Harnessing artificial intelligence for the assessment of liver fibrosis and steatosis via multiparametric ultrasound

Abstract

Artificial intelligence (AI) is revolutionizing medical imaging, particularly in chronic liver diseases assessment. AI technologies, including machine learning and deep learning, are increasingly integrated with multiparametric ultrasound (US) techniques to provide more accurate, objective, and non-invasive evaluations of liver fibrosis and steatosis. Analyzing large datasets from US images, AI enhances diagnostic precision, enabling better quantification of liver stiffness and fat content, which are essential for diagnosing and staging liver fibrosis and steatosis. Combining advanced US modalities, such as elastography and doppler imaging with AI, has demonstrated improved sensitivity in identifying different stages of liver disease and distinguishing various degrees of steatotic liver. These advancements also contribute to greater reproducibility and reduced operator dependency, addressing some of the limitations of traditional methods. The clinical implications of AI in liver disease are vast, ranging from early detection to predicting disease progression and evaluating treatment response. Despite these promising developments, challenges such as the need for large-scale datasets, algorithm transparency, and clinical validation remain. The aim of this review is to explore the current applications and future potential of AI in liver fibrosis and steatosis assessment using multiparametric US, highlighting the technological advances and clinical relevance of this emerging field.

Key Words: Artificial intelligence; Multiparametric ultrasound; Liver; Fibrosis; Steatosis; Shear wave elastography; Attenuation imaging; Machine learning; Deep learning

Core Tip: The emergence of artificial intelligence has led to its application across various fields, including hepatology and medical imaging. Its enormous potential has already been recognized and documented in numerous studies. This review explores the current application and future potential of artificial intelligence in ultrasound imaging, emphasizing its role in chronic liver disease early diagnosis and follow-up.

INTRODUCTION

Liver diseases encompass a broad spectrum of conditions marked by hepatocyte damage, inflammatory cells infiltration and hepatic stellate cells activation, leading to impaired liver function and structural disruption resulting from the progressive replacement of hepatic parenchyma with fibrotic tissue[1]. Approximately, one million deaths each year are attributed to chronic liver disease (CLD)[2] and cirrhosis ranks among the leading twenty causes of death worldwide[3]. While viral aetiologies continue to be a significant element, the landscape of liver cirrhosis is shifting with a growing proportion now accounting to metabolic causes[4]. Early detection of liver steatosis and treatment of its reversible causes can prevent progression and improve outcomes[5]. Artificial intelligence (AI), through machine learning (ML) and deep learning (DL) algorithms, is rapidly emerging as a transformative tool to face these challenges[6]. The recent reclassification[7] highlights liver steatosis as a central condition with multiple aetiologies, most commonly alcoholic and metabolic. Both conditions share key mechanisms steatosis, inflammation, fibrosis whose interaction may synergically accelerate disease progression to steatohepatitis and cirrhosis, increasing complication rates and mortality[8]. The significance of liver steatosis is still uncertain; nevertheless, the same risk factors contributing to hepatic fat accumulation, may also promote disease progression[8]. Non-liver-related causes particularly cardiovascular disease and extrahepatic cancers represent the leading causes of death in patients with steatotic liver disease (SLD), especially in metabolic dysfunction-associated steatotic liver disease (MASLD) ones[9]. Shi et al[10] demonstrated that the severity of liver steatosis is an independent risk factor for the development of major adverse cardiovascular events, with patients showing moderate to severe steatosis experiencing worse outcomes compared to mild ones. However, liver fibrosis severity is more directly associated with cardiovascular disease risk and remains the strongest predictor of liver-related outcomes across all disease stages[11,12]. Given this close relationship between hepatic injury and systemic complications, accurate assessment of liver disease severity is essential for both prognostic and therapeutic purposes. Hence, liver histology remains the gold standard for staging fibrosis (F0-F4) and grading steatosis (S0-S3)[7,13,14]. Nevertheless, due to its invasiveness, cost, risk of potentially life-threatening complications, sampling bias, variability in histological interpretation and limited role in longitudinal follow-up, the development of non-invasive alternatives for liver assessment has become essential[15].

Non-invasive serum tests, such as the fibrosis-4 (FIB-4) index, the nonalcoholic fatty liver disease (NAFLD) fibrosis score or enhanced liver fibrosis test, are increasingly used in clinical practice despite some well-known limitations[13,16]. Elastography techniques are already reliable tools for non-invasive liver fibrosis staging, supported by Baveno VI[17]/VII[18] and World Federation for Ultrasound in Medicine and Biology guidelines[19]. Magnetic resonance elastography (MRE) and magnetic resonance imaging proton density fat fraction (MRI-PDFF) offer excellent accuracy for fibrosis and steatosis assessment; nevertheless, high costs and limited availability restrict their application to specialized centers, mostly for research purposes[20,21]. B-mode ultrasound (US) can provide an initial assessment of hepatic steatosis, but sensitivity is limited when fat infiltration is below 20%-30%[14]. New techniques like attenuation imaging (ATI) improve non-invasive quantification and correlate well with MRI-PDFF. ATI measures US energy loss, which increases with intracellular triglyceride accumulation, providing a quantitative estimate of liver fat[22]. Two main methods exist: Two-dimensional ATI, integrated into US scanners, and controlled attenuation parameter (CAP), an add-on to transient elastography (TE). Although based on the same principle, differences in algorithms and protocols highlight the need for standardized acquisition[21]. Therefore, integrating AI into clinical hepatology is not only an innovative advancement but also a necessary evolution to meet increasing diagnostic demands and to deal with the inherent complexity of liver disease pathophysiology. In this review, we explore the current applications and future potential of AI in liver fibrosis and steatosis assessment using multiparametric US, highlighting the technological advances and clinical relevance of this emerging field.

This topic is of particular clinical relevance to hepatology, as the integration of AI-assisted multiparametric US directly supports the non-invasive evaluation and management of CLDs. By bridging advanced imaging technologies with routine hepatology practice, these approaches have the potential to enhance diagnostic accuracy, standardize longitudinal monitoring, and ultimately improve patient outcomes.

