepatic stellate cells and turn them into myofibroblasts, yielding large amounts of extracellular matrix proteins and excessive sedimentation. Common causes of liver fibrosis include NAFLD, ALD, chronic viral hepatitis, drug-induced liver injury, autoimmune diseases, and metabolic disorders. If the primary cause is not treated promptly, liver fibrosis may progress to cirrhosis, which can lead to portal hypertension and related complications, such as varicose bleeding, hepatic encephalopathy, hydrops, hepatorenal syndrome, and even hepatocellular carcinoma [15-18]. Early treatment of the underlying factors of liver fibrosis can stabilize or even reverse the disease [19]. Accurate diagnosis and staging of hepatic fibrosis is the key to improving prognosis, determining a treatment plan, and monitoring the progress of the disease. Radiological features and bioinformatics could have a potential role in assessing the severity of fibrosis. IDEAL-IQ magnetic resonance imaging (MRI) and computed tomography (CT) images of liver fibrosis are shown in Figure 1.

Budd–Chiari syndrome (BCS) is caused by thrombosis of the hepatic veins, inferior vena cava, or both, resulting in impaired hepatic venous outflow [20]. One or more prothrombotic conditions are often found, for instance, myeloproliferative disorders. The main complications of BCS are portal hypertension and the development of HCC, but the latter is less common [20]. The underlying disease may develop negatively in about 10% of cases, as is the case for prothrombotic factors, thus affecting the prognosis of BCS. Additional concerns are bleeding and thrombosis of other organs [21]. Although commendable efforts have been made to establish the background of diagnostic methods for BCS, there are currently no universal diagnostic criteria. Clearly, further effort is needed to find satisfactory guidelines for diagnosing BCS.

Hepatic iron deposition is another common liver pathology for which the etiology includes primary and secondary hemochromatosis. Hereditary hemochromatosis is the most common primary hemochromatosis where changes in iron absorption lead to iron overload [22]. Secondary causes of hemochromatosis include cirrhosis, thalassemia, myelodysplastic syndrome, and multiple blood transfusions [23, 24]. Advanced hemochromatosis may progress to arthritis, involving multiple organs such as the liver, heart, and endocrine systems. Iron accumulation can lead to tissue damage and, if left untreated, can cause cirrhosis, HCC, and cardiovascular and endocrine disorders [25]. Early detection and treatment of hemochromatosis were shown to be crucial to completely prevent excess mortality caused by iron overload and also to largely prevent adverse consequences such as liver cancer, depending on the amount and duration of iron excess [26, 27]. Thus, providing a new basis for clinical practice to further investigate the mechanism of iron deposition and to follow up monitoring of the disease course will be hugely beneficial.

Hepatic amyloidosis is characterized by the deposition of amyloid fibers in the Disse space, usually starting in the area around the portal vein, but occasionally starting in the center of lobules or being deposited in the hepatic vascularization system [28]. In severe cases, amyloid deposition causes the pressure on liver cells to atrophy, which interferes with bile passage leading to cholestasis or blocks the sinuses leading to portal hypertension [29]. Hepatic amyloidosis is usually clinically silent, with hepatomegaly and mildly elevated serum alkaline phosphatase being the most common clinical manifestations. Severe cases may be associated with jaundice, portal hypertension, liver failure, and in rare cases, spontaneous rupture of the liver [30]. Although hepatic amyloidosis rarely presents clinically, it is associated with poor prognosis and reduced survival in untreated patients. Therefore, early diagnosis is the key to effective clinical management.

2.2 Neoplastic diffuse liver diseases

HCC is the fourth leading cause of cancer-related mortality globally. Chronic hepatitis B and C, alcohol addiction, metabolic liver diseases (notably nonalcoholic fatty liver disease), as well as exposure to dietary toxins such as aflatoxins and aristolochic acid, are established risk factors for HCC [31, 32]. Some liver diseases can imitate the invasive presentation of HCC, such as focal confluent fibrosis, hepatic steatosis, hepatic microabscesses, intrahepatic cholangiocarcinoma, and widespread metastatic disease (pseudocirrhosis) [33]. Thus, the diagnosis of invasive HCC is greatly important for treating patients with this disease.

