Artificial intelligence-driven risk estimation, early detection, diagnostic characterisation and prognosis in hepatitis b-related hepatocellular carcinoma

Abstract

Chronic hepatitis B virus (HBV) infection remains the leading global antecedent of hepatocellular carcinoma (HCC), yet the surveillance framework on which early detection depends is constrained at both of its load-bearing points: The modest sensitivity of biannual ultrasonography for early-stage tumours, and the bounded discrimination of risk scores that compress a nonlinear, time-evolving process into a small set of additively weighted covariates. Artificial intelligence (AI) has been advanced as a corrective, but the term is applied loosely and the supporting evidence is drawn from heterogeneous populations and from clinical tasks that are frequently conflated. This narrative review appraises AI across four separable tasks in the HBV-HCC pathway: Risk estimation in chronic hepatitis B, early detection during surveillance, diagnostic characterisation of established lesions, and prognosis after diagnosis. We state what is and is not counted as AI, describe the search strategy and selection principles, and categorise every cited study according to whether its evidence is HBV-specific, mixed-aetiology, hepatitis C virus-derived, drawn from general HCC populations, or confined to recurrence and prognosis. Where machine-learning models built on routine clinical variables have been compared with established scores within a single validation cohort, they have discriminated more accurately. Convolutional networks have improved ultrasonographic lesion detection, serological and cell-free DNA methylation models have been applied to early-stage detection beyond alpha-fetoprotein, and multi-omics integration has been used for molecular subclassification. These findings rest on a literature that is predominantly retrospective and single-centre, in which external validation is the exception, calibration is reported inconsistently, explainability is uneven, and meta-analysts repeatedly judge the underlying studies to be at high risk of bias. Evidence drawn specifically from hepatitis B cohorts remains a minority of the published work, and no study has demonstrated that an AI-guided strategy improves stage at diagnosis or survival. Prospective multicentre validation, transparent calibration, adherence to TRIPOD+AI and PROBAST+AI, and deliberate inclusion of African and South Asian populations are therefore prerequisites rather than refinements.

Key Words: Hepatitis B; Hepatocellular carcinoma; Artificial intelligence; Deep learning; Risk prediction; Liquid biopsy; Surveillance

Core Tip: Artificial intelligence is applied to four distinct problems in chronic hepatitis B, namely risk estimation, early detection, diagnostic characterisation and prognosis, and this review appraises each separately, reframing hepatocellular carcinoma risk from a static, population-derived score into a dynamic, individualised probability. The discriminative advantage over legacy instruments is reproducible, yet it rests overwhelmingly on retrospective, single-centre, internally validated work in which hepatitis B-specific evidence is a minority. The field’s maturation will be measured not by further increments in the area under the curve but by the rigour of external validation, the candour of calibration, and the equity of deployment.

INTRODUCTION

Hepatocellular carcinoma (HCC) ranks among the foremost causes of cancer-related death worldwide, and chronic hepatitis B virus (HBV) infection, affecting more than 250 million people, remains its single most important cause[1,2]. Prognosis is dictated almost wholly by stage at detection: Tumours found within curative thresholds permit resection, ablation or transplantation, whereas symptomatic presentation forecloses these options. The rationale for surveillance therefore rests on a single proposition, that earlier detection prolongs survival, and international guidance accordingly recommends biannual ultrasonography, with or without alpha-fetoprotein, in those at elevated risk[3,4]. What has changed in the current iteration of that guidance is the assumption of uniformity. The 2025 European Association for the Study of the Liver guidelines identify surveillance personalised to individual risk as a principal departure from their 2018 predecessor, and the 2023 American Association for the Study of Liver Diseases guidance similarly treats stratification, rather than a single schedule applied to all, as the organising principle[3,4].

The framework is fragile at both load-bearing points. Meta-analytic synthesis places the sensitivity of ultrasonography for early-stage HCC well below the level clinicians intuitively assume, with performance degrading further in obesity and steatosis[5]. The complementary task of deciding whom to surveil has been delegated to a generation of risk scores that compress a nonlinear, time-evolving biological process into a small set of additively weighted covariates. It is precisely at this junction, between an undifferentiated surveillance schedule and the need for individualised stratification, that artificial intelligence (AI) has been advanced as a corrective[6].

The published literature on that proposition is wider than the surveillance question alone, and reviews that present it otherwise create a false impression of directness. A model that predicts microvascular invasion in a resected tumour says nothing about whom to enrol in surveillance, and evidence generated for one purpose cannot be transferred to another without argument. This review therefore examines AI across four separable tasks: Risk estimation in chronic hepatitis B, early detection during surveillance, diagnostic characterisation of established lesions, and prognosis after diagnosis, including recurrence. Prediction of response to specific therapies and selection between treatment modalities lie outside our scope. Throughout, the aim is to distinguish what has been demonstrated from what has been asserted.

REVIEW METHODOLOGY

PubMed/MEDLINE, EMBASE, Scopus and Web of Science were searched for articles published from January 2015 to March 2026. The search combined disease terms (“hepatitis B” OR HBV) AND (“hepatocellular carcinoma” OR HCC OR “liver cancer”) with method terms (“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network” OR radiomics OR “random forest” OR “gradient boosting” OR “support vector machine”) and task terms (“risk prediction” OR surveillance OR “early detection” OR “liquid biopsy” OR methylation OR “multi-omics” OR prognosis OR recurrence). Reference lists of retrieved reviews and meta-analyses were screened by hand.