ROLE OF AI IN LIVER DISEASE EVALUATION

AI is defined as the set of algorithms able to process and interpret large volumes of data, enabling machines to learn, recognize patterns and make decisions in a human-like manner[23,24]. By swiftly analyzing complex datasets, AI can uncover hidden relationships and support both individual diagnoses and large-scale predictive analyses[24]. As a result, it is becoming increasingly vital in numerous fields, particularly in healthcare, where it has been employed especially in medical imaging analysis and diagnostics, but also in disease outbreak forecasting, patient monitoring, clinical decision-making, drug development, robotic surgery and beyond[25].

AI application in medical imaging analysis and diagnostic is based mainly on ML and DL[23]. ML is an AI branch allowing computers to learn from data without a specific programmed planning. It consists in complex algorithms, raised on trained specific datasets with predefined features[26]. Algorithms are translated in code, which provides instructions for quickly analyzing the input to produce the output, in other words the results[25,26]. ML may be categorized in three kinds of learning: Supervised learning, unsupervised learning and reinforced learning[25]. In supervised learning, the model is exercised on datasets with already known results, aimed to acquire correct output in response to certain input[25,27]. In unsupervised learning instead, the system analyses the data to uncover hidden structures or patterns by itself without any predefined categories[25,27]. Finally, the reinforcement learning is a different approach in which the machine learns to make decisions by interacting with its environment through a system of rewards and penalties to achieve a specific aim[25]. In a different manner DL is a subfield of ML, which leverages deep neural networks (DNN) to automatically learn hierarchical feature representations from data at multiple levels of abstraction. Consequently, it requires a great amount of data to train models efficaciously[25]. Artificial neural networks (ANNs) and convolutional neural networks (CNNs) belong to DL. ANN is a system composed of interconnected simulated neurons, similarly to biological ones, whose capabilities and advantages stem from distinctive features such as nonlinear behavior, adaptability and parallel processing that allows input elaboration and output producing. ANN is composed of three neuron layers: An input layer, some hidden layers and an output layer. Although more layers can be used to grow network capacity, too many layers can prejudicial its functioning, so three layers are generally sufficient[28]. CNNs are a category of ANN and are widely regarded as the most suitable DL architecture for analyzing medical images because they include deep architectures composed by specialized layers. A typical CNN is structured with an input layer, several hidden layers including convolutional layers, pooling layers, fully connected layers and normalization layers and an output layer. Convolutional and pooling layers play a key role in extracting and consolidating relevant features from the input data, which are then passed on to the output layer[26].

The hepatology scientific community has quickly recognized AI potentialities to enhance liver disease diagnosis, staging and prognosis, even extending to transplantation and post-transplant care[29-32]. AI has been applied for the diagnosis of both focal and diffuse liver diseases not only in the field of imaging diagnostics including US, computed tomography and MRI but also for the histological examinations[22,33,34]. However, the implementation of AI in US imaging could be crucial for achieving an early, accurate, and non-invasive diagnosis, potentially avoiding the need for liver biopsy.

Although a broad spectrum of AI applications in liver US has been reported, few investigations have undertaken a systematic comparison of the various algorithmic approaches. A clearer understanding of the relative strengths and weaknesses of ML and DL methods across different clinical settings is essential to inform their optimal selection and application in practice.

Comparative analyses indicate that CNN-based architectures generally achieve superior performance to traditional ML algorithms in image-based feature extraction and fibrosis staging, due to their capacity to autonomously learn multi-level hierarchical representations from US data. In contrast, support vector machines (SVM) and random forest models tend to perform more robustly in studies with limited sample sizes or when combining imaging-derived parameters with clinical and biochemical data. Generative adversarial networks (GAN), on the other hand, provide unique advantages for data augmentation and synthetic image generation, thereby enhancing model generalizability and image quality. The complementary nature of these approaches highlights the potential of hybrid and multimodal frameworks that integrate their respective strengths to improve diagnostic accuracy and reproducibility in liver disease assessment. Figure 1 summarize the potential application of AI in liver disease evaluation.

Figure 1 Artificial intelligence-powered multiparametric ultrasound for the assessment of liver fibrosis and steatosis. Schematic representation of the integration of different imaging and diagnostic modalities, including conventional ultrasound, color doppler, shear wave elastography, attenuation imaging, magnetic resonance, and liver biopsy as input sources for artificial intelligence analysis. Machine learning and deep learning algorithms process multiparametric data to improve the assessment of liver fibrosis and steatosis, leading to enhanced prognosis stratification, clinical monitoring, precision medicine, and risk stratification. DL: Deep learning; ML: Machine learning.

AI US TECHNIQUES FOR LIVER FIBROSIS EVALUATION

Liver fibrosis is the main prognostic risk factor in CLD, so having early and accurate assessment is essential[13,35]. However, a biopsy cannot be performed on all patients. US and elastography are often the first diagnostic approach for these patients. Numerous studies have highlighted AI potentialities applied to these techniques for staging liver fibrosis[23], as reported in Table 1. B-mode and doppler US were among the earliest imaging techniques to incorporate the first types of AI. As early as 1999, Badawi et al[36] used one of the first “algorithms” to differentiate among cirrhotic, fatty and normal liver using parameters such as mean grey level and attenuation, achieving good sensitivity and specificity.

Table 1 Studies that evaluated the artificial intelligence ability to stage liver fibrosis[40,42-57,110,111].