Hepatic metastasis is a frequent manifestation in the metastatic cascade of primary neoplasms, connoting an advanced stage of disease progression [34]. The liver has a dual blood supply from the portal vein and the hepatic artery and is prone to metastasizing extrahepatic malignancies. The highest probability of metastasis is in tumors of the digestive tract. Colorectal cancer is one of the most common malignant tumors of the alimentary canal system worldwide, and approximately half of the patients either present with metastasis at the time of initial diagnosis or subsequently develop metastasis during the course of their illness [35]. Early detection and accurate localization of liver metastasis are greatly valuable for selecting appropriate treatments and improving patient survival.

Hepatic lymphomas (HLs), such as non-Hodgkin's lymphoma, affect the liver in the advanced stages of systemic disease. Clinical presentation varies widely from asymptomatic to fulminant liver failure with rapid progression to coma and death. A frequently encountered manifestation is abdominal discomfort or pain observed in 39%–70% of patients; other symptoms include weight loss, night sweats, fever, fatigue, and anorexia [36]. The significance of texture analysis was corroborated based on positron emission tomography/computed tomography (PET/CT) images for the detection of distinct pathological forms of cancer. As a consequence, PET/CT has been posited as a novel approach to discriminate between HCC and HL. Although both image-based and textural parameters have the capacity to distinguish HCC from HL, textural parameters are more efficacious. Furthermore, synergistic integration of the two parameters enhanced the diagnostic accuracy of HCC and HL [37]. Therefore, radiomics has potential value to distinguish these two diseases.

Liver biopsy is the current “gold standard” for diagnosis of diffuse liver diseases, but its use is limited owing to drawbacks such as its invasive nature, high cost, and other complications [38, 39]. Serum markers are currently not specific enough for diagnosing diffuse liver diseases. Traditional imaging modalities such as ultrasound, CT, MRI, and PET/CT provide noninvasive options to evaluate liver disease. Ultrasound is considered a first-line screening tool for identifying liver disease, but it relies heavily on the subjective judgment of the examiner [40]. CT is frequently used for the diagnosis of liver disease; however, the diagnostic efficacy of CT for quantitatively assessing macrovesicular steatosis has been deemed clinically inadequate. Nevertheless, unenhanced CT has shown substantial diagnostic accuracy in qualitatively identifying macrovesicular steatosis that exceeds the threshold of 30% [41]. MRI is not widely used because of its high price and contraindications for patients with metal implants and devices in their bodies [42]. Artificial intelligence (AI) radiomics and deep learning methods are increasingly being widely applied in medical studies and have shown excellent performance in diffuse liver disease recognition, classification, and segmentation [43-45]. Nonetheless, both radiomics models and deep learning algorithms are subject to overfitting because they are based on large numbers of image-derived parameters, and therefore, all radiomics and deep learning results require rigorous clinical validation.

4.1 Liver steatosis

4.2 Liver fibrosis

4.3 Hepatitis

Imaging manifestations of hepatitis are not specific, and therefore it is impossible for the radiologist to identify hepatitis with the naked eye. AI has been applied to identify hepatitis. Tana et al. [45] used texture analysis to build measurable CT-based imaging features to predict clinical severity in hepatitis. Naganawa et al. [95] were the first to report that texture analysis of non-contrast-enhanced CT images effectively predicted NASH. Subsequent efforts to apply machine learning algorithms to NASH histology have demonstrated the feasibility of machine learning approaches in small cohorts [96-98]. Taylor-Weiner et al. [99] described a machine learning approach to assess liver histology that accurately characterized disease severity as well as heterogeneity and sensitively quantified the treatment response in NASH. These findings demonstrated the power of machine learning in advancing the understanding of NASH disease heterogeneity in affected patients by risk stratification and in facilitating treatment development.