Eligible articles were peer-reviewed studies in humans, published in English, that developed or evaluated a model addressing risk estimation, early detection, diagnostic characterisation or prognosis in HCC, and that reported at least one quantitative measure of performance. Conference abstracts without a corresponding full paper, editorials, commentaries, and technical reports of image segmentation without a clinical endpoint were excluded, as were studies reporting no measure of discrimination or accuracy.

This is a narrative review. Citation is representative rather than exhaustive, and where several studies addressed the same question, selection followed four principles applied in order: Cohorts specific to hepatitis B were preferred to indirect evidence; externally validated models were preferred to internally validated ones; larger cohorts were preferred to smaller; and pooled estimates were taken from meta-analyses in preference to individual studies. Quantitative synthesis was not undertaken, and formal risk-of-bias appraisal of individual studies was outside the scope of this review, although risk-of-bias judgements reached by the meta-analyses cited here are reported.

Terminology and analytical scope

AI spans methods that differ in their data requirements, interpretability and readiness for clinical use, and the term is applied inconsistently across this literature. Machine learning refers to algorithms that learn a mapping from curated, largely tabular features without a pre-specified functional form, including random forests, gradient-boosting machines, support vector machines and naive Bayes classifiers. Deep learning refers to multilayered neural networks that learn hierarchical representations from raw or minimally processed input, the convolutional network being the archetype for images and the recurrent network for longitudinal sequences. Radiomics is a feature-generation step rather than an algorithm: It yields quantitative intensity, shape and texture descriptors, and qualifies as AI only where those descriptors are subsequently modelled by machine or deep learning. Conventional statistical models, in which investigator-selected covariates enter a logistic or proportional-hazards equation with additive weights, fall outside this definition, as do biomarker-based scores such as gender, age, alpha-fetoprotein-L3 fraction, alpha-fetoprotein and des-gamma-carboxyprothrombin score (GALAD) and age-male-albumin-bilirubin-platelets, which are regression-derived instruments frequently but incorrectly grouped with AI. Penalised regression occupies the boundary, and is treated here as conventional where it selects covariates for a subsequent equation and as machine learning where embedded in an automated selection, tuning and validation pipeline.

Categories of evidence

Because the volume of publication in this field exceeds the volume of HBV-specific evidence, each study tabulated in Tables 1, 2, and 3 carries one of six labels: HBV-specific, where the cohort comprised patients with chronic HBV infection or HBV-related disease; mixed-aetiology, where HBV formed part of a broader viral or chronic liver disease cohort; hepatitis C virus-derived, where the cohort was hepatitis C; general HCC, where aetiology was unselected or unreported; recurrence-related; and prognostic. Studies outside the first category are retained where they establish a methodological principle, and are presented as indirect evidence.

Table 1 Artificial intelligence models for risk estimation in chronic hepatitis B.

Model/Ref.	Design, region	Population, aetiology and evidence category (n)	Endpoint	Algorithm	Validation	Performance	Calibration and explainability	Key limitation
PLAN-B[17]	Retrospective, multinational (Korea, Europe)	Antiviral-treated CHB; HBV-specific (> 13000)	Incident HCC	Gradient boosting	External (independent Korean and Caucasian cohorts)	Superior stratification to legacy scores; minimal-risk group identified	Not reported in detail	Retrospective derivation; treated patients only
Nam et al[18]	Retrospective, Korea	HBV-related cirrhosis on entecavir; HBV-specific (424 derivation, 316 external)	Incident HCC	Deep neural network	Internal and external	C-index 0.719 internal, 0.782 external; six comparator scores all < 0.60 in same cohort	Explainability not reported	Small derivation cohort; single-country
Singal et al[16]	Retrospective cohort	Cirrhosis; general HCC/non-HBV-specific	Incident HCC	Multiple ML algorithms vs regression	Internal	ML superior to conventional regression	Not reported	Indirect evidence for HBV; early-generation methods
Ioannou et al[19]	Retrospective, United States (Veterans Affairs)	HCV-related cirrhosis; HCV-derived (48151)	3-year HCC incidence	Recurrent neural network	Internal	AUC: 0.759 overall; 0.806 in sustained virological response	Not reported	Aetiology not HBV; internal validation only
Phan et al[20]	Retrospective, Taiwan (national insurance)	Viral hepatitis; mixed-aetiology (approximately 1000000 records)	Incident liver cancer	Convolutional neural network	Internal	AUC: 0.886	Not reported	Administrative data; no laboratory granularity
Wu et al[15]	Case-control, China	CHB; HBV-specific	Incident HCC	Five algorithms with LASSO feature selection	Internal	AUC commonly > 0.80	SHAP-based interpretation reported	Case-control design; no external validation
Li et al[21]	Retrospective, single centre, China	HBV-related cACLD; HBV-specific (1051; 736 training, 315 validation)	Incident HCC (103 events, 9.8%)	RF, XGBoost, SVM, LR, NB	Internal (random split)	RF AUC: 0.979 training, 0.956 validation; LSM alone 0.745 and 0.826	Calibration plots reported; SHAP reported	Single centre; no external validation; authors attribute high AUC partly to RF overfitting
Kuaaroon et al[22]	Retrospective, Thailand	Chronic HBV; HBV-specific	Short-term HCC development	Extreme gradient boosting	Internal with independent cohort	Models developed to tailor surveillance interval	Not reported in detail	Limited external generalisability