Ref.	Type of study	Population	Number of patients	AI technique employed	Main results
Ruan et al[43]	Retrospective, multi-center	HBV	508	MSTNet DL	High accuracy in detecting both moderate (≥ F2) and advanced (F4) liver fibrosis, outperforming conventional clinical tools (APRI, FIB-4 and Forns) and human sonographers
Song et al[42]	Retrospective, single-center	HBV	93	ANNs DL	Excellent predictive capability to stage liver fibrosis and superior to serum fibrosis tests
Zhang et al[44]	Retrospective, multi-center	HBV	1500	CNNs DL	High-frequency images outperformed low-frequency ones across all trained CNNs models, as well as FIB-4, APRI and SWE in staging liver fibrosis
Duan et al[45]	Retrospective, two-center	CLD	434	GAN model DL	Good performances in staging liver fibrosis. Good predictive accuracy in identifying liver cirrhosis
Miura et al[55]	Retrospective, single-center	CLD	517	CNNs DL	Higher diagnostic accuracy than human scoring for detecting significant fibrosis (≥ F2)
Li et al[40]	Prospective, single-center	Chronic HBV infection	144	Adaptive boosting, random forest, SVM ML	ML algorithms improve the accuracy of liver fibrosis assessment. Combining conventional radiomics, ORF and CEMF data with ML algorithms enhances accuracy in detecting significant liver fibrosis
Durot et al[46]	Retrospective	CLD or elevated liver enzymes	204	SVM ML	SVM ML algorithm demonstrated excellent diagnostic accuracy in distinguishing significant liver fibrosis (≥ F2) when applied to both p-SWE and 2D-SWE data from two different systems, compared with MRE
Gatos et al[47]	Prospective, single-center	54 healthy patients, 31 with CLD	85	ML	Good accuracy in distinguishing healthy individuals from patients with CLD
Gatos et al[48]	Retrospective	56 healthy patients, 70 with CLD	126	ML	Good accuracy in distinguishing healthy individuals from patients with CLD combining different cluster features
Destrempes et al[49]	Retrospective, cross-sectional	CLD (HBV, HCV, NAFLD, AIH)	82	ML	Combining QUS and p-SWE in an ML model enhanced accuracy in staging fibrosis, inflammation and steatosis
Wang et al[51]	Prospective, multi-center	HBV	398	Dlre	Dlre outperformed 2D-SWE in detecting cirrhosis and advanced fibrosis. It was more reliable than biomarkers (FIB-4, APRI) to identify all fibrosis stages
Lu et al[52]	Retrospective, multi-center	CLD	807	Dlre2.0	Dlre2.0 achieved a higher AUC than Dlre for significant fibrosis, but without statistical significance
Kagadis et al[50]	Retrospective	88 healthy individuals, 112 with CLD	200	GoogLeNet, AlexNet, VGG16, ResNet50, DenseNet201 DL	All pre-trained DL networks achieved good to excellent performance in staging liver fibrosis, outperforming radiologists. ResNet50 and DenseNet201 showed high accuracy across all fibrosis stages
Xue et al[54]	Retrospective	Local liver lesions treated by partial hepatectomy	466	Inception-V3 network (DL), TL	Gray scale US images and 2D-SWE images analyzed with Inception-V3 (DL) using the TL achieved excellent performance in staging liver fibrosis
Brattain et al[53]	Retrospective	NAFLD	328	Random forest, SVM ML; CNN DL	CNN demonstrated the highest performance in distinguishing liver fibrosis as significant or not
Zhou et al[57]	Retrospective	94 patients with liver fibrosis; 143 patients with liver fibrosis and liver steatosis	237	iANN, DL	Radiomics with iANN-based homodyned-K US imaging outperformed both the standalone iANN method and radiomics on uncompressed US data for liver fibrosis assessment
Park et al[110]	Retrospective, multi-center	Patients underwent to liver biopsy or hepatectomy	933	DL (VGGNet, ResNet, DenseNet, EfficientNet, ViT)	Deep CNNs accurately staged liver fibrosis by METAVIR score from B-mode US images. EfficientNet showed the best performance among models
Lee et al[111]	Retrospective, multi-center	Healthy individuals and patients with CLD	838	DCNN, DL	DCNN accurately assessed METAVIR score from US images and outperformed radiologists in diagnosing cirrhosis in simulated US examination

AI: Artificial intelligence; HBV: Hepatitis B virus; CLD: Chronic liver disease; HCV: Hepatitis C virus; NAFLD: Nonalcoholic fatty liver disease; AIH: Auto-immune hepatitis; MSTNet: Multi-scale texture network; DL: Deep learning; ML: Machine learning; ANN: Artificial neural network; CNN: Convolutional neural network; GAN: Generative adversarial network; SVM: Support vector machines; Dlre: Deep learning-based radiomics of two dimensional shear wave elastography; GoogLeNet: Google inception network; AlexNet: AlexNet convolutional neural network; VGG: Visual geometry group; ResNet50: Residual network 50 layers; DenseNet201: Densely connected convolutional networks 201; TL: Transfer learning; iANN: Improve artificial neural network; ViT: Vision transformer; DCNN: Deep convolutional neural network; APRI: Aspartate aminotransferase to platelet ratio index; SWE: Shear wave elastography; FIB-4: Fibrosis-4; CEMF: Contrast-enhanced micro-flow; ORF: Original radiofrequency; p-SWE: Point shear wave elastography; 2D-SWE: Two dimensional shear wave elastography; MRE: Magnetic resonance elastography; QUS: Quantitative ultrasound; AUC: Area under the curve; UC: Ultrasound.

The onset of ML

The rise of ML led to the integration of different algorithms into US applications, such as SVM and decision trees, both of which are commonly used supervised ML algorithms[37,38]. In 2003, Yeh et al[37] described a SVM model based on B-mode US-image features extracted using gray-level concurrence and non-separable wavelet transforms, to classify liver fibrosis in 20 resected human liver. Histological grading served as the reference standard and the model showed good correlation, although classification accuracy declined as the number of fibrosis classes increased. Recently, a model based on quantitative US (QUS) texture features collected from rats identified 20 parameters able to distinguish early from advanced fibrosis with high accuracy; notably, performance was further enhanced when combined with ML techniques[39].

ML main feature lies in the ability to analyze large amounts of data. Li et al[40] extracted a large quantity of numerical data from US images from hepatitis B patients using conventional radiomics, original radiofrequency and contrast-enhanced micro-flow features. Then they applied ML algorithms to these data, improving the accuracy of liver fibrosis assessment and achieving better performance in diagnosing significant fibrosis by combining multiple data sources.

DL revolution

Following the breakthrough of DL, AI applications in medical imaging have advanced significantly, particularly in tasks such as fibrosis staging and lesion detection. An ANNs optimizing parameters acquisition model was tested employing B-mode and doppler parameters for staging liver fibrosis getting an excellent diagnostic accuracy enrolling 5 parameters: Liver parenchyma, spleen, hepatic artery pulsatile index, portal vein velocity and hepatic vein damping index[41]. Another ANN-based model, trained using four input indices liver parenchyma, liver margin, spleen size and portal vein blood flow identified as relevant in prior statistical analysis, also demonstrated excellent capability to stage liver fibrosis, outperforming serum fibrosis test such as aspartate aminotransferase to platelet ratio index (APRI), FIB-4, gamma-glutamyl transpeptidase to platelet radio index, suggesting how a combination of ANN models serum fibrosis test could further improve non-invasive diagnostic performance[42]. However, DL advantages lie in “automatized” US image analysis to evaluate fibrosis. Multi-scale texture network is a DL approach based on image pyramid patches extracted from US scans with a distribute-gather attention framework organized in several stages, each containing several distribute-gather attention blocks. Its performance in patients with chronic hepatitis B virus (HBV) infection showed high accuracy in detecting both moderate (≥ F2) and advanced (F4) liver fibrosis, outperforming conventional clinical tools, such as APRI, FIB-4 and Forns index, as well as the assessments made by human sonographers[43].