In general, AI techniques have shown promising results in diffuse liver diseases.

5.1 Genomics

After the publication of the human genome sequence in 2003, genomics has become one of the most developed omics fields. Growing numbers of genome-wide association studies (GWAS) have identified specific genes involved in the development of many complex diseases and millions of genetic variations have been discovered [102]. The current National Human Genome Research Institute (NHGRI)-GWAS catalog includes 24 GWAS studies that have reported >100 gene variants relevant to NAFLD [103].

5.2 Metagenomics

Metagenomics focuses on the genome sequences of organisms that live in a common environment, whereas genomics focuses on the entire genetic makeup of an organism [104]. DNA from microbial communities has been analyzed by metagenomic methods, which until now have been used mainly to study microbial communities [104]. The human microbiome and its metabolites have been suggested to play a role in NAFLD pathogenesis [105]. A large metabolomic and metagenomic study showed that 6 out of 56 altered metabolites were gut derived and demonstrated sharing of gene effects with liver steatosis and fibrosis in NAFLD [106]. Recent metagenomic studies suggested that some microbiota members may serve as biomarkers for NAFLD diagnosis and prognosis [107-109].

5.3 Pharmacogenomics

Pharmacogenomics uses genetic variants to predict treatment responses, which can lead to the development of individualized treatments based on genetic characteristics. Besides extrinsic causes such as environmental chemicals, diet, drug–drug interactions, and alcohol, genetic factors also play a significant role in NAFLD [110]. Indeed, pharmacogenomic studies are underway to identify potential drugs for treating NAFLD and metabolic syndrome on a more individual level.

5.4 Transcriptomics

Besides environmental factors, other factors can alter gene expression, which can affect an individual's phenotype and their risk of developing a metabolic disease. Most transcriptome studies of metabolic diseases have focused on islets and peripheral tissues such as the liver, skeletal muscle, and adipose tissue. Baselli et al.[111] identified 320 genes that were differentially expressed in severe NAFLD compared with their expression in nonsevere NAFLD and the normal liver, and 16 of these genes were deregulated in carriers of the PNPLA3 rs738409 variant. Liver transcriptome profiles were found to be similar in healthy individuals but were clearly different from the profiles of NAFLD or NASH patients who were obese [112]. Despite the observational nature of these transcriptomic studies of NAFLD, the findings have provided insight into the treatment of this disease with many benefits for precision medicine.

5.5 Epigenomics

Epigenetics is the study of changes in gene expression and phenotypic variations that are not the result of changes in DNA sequences [113]. Methylation of DNA and modification of histones are among the most studied epigenetic modifications [114]. DNA methylation is a significant epigenetic modification of cytosine bases in cytosine-phospho-guanine (CpG) dinucleotides, called CpG islands [115]. There is a link between the methylation status of CpG islands and gene suppression, especially when methylation is in the regulatory and promoter regions of genes [115]. Some genes of patients with NAFLD were found to have different methylation levels than the same genes of healthy individuals [114], and a next-generation sequencing analysis of liver samples showed that patients with NAFLD had significantly lower levels of global DNA methylation than those with healthy controls [115]. In addition, changes in DNA methylation levels were found to be associated with NAFLD progression from mild to advanced and may be highly predictive of fibrosis [116]. Histone modifications are crucial posttranslational epigenetic changes that include acetylation, methylation, phosphorylation, and ubiquitylation. Chromatin structure can be altered by these modifications to regulate transcriptional activity [117]. The altered acetylation profile of histones that has been observed in patients with NAFLD has received more attention in recent years. For example, acetylation of the carbohydrate-responsive element-binding protein (ChREBP) by a histone was found to be involved in NAFLD modulation [118].