CHB: Chronic hepatitis B; HBV: Hepatitis B virus; HCC: Hepatocellular carcinoma; C-index: Harrell concordance index; ML: Machine learning; HCV: Hepatitis C virus; AUC: Area under the receiver operating characteristic curve; LASSO: Least absolute shrinkage and selection operator; SHAP: Shapley additive explanations; cACLD: Compensated advanced chronic liver disease; RF: Random forest; XGBoost: Extreme gradient boosting; SVM: Support vector machine; LR: Logistic regression; NB: Naive Bayes; LSM: Liver stiffness measurement.

Table 2 Artificial intelligence for early detection during surveillance.

Ref.	Design, region	Population, aetiology and evidence category (n)	Endpoint	Approach	Validation	Performance	Calibration and explainability	Key limitation
Chaiteerakij et al[23]	Retrospective image cohort, Thailand	Patients undergoing hepatic ultrasonography; general HCC (26288 images, 5444 patients)	Focal lesion detection and malignancy	YOLOv5 object detection	Internal	Detection 84.8%; sensitivity for malignancy 97.0%; maintained at lesions ≤ 1 cm	Not reported	Image-level rather than outcome-level endpoint
Rhyou and Yoo[24]	Retrospective, single centre	Ultrasound dataset; general HCC	HCC classification	Wavelet-transform deep network (HCC-Net)	Internal	Accuracy: 96.2%; sensitivity: 97.6%; specificity 98.1%	Not reported	Single-centre dataset; no external validation
Brehar et al[25]	Retrospective	Ultrasound images; general HCC	Tumour vs cirrhotic parenchyma	CNN vs texture-based ML	Internal	CNN superior to conventional ML	Not reported	Small dataset; no clinical endpoint
Dana et al[26]	Retrospective multicentre	Cirrhosis under surveillance; mixed-aetiology	Risk stratification and early HCC detection	Ultrasound-based deep learning	Internal and multicentre	Improved stratification and detection of early HCC	Not reported in detail	Retrospective; no prospective outcome data
Schotten et al[27]	Retrospective, Europe	Chronic hepatitis B and C; mixed-aetiology	Early-stage HCC detection	GALAD (logistic regression; conventional comparator)	Internal	AUC around 0.95 in selected cohorts	Not applicable	Not an AI model; comparator instrument
Marsh et al[28]	Phase 3 biomarker validation, United States	Cirrhosis; mixed-aetiology	HCC detection within 12 months of diagnosis	GALAD (conventional comparator)	Prospective phase 3 design	Improved detection in the 12 months preceding clinical diagnosis	Not applicable	Not an AI model; comparator instrument
Johnson et al[29]	Retrospective, surveillance setting	Surveillance cohort; mixed-aetiology	HCC detection	Random forest applied to GALAD biomarkers	Internal	Outperformed original logistic GALAD in discrimination and F1 score	Not reported	Internal validation only
Tayob et al[30]	Community-based cohort, United States	Cirrhosis of multiple aetiologies; mixed-aetiology	Early HCC detection	HES V2.0 algorithm	External validation in community cohort	Validated in community-based cirrhosis cohort	Not reported in detail	Aetiology heterogeneous; limited HBV representation
Sun et al[31]	Case-control, China	Healthy, HBV-infected, cirrhotic and HCC participants; mixed-aetiology with HBV subgroup (374)	All-stage and early-stage HCC detection	Targeted bisulfite sequencing with ML	Internal	All-stage sensitivity 96.7% at specificity 93.7%; early-stage sensitivity 90.0%	Not reported	Case-control design including healthy controls; spectrum bias inflates apparent sensitivity
Xu et al[32]	Case-control, China	HCC patients and normal controls; general HCC (1098 HCC, 835 controls)	Diagnosis and prognosis	ctDNA methylation markers with prediction modelling	Internal	Diagnostic and prognostic models correlating with tumour burden, stage and treatment response	Not reported	Healthy comparator group; not a surveillance population

HCC: Hepatocellular carcinoma; YOLOv5: You only look once, version 5; CNN: Convolutional neural network; ML: Machine learning; GALAD: Gender, age, alpha-fetoprotein-L3 fraction, alpha-fetoprotein and des-gamma-carboxyprothrombin score; AUC: Area under the receiver operating characteristic curve; AI: Artificial intelligence; HES: Hepatocellular carcinoma early detection screening algorithm; HBV: Hepatitis B virus; ctDNA: Circulating tumour DNA.

Table 3 Artificial intelligence for diagnostic characterisation and prognosis.

