Jian-Xu Rao
Ying Li
Kai Leng- 1Department of Hepatopancreatobiliary Surgery, The First People’s Hospital of Zunyi (The Third Affiliated Hospital of Zunyi Medical University), Zunyi, Guizhou, China
- 2Department of Gastroenterology and Hepatology, West China Hospital, Sichuan University, Chengdu, Sichuan, China
Hepatocellular carcinoma (HCC) is the predominant histological subtype of primary liver cancer, with a 5-year survival rate of approximately 18%. Early detection of HCC is critical for guiding treatment selection and improving patient survival outcomes. The effectiveness of conventional screening methods is decreased due to their inherent limitations and individual variability. Artificial intelligence (AI) has advanced rapidly in medical practice and has played a significant role in increasing the early detection rates of HCC by replacing manual tasks and accessing hidden information in routinely available clinical data. However, numerous challenges, such as ethical concerns, model instability, and generalizability, must be overcome before their full clinical implementation. This article reviews recent studies that describe AI-based models for the early diagnosis of HCC, focuses on the current applications and persistent challenges of AI in HCC screening and discusses its perspectives. We aim to provide a critical evaluation of the potential of AI for enhancing early HCC detection and improving patient prognosis.
1 Introduction
Primary liver cancer ranks as the sixth most common malignancy and the third leading cause of cancer-related mortality worldwide, representing a significant global health burden (1, 2). It primarily includes hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and combined hepatocellular-cholangiocarcinoma (HCC-ICC). HCC accounts for approximately 90% of all primary liver cancers and is associated with a 5-year survival rate of approximately 18% (3, 4). Early detection is critical for both therapeutic decision-making and patient prognosis (5).
Screening is important for the early detection of HCC, but its effectiveness is limited by multiple barriers (6, 7). Patient-related obstacles include misconceptions about screening, financial constraints, limited healthcare access, difficulties in scheduling appointments, long intervals between visits and ultrasounds, and transportation issues (8–10). Provider-related barriers include insufficient knowledge of guidelines, failure to identify at-risk patients, time constraints, and resource limitations (11, 12). These factors collectively undermine suboptimal screening adherence. Furthermore, current HCC screening strategies are limited by shifting etiologies, notably, the rise of metabolic dysfunction-associated steatotic liver disease (MASLD) and alcohol-related liver disease (ALD) against a background of persistent viral hepatitis (13, 14). This epidemiological shift, characterized by a large at-risk population, a lower incidence of HCC in non-cirrhotic individuals, and ongoing cost-effectiveness concerns, substantially complicates the justification for widespread screening (15, 16). The screening modalities used for HCC patients also face significant limitations. However, tumor biomarkers for accurate early detection are still lacking. The available data show that biomarkers are suboptimal in terms of cost-effectiveness for routine surveillance of early HCC (6). Ultrasound has several inherent limitations, including operator dependence, suboptimal visualization in obese patients, and the risks associated with false-positive or indeterminate findings, potentially compromising its efficacy (17). Although contrast-enhanced computed tomography (CT) provides high diagnostic accuracy, its routine use is constrained by radiation exposure. Magnetic resonance imaging (MRI) demonstrates superior sensitivity for small lesions but is less feasible for population-level screening because of longer scan times, high cost, and complex interpretation.
Artificial intelligence (AI) is defined as a broad concept involving the use of computer systems to perform complex functions that could require human intelligence, quantifying uncertainty, probabilistic reasoning, problem-solving behaviors and knowledge representation. The field of AI encompasses various methodologies and technologies, such as machine learning and its subset deep learning, which employ algorithms including artificial neural networks (ANNs) and convolutional neural networks (CNNs) to enable applications in areas like image processing and computer vision (18). The implementation of AI-powered diagnostic technologies that integrate multifactorial analysis is essential to achieve precise diagnostics and advance personalized medicine. Given the heterogeneity of HCC, AI can extract critical information and leverage massive amounts of multiparametric data (Figure 1). The application of AI in HCC screening has significantly improved the early detection rate of HCC.
Figure 1. Flowchart of an AI-assisted precision surveillance strategy for HCC. The framework integrates readily available patient data—including demographics, etiology, liver disease status, and metabolic factors—to generate a personalized risk score. Based on this score, patients are stratified into distinct risk categories. For each category, a tailored screening strategy (e.g., no surveillance, semi-annual ultrasound with or without AFP…) is selected and implemented via a dedicated AI-assisted model, enabling precision in early detection. HBV, hepatitis B virus; HCV, hepatitis C virus; BMI, body mass index; HCC, hepatocellular carcinoma; AI, artificial intelligence; US, ultrasound; AFP, Alpha-fetoprotein.
In the management of HCC, it is crucial to differentiate between the application of AI for screening and diagnostic classification. Screening involves the use of AI tools in asymptomatic, high-risk populations with the primary objective of detecting potential HCC at an early stage. It aims to achieve high sensitivity and minimize the rate of missed diagnoses (19). In contrast, diagnostic classification entails the use of AI models to assist in the characterization of identified focal liver lesions (FLLs). These models aim to differentiate between benign and malignant nodules or to further classify specific HCC subtypes. Consequently, this application prioritizes high specificity and overall diagnostic accuracy (20, 21). Studies have demonstrated that developed AI models exhibit excellent performance in the differential diagnosis of HCC, achieving area under the curve (AUC) values exceeding 0.90. Furthermore, a variety of AI-based tools have been widely integrated into clinical workflows for both HCC screening and diagnostic characterization, with certain technologies demonstrating potential applicability across both domains (21, 22). Nevertheless, clearly delineating the intended application scenario of each model remains crucial for guiding its appropriate clinical implementation and validation.
In this study, we reviewed the current studies that describe AI-based models for screening HCC and focused on the applications, challenges and future perspectives of AI in HCC screening. We aimed to comprehensively evaluate the role and limitations of AI in HCC screening to improve early HCC detection rates and patient prognosis.
2 AI for HCC screening
A comprehensive search of the literature was conducted in English on human subjects via four major databases: PubMed, Embase, Cochrane and Web of Science. The search aimed to identify relevant articles published from inception until November 14th, 2025. The search strategy employed a combination of the following key terms and their variations: “hepatocellular carcinoma,” “hepatic carcinoma,” “HCC,” “artificial intelligence,” “machine intelligence,” “computational intelligence,” “cancer screening,” “cancer early detection,” and “early diagnosis of cancer.” The initial search identified 449 papers: 167 from PubMed, 136 from Embase, 30 from the Cochrane Library and 116 from the Web of Science. After excluding duplications, 389 article titles and abstracts were analyzed. The full texts of 54 articles, which were relevant only to AI in HCC screening, were assessed for eligibility. Finally, 43 articles that met our inclusion criteria were included in this review (Figure 2). A summary of these studies is shown in Supplementary Table 1.
Figure 2. Flowchart of the literature search and study selection.
2.1 Serum biomarkers
Serum biomarker testing offers a non-invasive and readily implementable approach. Its feasibility and effectiveness for the early detection of HCC are well established (23). Alpha-fetoprotein (AFP) remains the most widely used serological marker for HCC screening. It typically stabilizes in normal adults by 8–12 months of age and is elevated in patients with liver disease, particularly primary liver tumors (hepatoblastoma in children and HCC in adults) (24). However, its sensitivity and specificity are suboptimal, particularly for early-stage HCC detection (25). The latest meta-analysis revealed the diagnostic performance of AFP for early HCC detection, with a sensitivity of only 49.1%, despite an 87.9% specificity (26). The sensitivity of AFP for HCC detection may be influenced by the underlying etiology of HCC. For example, patients with HCC associated with hepatitis B virus (HBV) infection typically present higher serum AFP levels than those with HCC caused by other factors (27). The low sensitivity of AFP diminishes its effectiveness in the early detection of HCC, thereby limiting the utility of serum biomarkers in HCC screening.
AI has become instrumental in bioinformatics, especially in managing large-scale datasets. Numerous studies have utilized machine learning algorithms combined with data from repositories such as the Gene Expression Omnibus (GEO) and the Cancer Genome Atlas (TCGA) to identify feature genes, resulting in the discovery of multiple clinically relevant biomarkers and improving the accuracy and efficiency of HCC screening (28–31). One study leveraged the flexibility of bioinformatic approaches by combining AI and bioinformatics to explore the relationship between diabetic kidney disease and HCC and to identify associated diagnostic markers (32). The integration of AI and bioinformatics enables more precise identification of typical genes associated with HCC of unusual etiology, thereby contributing to the advancement of precision screening and personalized therapeutic strategies.
Cancer-associated autoantibodies (AAbs) can arise early in carcinogenesis and are triggered by the expression of tumor-associated antigens in premalignant or malignant tissues. The ability of the immune system to amplify and sustain specific immune responses makes AAbs promising biomarkers for early cancer detection. Zhang et al. (33) employed human proteome microarrays and an ANN to identify a 7-autoantibody panel for HCC detection. This panel demonstrated superior performance to AFP, particularly in AFP-negative and early-stage (BCLC 0/A) HCC patients. Sato et al. (34) established a novel machine learning-based predictive model for HCC. Their framework employed a grid-search method to optimize classifiers and their hyperparameters, achieving a predictive accuracy of 87.34% and an AUC of 0.94. This performance was superior to that of the standard serum biomarkers AFP, des-gamma-carboxyprothrombin (DCP), and AFP-L3.