5.6 Proteomics

Proteomic analysis includes the identification and quantification of all the proteins present in a cell, tissue, or organism [119]. Proteomics has made remarkable progress owing to technological advances and improved tools for bioinformatics analysis. Protein signatures related to metabolic-associated fatty liver disease have been identified in several studies, which could aid in the development of noninvasive biomarkers [120-122].

5.7 Metabolomics

Metabolomics is the study of the metabolome in cells, tissues, or fluids of an organism [123]. Metabolomics provides an integrated view of pathophysiology and complements other omics studies. High-throughput analytical technologies have led to a rise in metabolomics studies to identify biomarkers for NAFLD and other liver diseases [123]. Metabolomics includes lipidomics and glycomics as major subfields. Lipidomics has made substantial progress in understanding metabolic pathways involved in NAFLD and NASH [124]. Glycomics involves the analysis of glycan structures both on their own and when they are bound to proteins [118]. Alterations in protein glycosylation were found to be associated with NAFLD [125]. In addition, fucosylation of proteins, such as haptoglobin, transferrin, alpha-1 antitrypsin, and ceruloplasmin, was found to contribute to the progression of NAFLD to NASH, and during the progression of NAFLD to liver fibrosis, high concentrations of fucosylated proteins have been observed [126, 127].

5.8 Integrating multiomics

Combining singleomics data can increase their power and provide a more comprehensive and thorough picture of individuals. The rapid evolution of sequencing technologies has led to an increase in the role of multiomics data in biology. Multiomics approaches have been used to comprehensively explore the biological mechanisms behind complex traits at the levels of DNA, RNA, proteins, and metabolites. Indeed, medicine and biology have been transformed by the availability of multiomics data, which has led to the development of system level. Advanced technologies, such as mass spectrometry and high-throughput sequencing, have dramatically reduced the cost of generating biological data. Consequently, multiomics profiling is increasingly being used in research related to health and disease [128-130]. Interpretation of multigraph data is crucial for the prediction, diagnosis, and treatment of NAFLD. Random forest models were built to predict NAFLD progression and early diagnosis based on multiomics data [131, 132]. Furthermore, Perakakis et al. [133] developed a noninvasive supervised learning model consisting of lipids, glycans, and hormones that diagnosed NAFLD with >90% accuracy. Pirola et al. [134] selected a list of genes linked to NAFLD and metabolites that were altered in NAFLD and NASH as part of an integrative research method.

Despite these advances, the use of multiomics in diffuse liver diseases is still in its infancy, and further work is needed before multiomics can be applied to precision medicine. Establishing a solid evidence base is essential for multiomics integration and precision medicine in diffuse liver diseases. To achieve this, research designs should include integrated rigorous high-dimensional data from multiple sources and computational approaches that have been developed for large datasets. Reduction in the cost of meta-analysis is also needed. The workflow of multiomics is shown in Figure 4.

6.1 Logistic regression

Many medical research tasks can be framed as dichotomous outcomes, such as whether hepatic steatosis occurs, whether hepatitis is present, or whether hepatic fibrosis develops. Linear regression models describe continuous outcomes, whereas logistic regression models convert binary outcomes into continuous outcomes that are log odds or logits of events, also known as logit functions [135].

Logistic regression models are versatile and have been applied in different fields of medical research. Such models have often been used to evaluate predictors and to adjust the confounding factors and/or interactions for diffuse liver diseases in clinical practice [136-138]. A confounder is a variable that is correlated with both the outcome and the independent variable of interest. When a confounder is omitted from a model, results will be biased, and therefore it is crucial to consider all confounders. Logistic regression models have also been used to create predictive algorithms that describe the probability of an event using nomograms or online calculators, such as assessment of fibrosis and prediction of survival of patients with HCC [139-141]. Alternatively, logistic regression models have been applied to calculate propensity scores, which can be used to adjust the imbalance of baseline confounders between two cohorts when the goal is to compare their identical outcomes [142, 143]. Furthermore, because the outcome variables are binary, logistic regression can be used for binary classification tasks. Therefore, logistic regression models can evaluate tasks such as whether there is significant hepatic inflammation [144], whether hepatic steatosis is present [136], or whether cirrhosis is absent [137].