Recently, a large retrospective, multicenter study[44] enrolled 1500 patients who underwent liver biopsy and US examination using both low-frequency convex array probe and high-frequency linear array probe to develop a DL model able to stage liver fibrosis in HBV patients using preexisting CNNs models. High-frequency images outperformed low-frequency images across all trained CNNs models employed, FIB-4, APRI and shear wave elastography (SWE) in staging liver fibrosis. Among all CNNs models employed, InceptionNeXt tiny architecture model exhibited high accuracy both with high-frequency and low-frequency examinations. Good performances in staging liver fibrosis were also achieved using a GAN model a DL approach based on the interaction between generation and discrimination networks while good predictive accuracy in identifying liver cirrhosis was obtained using DL radiomics monogram combining imaging and clinical data[45].

AI for advancing SWE and new fields of application

Elastography’s growing interest in liver evaluation has also supported its AI implementation to refine fibrosis evaluation and staging. A ML approach, particularly using SVM, applied to US-SWE [point SWE (p-SWE) and two dimensional SWE (2D-SWE)] from different vendors demonstrated excellent diagnostic accuracy in distinguishing significant liver fibrosis (≥ F2), comparable to MRE. This improvement over conventional SWE analysis without AI support is due to the ability of ML models to capture non-linear patterns and complex relationships within the data, rather than relying solely on simple statistical summaries[46]. A similar approach, based on color map SWE images use converted into numerical stiffness values, was proposed by Gatos et al[47,48]. Their system analyses these types of datasets to find important characteristics, and then SVM ML algorithm was able to discriminate between healthy and cirrhotic liver with good accuracy.

AI ability to analyze a large data amount enables the integration of inputs from multiple US techniques. Destrempes et al[49] proposed a ML model combining QUS parameters and p-SWE to assess steatosis, inflammation and fibrosis in CLD patients. This strategy significantly improved diagnostic performance not only for fibrosis staging, but also for the grading of steatosis and inflammation. The study also highlighted that the simultaneous presence of inflammation and steatosis can affect stiffness values. Particularly, it pointed out associations between inflammation and fibrosis, both contributing to increased liver stiffness, suggesting that a multiparametric model is essential to differentiate between these overlapping pathological components.

DL applied to SWE also yielded excellent accuracy results in the staging of liver fibrosis. Kagadis et al[50] applied pre-trained DL networks [Google inception network, AlexNet CNN, visual geometry group 16, residual network 50 layers (ResNet50), densely connected convolutional networks 201 (DenseNet201)] on SWE imaging, all achieving good performance, although ResNet50 and DenseNet201 consistently showed better diagnostic accuracy across all fibrosis stages. A multi-center study demonstrated that DL-based radiomics of 2D-SWE (Dlre) outperformed 2D-SWE in detecting cirrhosis and advanced fibrosis. Moreover, Dlre also outperformed biomarkers employed (FIB-4, APRI) to identify all three stages of liver fibrosis in patients with chronic hepatitis B and its diagnostic accuracy increased obtaining more 2D-SWE images for each patient[51]. However, to further improve the detection of significant fibrosis, an updated model Dlre2.0 was developed from the original using a transfer learning (TL) approach. It showed a higher area under the curve (AUC), although the improvement was not statistically significant[52]. DL accuracy was compared with ML ones. Specifically, Brattain et al[53] developed SWE-assist, a model capable of assessing SWE image quality and selecting the region of interest (ROI) for analysis, using random forest, SVM and a CNN to classify fibrosis as significant or not (above or below stage F2). Among the three methods, the CNN demonstrated the best performance with an AUC of 0.89.

Interesting results were obtained by combining gray scale modality and 2D-SWE images, analyzed using Inception-V3, a DL model with TL from a pre-trained model. The model combining gray scale and elastogram modalities with TL proved to be the most accurate overall in staging liver fibrosis. It outperformed all other methods, including each modality used individually, blood tests (APRI and FIB-4) and the combination of gray scale and elastogram modalities without TL[54].

Finally, two recent studies have demonstrated innovative approaches to non-invasive liver fibrosis assessment. The first study explored stacked microvascular imaging (SMVI), an advanced US technique capable of evaluating hepatic micro-vascularization by detecting early morphological alterations of the vasculature. This modality appears particularly promising for the detection of early-stage fibrosis[55]. Notably, applying DL algorithms to SMVI data achieved higher diagnostic accuracy than conventional human scoring in identifying significant fibrosis (≥ F2), although their performance in differentiating specific fibrosis stages remained somewhat limited. Nevertheless, SMVI, especially when combined with conventional US and elastography, may serve as a valuable screening tool for the early detection of significant fibrosis[56]. The second study introduced a novel fibrosis staging approach based on radiomics applied to improve ANN (iANN)-based homodyned-K imaging. This technique extracts features from US backscattered radiofrequency signals and outperforms both the iANN-based homodyned-K US imaging alone and the radiomics method applied to uncompressed US backscattered envelope images in the liver fibrosis assessment. Notably, it outperformed conventional radiomics approaches based on envelope images, regardless of the presence of hepatic steatosis[57].

AI TECHNIQUES FOR LIVER STEATOSIS EVALUATION

MASLD represents one of the main diagnostic categories within the overarching framework of SLD[7]. It represents a newly proposed nomenclature that encompasses a heterogeneous population characterized by hepatic steatosis and at least one cardiometabolic risk factor[7,58]. Multiparametric US assessment in patients with MASLD allows for the non-invasive detection not only of hepatic steatosis, but also of its degree of severity and, in some cases, the evaluation of its spatial distribution.

In steatotic livers, the hepatic parenchyma typically exhibits increased echogenicity. Conventionally, liver echogenicity is assessed by comparison with the echogenicity of the renal cortex of the right kidney[59]. In patients with hepatic steatosis, the US beam is attenuated as it penetrates deeper tissues, which alters the backscatter pattern, as evidenced by reduced visualization of the diaphragm. In addition, vascular structures and any focal lesions may appear less distinct or poorly delineated[60,61]. Based on the interpretation of these imaging features, hepatic steatosis is graded as absent, mild, moderate, or severe.