Ref.	Design, region	Population, aetiology and evidence category (n)	Endpoint	Approach	Validation	Performance	Calibration and explainability	Key limitation
Wei et al[33]	Retrospective, China	Resected HCC; recurrence-related (434)	Early recurrence after treatment	MRI radiomics with 3D U-net automated segmentation	Internal validation cohort	C-index 0.706; exceeded BCLC and CNLC staging	Not reported	Post-diagnosis endpoint; single-country
Ji et al[34]	Retrospective, multi-institutional, China	Resected HCC; recurrence-related	Post-resection recurrence	Machine learning on contrast-enhanced CT radiomics	Multi-institutional	Predicted recurrence across institutions	Not reported	Retrospective; feature instability across scanners
Zhao et al[35]	Retrospective, China	HCC after curative resection; recurrence-related	Early recurrence	Deep-learning radiomics on contrast-enhanced MRI	Internal	Hybrid models outperformed single-stream inputs	Not reported	Single-centre; no external validation
Chaudhary et al[36]	Retrospective, public cohorts	Liver cancer; prognostic	Survival and molecular subgroup	Deep-learning multi-omics integration	External replication in independent cohorts	Reproducible prognostic molecular subgroups	Not reported	Omics assays costly and unavailable in high-burden settings
Kucukakcali et al[37]	Case-control, open-access gene expression data	HBV-related HCC vs HBV infection alone; HBV-specific (17 vs 36)	Discrimination of HBV-HCC from HBV alone	Extreme gradient boosting with 10-fold cross-validation	Internal cross-validation	18 genes selected; candidate biomarkers nominated	Not reported	Very small sample; reported accuracy not an estimate of real-world performance

HCC: Hepatocellular carcinoma; MRI: Magnetic resonance imaging; U-net: Convolutional network architecture for biomedical image segmentation; C-index: Harrell concordance index; BCLC: Barcelona Clinic Liver Cancer; CNLC: China Liver Cancer staging system; CT: Computed tomography; HBV: Hepatitis B virus.

THE CEILING OF CONVENTIONAL RISK STRATIFICATION

A lineage of risk scores has matured over two decades. Instruments derived in untreated cohorts include the Guide with Age, Gender, HBV DNA, Core Promoter Mutations and Cirrhosis-HCC score combining age, sex, HBV DNA, core promoter mutation and cirrhosis; the Chinese University HCC score; the risk estimation for hepatocellular carcinoma in chronic hepatitis B score; and the liver stiffness-augmented liver stiffness measurement-HCC score[7-10]. As nucleos(t)ide analogue therapy reshaped the natural history and attenuated the predictive weight of viraemia, a second generation emerged for treated populations, notably platelets-age-gender-hepatitis B (PAGE-B), its albumin-incorporating derivative modified PAGE-B, and the age-male-albumin-bilirubin-platelets score combining age, sex, albumin-bilirubin and platelets[11-13].

These instruments are clinically serviceable, and several achieve high negative predictive value, identifying low-risk subgroups in whom surveillance may be safely de-intensified[13]. Their discrimination, however, is bounded. Across validation cohorts the area under the curve for the better instrument clusters between roughly 0.72 and 0.84, and performance attenuates appreciably in the cirrhotic subgroup in whom accurate stratification is most consequential[13,14]. The development populations also differ in ways that preclude direct comparison of these figures, as Table 4 sets out: PAGE-B was derived in Caucasian patients receiving antiviral therapy, modified PAGE-B in Asian patients, and risk estimation for hepatocellular carcinoma in chronic hepatitis B and Guide with Age, Gender, HBV DNA, Core Promoter Mutations and Cirrhosis-HCC in untreated cohorts, so a higher reported area under the curve for one score over another may reflect the population in which it was built rather than any intrinsic superiority.

Table 4 Conventional risk scores for hepatocellular carcinoma in chronic hepatitis B, with development and validation populations.

Score	Development population	Validation populations	Component variables	Reported discrimination
GAG-HCC	Untreated CHB, Chinese (Hong Kong)	Asian CHB cohorts, largely untreated	Age, sex, HBV DNA, core promoter mutation, cirrhosis	C-statistic up to approximately 0.89[7,10]
CU-HCC	CHB carriers, Chinese (Hong Kong)	Asian CHB cohorts; treated and untreated	Age, albumin, bilirubin, HBV DNA, cirrhosis	AUC approximately 0.73-0.85[8]
REACH-B	Untreated CHB, Taiwanese (REVEAL-HBV)	Asian cohorts; attenuated performance in treated patients	Sex, age, ALT, HBeAg, HBV DNA	AUC approximately 0.64-0.79[9]
LSM-HCC	CHB with transient elastography, Chinese (Hong Kong)	Asian CHB cohorts with elastography	Liver stiffness, age, albumin, HBV DNA	AUC approximately 0.77-0.83[10]
PAGE-B	Caucasian CHB on 5-year antiviral therapy (European)	European treated cohorts; reduced performance in Asian populations	Platelets, age, sex	AUC approximately 0.78-0.80[11]
mPAGE-B	Asian CHB on antiviral therapy (Korean)	Asian treated cohorts	PAGE-B variables plus albumin	AUC approximately 0.80-0.82[12]
aMAP	Mixed CHB and chronic hepatitis, multinational (Asian and Caucasian)	Multinational; validated in HBV cirrhosis on ETV or TDF	Age, sex, albumin-bilirubin, platelets	AUC/C-index approximately 0.72-0.87[13,14]