AI also aids in the early detection of HCC in patients with diverse characteristics. In patients with HCV-related cirrhosis, the incidence of HCC has decreased markedly among those who achieve a sustained virologic response (SVR) (35, 36). Traditional scoring systems based on routine clinical features overlook population heterogeneity or predictors related to viral eradication (37). Audureau et al. (38) developed novel machine learning-based prognostic models to predict the risk of HCC in patients with compensated HCV-related cirrhosis according to virological status. The key predictors differed between groups: in non-SVR patients, the platelet count and the γ-glutamyl transpeptidas (GGT), AFP, and albumin levels were the most influential, whereas in SVR patients, the prothrombin time, Alanine Aminotransferase (ALT) level, age, and platelet count were predominant. These predictors stratify patients into high- and low-risk populations, facilitating triage for modality-appropriate screening and thereby optimizing the effectiveness and cost efficacy of HCC surveillance. Furthermore, HCC is a major complication of hereditary tyrosinemia type 1 (HT-1), a rare inborn error of metabolism. Since AFP has limited sensitivity for early-stage HCC, Fuenzalida et al. (39) applied a machine learning approach to multidimensional data from HT-1 cohorts. They identified ALT as a promising biomarker that may precede AFP elevation, suggesting a cutoff of 29 U/L to improve the specificity of early HCC detection.
These AI-discovered biomarkers aid in early HCC detection, either directly through risk stratification or by refining AFP-based screening, collectively improving diagnostic sensitivity and specificity to inform early treatment strategies.
2.2 Digital biomarkers
Recent advances in next-generation AI, particularly deep neural network (DNN) algorithms, have significantly greatly enhanced biopattern recognition (40). The identification of digital biomarkers via serum surface-enhanced Raman spectroscopy (SERS) has positioned antibody-free SERS detection as a promising tool for cancer screening. Cheng et al. (41) first developed a nanoplasmonic biosensing chip (NBC) for SERS-based, one-drop blood tests. Next, they integrated DNN modeling with this platform to define a digital HCC biomarker, achieving automated cancer identification within minutes with >92% accuracy and AUC values >0.93 (42).
Urine is well suited for large-scale screening owing to its non-invasive collection, cost effectiveness, and ease of storage (43). As the primary organ for amino acid metabolism, liver damage induces intracellular metabolic dysregulation, consequently increasing serum amino acid concentrations and their loss in urine (44, 45). Alterations in the urinary levels of porphyrin derivatives and bilirubin were observed in patients with HCC. Furthermore, 2,3-Dinor-6-keto-PGF1α, the primary urinary metabolite of systemic prostacyclin, is elevated in cirrhosis. Phosphocreatine and N-glycolylneuraminic acid are upregulated during HCC progression. Urine testing may provide a convenient and rapid method for HCC screening. Dawuti et al. (46) demonstrated that the diagnostic sensitivity for HCC was significantly greater when urinary SERS spectra were analyzed with a support vector machine (SVM) algorithm than when serum AFP was used (85.5% vs. 34.5%). Dou et al. (47) also reported that the combination of urine fluorescence spectroscopy and the SVM algorithm has great potential for non-invasive screening of HCC (diagnostic accuracy was 83.42%, sensitivity was 93.55%, and specificity was 88.00%). Shi et al. (48) developed multiscale trimetal oxide heterojunctions for high-quality urinary metabolic fingerprinting. By employing a machine learning algorithm and mass spectrometry, the platform was trained to identify HCC with an accuracy of 93.3% and was also effective in detecting early-stage, AFP-negative HCC.
Digital biomarkers can also be extracted from medical images. Currently, such imaging-based digital biomarkers are employed primarily for predicting treatment efficacy and facilitating therapy stratification (49–51). However, a recent study utilized AI techniques to develop an HCC prediction model for patients with chronic hepatitis B on the basis of CT-derived digital biomarkers (52). This model demonstrated superior performance compared with conventional models relying solely on clinical and demographic data, thereby offering a novel approach to personalized monitoring and screening strategies. Integrated models that combine imaging features with clinical and demographic data may offer a viable strategy for enabling both large-scale screening and refined risk stratification. Current studies have explored the use of ultrasonic biomarkers for the assessment of Achilles tendon integrity and the prediction of breast cancer recurrence (53, 54). Further research on digital biomarkers related to HCC is anticipated.
2.3 Liquid biopsy
Liquid biopsy primarily comprises circulating tumor cells (CTCs), cell-free DNA (cfDNA), and extracellular vesicles (EVs). As a minimally invasive procedure utilizing readily accessible biofluids such as blood or urine, it offers significant advantages over tissue biopsy for HCC screening. However, each component faces distinct technical hurdles. CTCs, which are tumor-derived cells in the circulation, are exceedingly rare (approximately one cell per 106–107 leukocytes), and their concentration correlates with disease burden, rendering them suboptimal for early-stage detection (55). Similarly, the tumor-derived fraction of cfDNA, known as ctDNA, often constitutes less than 1% of total cfDNA, with levels decreasing to approximately 0.01% in early-stage cancers, posing a paramount challenge for analytical sensitivity (56). Moreover, EV-based applications are confounded by insufficient tissue specificity and purity, complicating the discrimination of tumor-derived EVs from their non-malignant counterparts (57). Consequently, these limitations have impeded their broad clinical translation. Among these, circulating nucleic acids, especially cfDNA, have shown considerable promise. Recent studies have indicated that tumor-specific genetic alterations in ctDNA can achieve superior sensitivity and specificity over AFP for diagnosing early-stage HCC (58, 59).
AI plays a pivotal role in enhancing diagnostic accuracy by enabling the identification of these tumor-specific genetic alterations. For example, Tao et al. (60) employed machine learning to develop a model based on somatic copy number aberrations for early HCC detection in HBV patients, which achieved impressive AUCs of 0.920 and 0.812 in two independent validation cohorts. Similarly, Samir et al. (61) leveraged machine learning to predict HCC by integrating long non-coding RNA (lncRNA) expression profiles with conventional laboratory biomarkers, attaining 100% sensitivity and 97% specificity. Furthermore, Yu et al. (62) identified nine recurrent fusion genes in HCC and developed a logistic regression model that combined two of these fusions with serum AFP levels, achieving 94.8% accuracy in the training cohort. These studies collectively underscore the transformative potential of AI-driven approaches in refining HCC diagnostics.
However, liquid biopsy still faces several challenges in the screening of HCC. Clonal hematopoiesis-related mutations may confound the identification of tumor-derived mutations in ctDNA analyses, necessitating careful bioinformatic filtering (63). The low fractional abundance of ctDNA in early-stage HCC poses significant technical challenges, as it approaches the detection limits of conventional polymerase chain reaction, making it difficult to distinguish true somatic variants from technical artifacts (64). Moreover, current research efforts remain largely confined to a single class of liquid biopsy biomarkers. Future studies should prioritize the integrated analysis of multiple biomarker classes from a single blood draw to unlock more comprehensive diagnostic information. Current research efforts integrating liquid biopsy and AI are predominantly concentrated on the domain of precise diagnosis, with a primary focus on enhancing specificity to reduce false-positive rates. In contrast, AI-driven models utilizing liquid biopsy for HCC screening must prioritize higher sensitivity to minimize false-negative results and prevent disease progression.
2.4 Ultrasound
Ultrasound is the most commonly used modality for HCC screening and detection, offering the advantages of being non-invasive and free from radiation exposure (65). According to the European Association for the Study of the Liver (EASL) guidelines, ultrasound is the recommended examination for HCC screening in all high-risk populations every 6 months (6). This recommendation is consistent with the American Association for the Study of Liver Diseases (AASLD) and Chinese guidelines (66, 67). It has high sensitivity for detecting liver space-occupying lesions >2 cm, but its sensitivity and specificity for early-stage HCC detection remain unsatisfactory (68). Furthermore, the diagnostic performance of ultrasound is significantly operator dependent and can be adversely affected by factors such as ascites and obesity (69).
Recent advances in AI technologies have substantially improved the diagnostic accuracy of ultrasound for HCC detection (70). Two retrospective studies developed AI-based systems for HCC identification among FLLs in ultrasound images, each demonstrating >80% diagnostic accuracy (20, 71). AI also significantly enhanced the diagnostic accuracy of non-expert operators for HCC. A randomized controlled trial demonstrated that non-expert operators using AI-assisted ultrasound not only achieved a significantly higher detection rate of HCC among FLLs but also maintained a comparable rate of false-positive findings relative to conventional ultrasound (72). Furthermore, since serum AFP levels are elevated in only one-third of HCC patients, AI plays a crucial role in assisting ultrasound in identifying biomarker-negative HCC. Zhang et al. (73) developed a deep learning model capable of identifying AFP-negative HCC among FLLs in high-risk patients, achieving an AUC of 93.68%. Contrast-enhanced ultrasound (CEUS) serves as a key diagnostic imaging modality for HCC, advancing conventional ultrasound from screening to definitive diagnosis. While it has 80%–94% sensitivity for early HCC and 63%–70% sensitivity for small lesions (≤2 cm), its performance remains operator dependent and susceptible to interpreter fatigue (74). AI enables rapid analysis and quantification of large datasets, accurate identification of subtle patterns and high-throughput diagnosis without fatigue (71). Ding et al. (75) developed a CEUS-based AI model for multiclass classification of FLLs, which demonstrated superior performance to that of junior radiologists and comparable accuracy to that of senior CEUS specialists. Feng et al. (76) introduced a neural network model that extracts perfusion features through a multiview learning process to differentiate HCC from other malignant conditions, showing strong potential for application in CEUS. Although CEUS is primarily utilized in the diagnostic evaluation of HCC, its high sensitivity for small lesions (≤2 cm) and early-stage HCC may offer potential utility in screening specific populations.