Typically, triglyceride accumulation in the liver is diffuse and homogeneous. However, in some patients with liver steatosis, fat deposition may present in a focal or heterogeneous pattern, with variability observed between different hepatic segments[62]. Fat accumulation tends to be more prominent in the posterior segments (VI-VII) and in the right hepatic lobe. In some cases, certain regions, such as the periportal area, may be relatively spared resulting in a so-called “geographic map” pattern of steatosis. These features are usually detectable via conventional B-mode US, although their identification can be challenging, particularly when fat is deposited in deep or poorly accessible segments.

The interpretation of such findings is inherently operator-dependent: The ability to recognize subtle grades of steatosis or accurately quantify its extent is largely influenced by the clinician’s expertise. Mild forms of hepatic steatosis may go undetected, and inter-operator variability in grading remains a well-documented limitation of conventional US imaging[60,62]. To reduce operator-dependent variability, some standardized tools for the quantification of liver fat content with US already available on several US systems are gaining increasing relevance. According to the latest guidelines published by the World Federation for Ultrasound in Medicine and Biology[21], the currently available tools that analyze the degree of attenuation of the US beam through the liver parenchyma are attenuation coefficient (AC), backscatter coefficient (BSC), and speed of sound (SoS) measurement.

AC measures the attenuation of the US beam as it traverses the hepatic parenchyma, which is frequency dependent and directly proportional to the degree of steatosis. As the US beam traverses the steatotic liver parenchyma, its intensity decreases more than when traversing the non-steatotic parenchyma, leading to an increased AC. CAP[63,64] is one of the algorithms available to estimate the AC: It is a tool available on the FibroScan system (TE) to estimate the amount of fat in the liver, and estimates the attenuation slope in dB/m within a range of 100-400 dB/m. It is generally used to monitor response to treatment in patients with MASLD, however, more in-depth studies are needed as values may be altered by confounding factors or the different aetiologies of liver disease.

BSC[65] reflects the amount of US energy backscattered from the hepatic tissue in response to fat accumulation. Since fat deposition alters the acoustic scattering properties of the parenchyma, BSC values tend to increase. This metric is emerging as a promising non-invasive parameter for hepatic steatosis assessment. To complete the set of quantitative tools, SoS[66] is another measurable physical parameter available in certain advanced US systems. It reflects the propagation speed of the acoustic wave through liver tissue, which varies depending on its composition. A lower SoS is typically associated with increased intrahepatic fat content.

Applications of AI in US-based detection of liver steatosis

In the evaluation of hepatic steatosis, the use of AI techniques particularly ML and DL algorithms such as CNNs is gaining increasing traction. These methods provide a standardized, non-invasive approach to liver fat quantification, reducing inter-operator variability in US image interpretation and enhancing diagnostic accuracy, with performances comparable to or even exceeding that of experienced sonographers.

By analyzing raw B-mode US data, these algorithms can generate a quantitative index of steatosis, classify its severity into clinical grades, and effectively support clinicians in diagnosis, risk stratification, and longitudinal follow-up[67,68].

To this end, several DL algorithms have been developed and “trained” on large-scale US datasets, often integrated with data from MRI-PDFF the current non-invasive gold standard[69,70] or with histological findings obtained via liver biopsy. These models aim to estimate the hepatic fat content at the parenchymal level.

All such tools should be considered in conjunction with elastography techniques, such as quantitative SWE, to assess the potential evolution of steatosis into hepatic fibrosis[49].

A single-center cross-sectional study conducted by Kwon et al[71] aimed to validate an AI-enhanced quantitative US algorithm for non-invasive assessment of MASLD severity, comparing its performance with that of MRI-PDFF. This algorithm is a DNN model designed to quantitatively estimate AC of the hepatic parenchyma from B-mode US data and detected echo signals (pulse-echo data). It includes multiple convolutional encoders that extract temporal features from envelope-detected pulse-echo signals acquired with five-angle plane-wave insonation. These features are then normalized through an adaptive layer guided by the B-mode image. Finally, a series of convolutional layers process the normalized data to output the quantitative AC value, enabling accurate non-invasive assessment of hepatic steatosis. The results obtained through this model were compared, in addition to MRI-PDFF values, to two conventional US techniques FibroTouch’s US attenuation parameter (UAP) and Canon’s ATI.

The study demonstrated a strong correlation between QUS-AC and MRI-PDFF, supporting the potential of QUS to standardize hepatic fat quantification and reduce human error. In contrast, ATI showed only moderate correlation, while UAP exhibited weak correlation with MRI-PDFF. These findings suggest that AI-enhanced QUS may serve as a reliable and non-invasive tool for the assessment of hepatic steatosis in patients with NAFLD.

A retrospective study conducted in 2021 by Rhyou and Yoo[72] proposed a fully automated cascade DL model, based on three CNNs, for the automated assessment of hepatic steatosis using B-mode US images. The algorithm consists of three main stages: (1) Automatic segmentation of the liver-kidney region using a DeepLabv3+ architecture; (2) Renal ring detection to ensure the correct identification of the ROI; and (3) Steatosis grading via a pre-trained CNN (SteatosisNet, based on Inception-V3). This pipeline enables standardized and reproducible image interpretation, reducing inter-operator variability and supporting accurate steatosis quantification. The experimental results demonstrated excellent diagnostic performance, with a sensitivity of 99.78%, specificity of 100%, positive predictive value of 100%, negative predictive value of 99.83%, and overall diagnostic accuracy of 99.91%. Similarly, excellent results were obtained using a one-dimensional CNN algorithm based on US radiofrequency signals for NAFLD diagnosis and fat quantification. This is particularly relevant because radiofrequency signals contain rich information on about liver composition, which may be lost or distorted when converted in B-mode images[73].

The applicability of these models has been demonstrated by numerous studies, which consistently report that the use of software based on DNNs or ML techniques achieves excellent sensitivity and specificity. Different grades of hepatic steatosis can be discriminated with areas under the AUCs generally exceeding 0.90, ensuring both diagnostic accuracy and data reproducibility[74,75] (Table 2).

Table 2 Studies that evaluated the artificial intelligence ability to stage liver steatosis[49,68,73-75,112-120].

