Performance of artificial intelligence in predicting hepatocellular carcinoma recurrence after thermal ablation: A systematic review

Abstract

BACKGROUND

Recurrence prediction of hepatocellular carcinoma (HCC) after thermal ablation represents a challenge that can impact patients' quality of life. Artificial intelligence (AI)-based radiomics models applied to various imaging modalities can improve recurrence prediction, therefore guiding therapeutic decisions.

AIM

To evaluate the effectiveness of AI-driven predictive models in predicting HCC recurrence.

METHODS

A systematic literature search in PubMed and Scopus was performed, and a total of ten studies were included in this systematic review. All studies included response prediction evaluation with AI models for patients who underwent thermal ablation for HCC. Deep learning and machine learning algorithms were utilized to evaluate the predictive performance and accuracy through metrics such as the area under the curve and concordance index.

RESULTS

The developed models demonstrated high accuracy in predicting local progression and recurrence, allowing a solid risk stratification. In particular, the integration of imaging data and clinical-laboratory variables optimized treatment selection, highlighting the superior ability of imaging models to predict therapeutic outcomes compared to clinical parameters alone. Furthermore, radiomic analysis of follow-up imaging enabled highly accurate detection of ablation site recurrence.

CONCLUSION

AI-driven predictive models based on multimodal radiomic analyses integrated with clinical data represent promising tools for predicting tumor recurrence after thermal ablation in HCC patients.

Key Words: Artificial intelligence; Hepatocellular carcinoma; Thermal ablation; Prediction; Tumor recurrence

Core Tip: Artificial intelligence can aid the prediction of post-thermal ablation treatment recurrence of hepatocellular carcinoma, improving patient's quality of life, tailoring the follow-up and avoiding unnecessary treatments while providing an early recognition of tumor recurrence.

INTRODUCTION

Hepatocellular carcinoma (HCC) is one of the most prevalent malignancies worldwide, most commonly affecting patients with underlying liver cirrhosis and significantly reducing life expectancy[1].

In early-stage HCC, defined by the Barcelona Clinic Liver Cancer guidelines as up to three lesions, each measuring no more than 3 cm, the treatment of choice is liver transplantation, which is considered the only curative option[2]. When transplantation is not feasible, surgical resection may be considered[3]. However, not all patients are eligible for surgery due to factors such as lesion location, comorbidities, or patient refusal. In such cases, heat-based thermal ablation plays a key role[4]. These techniques, including radiofrequency ablation (RFA) and microwave ablation (MWA), provide localized tumor control and can be performed percutaneously or laparoscopically[5,6].

Despite their minimally invasive nature, these techniques have limitations, especially in case of larger or multiple lesions, where complete response rates may be suboptimal. Moreover, even after an initial complete response, recurrence-free survival may be suboptimal due to early recurrence, often driven by undetected microsatellite nodules, thereby worsening the patient's prognosis[7].

Predicting early recurrence after thermal ablation is a major focus of research since the introduction of these techniques. The ability to anticipate treatment response and recurrence risk would allow clinicians to better tailor therapeutic strategies, optimize follow-up protocols, and avoid unnecessary procedures, improving patient outcomes[8]. Identifying non-responders to ablative treatments could shift the current paradigm in treating and following-up HCC patients.

In recent years, the rise of radiomics and artificial intelligence (AI)-based models [deep learning (DL) models and machine learning (ML) ones] has transformed predictive oncology, showing promise in forecasting treatment response and tumor recurrence across a wide range of malignancies[9].

The aim of this systematic review is to evaluate the performance of AI-based models (both DL and ML) in predicting early recurrence of HCC following thermal ablation, with a particular focus on metrics such as the area under the curve (AUC), sensitivity, specificity, positive and negative predictive values, concordance index (C-index), and accuracy.

MATERIALS AND METHODS

Eligibility criteria

This systematic review was registered in PROSPERO (CRD420251120913) and was designed using the Population, Intervention, Comparator, Outcome framework to define the clinical question and inclusion criteria[10]. The framework was constructed as follows:

Population: Adult patients diagnosed with HCC treated with curative intent using heat-based thermal ablation techniques (RFA and MWA), with follow-up imaging performed using computed tomography (CT) or magnetic resonance imaging (MRI).

Intervention: Application of AI, ML, or DL models to predict local recurrence following thermal ablation, integrating imaging and clinical information or using imaging alone.

Comparator: Conventional prediction models, follow-up examinations, or no comparator in cases where the study was purely predictive without controls.

Outcome: Diagnostic performance of AI-based models in predicting recurrence, evaluated using metrics such as AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), C-index, and accuracy.

Studies eligible for inclusion comprised randomized controlled trials (RCTs) and prospective and retrospective cohort studies.

Sources and search strategy

The PRISMA standards were used to conduct this systematic review[11]. A comprehensive literature search was performed in three electronic databases: PubMed, Scopus, and Web of Science. The search aimed to identify all relevant studies evaluating the use of AI, ML, or DL models to predict local recurrence after heat-based thermal ablation (RFA or MWA) in patients with HCC.