The integration of AI into ultrasound is limited by the inherently dynamic nature of image acquisition and interpretation. Early studies by Shmauch et al. (77) and Yamakawa et al. (78) demonstrated feasibility but were limited by imprecise lesion localization or required manual input, falling short of full automation. A significant advancement was achieved by Tiyarattanachai et al. (79), Xu et al. (80), and Rhyou et al. (81), who utilized video data and object detection models such as YOLOv5 to develop systems capable of fully automated interpretation, outperforming human operators in sensitivity and specificity. While these studies focus on diagnostic accuracy, detection speed—critical for capturing transient anatomical details in real-time scanning—remains underexplored. Furthermore, current automation is largely confined to image analysis; the physical scanning process still relies on human operators. Future research must pursue end-to-end automation, integrating AI-driven robotic scanning. Such a system would drastically reduce costs, increase accessibility in remote areas, and improve early detection rates. To ensure reliability, AI interpretation should be coupled with a secondary verification AI model, with human arbitration for discrepant results, establishing a robust, supervised screening workflow. In cases of discrepancy, human intervention is required for further identification. Given the success of fully automated ultrasound systems, we anticipate the development and validation of fully automated CEUS.
2.5 CT
While contrast-enhanced CT is predominantly considered a diagnostic modality, its evolving role within the HCC screening-to-diagnosis continuum is increasingly recognized as crucial. In a comprehensive HCC screening program, CT fulfills two critical functions: (1) serving as a secondary or confirmatory screening tool for high-risk individuals with suboptimal ultrasonography results or elevated clinical suspicion (82, 83), and (2) providing a rapid triage mechanism for the characterization of indeterminate nodules identified during initial ultrasound surveillance (82). The integration of AI into CT workflows substantially enhances the feasibility and accuracy of its screening-related applications, potentially reducing time-to-diagnosis and optimizing resource utilization. Studies indicate that deep learning models utilizing multiphase contrast-enhanced CT and clinical data can effectively extract and integrate radiomic features relevant to diagnosis, achieving diagnostic performance comparable to expert consensus. When deployed as an assistive tool, these models have been shown to improve clinician sensitivity in detecting tumors such as ICC (82). Furthermore, deep learning algorithms trained on contrast-enhanced CT images have demonstrated success in predicting microvascular invasion and histopathological grade in HCC, enabling non-invasive assessment of tumor biology (83). Therefore, CT not only remains a cornerstone for diagnosis and treatment response prediction but, through AI integration, also holds significant promise for screening and early risk stratification.
The sensitivity of CT alone for detecting HCC lesions is 65%, which decreases to 40% for lesions <2 cm (84). By automatically capturing high-level imaging features, AI enhances the diagnostic accuracy for HCC and streamlines the diagnostic workflow (85). Lysdahlgaard et al. (86) employed machine learning algorithms to analyze and classify radiomic features derived from both hepatic tumors and normal liver tissue in a limited CT dataset, achieving classification accuracies ranging from 94% to 100%. Şahin et al. (87) developed a deep learning model using CT images from 122 patients, which achieved exceptional diagnostic accuracy (95.35%) for HCC and significantly outperformed conventional diagnostic methods. Wang et al. (88) employed an AI-based methodology utilizing early CT radiomics features to predict hyperspectral imaging characteristics, enabling non-invasive tumor prediction and early screening with >90% classification accuracy. Owing to overlapping imaging features between small HCC (≤2 cm) and benign cirrhotic nodules, the direct application of deep learning algorithms to multiparametric images for small HCC detection remains challenging. Zheng et al. (89) developed an integrated pattern-matching and deep learning model that achieved a sensitivity of 89.74% and a positive predictive value of 85.00% in small HCC diagnosis.
Recently, AI has also been employed to detect subtle patterns on non-contrast CT scans that are imperceptible to human observers, enabling highly sensitive and specific cancer detection. Cao et al. (90) developed a deep learning model that identified pancreatic cancer on non-contrast CT with 92.9% sensitivity and 99.9% specificity, significantly outperforming radiologists. Similarly, Peng et al. (91) created an AI algorithm for HCC detection via non-contrast CT, achieving an AUC value of 0.789. The development of accurate AI algorithms for HCC detection enables the application of non-contrast CT for opportunistic screening by reducing both radiation exposure and costs. The combination of non-contrast CT and AI plays a vital role in screening special populations, such as patients with contrast agent allergies.
The clinical implementation of AI-assisted CT models faces several limitations, such as independence from high-quality imaging data and heterogeneity in scanning protocols across institutions. Notably, the false-negative rate for detecting HCCs < 1 cm remains substantial. Advancing the detection of subcentimeter HCCs through combined CT and AI approaches represents a critical research priority to reduce interval tumor progression. While CT offers high diagnostic accuracy and remains a cornerstone in HCC diagnosis, its inherent limitations—such as ionizing radiation exposure and considerable cost—preclude its adoption for routine surveillance. Nevertheless, AI-enhanced CT holds distinct, actionable value within a risk-stratified screening paradigm. Specifically, it can serve as (1) a problem-solving modality for indeterminate ultrasound findings, (2) a high-precision screening tool for selected high-risk subgroups, and (3) an opportunistic screening resource when non-contrast CT is performed for other indications. Future screening algorithms may strategically integrate AI-CT at key decision points to enhance the efficiency and accuracy of the screening cascade, thereby transcending the conventional dichotomy between screening and diagnostic imaging.
3 Challenges of AI applications in HCC screening
3.1 Limited generalizability of the models
A key limitation of AI applications in HCC screening is the poor generalizability of models, primarily due to interpatient variability and the heterogeneous etiologies of HCC. The majority of studies have been performed in patients with chronic viral hepatitis B. However, metabolic-associated liver disease has become the most common cause of liver disease worldwide and is the leading cause of non-cirrhotic HCC, with approximately 25%–30% of patients developing HCC in this context (92, 93). This change has not only challenged the diagnostic capabilities of AI models, as previously validated radiographic systems are not applicable in this different cohort, but has also challenged HCC screening and surveillance practices. Furthermore, overfitting is also a main reason why a model demonstrates high accuracy on the training data but fails to generalize to unseen test datasets (94). This occurs when the model memorizes dataset-specific noise and patterns that are not generalizable, a situation often exacerbated by limited training data. Moreover, modest sample sizes also increase the risk of overfitting. To mitigate overfitting and enhance the generalization performance, techniques such as input feature reduction and increased regularization can be employed (Dropout: a simple way to prevent neural networks from overfitting). AI models from academic prototypes to clinical implementation need to be as rigorously assessed.
3.2 Ethical challenges
The key ethical issues central to ongoing ethical debates encompass data privacy and security, the assignment of moral accountability and liability for AI errors, and the mitigation of inherent biases and algorithmic inequalities within AI-driven HCC screening.
A trusted data space is critically important for ensuring privacy, data security, and responsible data processing and storage. The establishment of such infrastructure not only accelerates the flow of data among data providers, service providers, and data users in AI applications but also ensures privacy and security in data usage. The National Data Administration of China has issued the Trusted Data Space Development Action Plan (2024–2028) (95), which outlines a structured framework to promote the development of trusted data spaces at the corporate, industrial, municipal, individual, and cross-border levels. This vision aligns closely with the AI-based HCC screening and surveillance strategy proposed in our work. By leveraging trusted data space platforms, it becomes feasible to develop dedicated AI systems for disease screening and monitoring. Primary care institutions can act as both data providers and end-users. Municipal and international data spaces can facilitate secure data transmission and storage while also supporting the training, maintenance, updating, and oversight of AI models. This infrastructure provides a robust foundation for scalable and ethically aligned AI-powered disease screening. Additionally, cancer diagnosis often involves complex ethical issues. Whether to disclose the diagnosis to the patient is a critical consideration in clinical practice, particularly for instant testing modalities such as ultrasound examinations. It is recommended that AI systems use ambiguous terms—such as “nodule” or “space-occupying lesion”—when interpreting reports and advise further hospital-based diagnostic assessment to avoid potential ethical complications.
The ethical implications, accountability, and attribution of responsibility in cases of errors committed by AI systems should be considered. Compared with reactions to comparable human errors, public responses to errors by AI systems tend to be more pronounced. Moreover, failures in a single AI system are often overgeneralized to imply systemic deficiencies across all AI technologies. As a result, the ethical consequences and erosion of trust attributable to AI errors can be substantially amplified (96). In light of these challenges, establishing a clear chain of accountability for AI and robotic systems, with explicitly defined roles and responsibilities for manufacturers, healthcare institutions, and medical practitioners, is essential. An incident reporting mechanism should be implemented to enable healthcare providers to promptly document errors and adverse events related to AI applications.
Bias in AI systems largely originates from the underlying data used to train these models (97, 98). Most current AI tools are developed on the basis of historical institutional datasets, which makes them prone to inheriting biases resulting from imbalanced representations of different demographic groups in disease datasets (99). For example, an AI tool designed to predict healthcare costs, when trained primarily on a White population, may perform poorly when applied to African American patients (100). Regional and ethnic heterogeneity in HCC risk factors introduces significant complexity to the development and validation of AI models for screening, challenging their performance stability and fairness across diverse populations. Globally, epidemiological patterns vary substantially. Asia bears approximately 70% of the worldwide HCC burden, with HBV infection representing the predominant etiological factor (101). Similarly, HBV remains the leading cause in Africa (102). In Europe, risk profiles display distinct regional patterns: hepatitis C virus (HCV) is the primary contributor in Western Europe, whereas ALD accounts for nearly half of HCC cases in Central and Eastern Europe. Overall, HBV is responsible for only approximately 16% of European cases (103, 104).
Therefore, rather than adopting a one-size-fits-all general model, a more refined strategy—developing population-specific AI models—may more effectively mitigate bias while preserving important biological differences. This approach entails well-annotated data collection, rigorous data curation and monitoring during the data acquisition phase, as well as the implementation of tailored modeling strategies and clear recommendations regarding the scope of application during the model design and deployment stages.