Ref.	Type of study	Population	Number of patients	AI technique employed	Main results
Fujii et al[112]	Prospective, cross-sectional	MASLD	486	DL (U-net)	DL-based segmentation reliably identified the surface irregularity of the liver
Drazinos et al[113]	Retrospective, monocentric	MASLD	112	DL (Inception-V3, MobileNetV2, ResNet50, DenseNet201 and NASNet mobile)	DenseNet201 achieved the highest overall performance, while Inception-V3 showed superior accuracy in the binary classification of steatosis
Chou et al[114]	Retrospective	Healthy patients and patients with liver steatosis	2070	DL	DL models achieved higher 88.7% sensitivity for mild steatosis and consistent accuracy across all grades (normal 91.8%, moderate 77.3% moderate, severe 84.4%)
Vianna et al[115]	Retrospective	Healthy patients and patients with liver steatosis	199	DL (VGG16, ResNet50 and Inception-V3)	DL–based analysis of B-mode US images demonstrated diagnostic performance comparable to expert human readers in both the detection and grading of hepatic steatosis
Vianna et al[116]	Retrospective, multi-center	Patients with suspected hepatic steatosis datasets	Not specified	DL	Diagnostic AUC for steatosis detection increased from 0.78 to 0.97. Test-time adaptation improved DL models robustness and generalizability B-mode US
Cao et al[117]	Prospective, cross-sectional	Healthy patients and patients with liver steatosis	240	DL	The methods showed a good ability (AUC > 0.7) to identify steatosis, particularly in distinguishing moderate from severe (AUC = 0.958)
Han et al[73]	Prospective	Healthy individuals and patients with NAFLD	204	CNN DL	Accurate diagnosis of NAFLD and fat quantification using US radiofrequency signals
Byra et al[74]	Prospective	Steatosis and/or obese patients	55	DL (Inception ResNet-v2)	The AI-based model performed best (AUC = 0.977) outperforming the hepatorenal sonographic index (not significant) and grey-level co-occurrence matrix (significant difference)
Constantinescu et al[75]	Retrospective	Healthy patients and patients with liver steatosis	60	DL (Inception-V3 and VGG-16)	DL algorithms demonstrated excellent diagnostic performance, achieving accuracy rates exceeding 90%
Jeon et al[68]	Prospective	Suspected steatosis	173	DL	DL algorithm combining QUS parametric maps with B-mode imaging accurately estimated hepatic fat fraction and reliably diagnosed hepatic steatosis
Gómez-Gavara et al[118]	Prospective	Livers from brain-dead donors, evaluated during the procurement phase	192 livers	ML	Integrating ML with liver texture and color analysis smartphone images enables highly accurate estimation of hepatic steatosis severity
Santoro et al[119]	Prospective, cross-sectional	Healthy patients and patients with liver steatosis	134	ML	AI application enhances both the diagnostic accuracy and efficiency of US in the assessment of hepatic steatosis
Kaffas et al[120]	Retrospective, single center	Healthy patients and patients with liver steatosis	403	DL	This DL algorithm achieved accurate estimation of hepatic fat fraction and reliable diagnosis of hepatic steatosis
Destrempes et al[49]	Prospective	CLD	82	ML (random forest)	Random Forest integration of QUS and SWE markedly enhanced diagnostic vs SWE alone, particularly for steatosis assessment, increasing AUC by 25%-50%

AI: Artificial intelligence; CLD: Chronic liver disease; NAFLD: Nonalcoholic fatty liver disease; MASLD: Metabolic dysfunction-associated steatotic liver disease; DL: Deep learning; ResNet50: Residual network 50 layers; DenseNet201: Densely connected convolutional networks 201; VGG: Visual geometry group; CNN: Convolutional neural network; ML: Machine learning; QUS: Quantitative ultrasound; AUC: Area under the curve; SWE: Shear wave elastography; UC: Ultrasound.

Beyond the individual studies summarized above, a comparative overview of the main AI models applied to liver US provides additional insight into their relative performance and suitability across different diagnostic contexts. Collectively, the studies summarized in Tables 1 and 2 illustrate the expanding role of AI in liver US across different clinical settings. At the same time, they reveal substantial heterogeneity in study design, data acquisition protocols, and the types of algorithms employed.

In comparative terms, CNN-based architectures typically outperform conventional ML models in image-based feature extraction and fibrosis staging, owing to their ability to autonomously learn hierarchical representations from US data. Conversely, SVMs and random forest models often show more stable performance in smaller datasets or when combining imaging-derived metrics with clinical and biochemical parameters. GANs provide distinctive advantages for data augmentation and synthetic image generation, thereby improving model generalizability and image quality. The complementary characteristics of these approaches support the development of hybrid and multimodal frameworks that integrate their respective strengths to enhance diagnostic precision and reproducibility in liver disease assessment. The key principles, advantages, and limitations of the most commonly employed AI techniques in liver US are summarized in Table 3.

Table 3 Comparative summary of the main artificial intelligence models applied to liver ultrasound, outlining their key features, strengths, limitations, and representative clinical applications.

Model type	Main features	Strengths	Limitations	Typical clinical applications
Convolutional neural networks	Deep-learning models extracting hierarchical image features from B-mode or SWE data	High accuracy in fibrosis staging; automatic feature extraction; excellent for large datasets	Require large training datasets; limited interpretability (“black box”)	Fibrosis staging, steatosis grading, lesion detection
Support vector machines	Supervised ML classifier using kernel-based separation of data	Robust for small datasets; interpretable decision boundaries	Lower performance for complex, high-dimensional data	Early fibrosis detection, ML radiomics, feature selection
Random forest	Ensemble ML algorithm combining multiple decision trees	Handles mixed data (imaging + clinical); resistant to overfitting	Limited ability to capture image texture; less suitable for pixel-level analysis	Integration of US features with clinical and laboratory data
Generative adversarial networks	DL models using generator-discriminator structure	Effective for data augmentation; improves synthetic image realism and model generalizability	Computationally demanding; risk of instability during training	Image synthesis, dataset expansion, quality enhancement
Hybrid/multimodal models	Combine DL image-based features with ML classifiers or clinical variables	Capture complementary information; improve diagnostic precision	Require harmonized data and complex implementation	Comprehensive multiparametric liver assessment (fibrosis + steatosis)

DL: Deep learning; ML: Machine learning; SWE: Shear wave elastography; UC: Ultrasound.