Scores were developed in populations differing by ethnicity, treatment status and stage of liver disease; reported discrimination values are therefore not directly comparable between rows. GAG-HCC: Guide with Age, Gender, Hepatitis B virus DNA, Core Promoter Mutations and cirrhosis-hepatocellular carcinoma; CHB: Chronic hepatitis B; HBV: Hepatitis B virus; CU-HCC: Chinese University hepatocellular carcinoma score; AUC: Area under the receiver operating characteristic curve; REVEAL-HBV: Risk Evaluation of Viral Load Elevation and Associated Liver Disease/Cancer-Hepatitis B Virus; REACH-B: Risk estimation for hepatocellular carcinoma in chronic hepatitis B; ALT: Alanine aminotransferase; HBeAg: Hepatitis B e antigen; LSM: Liver stiffness measurement; HCC: Hepatocellular carcinoma; PAGE-B: Platelets-age-gender-hepatitis B; mPAGE-B: Modified platelets-age-gender-hepatitis B; aMAP: Age-male-albumin-bilirubin-platelets; ETV: Entecavir; TDF: Tenofovir disoproxil fumarate.

The limitation is structural rather than incidental. Linear and proportional-hazards formulations cannot accommodate the higher-order interactions and collinearity that characterise hepatocarcinogenesis, in which the effect of any variable is conditioned on the others; and a baseline score discards the prognostic signal embedded in the longitudinal trajectory of platelets, transaminases, viral suppression and elastography[15]. It is this representational ceiling that motivates a turn toward methods designed to learn nonlinear, temporal structure directly from data.

TASK 1: RISK ESTIMATION IN CHRONIC HEPATITIS B

Risk estimation asks who, among people without HCC, will develop it. The population is chronic hepatitis B, the output governs surveillance eligibility and intensity, and the comparator is the set of conventional scores described above. This is the task on which the AI evidence is most mature, and the only one on which AI competes with the legacy instruments on their own terrain.

An early demonstration of the principle came from a cirrhosis cohort in which machine-learning algorithms applied to routine clinical variables outperformed conventional regression models for predicting the development of HCC[16]. That study was not HBV-specific, and it is cited here for the methodological point rather than as direct evidence.

The PLAN-B model was derived using a gradient-boosting algorithm in more than thirteen thousand antiviral-treated patients and validated in independent Korean and Caucasian cohorts. Integrating ten baseline parameters, it stratified risk more accurately than the established scores and isolated a minimal-risk group plausibly suited to reduced-intensity surveillance[17]. A deep neural network developed by Nam et al[18] in patients with HBV-related cirrhosis on entecavir, with internal and external validation, achieved Harrell concordance indices of 0.719 and 0.782 respectively, against PAGE-B, Chinese University-HCC, HCC-RESCUE, ADRESS-HCC, modified PAGE-B and the Toronto HCC risk index, all of which fell below 0.60 in the same validation cohort. This last comparison is the strongest single piece of evidence in the field, precisely because model and comparators were evaluated in one cohort rather than across separate publications.

The temporal dimension has proven fruitful. In a large Veterans Health Administration cohort of hepatitis C-related cirrhosis, a recurrent network trained on baseline and serial longitudinal variables predicted three-year HCC incidence with an area under the curve of 0.759 overall and 0.806 among those with sustained virological response, significantly surpassing logistic regression; the aetiology is not hepatitis B, so this furnishes proof of principle for sequence-aware modelling rather than direct evidence[19]. A convolutional approach applied to approximately one million records from the Taiwanese national insurance database, in which each patient history was rendered as a matrix, predicted liver cancer in a viral hepatitis cohort of mixed aetiology with an area under the curve of 0.886[20].

More recent contributions have coupled performance with interpretability and disease-stage specificity. An interpretable model spanning five algorithms with regularised feature selection was developed for chronic hepatitis B[15]. In HBV-related compensated advanced chronic liver disease, five algorithms were compared in 1051 patients from a single centre, the random forest achieving an area under the curve of 0.979 in the training set and 0.956 in the validation set, with liver stiffness measurement the dominant predictor on Shapley analysis and calibration reported[21]. Those figures warrant caution rather than celebration: Validation was by random split of a single cohort rather than an external population, the outcome was uncommon at 9.8 percent, and the investigators themselves attribute the unusually high discrimination in part to random-forest overfitting. Short-term models have also been proposed explicitly to tailor the surveillance interval to individual risk[22]. Table 1 sets these models against the conventional scores.

TASK 2: EARLY DETECTION DURING SURVEILLANCE

Early detection asks whether a tumour already present but subclinical can be identified in a person under surveillance. If risk estimation addresses whom to watch, early detection addresses what the watching finds, and the operator-dependent ultrasonography on which surveillance rests is the most clinically pertinent target.