Disparities in the number of AI models across countries with different income levels contribute to AI inequality (105). Typically, only researchers in well-resourced settings are able to generate or collect sufficient data and implement novel AI methodologies (106, 107). Algorithms trained on data from high-income countries are deployed in real-world contexts in low- and middle-income countries (LMICs). However, without adequate consideration of the unique local conditions and needs of LMICs, these systems may lack generalizability and broad effectiveness (108, 109). Moreover, infrastructural limitations in LMICs restrict the applicability of AI models, further exacerbating AI inequality. Recent empirical evidence indicates that variations in technical parameters significantly influence the performance of AI algorithms developed from medical imaging data (110–113). For example, AI models developed on the basis of 3T MRI equipment may not be suitable for many African nations and other LMICs, where scanners typically operate at ≤1.5 T. Finally, differences in available data modalities across income settings also contribute to AI inequality. To promote more equitable AI, findings from Jenny Yang et al. (114) were obtained suggest that transfer learning and fine-tuning models with LMIC data can enhance the generalization capabilities of AI systems. Nonetheless, the study emphasizes the importance of developing resilient AI tools specifically designed for healthcare systems in LMICs. Therefore, greater attention should be given to creating algorithms compatible with more accessible data types and increasing the number of AI models trained on data from LMICs, thereby improving the applicability and relevance of AI technologies in these settings (115).
4 Translational feasibility and implementation pathways
Based on the synthesized evidence, we evaluate the translational readiness of various AI applications and their enabling infrastructures for HCC screening, focusing on technical validation, integration potential, and preliminary implementation evidence.
4.1 Most promising for near-term adoption: AI-assisted ultrasound detection and risk stratification tools
AI-assisted ultrasound: This approach constitutes the most direct and pragmatically viable pathway for near-term integration. Substantial evidence indicates that AI can significantly enhance the sensitivity of HCC detection by both expert and non-expert operators during real-time or retrospective image analysis (71, 72). Its principal advantages include the following: (1) Workflow compatibility augments the established first-line screening standard (biannual ultrasound) without necessitating fundamental changes to the clinical care pathway. (2) Targeting a critical gap: It directly addresses the well-recognized operator-dependent variability of ultrasonography, a major impediment to screening effectiveness. (3) Phased deployment: Implementation can commence with AI serving as a “second reader” or decision-support system, thereby minimizing workflow disruption and facilitating clinical adoption and trust.
Clinical data-driven AI risk stratification models: These models integrate readily available clinical and laboratory variables to calculate individual HCC risk, representing a highly implementable tool (38). Their key strengths are as follows: (1) Low incremental resource burden: They leverage data routinely captured within electronic health records. (2) Actionable clinical output: They enable personalized screening strategies, such as intensifying surveillance for high-risk patients or safely extending intervals for low-risk individuals, thereby optimizing resource allocation. (3) Streamlined regulatory pathway: Classified as Software as a Medical Device (SaMD) that does not involve novel hardware or diagnostic procedures, these models may navigate a relatively more defined regulatory approval process.
4.2 Promising but with specific hurdles: integrated serum biomarker panels, CT-based radiomics and trusted data spaces
AI-discovered/integrated serum biomarker panels: Panels such as the 7-autoantibody assay or models combining AFP with other markers have demonstrated superior diagnostic performance compared to AFP alone (33, 34). Their feasibility hinges on: (1) Standardization and regulatory approval: Transitioning from research-based assays to standardized, FDA/CE-approved in vitro diagnostics represents a critical development step. (2) Demonstration of clinical utility: Prospective trials are needed to confirm that their application improves early detection rates and patient-relevant outcomes compared to current surveillance standards.
CT-based radiomics/AI models: Although these models achieve high diagnostic accuracy, their application in population-wide screening is constrained by the inherent limitations of CT (86, 87). Their near-term implementation is more feasible in targeted contexts, such as surveillance of subpopulations with indeterminate nodules or as a triage tool following inconclusive ultrasound findings.
Trusted data spaces: The establishment of secure, interoperable, and ethically governed data platforms is a foundational enabler for advanced AI development. While not a direct screening tool, creating such trusted data spaces at institutional or regional levels is a tangible near-to-mid-term objective but faces distinct non-technical challenges. These include defining data sovereignty and access frameworks, ensuring compliance across legal jurisdictions, and establishing viable governance and sustainability models. Although pilot projects in related fields have demonstrated technical feasibility—and initiatives such as China’s plan to establish over 100 trusted data spaces by 2028 aim to provide solutions and best practices—their broad implementation for HCC screening will require coordinated policy alignment and dedicated investment.
4.3 Future horizons: liquid biopsy with AI, fully automated ultrasound systems, and global AI-integrated surveillance ecosystems
AI-enhanced liquid biopsy (cfDNA/EVs): Despite promising proof-of-concept studies (60–62), detecting the low circulating tumor DNA fraction in early-stage HCC remains technically challenging. Broader clinical adoption awaits the development of more robust, cost-effective, and standardized detection platforms.
A mid- to long-term objective: While conceptually promising for expanding screening access, such systems require substantial advances in robotic scanning technology and real-time AI interpretation of dynamic image data. Their development and rigorous clinical validation represent a mid- to long-term objective.
Global AI-integrated surveillance ecosystems: The vision of a globally interoperable AI-driven surveillance network represents a long-term, transformative goal. Its realization depends on the maturation and integration of multiple components: validated AI tools, secure and scalable data infrastructures, and automated screening technologies. Key challenges include ensuring technical interoperability across diverse AI systems, harmonizing international clinical and regulatory standards, and addressing global equity to ensure accessibility in low-resource settings. Progress will depend less on technological feasibility and more on sustained international collaboration, governance alignment, and political commitment.
The integration of AI into HCC screening is expected to follow a staged pathway. In the near term, prioritizing the implementation of AI-augmented ultrasound and clinical risk stratification tools within existing workflows offers the most pragmatic and impactful strategy. Success in these initial applications will establish essential clinical trust and infrastructural foundations to support the future adoption of more advanced, multimodal AI systems.
5 Future perspectives
5.1 Risk-stratified HCC screening
The developmental pathogenesis of HCC remains incompletely understood and exhibits pronounced heterogeneity across ethnicities and etiologies. Current screening paradigms, which predominantly target broad populations, may lack the precision offered by risk-stratified strategies capable of optimizing early detection in high-risk subgroups. Indeed, the risk of HCC is not uniform; it is substantially elevated in patients with cirrhosis, who face an annual incidence of 2%–4% (116). Consequently, both the AASLD and EASL guidelines endorse routine surveillance for this demographic, an approach proven to improve overall survival (117). With the evolving landscape of HCC etiology, MASLD and alcohol-associated liver disease are projected to become major sources of HCC in the future. The large population base and relatively low incidence rate associated with these conditions impose a substantial burden and create significant barriers to effective screening and surveillance. Consequently, there is a pressing need to develop and validate AI-assisted diagnostic models for HCC that are applicable across diverse clinical scenarios. Risk-stratified models tailored to distinct high-risk populations should be implemented. For individuals at lower risk levels, fully automated end–end ultrasound should serve as the foundational screening and surveillance modality, supplemented by clinical features such as age, sex, and liver disease status, to guide specific screening strategies. As risk levels increase, the protocol should selectively incorporate additional tests—such as serum biomarkers, liquid biopsy, gut microbiome analysis, CT, and MRI—based on patient stratification to increase diagnostic accuracy. For individuals at intermediate to high risk (e.g., cirrhosis with prior indeterminate lesions), screening may commence with or quickly escalate to AI-assisted CT, balancing the higher sensitivity for small lesions against the increased cost and radiation exposure. This targeted approach maximizes the value of CT within a screening context. This stratified approach is expected to significantly improve screening efficiency, enhance cost-effectiveness, optimize the allocation of finite healthcare resources, and ultimately improve HCC outcomes.
The establishment of an AI-based precision screening system for HCC is proposed (Figure 3). From an organizational perspective, a multitiered AI infrastructure spanning primary care and municipal, national, and global levels should be constructed. At the primary care level, AI should assist healthcare providers in risk stratification, formulation and interpretation of stratum-specific screening and surveillance strategies, and patient education. At the municipal level, trusted data spaces should be established and expanded to host AI-assisted screening and surveillance systems, which are responsible for local model maintenance, human oversight, and data storage. At the national or regional level, AI-based platforms should be developed to integrate and analyze regional data, train and update models, and implement operational protocols and ethical frameworks tailored to local contexts. At the global level, an international platform should be established to issue guidelines, aggregate compatible AI models, and enhance generalizability across populations.
Figure 3. Multi-tiered architecture for implementing an AI-assisted precision surveillance strategy. The schematic depicts a hierarchical framework spanning from primary care to global levels. At the primary care level, AI supports risk stratification, strategy formulation, and patient education. Municipal levels host trusted data spaces for system maintenance and oversight. National/regional platforms are responsible for data integration, model training, and implementing context-specific protocols. A global tier issues guidelines and aggregates compatible AI models to enhance generalizability, collectively enabling a scalable and ethically governed precision screening ecosystem. CT, computed tomography; MRI, magnetic resonance imaging; AI, artificial intelligence.
5.2 Accessible HCC screening
The majority of patients do not adhere to recommended HCC screening strategies. Davila et al. (118) reported that among patients diagnosed with HCC, no more than 28% had undergone at least one ultrasound examination within the 3 years preceding diagnosis. The same research group, in a study involving a larger at-risk population, also confirmed low surveillance rates, with only 12% and 59% of patients receiving consistent and inconsistent surveillance, respectively. Both studies were conducted in a developed country, and the situation is likely more severe in developing or resource-limited regions (119). This may partly explain the current low rate of early HCC diagnosis, underscoring the urgent need for improved screening strategies. Some studies have employed methods such as mailed outreach and primary care-based clinical reminders to increase the degree of participation in HCC screening, with positive outcomes reported (120, 121). However, these approaches are not cost effective and rely heavily on patient health literacy, making them difficult to implement in underserved areas.