The use of hybrid DL and ML models further enhances diagnostic accuracy[76]. By “hybrid models” we refer to the integration of DL architectures such as CNNs, which perform visual analysis by extracting features from US images with the robustness of ML algorithms (e.g., random forest, SVM, eXtreme Gradient Boosting), which analyze the extracted data and generate the final prediction.

But US alone is insufficient for risk stratification in patients with MASLD: To provide clinically meaningful decision support, it is necessary to combine imaging data with clinical and laboratory information, including anthropometric measures (e.g., weight, height), liver function indices (e.g., transaminases, albumin, international normalized ratio), and metabolic markers (e.g., fasting glucose, insulin levels, total and high density lipoprotein cholesterol, triglycerides).

Some predictive AI models capable of integrating clinical and laboratory data already exist: For example, NASHmap^TM[77,78] is a tool developed in collaboration with the Mayo Clinic, which analyzes 14 easily accessible clinical and laboratory variables to predict the likelihood of histologically confirmed non-alcoholic steatohepatitis.

Despite these improvements, there is not yet an AI model that integrates data obtained through US investigation or the study of US beam behavior with laboratory, anthropometric or other imaging data. These data could be integrated through ML algorithms such as random forests or gradient boosting machines. The resulting predictive models could be used to estimate the likelihood of disease progression toward fibrosis or cirrhosis, or to monitor improvement following appropriate dietary and pharmacological interventions.

AI models for steatosis distribution

As previously mentioned, steatosis distribution within the liver is not always homogeneous. In some patients, segmental steatosis accumulation may occur, and certain studies suggest that this may be associated with reduced portal venous flow[79] and a higher likelihood of progression to hepatic fibrosis.

In the current state of research, there is no AI model capable of recognizing a non-uniform distribution of steatosis in the liver parenchyma on US, including both fat accumulation and fat-sparing regions. However, some CNNs developed through AI, if further developed, could prove very promising in this field.

For instance, among CNNs, U-Net is a model trained to perform pixel-wise segmentation of radiological images, allowing the recognition of anatomical structures[80,81]. U-Net is currently used for liver segmentation (the automated delineation of liver boundaries) on US, MRI or TC, to isolate the liver parenchyma from surrounding anatomical structures. In a future development, such models could potentially be used to recognize segmental parenchymal alterations. The development of such tools could represent a promising frontier for assessing heterogeneous liver parenchyma, enabling more accurate fibrosis risk stratification and further personalization of patient care.

CHALLENGES AND LIMITATIONS OF AI IN LIVER DISEASE EVALUATION

Data quality and availability

The analysis of US images using AI models is a rapidly expanding and evolving field. However, a major barrier to its broader implementation lies in the limited number of multicenter studies, as well as the current lack of large, high-quality annotated datasets based on sufficiently sized cohorts to support standardized clinical applications.

A key methodological limitation of current research lies in the predominance of single-center, retrospective studies involving relatively homogeneous populations. The limited heterogeneity of training datasets and the absence of prospective, multicenter validation reduce model generalizability and increase the risk of overfitting, ultimately limiting the reproducibility and clinical applicability of AI-based liver US approaches.

To ensure the reliability of such datasets, the images must be assessed by expert operators and compared to gold standard diagnostic references, such as liver biopsy or MRI-PDFF. This process is not always feasible in routine clinical practice. Moreover, data from different studies are often acquired using heterogeneous imaging equipment, with varying price ranges, image resolution, and acquisition protocols.

Although recent studies applying AI to liver US have shown encouraging results, most remain at a preliminary stage of development. Their methodologies are frequently constrained by small, single-center cohorts, heterogeneous imaging protocols, and a lack of external or prospective validation, all of which limit the reproducibility and generalizability of the reported findings. In addition, the interpretability of DL models continues to represent a major challenge, as many algorithms operate as “black boxes”, thereby hindering clinical transparency, user confidence, and seamless integration into routine practice.

To address these issues, several unsupervised domain adaptation techniques are currently under investigation. These methods aim to enhance the robustness and accuracy of AI models in assigning diagnostic features across different imaging domains and patient populations[82].

Interpretability and transparency

Some AI models, particularly those based on DL, exhibit a so-called “black-box” nature[83], meaning that it is difficult to clearly understand how and why a specific diagnostic conclusion is reached. These models operate through opaque decision-making processes, which are not easily traceable or interpretable. This poses a significant challenge in the medical field, as it hinders clinicians from following the model’s clinical reasoning and, consequently, from effectively communicating AI-assisted findings to patients raising important ethical concerns.

In the context of multiparametric liver US, for instance, it is often unclear which areas of the parenchyma were evaluated by AI algorithms to generate a given classification. In order to enhance both clinical trust and scientific validation of AI-based systems, efforts are currently being directed toward the development of explainable AI (XAI) frameworks[84-86], which aim to make AI model interpretations more transparent and comprehensible for both physicians and patients.

Generalization and overfitting

Two additional challenges in the medical application of AI are related to the issues of generalization and overfitting.

Many AI models are trained on monocentric cohorts, meaning that the examinations are often conducted on homogeneous populations, for instance, individuals from the same geographic area or with similar demographic characteristics. As a result, applying the same model to patients with different characteristics such as age, body mass index, or ethnicity may reduce the model’s clinical applicability due to its limited ability to “generalize” findings across diverse clinical contexts.

Another significant concern is overfitting[87,88]. When an AI model is trained using data from a single type or brand of US machine, it may learn very specific details, including minor artifacts or patterns related to that device. This can lead to excellent performance on the training dataset but significantly reduced accuracy when applied to images acquired with different equipment or in different clinical settings, due to the model’s inability to generalize to unseen data.

To develop robust and reliable models, it is therefore crucial to train AI systems on diverse and heterogeneous datasets and to incorporate computational strategies aimed at improving model performance and generalizability such as domain adaptation, data augmentation, early stopping, and regularization[89-91].

Regulatory and ethical considerations

Several ethical and regulatory considerations must be addressed when applying AI to clinical practice[92-94]. First, the lack of standardized international guidelines for the clinical implementation and validation of AI systems introduces significant regulatory uncertainty.

Secondly, the processing of sensitive patient health information (such as medical history, blood test results, and clinical records) raises substantial concerns regarding privacy, particularly as these data are often stored within digital archives or cloud-based systems[95]. From a social perspective, the use of AI may also pose challenges in terms of equity: Existing disparities can be further exacerbated if underrepresented populations such as ethnic minorities, older individuals, or children are not adequately included in training datasets.