Ultrasonographic detection

A YOLOv5 detection system trained on 26288 ultrasound images from 5444 patients achieved an overall focal-lesion detection rate of 84.8 percent with sensitivity for malignancy of 97.0 percent, and performance was maintained for lesions at and below one cm[23]. A wavelet-based hierarchical network, HCC-Net, reported accuracy of 96.2 percent, sensitivity of 97.6 percent and specificity of 98.1 percent on a single-centre dataset[24], and lightweight convolutional architectures have outperformed conventional texture-based machine learning for discriminating tumour from cirrhotic parenchyma[25]. These advances address directly the variability and obesity-related failure that limit human ultrasonography[26].

Serological detection

Blood-based detection offers non-invasiveness and scalability, and the established instrument in this space is not an AI model at all. The GALAD score, combining sex, age, alpha-fetoprotein, its L3 isoform and des-gamma-carboxyprothrombin, is a logistic regression instrument, and it is against this conventional comparator that machine learning has been tested. GALAD has shown high accuracy for HCC, with reported areas under the curve around 0.95 and superiority over ultrasonography in selected cohorts, and a phase 3 validation demonstrated improved detection within the twelve months preceding clinical diagnosis[27,28]. A random-forest reformulation of the same GALAD biomarkers outperformed the original logistic model in both discrimination and F1 score within a standard surveillance setting[29], and the updated HCC Early Detection Screening algorithm, HES V2.0, has been validated in a community-based cirrhosis cohort of multiple aetiologies[30].

Liquid biopsy

Cell-free DNA methylation is the most rapidly advancing strand of this literature. A targeted bisulfite-sequencing assay coupled with machine learning, evaluated in a cohort of 374 participants that included healthy individuals, HBV-infected and cirrhotic patients, and patients with HCC, discriminated all-stage HCC with sensitivity of 96.7 percent at specificity of 93.7 percent and early-stage disease with sensitivity of 90.0 percent, while nominating GJA4 promoter methylation as a candidate biomarker of HBV-related hepatocarcinogenesis[31]. In a larger cohort of 1098 patients with HCC and 835 normal controls, circulating tumour DNA methylation markers supported both a diagnostic and a prognostic prediction model, with results correlating with tumour burden, stage and treatment response[32].

Several features of this evidence temper the headline figures. The cohorts are small relative to the dimensionality of the data, a setting in which discrimination is characteristically optimistic unless validation is external. The comparison groups matter as much as the assay: Where healthy controls are included alongside patients with established cancer, the case-control contrast is wider than the contrast a surveillance test must actually resolve, which is between cirrhosis with and without an early tumour, so reported sensitivity overstates what would be achieved in the population where the test would be deployed. Assays remain unstandardised across laboratories, with no agreed thresholds or reference materials. Cost is a further constraint, and it falls most heavily on the high-burden regions where HBV is concentrated and where the marginal expense of a sequencing-based assay is measured against the cost of the ultrasound it would supplement.

TASK 3: DIAGNOSTIC CHARACTERISATION

Diagnostic characterisation asks what a lesion already identified actually is: Its grade, its vascular behaviour, its molecular class. The patient has a tumour. Nothing in this section bears on whom to place under surveillance, and evidence drawn from it should not be read as supporting the tasks above.

In cross-sectional imaging, radiomics extracts high-dimensional descriptors of intensity, shape and texture that escape qualitative reading. Extraction alone is not AI; the studies below qualify because those descriptors were subsequently modelled by machine-learning classifiers, whereas convolutional networks bypass the extraction step and learn discriminative features directly.

Deep-learning automated segmentation has begun to displace laborious manual delineation. A magnetic resonance study in 434 patients demonstrated that a radiomic signature incorporating a defined tumour-border extension, analysed with three-dimensional U-net segmentation, achieved a validation concordance index of 0.706 and exceeded both the Barcelona Clinic and China Liver Cancer staging systems[33]. Contrast-enhanced computed tomography radiomics analysed by machine learning predicted post-resection recurrence across institutions[34], and hybrid models fusing deep-learning, radiomic and clinical-radiological streams outperformed any single input[35]. The frontier is moving from tumour characterisation toward risk estimation in the pre-malignant liver, exemplified by integration of quantitative computed tomography signatures of liver and spleen with a clinical score to refine HCC risk in cirrhosis, exploiting splenic morphology as a surrogate of portal hypertension[6]. The persistent liabilities are scanner and protocol heterogeneity, which destabilise radiomic features across sites, and a preponderance of small, retrospective, single-centre designs.

TASK 4: PROGNOSIS AFTER DIAGNOSIS

Prognosis asks what will happen to a patient whose tumour is known. It is the task furthest removed from surveillance, and it accounts for a substantial share of the AI literature in HCC.

Recurrence prediction

Recurrence after curative-intent treatment is the endpoint most frequently modelled. The magnetic resonance and computed tomography radiomics studies described above were developed for precisely this purpose[33-35], and are counted here rather than among detection studies. Their clinical value lies in planning follow-up intensity and adjuvant strategy, not in identifying candidates for surveillance.