Therefore, we propose the development of a fully automated end-to-end ultrasound examination system. This strategy would involve primary healthcare providers or community health workers as the main actors in HCC screening, implementing a grid-based approach to cover at-risk populations within their jurisdictions. Using ultrasound as the foundational modality, these primary care personnel—who have the most frequent contact with high-risk groups—can conduct screenings more efficiently than specialists, whose coverage is often broader but less focused. The implementation of a fully automated ultrasound process would significantly reduce technical personnel costs, allow flexible scheduling for patients, and greatly decrease the cost per screening session. Consequently, the primary role of grassroots healthcare workers would shift to reminding and supervising the effective execution of HCC screening, thereby alleviating their workload. This approach is expected to not only increase the coverage of HCC screening but also reduce the overall cost of screening.
5.3 Cost-effective HCC screening
Cost-effectiveness is a critical determinant for the practical application of AI in clinical settings. Multiple studies have demonstrated that screening and surveillance for HCC in high-risk populations, particularly among patients with cirrhosis, is cost-effective (122–124). However, other research has indicated that in countries with a low incidence of HCC, such screening may not be cost-effective (125). In contrast to risk-stratified monitoring, Kao et al. demonstrated that a precision surveillance strategy represents the most cost-effective approach to HCC monitoring. This strategy integrates four patient-specific factors within risk-defined cohorts and evaluates four potential surveillance protocols, thereby providing a framework for individualized monitoring. Future surveillance stratification and strategies should adopt greater flexibility to accommodate diverse population characteristics. Furthermore, the integration of AI into HCC screening is anticipated to increase cost-effectiveness (126).
6 Conclusion
AI technologies are transforming HCC screening paradigms. These innovative approaches leverage comprehensive data mining from medical records, either as stand-alone tools or in combination with conventional serological markers and imaging modalities, to achieve superior diagnostic sensitivity and specificity, significantly reducing the technical barriers associated with operator dependency and image interpretation variability. This synergy may establish real-time AI-assisted HCC screening as a groundbreaking paradigm in early cancer detection. These advancements fundamentally enhance HCC management while aligning with global health priorities to reduce cancer mortality in resource-limited settings. However, owing to the current limitations of AI models in HCC screening due to their poor generalizability and ethical issues, the clinical implementation of AI models poses substantial challenges. Before AI-assisted HCC detection can be incorporated into routine screening protocols, it must undergo rigorous validation in large-scale, high-quality multicenter studies. Furthermore, a critical resolution of the accompanying ethical dilemmas is imperative.
Author contributions
J-XR: Software, Conceptualization, Writing – original draft, Formal analysis, Writing – review & editing, Methodology. YL: Methodology, Conceptualization, Writing – review & editing, Formal analysis, Writing – original draft. KL: Writing – original draft, Conceptualization, Methodology, Supervision.
Funding
The author(s) declared that financial support was not received for this work and/or its publication.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmed.2025.1713887/full#supplementary-material
Abbreviations
HCC, hepatocellular carcinoma; AI, artificial intelligence; ICC, intrahepatic cholangiocarcinoma; HCC-ICC, hepatocellular-cholangiocarcinoma; CT, computed tomography; MRI, magnetic resonance imaging; ANNs, artificial neural networks; CNN, convolutional neural network; AFP, Alpha-fetoprotein; GEO, Gene Expression Omnibus; TCGA, the Cancer Genome Atlas; AAbs, autoantibodies; DCP, des-gamma-carboxyprothrombin; SVR, sustained virologic response; GGT, γ-glutamyl transpeptidas; HT-1, hereditary tyrosinemia type 1; ALT, Alanine Aminotransferase; SERS, surface-enhanced Raman spectroscopy; NBC, nanoplasmonic biosensing chip; DNN, deep neural network; SVM, support vector machine; CTCs, circulating tumor cells; cfDNA, cell-free DNA; EVs, extracellular vesicles; LncRNA, long non-coding RNA; HBV, hepatitis B virus; HCV, hepatitis C virus; EASL, European Association for the Study of the Liver; AASLD, American Association for the Study of Liver Diseases; MASLD, metabolic dysfunction-associated steatotic liver disease; FLLs, focal liver lesions; CEUS, contrast-enhanced ultrasound; AUC, area under the curve; LMICs, low and middle income countries; ALD, alcohol-related liver disease; SaMD, Software as a Medical Device.
References
1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 Cancers in 185 countries. CA Cancer J Clin. (2021) 71:209–49. doi: 10.3322/caac.21660
2. Rumgay H, Ferlay J, de Martel C, Georges D, Ibrahim AS, Zheng R, et al. Global, regional and national burden of primary liver cancer by subtype. Eur J Cancer. (2022) 161:108–18. doi: 10.1016/j.ejca.2021.11.023
3. European Association for the Study of the Liver EASL Clinical Practice Guidelines on the management of hepatocellular carcinoma. J Hepatol. (2025) 82:315–74. doi: 10.1016/j.jhep.2024.08.028
4. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. (2023) 73:17–48. doi: 10.3322/caac.21763
5. Tsuchiya N, Sawada Y, Endo I, Saito K, Uemura Y, Nakatsura T. Biomarkers for the early diagnosis of hepatocellular carcinoma. World J Gastroenterol. (2015) 21:10573–83. doi: 10.3748/wjg.v21.i37.10573
6. European Association for the Study of the Liver. EASL clinical practice guidelines: management of hepatocellular carcinoma. J Hepatol. (2018) 69:182–236. doi: 10.1016/j.jhep.2018.03.019
7. Harris PS, Hansen RM, Gray ME, Massoud OI, McGuire BM, Shoreibah MG. Hepatocellular carcinoma surveillance: an evidence-based approach. World J Gastroenterol. (2019) 25:1550–9. doi: 10.3748/wjg.v25.i13.1550
8. Zhang W, Chao S, Chen S, Rao H, Huang R, Wei L, et al. Awareness and knowledge of nonalcoholic fatty liver disease among office employees in Beijing. China. Dig Dis Sci. (2019) 64:708–17. doi: 10.1007/s10620-018-5389-5
9. Rodriguez Villalvazo Y, McDanel JS, Beste LA, Sanchez AJ, Vaughan-Sarrazin M, Katz DA. Effect of travel distance and rurality of residence on initial surveillance for hepatocellular carcinoma in VA primary care patient with cirrhosis. Health Serv Res. (2020) 55:103–12. doi: 10.1111/1475-6773.13241
10. Singal AG, Tiro JA, Murphy CC, Blackwell JM, Kramer JR, Khan A, et al. Patient-Reported Barriers Are Associated With Receipt of Hepatocellular Carcinoma Surveillance in a Multicenter Cohort of Patients With Cirrhosis. Clin Gastroenterol Hepatol. (2021) 19:987–95.e1. doi: 10.1016/j.cgh.2020.06.049.
11. Simmons OL, Feng Y, Parikh ND, Singal AG. Primary care provider practice patterns and barriers to hepatocellular carcinoma surveillance. Clin Gastroenterol Hepatol. (2019) 17:766–73. doi: 10.1016/j.cgh.2018.07.029
12. Teerasarntipan T, Sritunyarat Y, Prathyajuta P, Pitakkitnukun P, Phathong C, Ariyaskul D, et al. Physician- and patient-reported barriers to hepatocellular carcinoma surveillance: a nationwide survey. Medicine. (2022) 101:e30538. doi: 10.1097/md.0000000000030538
13. Vitale A, Svegliati-Baroni G, Ortolani A, Cucco M, Dalla Riva GV, Giannini EG, et al. Epidemiological trends and trajectories of MAFLD-associated hepatocellular carcinoma 2002-2033: the ITA.LI.CA database. Gut. (2023) 72:141–52. doi: 10.1136/gutjnl-2021-324915
14. Riazi K, Azhari H, Charette JH, Underwood FE, King JA, Afshar EE, et al. The prevalence and incidence of NAFLD worldwide: a systematic review and meta-analysis. Lancet Gastroenterol Hepatol. (2022) 7:851–61. doi: 10.1016/s2468-1253(22)00165-0
15. Orci LA, Sanduzzi-Zamparelli M, Caballol B, Sapena V, Colucci N, Torres F, et al. Incidence of hepatocellular carcinoma in patients with nonalcoholic fatty liver disease: a systematic review, meta-analysis, and meta-regression. Clin Gastroenterol Hepatol. (2022) 20:283–92.e10. doi: 10.1016/j.cgh.2021.05.002.
16. Cucchetti A, Cescon M, Erroi V, Pinna AD. Cost-effectiveness of liver cancer screening. Best Pract Res Clin Gastroenterol. (2013) 27:961–72. doi: 10.1016/j.bpg.2013.08.021
17. Del Poggio P, Olmi S, Ciccarese F, Di Marco M, Rapaccini GL, Benvegnù L, et al. Factors that affect efficacy of ultrasound surveillance for early stage hepatocellular carcinoma in patients with cirrhosis. Clin Gastroenterol Hepatol. (2014) 12:1927–33.e2. doi: 10.1016/j.cgh.2014.02.025.