However, nonlinear problems cannot be evaluated by logistic regression because of its linear decision surface and because the independent variables need to be averaged among themselves or without multicollinearity. More importantly, logistic regression cannot tackle multiple classification tasks. Therefore, it is difficult to obtain complex relationships using logistic regression, and more powerful and compact algorithms, such as deep learning, are needed [145]. Logistic regression is the parametric statistic for the predictive model of classification, whereas the semiparametric statistic of Cox regression can be considered for survival analysis.

6.2 Nonparametric statistics

SVM, random forest, and decision tree are supervised machine learning models that can be used for classification and regression. A schematic diagram of the three machine learning algorithms is given in Figure 5.

SVM is a supervised machine learning model that can be used for classification and regression. The classification principle of SVMs is as follows. If a dataset contains M objects and N variables, the objects can be divided into two categories. If an object is represented as points in a variable space, an n-dimensional space is obtained. SVM aims to find a hyperplane in the multidimensional space, which can divide all objects into two optimal categories with the margin of the closest points in the two categories being as large as possible. These points are called support vectors on the spacing boundary, and they determine the spacing as well as the segmented hyperplane in the middle of the spacing. Therefore, an optimal hyperplane of n−1th dimension can be obtained in an n-dimensional space (corresponding to N variables) [146].

Random forest is a classifier that contains multiple decision trees, and its output category is determined by the mode of individual trees [147]. The classification process of random forest is as follows. Assume N represents the number of training samples and M represents the number of features. Then, the number of input features, m, is used to determine the decision result of a node in the decision tree, where m is much lower than M. A training set (bootstrap sampling) is formed by sampling N times from training samples by put-back sampling. The unsampled samples are used for prediction and error assessment. For each node, m features are selected randomly, and the decision of each node in the decision tree is determined based on the selected features. Then, the optimal splitting mode is calculated using these m features. Each tree grows intact without pruning, although pruning can be applied after building a normal tree classifier.

Decision tree is a commonly used classification method. It has a tree-like structure in which the internal node represents a judgment on an attribute, the branch represents the output of a judgment result, and the leaf node represents a final classification result. Decision tree requires supervised learning, where a group of samples and each sample have a set of attributes as well as classification result [148]. Because the classification result is known, a decision tree can be obtained by learning the results of these samples and applying this knowledge to correctly classify new data.

Machine learning algorithms are powerful tools for processing and modeling large amounts of omics data. Indeed, the application of machine learning to radiomics has produced gratifying results for diffuse liver diseases, such as diagnosis of liver fibrosis staging [149], inflammatory activity grading [150], and differentiation of a healthy liver from hepatic steatosis, cirrhosis, amiodarone deposition, and iron overload [87]. Recent studies have proposed targeted metabolomics with machine learning to identify metabolites that distinguish the progression of NAFLD [151, 152]. Findings of such studies can be used to assess metabolic changes in NASH and will contribute to an understanding of the pathophysiological significance of metabolite profiles in relation to NAFLD progression. Bioinformatics analysis has been increasingly used in studies of NAFLD and has provided useful information for exploring potential candidate biomarkers for the diagnosis, prognosis, and drug targets of NAFLD [5, 6, 153]. However, studies that combine bioinformatics, machine learning, and radiomics data have not yet been reported. We believe that radiobioinformatics will provide new insight into diffuse liver diseases.