Molecular subclassification and survival

Deep-learning integration of multi-omics data, combining transcriptomic, methylation and other layers, has predicted survival in liver cancer and defined reproducible molecular subgroups recurring across external cohorts, a stratification clinical variable alone cannot supply[36]. In the HBV-specific setting, extreme gradient boosting applied to open-access gene-expression data discriminated HBV-related HCC from chronic HBV infection without HCC and nominated candidate transcriptomic biomarkers of malignant transformation; the case-control design comprised 17 patients with HBV-related HCC and 36 with HBV infection alone, a sample far too small for the reported accuracy to be taken as an estimate of real-world performance[37]. The promise is also the fragility: Multi-omics cohorts are characteristically small and high-dimensional, the assays costly and unevenly available in the high-burden regions where the epidemic is concentrated, leaving these models at considerable distance from routine access.

QUANTITATIVE SYNTHESIS FROM META-ANALYSES

Pooled estimates lend the field a more disciplined frame than any single study, provided the task each synthesis addresses is kept in view. For diagnosis of HCC from medical images, a meta-analysis of deep-learning methods, in which 30 of 48 eligible studies were quantitatively synthesised, reported a pooled sensitivity of 89 percent, a specificity of 90 percent and an area under the curve of 0.95[38]. These cohorts were general HCC populations, and the task was recognition of disease already visible on imaging rather than prediction of its future occurrence.

For diagnostic characterisation, a meta-analysis of computed tomography radiomics for preoperative prediction of microvascular invasion included eleven studies, nine of which were pooled, and reported sensitivity of 0.82, specificity of 0.79 and an area under the summary receiver operating characteristic curve of 0.87; subgroup analysis found deep-learning models superior to non-deep-learning approaches, with areas under the curve of 0.97 against 0.83, and three-dimensional segmentation superior to two-dimensional, 0.93 against 0.83[39]. A separate synthesis of imaging-based AI for the same endpoint across sixteen studies and 4759 cases reported a pooled area under the curve of 0.90 for deep learning against 0.82 for non-deep-learning methods[40]. For post-hepatectomy recurrence, AI models reached a pooled area under the curve of 0.89, with time-dependent discrimination of 0.83 at two years declining to 0.68 at five[41].

Two conclusions follow. AI confers a discriminative advantage that is reproducible across syntheses, and that advantage is reported within a literature that meta-analysts repeatedly judge to be at high risk of bias. It should also be noted that none of these pooled estimates addresses risk estimation in chronic hepatitis B, the task with which this review opened.

LIMITATIONS AND BARRIERS TO TRANSLATION

Validation, calibration, explainability and reporting

The literature surveyed shares a structural vulnerability no headline metric can disguise. The dominant designs are retrospective and single-centre, and external validation, in which a model is tested on a temporally or geographically distinct population, remains the exception[6,15,39]. Internal validation certifies only consistency with the data that produced it; it cannot establish transportability to a different scanner, laboratory, ethnicity or era. In a field whose methods are intrinsically prone to overfit, the asymmetry between abundant internal and scarce external validation is the central evidentiary weakness.

Performance metrics reported in different studies are not directly comparable. Cohorts differ in aetiology, treatment status, endpoint definition, follow-up duration and validation strategy, and an area under the curve obtained in one population carries no fixed relationship to one obtained in another. Claims of superiority over conventional scores are therefore credible only where model and comparator were evaluated in the same validation cohort, a condition met by a minority of the studies surveyed here.

Two further deficits compound the problem. The first is calibration: Discrimination quantifies ranking, but calibration, the agreement between predicted and observed risk, determines whether a probability can guide a surveillance decision, and it is reported far less consistently than the area under the curve. The second is opacity. Where complex models are deployed, post-hoc explainability methods such as Shapley additive explanations render feature contributions interpretable, yet such reporting is uneven, and a prediction that cannot be interrogated is one a clinician cannot reasonably act upon or defend[15,21]. The reporting and bias-appraisal instruments developed for this purpose, TRIPOD+AI and PROBAST+AI, exist precisely to standardise these disclosures and to surface the dependence on internal validation that bias assessments repeatedly identify as the predominant source of high risk of bias[42,43]. Their uneven uptake is itself a marker of the field’s immaturity.

Barriers to translation

Beyond evidentiary rigour lie obstacles of implementation. Data heterogeneity across imaging vendors, laboratory platforms and record systems undermines portability, and the absence of harmonised standards forecloses the large federated datasets robust training requires. Equity is a non-trivial concern: Models trained predominantly in East Asian or European cohorts may not generalise to the African and South Asian populations bearing a substantial share of the global HBV burden, and uncritical deployment risks encoding rather than correcting disparity. Regulatory frameworks for adaptive, continuously learning algorithms remain in evolution, and questions of liability, data governance and consent are unsettled. Above all, the field lacks the evidence that would most advance it: Prospective demonstration that an AI-guided strategy improves stage at diagnosis and ultimately survival relative to standard care.