18. Mintz Y, Brodie R. Introduction to artificial intelligence in medicine. Minim Invasive Ther Allied Technol. (2019) 28:73–81. doi: 10.1080/13645706.2019.1575882
19. Chierici A, Lareyre F, Iannelli A, Salucki B, Goffart S, Guzzi L, et al. Applications of artificial intelligence in liver cancer: a scoping review. Artif Intell Med. (2025) 169:103244. doi: 10.1016/j.artmed.2025.103244
20. Chaiteerakij R, Ariyaskul D, Kulkraisri K, Apiparakoon T, Sukcharoen S, Chaichuen O, et al. Artificial intelligence for ultrasonographic detection and diagnosis of hepatocellular carcinoma and cholangiocarcinoma. Sci Rep. (2024) 14:20617. doi: 10.1038/s41598-024-71657-z
21. Chatzipanagiotou OP, Loukas C, Vailas M, Machairas N, Kykalos S, Charalampopoulos G, et al. Artificial intelligence in hepatocellular carcinoma diagnosis: a comprehensive review of current literature. J Gastroenterol Hepatol. (2024) 39:1994–2005. doi: 10.1111/jgh.16663
22. Lu RF, She CY, He DN, Cheng MQ, Wang Y, Huang H, et al. AI enhanced diagnostic accuracy and workload reduction in hepatocellular carcinoma screening. NPJ Digit Med. (2025) 8:500. doi: 10.1038/s41746-025-01892-9
23. Cohen JD, Li L, Wang Y, Thoburn C, Afsari B, Danilova L, et al. Detection and localization of surgically resectable cancers with a multi-analyte blood test. Science. (2018) 359:926–30. doi: 10.1126/science.aar3247
24. Galle PR, Foerster F, Kudo M, Chan SL, Llovet JM, Qin S, et al. Biology and significance of alpha-fetoprotein in hepatocellular carcinoma. Liver Int. (2019) 39:2214–29. doi: 10.1111/liv.14223
25. Giannini EG, Marenco S, Borgonovo G, Savarino V, Farinati F, Del Poggio P, et al. Alpha-fetoprotein has no prognostic role in small hepatocellular carcinoma identified during surveillance in compensated cirrhosis. Hepatology. (2012) 56:1371–9. doi: 10.1002/hep.25814
26. Singal AG, Haaland B, Parikh ND, Ozbay AB, Kirshner C, Chakankar S, et al. Comparison of a multitarget blood test to ultrasound and alpha-fetoprotein for hepatocellular carcinoma surveillance: results of a network meta-analysis. Hepatol Commun. (2022) 6:2925–36. doi: 10.1002/hep4.2045
27. Llovet JM, Montal R, Sia D, Finn RS. Molecular therapies and precision medicine for hepatocellular carcinoma. Nat Rev Clin Oncol. (2018) 15:599–616. doi: 10.1038/s41571-018-0073-4
28. Zhang R, Ye J, Huang H, Du X. Mining featured biomarkers associated with vascular invasion in HCC by bioinformatics analysis with TCGA RNA sequencing data. Biomed Pharmacother. (2019) 118:109274. doi: 10.1016/j.biopha.2019.109274
29. Camaggi CM, Zavatto E, Gramantieri L, Camaggi V, Strocchi E, Righini R, et al. Serum albumin-bound proteomic signature for early detection and staging of hepatocarcinoma: sample variability and data classification. Clin Chem Lab Med. (2010) 48:1319–26. doi: 10.1515/cclm.2010.248
30. Yang T, Huang L, He J, Luo L, Guo W, Chen H, et al. Establishment of diagnostic model and identification of diagnostic markers between liver cancer and cirrhosis based on multi-chip and machine learning. Clin Exp Pharmacol Physiol. (2024) 51:e13907. doi: 10.1111/1440-1681.13907
31. Wu J, Wang P, Han Z, Li T, Yi C, Qiu C, et al. A novel immunodiagnosis panel for hepatocellular carcinoma based on bioinformatics and the autoantibody-antigen system. Cancer Sci. (2022) 113:411–22. doi: 10.1111/cas.15217
32. Chen C, Xie Z, Ni Y, He Y. Screening immune-related blood biomarkers for DKD-related HCC using machine learning. Front Immunol. (2024) 15:1339373. doi: 10.3389/fimmu.2024.1339373
33. Zhang S, Liu Y, Chen J, Shu H, Shen S, Li Y, et al. Autoantibody signature in hepatocellular carcinoma using seromics. J Hematol Oncol. (2020) 13:85. doi: 10.1186/s13045-020-00918-x
34. Sato M, Morimoto K, Kajihara S, Tateishi R, Shiina S, Koike K, et al. Machine-learning approach for the development of a novel predictive model for the diagnosis of hepatocellular carcinoma. Sci Rep. (2019) 9:7704. doi: 10.1038/s41598-019-44022-8
35. Nahon P, Layese R, Bourcier V, Cagnot C, Marcellin P, Guyader D, et al. Incidence of hepatocellular carcinoma after direct antiviral therapy for HCV in patients with cirrhosis included in surveillance programs. Gastroenterology. (2018) 155:1436–50.e6. doi: 10.1053/j.gastro.2018.07.015.
36. Beste LA, Leipertz SL, Green PK, Dominitz JA, Ross D, Ioannou GN. Trends in burden of cirrhosis and hepatocellular carcinoma by underlying liver disease in US veterans, 2001-2013. Gastroenterology. (2015) 149:1471–82.e5; quiz e17–8. doi: 10.1053/j.gastro.2015.07.056.
37. Zhou Y, McArdle JJ. Rationale and applications of survival tree and survival ensemble methods. Psychometrika. (2015) 80:811–33. doi: 10.1007/s11336-014-9413-1
38. Audureau E, Carrat F, Layese R, Cagnot C, Asselah T, Guyader D, et al. Personalized surveillance for hepatocellular carcinoma in cirrhosis - using machine learning adapted to HCV status. J Hepatol. (2020) 73:1434–45. doi: 10.1016/j.jhep.2020.05.052
39. Fuenzalida K, Leal-Witt MJ, Acevedo A, Muñoz M, Gudenschwager C, Arias C, et al. Integrating machine learning and follow-up variables to improve early detection of hepatocellular carcinoma in tyrosinemia Type 1: a multicenter study. Int J Mol Sci. (2025) 26:3839. doi: 10.3390/ijms26083839
40. Jiao W, Atwal G, Polak P, Karlic R, Cuppen E, Danyi A, et al. A deep learning system accurately classifies primary and metastatic cancers using passenger mutation patterns. Nat Commun. (2020) 11:728. doi: 10.1038/s41467-019-13825-8
41. Cheng N, Fu J, Chen D, Chen S, Wang H. An antibody-free liver cancer screening approach based on nanoplasmonics biosensing chips via spectrum-based deep learning. NanoImpact. (2021) 21:100296. doi: 10.1016/j.impact.2021.100296
42. Cheng N, Lou B, Wang H. Discovering the digital biomarker of hepatocellular carcinoma in serum with SERS-based biosensors and intelligence vision. Colloids Surf B Biointerfaces. (2023) 226:113315. doi: 10.1016/j.colsurfb.2023.113315
43. Hu X, Wang Z, Chen H, Zhao A, Sun N, Deng C. Diagnosing, typing, and staging of renal cell carcinoma by designer matrix-based urinary metabolic analysis. Anal Chem. (2022) 94:14846–53. doi: 10.1021/acs.analchem.2c01563
44. Pham ATT, Wallace A, Zhang X, Tohl D, Fu H, Chuah C, et al. Optical-Based biosensors and their portable healthcare devices for detecting and monitoring biomarkers in body fluids. Diagnostics. (2021) 11:1285. doi: 10.3390/diagnostics11071285
45. De Matteis S, Bonafè M, Giudetti AM. Urinary metabolic biomarkers in Cancer patients: an overview. Methods Mol Biol. (2021) 2292:203–12. doi: 10.1007/978-1-0716-1354-2_18
46. Dawuti W, Zheng X, Liu H, Zhao H, Dou J, Sun L, et al. Urine surface-enhanced Raman spectroscopy combined with SVM algorithm for rapid diagnosis of liver cirrhosis and hepatocellular carcinoma. Photodiagnosis Photodyn Ther. (2022) 38:102811. doi: 10.1016/j.pdpdt.2022.102811
47. Dou J, Dawuti W, Zheng X, Zhang R, Zhou J, Lin R, et al. Urine fluorescence spectroscopy combined with machine learning for screening of hepatocellular carcinoma and liver cirrhosis. Photodiagnosis Photodyn Ther. (2022) 40:103102. doi: 10.1016/j.pdpdt.2022.103102
48. Shi F, Ning L, Sun N, Yao Q, Deng C. Multiscale structured trimetal oxide heterojunctions for urinary metabolic phenotype-dependent screening of early and small hepatocellular carcinoma. Small Methods. (2024) 8:e2301634. doi: 10.1002/smtd.202301634
49. Wei H, Yang T, Chen J, Duan T, Jiang H, Song B. Prognostic implications of CT/MRI LI-RADS in hepatocellular carcinoma: state of the art and future directions. Liver Int. (2022) 42:2131–44. doi: 10.1111/liv.15362
50. Dana J, Sutter O. Prognostic stratification in early-stage hepatocellular carcinoma: imaging biomarkers are needed. Liver Int. (2024) 44:881–3. doi: 10.1111/liv.15869
51. Sahani DV, Jiang T, Hayano K, Duda DG, Catalano OA, Ancukiewicz M, et al. Magnetic resonance imaging biomarkers in hepatocellular carcinoma: association with response and circulating biomarkers after sunitinib therapy. J Hematol Oncol. (2013) 6:51. doi: 10.1186/1756-8722-6-51
52. Shin H, Hur MH, Song BG, Park SY, Kim GA, Choi G, et al. AI model using CT-based imaging biomarkers to predict hepatocellular carcinoma in patients with chronic hepatitis B. J Hepatol. (2025) 82:1080–8. doi: 10.1016/j.jhep.2024.12.029
53. Hamann MF, Meyer D, Knüpfer S, Fuchs J, Jünemann KP, Naumann CM. Application of ultrasound imaging biomarkers (HistoScanning™) improves staging reliability of prostate biopsies. BMC Res Notes. (2017) 10:579. doi: 10.1186/s13104-017-2896-y
54. Lalumiere M, Larivière C, Nadeau MJ, Paquette P, Lamontagne M, Desmeules F, et al. Proposing a minimal data set of musculoskeletal ultrasound imaging biomarkers to inform clinical practice: an analysis founded on the achilles tendon. Ultrasound Med Biol. (2020) 46:2222–35. doi: 10.1016/j.ultrasmedbio.2020.04.024
55. Diehl F, Li M, Dressman D, He Y, Shen D, Szabo S, et al. Detection and quantification of mutations in the plasma of patients with colorectal tumors. Proc Natl Acad Sci U S A. (2005) 102:16368–73. doi: 10.1073/pnas.0507904102