In black-box models, as previously discussed, the inability to trace the step-by-step reasoning behind a diagnostic output introduces further issues regarding transparency. This lack of interpretability raises medico-legal concerns, particularly with respect to accountability and liability in the event of diagnostic errors[96,97].

CLINICAL APPLICATIONS AND FUTURE PERSPECTIVES

Current clinical applications

Recent years have seen the development and refinement of several AI models aimed at supporting physicians in their clinical practice. Specifically, in the hepatology field, these tools have been trained to perform non-invasive assessments of hepatic steatosis and fibrosis, with the goal of standardizing the diagnosis and grading of these conditions.

These systems, which are still undergoing continuous development, may be integrated into routine clinical settings even at early stages, particularly in the screening of patients with risk factors for developing steatosis or fibrosis. In this context, AI may assist in the early detection of liver pathology.

Some of the latest-generation US systems already include integrated AI tools; in other cases, images must be processed using external software platforms. To achieve even more accurate classification of steatosis and fibrosis, the output from advanced AI-based image analysis algorithms should ideally be combined with results from other non-invasive tools, such as TE[98] and MRE[99], as well as with clinical, laboratory, and anthropometric data.

Although AI-based liver US techniques have advanced rapidly, their widespread clinical adoption remains limited. Major obstacles include the absence of standardized acquisition protocols across different US platforms, limited interoperability among vendors, and the current lack of transparent, explainable algorithms that clinicians can readily interpret and trust. In addition, regulatory pathways for AI validation and approval are still in development, and the paucity of large-scale, multicenter prospective studies continues to hinder their clinical integration.

From a practical standpoint, AI-assisted multiparametric US could be incorporated at several stages of the diagnostic workflow. It may act as a non-invasive complement or potential alternative to liver biopsy for fibrosis staging, facilitate early screening and risk stratification in patients with metabolic dysfunction, and enable consistent longitudinal follow-up through reproducible quantitative biomarkers. Integrating AI-derived US metrics into existing diagnostic algorithms alongside TE and MRE could enhance diagnostic accuracy, minimize operator dependence, and promote a more efficient use of healthcare resources in everyday hepatology practice.

AI in disease monitoring and prognostication

The application of AI-based software enables not only the early diagnosis of hepatic pathologies, but also longitudinal monitoring of disease progression in response to therapies, bariatric interventions, or lifestyle modifications. By integrating all available data US and other radiological imaging, laboratory and biometric parameters, as well as genomic, epigenomic, and metabolomic information AI-assisted models allow for a comprehensive and personalized longitudinal analysis tailored to the individual patient, thus advancing the paradigm of precision medicine.

Novel AI models can track real-time changes in liver parenchymal texture, including subtle alterations that may not be visible to the human eye, and monitor disease evolution such as fibrotic progression in patients with MASLD and the associated risk of hepatocellular carcinoma[100,101]. These data, stored in electronic medical records, facilitate continuous patient surveillance.

This system optimizes clinical resource allocation, supports the design of effective surveillance strategies, and aids in selecting the most appropriate therapeutic approach for each patient. Rather than replacing the clinician, it serves as a valuable decision-support tool[95].

Future trends and research directions

Increasingly reliable and sophisticated algorithms represent the future of AI applications in medicine, driven by the advancement of techniques such as reinforcement learning and federated learning, which can address and overcoming several critical limitations.

Reinforcement learning[102,103] is a branch of ML that enables AI models to learn effective monitoring and treatment strategies through a trial-and-error mechanism, interacting with the environment and receiving feedback to optimize outcomes.

Federated learning[104-106], on the other hand, is a training approach that allows AI algorithms to be developed on large datasets originating from multiple centers or institutions without the need to share sensitive data, thereby preserving patient privacy and complying with data protection regulations.

Most current studies are retrospective and monocentric, limiting the generalizability of findings. To further validate the performance of AI models, large-scale, prospective, international, and multicenter studies are required, involving heterogeneous populations across different age groups, ethnicities, disease stages, and comorbidities.

In the future, a tighter integration with genomic profiling and biomarker analysis is anticipated, enabling earlier identification of at-risk individuals and timely intervention. Additionally, wearable devices (e.g., smartwatches, biosensors) may provide continuous monitoring of parameters such as physical activity and dietary habits, supporting the design of personalized care strategies[107]. Dunn et al[108] employed wearable activity trackers in liver transplant candidates during corona virus disease 2019 pandemic, hypothesizing that a sudden decline in physical activity could serve as an early warning signal of liver disease complications. Integrating AI with wearables could help to identify and forecast future clinical events[109]. Moreover, wearables can also generate large volumes of data to analyze with AI algorithms to uncover novel, non-linear relationships between these parameters and liver disease[107,109].

Future research should focus on developing explainable and interpretable AI frameworks (XAI) capable of offering greater transparency in diagnostic decision-making. The integration of imaging data with clinical, laboratory, and molecular information through hybrid or multimodal AI models represents a crucial step toward achieving clinically meaningful and personalized applications. In parallel, federated learning approaches may enable large-scale, multi-institutional collaborations while maintaining data privacy and enhancing algorithmic robustness across diverse populations and imaging platforms. Collectively, these directions delineate the current knowledge gaps and point to the most promising frontiers in AI-assisted liver US.

CONCLUSION

AI consists of algorithms capable of making decisions in a human-like manner interpreting large volumes of data from various sources, such as clinical, laboratory and imaging data. It provides valuable tools to support early diagnosis, management and surveillance of CLD and SLD, serving as powerful resources to sustain the clinical expertise. Relying solely on a machine’s judgment without appropriate human oversight, could be unwise. It’s important to keep in mind that, despite its vast potential, AI should serve as a tool to support not replace clinical judgment. AI improves both US and elastography image analysis, particularly by reducing inter-operator variability for the former and enhancing the agreement among methods implemented on equipment of different vendors for the latter. The next step will be including AI algorithms in diagnostic flowchart to standardize their use; however large, multi-center studies are warranted to validate these methods. From a hepatological perspective, the integration of AI-based multiparametric US into clinical practice holds great promise for improving early detection, refining risk assessment, and enabling personalized follow-up strategies in patients with CLD. Such developments highlight the translational impact of these technologies and their growing importance in the field of clinical hepatology.