AN INTEGRATED CLINICAL FRAMEWORK

The four tasks examined separately above meet in a single patient pathway, and Figure 1 sets out how AI might enter it. A chronic HBV population undergoes baseline risk stratification using conventional scores or clinical-variable machine learning; longitudinal data then permit dynamic risk estimation that updates as platelets, viral suppression and elastography evolve; the resulting probability determines an individualised surveillance strategy rather than a uniform biannual schedule; AI-assisted ultrasonography, cross-sectional imaging or biomarker and liquid-biopsy testing performs the detection; a tumour so identified proceeds to diagnostic characterisation and prognostic assessment; and the output informs multidisciplinary clinical decision-making. Each transition in that sequence carries the evidentiary caveats set out above, and the figure should be read as a framework for investigation rather than a description of validated practice.

Figure 1 Proposed integration of artificial intelligence into the hepatitis B-related hepatocellular carcinoma care pathway. A chronic hepatitis B population undergoes baseline risk stratification using conventional scores or clinical-variable machine learning. Longitudinal laboratory, virological and elastographic data permit dynamic risk estimation, which determines an individualised surveillance strategy in place of a uniform biannual schedule. Detection is performed by artificial intelligence-assisted ultrasonography, cross-sectional imaging, or serological and liquid-biopsy testing. A tumour so identified proceeds to diagnostic characterisation and prognostic assessment, and the output informs multidisciplinary clinical decision-making. The lower band lists the methodological conditions that apply to every stage. The figure represents a framework for investigation and not a description of validated practice. HBV: Hepatitis B virus; ML: Machine learning; US: Ultrasonography; CT: Computed tomography; MRI: Magnetic resonance imaging; HCC: Hepatocellular carcinoma; AI: Artificial intelligence.

FUTURE DIRECTIONS

The corrective agenda follows from the deficits. The priority is prospective, multicentre validation embedded within pragmatic or randomised designs tested on clinically meaningful endpoints. Federated learning, training across institutions without centralising patient data, offers a route to the scale and diversity that generalisability demands while respecting privacy. Dynamic, sequence-aware models that ingest the longitudinal trajectory of laboratory, biomarker and elastographic data are conceptually better matched to a disease that evolves than any baseline snapshot[19]. Multimodal fusion of clinical, imaging, serological and molecular streams within a single calibrated framework is likely to outperform any single modality, provided the resulting complexity is matched by commensurate validation and transparency. Finally, adherence to TRIPOD+AI and PROBAST+AI should become a condition of publication, so that external validation, calibration and explainability are reported as standard[42,43]. Table 5 aligns these priorities against present-day gaps.

Table 5 Present-day gaps and translational priorities.

Domain	Current limitation	Recommended priority
Study design	Predominantly retrospective, single-centre	Prospective, multicentre, outcome-based designs
Aetiological specificity	HBV-specific evidence a minority of published work	Dedicated HBV cohorts rather than extrapolation from mixed or HCV populations
Validation	Internal validation dominant; external scarce	Mandatory external, temporal and geographic validation
Comparison	Superiority inferred across dissimilar studies	Head-to-head comparison against named scores within a single validation cohort
Calibration	Inconsistently reported	Routine calibration alongside discrimination
Explainability	Variable use of interpretability methods	Standardised Shapley-based explanation
Reporting	Uneven adherence to standards	TRIPOD+AI and PROBAST+AI as publication conditions
Equity	Cohorts skewed to East Asian and European data	Inclusion of African and South Asian HBV populations
Data scale	Siloed institutional datasets	Federated learning across centres

HBV: Hepatitis B virus; HCV: Hepatitis C virus; PROBAST+AI: Prediction model risk of bias assessment tool, artificial intelligence extension; TRIPOD+AI: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis, artificial intelligence extension.

KEY MESSAGES FOR CLINICIANS

For the clinician managing chronic hepatitis B today, the practical position is as follows. Machine-learning models built on routine clinical and laboratory variables represent the most credible application, and where they have been compared with conventional scores in a single validation cohort they have stratified risk more accurately; PLAN-B and the Nam deep neural network are the instruments with the strongest supporting evidence. AI applied to ultrasonography addresses a genuine weakness in current surveillance and may in time improve detection of small lesions, but no such system has yet been shown to improve stage at diagnosis in practice. Liquid-biopsy and multi-omics approaches remain investigational, and their reported sensitivities are derived from case-control designs that flatter performance relative to a surveillance setting. Radiomic models for microvascular invasion, recurrence and prognosis address questions that arise after a tumour is found and have no bearing on surveillance decisions. No AI model should at present override guideline-directed surveillance, and none has been prospectively validated against clinical outcomes. Where such a model is used, the clinician should ask three questions of it: Was it validated in a population resembling the patient in front of them, was its calibration reported rather than only its discrimination, and can the basis of an individual prediction be interrogated.

CONCLUSION

AI has enriched the assessment of HBV-related HCC across risk estimation, early detection, diagnostic characterisation and prognosis, learning the nonlinear, interactive and increasingly multimodal and temporal structure that conventional scores cannot represent. The honest reading of the evidence is that these gains remain provisional, built largely on retrospective, single-centre, internally validated work with inconsistent calibration and explainability, an evidence base in which genuinely hepatitis B-specific studies are a minority, and an essential absence of prospective outcome data. The field’s eventual clinical authority will be earned in prospective study rather than asserted in retrospective performance. On those terms, AI is best regarded not as a replacement for the surveillance architecture of chronic hepatitis B but as its disciplined and individualised refinement.