56. Qu C, Wang Y, Wang P, Chen K, Wang M, Zeng H, et al. Detection of early-stage hepatocellular carcinoma in asymptomatic HBsAg-seropositive individuals by liquid biopsy. Proc Natl Acad Sci U S A. (2019) 116:6308–12. doi: 10.1073/pnas.1819799116
57. Grölz D, Hauch S, Schlumpberger M, Guenther K, Voss T, Sprenger-Haussels M, et al. Liquid biopsy preservation solutions for standardized pre-analytical workflows-venous whole blood and plasma. Curr Pathobiol Rep. (2018) 6:275–86. doi: 10.1007/s40139-018-0180-z
58. Schwarzenbach H, Hoon DS, Pantel K. Cell-free nucleic acids as biomarkers in cancer patients. Nat Rev Cancer. (2011) 11:426–37. doi: 10.1038/nrc3066
59. Liang Y, Lehrich BM, Zheng S, Lu M. Emerging methods in biomarker identification for extracellular vesicle-based liquid biopsy. J Extracell Vesicles. (2021) 10:e12090. doi: 10.1002/jev2.12090
60. Tao K, Bian Z, Zhang Q, Guo X, Yin C, Wang Y, et al. Machine learning-based genome-wide interrogation of somatic copy number aberrations in circulating tumor DNA for early detection of hepatocellular carcinoma. EBioMedicine. (2020) 56:102811. doi: 10.1016/j.ebiom.2020.102811
61. Samir A, Abdeldaim A, Mohammed A, Ali A, Alorabi M, Hussein MM, et al. Analysis of four long non-coding RNAs for hepatocellular carcinoma screening and prognosis by the aid of machine learning techniques. Sci Rep. (2024) 14:29582. doi: 10.1038/s41598-024-80926-w
62. Yu YP, Liu S, Geller D, Luo JH. Serum fusion transcripts to assess the risk of hepatocellular carcinoma and the impact of Cancer treatment through machine learning. Am J Pathol. (2024) 194:1262–71. doi: 10.1016/j.ajpath.2024.02.017
63. Razavi P, Li BT, Brown DN, Jung B, Hubbell E, Shen R, et al. High-intensity sequencing reveals the sources of plasma circulating cell-free DNA variants. Nat Med. (2019) 25:1928–37. doi: 10.1038/s41591-019-0652-7
64. Heitzer E, Haque IS, Roberts CES, Speicher MR. Current and future perspectives of liquid biopsies in genomics-driven oncology. Nat Rev Genet. (2019) 20:71–88. doi: 10.1038/s41576-018-0071-5
65. Ahn JC, Lee YT, Agopian VG, Zhu Y, You S, Tseng HR, et al. Hepatocellular carcinoma surveillance: current practice and future directions. Hepatoma Res. (2022) 8:10. doi: 10.20517/2394-5079.2021.131
66. Xie DY, Ren ZG, Zhou J, Fan J, Gao Q. 2019 Chinese clinical guidelines for the management of hepatocellular carcinoma: updates and insights. Hepatobiliary Surg Nutr. (2020) 9:452–63. doi: 10.21037/hbsn-20-480
67. Heimbach JK, Kulik LM, Finn RS, Sirlin CB, Abecassis MM, Roberts LR, et al. AASLD guidelines for the treatment of hepatocellular carcinoma. Hepatology. (2018) 67:358–80. doi: 10.1002/hep.29086
68. Kudo M, Ueshima K, Osaki Y, Hirooka M, Imai Y, Aso K, et al. B-Mode ultrasonography versus contrast-enhanced ultrasonography for surveillance of hepatocellular carcinoma: a prospective multicenter randomized controlled trial. Liver Cancer. (2019) 8:271–80. doi: 10.1159/000501082
69. Byra M, Styczynski G, Szmigielski C, Kalinowski P, Michałowski Ł, Paluszkiewicz R, et al. Transfer learning with deep convolutional neural network for liver steatosis assessment in ultrasound images. Int J Comput Assist Radiol Surg. (2018) 13:1895–903. doi: 10.1007/s11548-018-1843-2
70. Rhyou SY, Yoo JC. Aggregated micropatch-based deep learning neural network for ultrasonic diagnosis of cirrhosis. Artif Intell Med. (2023) 139:102541. doi: 10.1016/j.artmed.2023.102541
71. Tangruangkiat S, Chaiwongkot N, Pamarapa C, Rawangwong T, Khunnarong A, Chainarong C, et al. Diagnosis of focal liver lesions from ultrasound images using a pretrained residual neural network. J Appl Clin Med Phys. (2024) 25:e14210. doi: 10.1002/acm2.14210
72. Tiyarattanachai T, Apiparakoon T, Chaichuen O, Sukcharoen S, Yimsawad S, Jangsirikul S, et al. Artificial intelligence assists operators in real-time detection of focal liver lesions during ultrasound: a randomized controlled study. Eur J Radiol. (2023) 165:110932. doi: 10.1016/j.ejrad.2023.110932
73. Zhang WB, Hou SZ, Chen YL, Mao F, Dong Y, Chen JG, et al. Deep learning for approaching hepatocellular carcinoma ultrasound screening dilemma: identification of α-fetoprotein-negative hepatocellular carcinoma from focal liver lesion found in high-risk patients. Front Oncol. (2022) 12:862297. doi: 10.3389/fonc.2022.862297
74. Chernyak V, Fowler KJ, Do RKG, Kamaya A, Kono Y, Tang A, et al. LI-RADS: looking back. looking forward. Radiology. (2023) 307:e222801. doi: 10.1148/radiol.222801
75. Ding W, Meng Y, Ma J, Pang C, Wu J, Tian J, et al. Contrast-enhanced ultrasound-based AI model for multi-classification of focal liver lesions. J Hepatol. (2025) 83:426–39. doi: 10.1016/j.jhep.2025.01.011
76. Feng X, Cai W, Zheng R, Tang L, Zhou J, Wang H, et al. Diagnosis of hepatocellular carcinoma using deep network with multi-view enhanced patterns mined in contrast-enhanced ultrasound data. Eng Appl Art Intell. (2023) 118:105635. doi: 10.1016/j.engappai.2022.105635
77. Schmauch B, Herent P, Jehanno P, Dehaene O, Saillard C, Aubé C, et al. Diagnosis of focal liver lesions from ultrasound using deep learning. Diagn Interv Imag. (2019) 100:227–33. doi: 10.1016/j.diii.2019.02.009
78. Yamakawa M, Shiina T, Nishida N, Kudo M, eds Computer aided diagnosis system developed for ultrasound diagnosis of liver lesions using deep learning. in Proceedings of the 2019 IEEE International Ultrasonics Symposium (IUS). Piscataway, NJ: IEEE (2019).
79. Tiyarattanachai T, Apiparakoon T, Marukatat S, Sukcharoen S, Yimsawad S, Chaichuen O, et al. The feasibility to use artificial intelligence to aid detecting focal liver lesions in real-time ultrasound: a preliminary study based on videos. Sci Rep. (2022) 12:7749. doi: 10.1038/s41598-022-11506-z
80. Xu Y, Zheng B, Liu X, Wu T, Ju J, Wang S, et al. Improving artificial intelligence pipeline for liver malignancy diagnosis using ultrasound images and video frames. Brief Bioinform. (2023) 24:bbac569. doi: 10.1093/bib/bbac569
81. Rhyou SY, Yoo JC. Automated ultrasonography of hepatocellular carcinoma using discrete wavelet transform based deep-learning neural network. Med Image Anal. (2025) 101:103453. doi: 10.1016/j.media.2025.103453
82. Gao R, Zhao S, Aishanjiang K, Cai H, Wei T, Zhang Y, et al. Deep learning for differential diagnosis of malignant hepatic tumors based on multi-phase contrast-enhanced CT and clinical data. J Hematol Oncol. (2021) 14:154. doi: 10.1186/s13045-021-01167-2
83. He X, Xu Y, Zhou C, Song R, Liu Y, Zhang H, et al. Prediction of microvascular invasion and pathological differentiation of hepatocellular carcinoma based on a deep learning model. Eur J Radiol. (2024) 172:111348. doi: 10.1016/j.ejrad.2024.111348
84. Yu NC, Chaudhari V, Raman SS, Lassman C, Tong MJ, Busuttil RW, et al. CT and MRI improve detection of hepatocellular carcinoma, compared with ultrasound alone, in patients with cirrhosis. Clin Gastroenterol Hepatol. (2011) 9:161–7. doi: 10.1016/j.cgh.2010.09.017
85. Li X, Chen H, Qi X, Dou Q, Fu CW, Heng PA. H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes. IEEE Trans Med Imag. (2018) 37:2663–74. doi: 10.1109/tmi.2018.2845918
86. Lysdahlgaard S. Comparing radiomics features of tumour and healthy liver tissue in a limited CT dataset: a machine learning study. Radiography. (2022) 28:718–24. doi: 10.1016/j.radi.2022.03.015
87. Şahin E, Tatar OC, Ulutaş ME, Güler SA, Şimşek T, Turgay NZ, et al. Diagnostic performance of deep learning applications in hepatocellular carcinoma detection using computed tomography imaging. Turk J Gastroenterol. (2024) 36:124–30. doi: 10.5152/tjg.2024.24538
88. Wang X, Wang T, Zheng Y, Yin X. Recognition of liver tumors by predicted hyperspectral features based on patient’s computed tomography radiomics features. Photodiagnosis Photodyn Ther. (2023) 42:103638. doi: 10.1016/j.pdpdt.2023.103638
89. Zheng R, Wang L, Wang C, Yu X, Chen W, Li Y, et al. Feasibility of automatic detection of small hepatocellular carcinoma (≤2 cm) in cirrhotic liver based on pattern matching and deep learning. Phys Med Biol. (2021) 66:085014. doi: 10.1088/1361-6560/abf2f8
90. Cao K, Xia Y, Yao J, Han X, Lambert L, Zhang T, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. (2023) 29:3033–43. doi: 10.1038/s41591-023-02640-w
91. Peng C, Yu PLH, Lu J, Cheng HM, Shen XP, Chiu KW, et al. Opportunistic detection of hepatocellular carcinoma using noncontrast CT and deep learning artificial intelligence. J Am Coll Radiol. (2025) 22:249–59. doi: 10.1016/j.jacr.2024.12.011
92. Cotter TG, Rinella M. Nonalcoholic fatty liver disease 2020: the state of the disease. Gastroenterology. (2020) 158:1851–64. doi: 10.1053/j.gastro.2020.01.052
93. Huang DQ, El-Serag HB, Loomba R. Global epidemiology of NAFLD-related HCC: trends, predictions, risk factors and prevention. Nat Rev Gastroenterol Hepatol. (2021) 18:223–38. doi: 10.1038/s41575-020-00381-6
94. Gerussi A, Scaravaglio M, Cristoferi L, Verda D, Milani C, De Bernardi E, et al. Artificial intelligence for precision medicine in autoimmune liver disease. Front Immunol. (2022) 13:966329. doi: 10.3389/fimmu.2022.966329
95. National Data Administration. Trusted data space development action plan (2024–2028) (2024). Available online at: https://www.gov.cn/zhengce/zhengceku/202411/content_6996363.htm (accessed September 26, 2025).