6.3 Potential value of CoAtNet in diffuse liver diseases

CoAtNet is a new architecture that combines depthwise convolution and attention models and vertically stacking convolution layers and attention layers, that have high generalization ability, capacity, and efficiency [154]. CoAtNet combines translation invariance, input-adaptive weighting, and a global receptive field based on mobile inverted bottleneck convolution (MBConv) blocks with a relative self-attention mechanism to steer the network to a completely different level of performance. The attention mechanism is a method of resource allocation that was first used in natural language processing [155] and was soon found to be applicable to the computer vision field. Vision Transformer (ViT) [156] was released in 2020 and quickly became the cornerstone for the development of subsequent transformer models in the computer vision field. ViT was found to have poor generalization performance when the dataset is insufficient. In trying to solve this problem, Dai et al. [154] eventually devised CoAtNet. The network illustrating the principle of CoAtNet is displayed in Figure 6. To incorporate the translation invariance of the convolution with the input-adaptive weights and the global receptive field of the self-attention mechanism, CoAtNet combines the global static convolution kernel with an adaptive attention matrix after or before softmax initialization, which exhibited better classification performance on an ImageNet [157] classification task than it did on other architectures (e.g., ResNet [158], ViT, or NFNet [159]).

CoAtNet is one of the state-of-the-art frameworks released in September 2021, and therefore not many applications of CoAtNet have been reported so far. However, because of its superior performance, CoAtNet has been applied in the medical domain. Wang et al. [160] demonstrated that CoAtNet was effective in classifying renal parenchymal tumor subtypes and performed better than ViT. Kvak [161] used CoAtNet to classify malignant and benign melanoma and obtained an accuracy of 0.901, which is better than the accuracies obtained with other state-of-the-art algorithms. Tripathi et al. [162] used CoAtNet for the cytomorphological classification of bone marrow cells and found that the CoAtNet model outperformed the EfficientNetV2 and ResNext50 models. Diffuse liver diseases are difficult to distinguish in the early stages by medical imaging and naked eye classification [163]. Early identification, diagnosis, and treatment of these diseases can significantly improve the prognosis and reduce the risk of progression to liver cirrhosis and liver cancer. Medical image classification is commonly performed using deep learning models such as ResNet and EfficientNet. Although there are few examples of CoAtNet used for diffuse liver disease discrimination, related studies have shown that the CoAtNet model outperforms other state-of-the-art algorithms for both general and medical image classification tasks. We believe that CoAtNet will also perform well for classifying diffuse liver diseases.

6.4 Deep learning for survival analysis (DeepSurv)

Cox proportional hazards models are widely used in survival analysis studies and to investigate the effects of different prognostic factors on tumor death and recurrence [164, 165]. Cox regression assumes that the covariates are linear, and therefore fitting Cox regression is challenging when survival data are nonlinear. Cox linear regression models also require in-depth feature engineering and prior medical knowledge for the deployment of individual treatment plan recommendation systems. Survival models built on deep learning can more effectively fit nonlinear survival data, and survival models based on deep learning and machine learning may be well suited for high-dimensional interaction terms in survival data. DeepSurv is a deep feed-forward neural network that predicts hazard rates by boosting the weights θ and patient covariates.

DeepSurv outperformed more conventional linear prediction models in numerous biomedical studies. DeepSurv has been used to predict the efficacy of chemotherapy, guide individualized treatment, and in clinical decision support systems. She et al. [166] suggested that the DeepSurv algorithm may have potential benefits for prognosis evaluation and treatment recommendation in nonsmall cell lung cancer. Kim et al. [167] showed that DeepSurv outperformed random survival forest and Cox proportional hazards models in survival prediction of patients with oral cancer. In a study of cardiovascular risk prediction using DeepSurv, Hathaway et al. [168] obtained independent accurate predictions of atherosclerotic cardiovascular disease risk and cardiovascular outcomes using clinical characteristics without imaging information and inflammatory factors. These findings demonstrated that DeepSurv had superior capabilities in guiding individualized therapy and disease risk monitoring when compared with conventional linear prediction models. The network illustrating the principle of DeepSurv is displayed in Figure 7. Although DeepSurv has not yet been applied to diffuse liver diseases, it has shown good predictive performance, implying that similar results may be obtained in diffuse liver diseases; however, further studies are required to confirm this.