96. Bonnefon J-F, Rahwan I, Shariff A. The moral psychology of artificial intelligence. Ann Rev Psychol. (2024) 75:653–75. doi: 10.1146/annurev-psych-030123-113559
97. European Society of Radiology [ESR] White paper on imaging biomarkers. Insights Imag. (2010) 1:42–5. doi: 10.1007/s13244-010-0025-8
98. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures. they are data. Radiology. (2016) 278:563–77. doi: 10.1148/radiol.2015151169
99. Celi LA, Cellini J, Charpignon ML, Dee EC, Dernoncourt F, Eber R, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities-A global review. PLoS Digit Health. (2022) 1:e0000022. doi: 10.1371/journal.pdig.0000022
100. Seyyed-Kalantari L, Zhang H, McDermott MBA, Chen IY, Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med. (2021) 27:2176–82. doi: 10.1038/s41591-021-01595-0
101. Goh GB, Chang PE, Tan CK. Changing epidemiology of hepatocellular carcinoma in Asia. Best Pract Res Clin Gastroenterol. (2015) 29:919–28. doi: 10.1016/j.bpg.2015.09.007
102. Kedar Mukthinuthalapati VVP, Sewram V, Ndlovu N, Kimani S, Abdelaziz AO, Chiao EY, et al. Hepatocellular carcinoma in sub-saharan africa. JCO Glob Oncol. (2021) 7:756–66. doi: 10.1200/go.20.00425
103. Konyn P, Ahmed A, Kim D. Current epidemiology in hepatocellular carcinoma. Expert Rev Gastroenterol Hepatol. (2021) 15:1295–307. doi: 10.1080/17474124.2021.1991792
104. Institute for Health Metrics and Evaluation (IHME). GBD compare data visualization. Seattle, WA: IHME, University of Washington (2025). Available online at: https://vizhub.healthdata.org/gbd-compare (accessed September 26, 2025).
105. Reddy H, Joshi S, Joshi A, Wagh V. A critical review of global digital divide and the role of technology in healthcare. Cureus. (2022) 14:e29739. doi: 10.7759/cureus.29739
106. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. (2022) 19:132–46. doi: 10.1038/s41571-021-00560-7
107. Shen Y, Shamout FE, Oliver JR, Witowski J, Kannan K, Park J, et al. Artificial intelligence system reduces false-positive findings in the interpretation of breast ultrasound exams. Nat Commun. (2021) 12:5645. doi: 10.1038/s41467-021-26023-2
108. Bulten W, Kartasalo K, Chen PC, Ström P, Pinckaers H, Nagpal K, et al. Artificial intelligence for diagnosis and Gleason grading of prostate Cancer: the PANDA challenge. Nat Med. (2022) 28:154–63. doi: 10.1038/s41591-021-01620-2
109. Diamant A, Chatterjee A, Vallières M, Shenouda G, Seuntjens J. Deep learning in head & neck cancer outcome prediction. Sci Rep. (2019) 9:2764. doi: 10.1038/s41598-019-39206-1
110. Cherezov D, Viswanathan VS, Fu P, Gupta A, Madabhushi A. Rank acquisition impact on radiomics estimation (AcquIRE) in chest CT imaging: a retrospective multi-site, multi-use-case study. Comput Methods Programs Biomed. (2024) 244:107990. doi: 10.1016/j.cmpb.2023.107990
111. Baxi V, Edwards R, Montalto M, Saha S. Digital pathology and artificial intelligence in translational medicine and clinical practice. Mod Pathol. (2022) 35:23–32. doi: 10.1038/s41379-021-00919-2
112. Mi H, Yuan M, Suo S, Cheng J, Li S, Duan S, et al. Impact of different scanners and acquisition parameters on robustness of MR radiomics features based on women’s cervix. Sci Rep. (2020) 10:20407. doi: 10.1038/s41598-020-76989-0
113. Yuan J, Xue C, Lo G, Wong OL, Zhou Y, Yu SK, et al. Quantitative assessment of acquisition imaging parameters on MRI radiomics features: a prospective anthropomorphic phantom study using a 3D-T2W-TSE sequence for MR-guided-radiotherapy. Quant Imag Med Surg. (2021) 11:1870–87. doi: 10.21037/qims-20-865
114. Yang J, Dung NT, Thach PN, Phong NT, Phu VD, Phu KD, et al. Generalizability assessment of AI models across hospitals in a low-middle and high income country. Nat Commun. (2024) 15:8270. doi: 10.1038/s41467-024-52618-6
115. Schwalbe N, Wahl B. Artificial intelligence and the future of global health. Lancet. (2020) 395:1579–86. doi: 10.1016/s0140-6736(20)30226-9
116. El-Serag HB. Hepatocellular carcinoma. N Engl J Med. (2011) 365:1118–27. doi: 10.1056/NEJMra1001683
117. Burke L, Hinkson A, Haghnejad V, Jones R, Parker R, Rowe IA. Hepatocellular carcinoma risk scores for non-viral liver disease: a systematic review and meta-analysis. JHEP Rep. (2025) 7:101227. doi: 10.1016/j.jhepr.2024.101227
118. Davila JA, Weston A, Smalley W, El-Serag HB. Utilization of screening for hepatocellular carcinoma in the United States. J Clin Gastroenterol. (2007) 41:777–82. doi: 10.1097/MCG.0b013e3180381560
119. Davila JA, Henderson L, Kramer JR, Kanwal F, Richardson PA, Duan Z, et al. Utilization of surveillance for hepatocellular carcinoma among hepatitis C virus-infected veterans in the United States. Ann Intern Med. (2011) 154:85–93. doi: 10.7326/0003-4819-154-2-201101180-00006
120. Singal AG, Tiro JA, Marrero JA, McCallister K, Mejias C, Adamson B, et al. Mailed outreach program increases ultrasound screening of patients with cirrhosis for hepatocellular carcinoma. Gastroenterology. (2017) 152:608–15.e4. doi: 10.1053/j.gastro.2016.10.042.
121. Beste LA, Ioannou GN, Yang Y, Chang MF, Ross D, Dominitz JA. Improved surveillance for hepatocellular carcinoma with a primary care-oriented clinical reminder. Clin Gastroenterol Hepatol. (2015) 13:172–9. doi: 10.1016/j.cgh.2014.04.033
122. Cadier B, Bulsei J, Nahon P, Seror O, Laurent A, Rosa I, et al. Early detection and curative treatment of hepatocellular carcinoma: a cost-effectiveness analysis in France and in the United States. Hepatology. (2017) 65:1237–48. doi: 10.1002/hep.28961
123. Parikh ND, Singal AG, Hutton DW, Tapper EB. Cost-Effectiveness of hepatocellular carcinoma surveillance: an assessment of benefits and harms. Am J Gastroenterol. (2020) 115:1642–9. doi: 10.14309/ajg.0000000000000715
124. Gurley T, Hernaez R, Cerda V, Thomas T, Narasimman M, Mittal S, et al. Cost-effectiveness of an outreach program for HCC screening in patients with cirrhosis: a microsimulation modeling study. EClinicalMedicine. (2025) 81:103113. doi: 10.1016/j.eclinm.2025.103113
125. Yuen MF, Lai CL. Screening for hepatocellular carcinoma: survival benefit and cost-effectiveness. Ann Oncol. (2003) 14:1463–7. doi: 10.1093/annonc/mdg400
Keywords: artificial intelligence, challenges, deep learning, future perspectives, hepatocellular carcinoma screening, machine learning, applications
Citation: Rao J-X, Li Y and Leng K (2026) Artificial intelligence in hepatocellular carcinoma screening: applications and challenges. Front. Med. 12:1713887. doi: 10.3389/fmed.2025.1713887
Received: 26 September 2025; Revised: 18 December 2025; Accepted: 30 December 2025;
Published: 16 January 2026.
Edited by:
Guangyan Zhangyuan, Shanghai Jiao Tong University, ChinaReviewed by:
Ting Li, National Center for Toxicological Research (FDA), United StatesAmirreza Khalaji, Emory University, United States
Copyright © 2026 Rao, Li and Leng. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Kai Leng, bGVuZ2thaTQ3NTdAcXEuY29t
†These authors share first authorship