The following keywords and medical subject headings terms were used alone and in multiple combinations: “HCC” “Hepatocellular carcinoma” “Recurrence” “Ablation” “RFA” “MWA” “AI” “Artificial Intelligence” “Radiomic” “Machine learning” “Deep learning” “Prediction”; in particular, the combinations used were: HCC recurrence ablation AI; AI recurrence HCC ablation; recurrence RFA HCC radiomic; recurrence ablation HCC radiomic; prediction ablation HCC AI; prediction ML model ablation HCC. The search was designed to maximize sensitivity and database coverage. No restrictions on the year of publication were applied, up to June 2025, and only articles published in English were considered. The following exclusion criteria were applied: Case reports, conference abstracts, reviews, technical notes, editorials, book chapters, surveys, and letters to the editor. Only peer-reviewed original articles were considered eligible.

Selection process

Two authors (Lippi M, Andreani EV) independently screened the titles and abstracts of all identified records to assess their eligibility for inclusion. Any disagreements regarding inclusion were resolved through a discussion with a third expert author (Barbieri P). Final inclusion decision and methodological consistency were independently verified and approved by a committee (Posa A, Iezzi R).

Studies were excluded if they focused on patients with malignancies other than HCC, included treatments other than heat-based thermal ablation, or assessed only clinical prediction models without AI or radiomics components.

Data items

The primary outcomes extracted from the included studies were: AUC, sensitivity, specificity, PPV, NPV, C-Index, and accuracy of the models in the context of predicting local recurrence of HCC after RFA or MWA.

Risk of bias and certainty of evidence

The GRADEpro Guideline Development Tool (GDT) was used to create table summaries of results in Cochrane systematic reviews, considering study limitations, imprecision, indirectness, inconsistency, and publication bias. Two authors independently assessed the risk of bias for each included study, evaluating limitations in study design, reporting quality, consistency of outcomes, and potential publication bias. Discrepancies were resolved through discussion until consensus was reached. In cases where consensus was not achieved, a third author acted as an arbiter.

RESULTS

Study selection

The literature search initially yielded 308 articles. After removal of duplicates, 138 articles remained for title and abstract screening. Based on the predefined inclusion and exclusion criteria, 113 records were excluded. A total of 25 full-text articles were assessed for eligibility. Among these, 3 studies were excluded for using contrast-enhanced ultrasound as imaging evaluation modality, 1 study was excluded for focusing exclusively on a clinical (non-imaging based) predictive model, 2 were reviews and thus excluded, 2 studies investigated treatment approaches other than RFA or MWA, 5 did not employ ML or DL models, 1 article included outcomes related to secondary liver lesions, 1 focused on post-treatment outcomes of other therapies. Ultimately, 10 studies met the inclusion criteria and were included in the final synthesis[12-21] (Table 1). The study selection process is illustrated in the PRISMA flow diagram (Figure 1).

Figure 1 PRISMA flow-chart.

Table 1 Characteristics of the included studies¹.

Ref.	Study type	Patient number	Model	AUC	Sensitivity	Specificity	C-index	PPV	NPV	Accuracy
Lim et al[12]	Retro	74	DL	0.99	96%	97%	N/A	0.91	N/A	97.6%
Sato et al[13]	Retro	1778	ML + C	N/A	N/A	N/A	0.67	N/A	N/A	N/A
Zhang et al[14]	Retro	90	ML + C	0.86	N/A	N/A	N/A	N/A	N/A	N/A
Tabari et al[15]	Retro	97	ML + R + C	0.83	82%	67%	N/A	0.69	0.80	N/A
Peng et al[16]	Retro	149	ML + R + C	N/A	N/A	N/A	0.72	N/A	N/A	N/A
Ren et al[17]	Retro	607	ML + C	0.89	N/A	85%	N/A	N/A	N/A	N/A
Chen et al[18]	Retro	417	DL + R + C	0.87	86%	91%	N/A	0.79	0.90	88%
Wang et al[19]	Retro	535	DL + R + C	0.79	72%	86%	N/A	N/A	N/A	78%
Kong and Li[20]	Retro	289	DL + R + C	0.74	71%	71%	N/A	0.39	0.91	71%
Li et al[21]	Retro	288	DL + R + C	0.98	N/A	N/A	N/A	0.94	0.87	91.6%

¹In case of more than one trained model, the best-performing one was reported.

Retro: Retrospective; DL: Deep learning; ML: Machine learning; R: Radiomic; C: Clinical; AUC: Area under the curve; PPV: Positive predictive value; NPV: Negative predictive value; N/A: Not applicable.

Study characteristics

The ten included studies, published between 2022 and 2025, involved a total of 4324 patients with HCC treated with heat-based thermal ablation techniques, either RFA or MWA. Specifically, all patients underwent RFA in 2 studies[13,14], all patients were treated with MWA in 1 study[17], both RFA and MWA were used in 4 studies[15,16,18,19], while the ablation modality was not specified in 3 studies[12,20,21]. All studies had a retrospective design, and 3/10 of them (30%) were multicentric[19-21]. Pre-procedural imaging (CT or MRI) was explicitly required in 6 studies, with imaging performed between 2 weeks and 3 months prior to treatment[14-16,18,19,21]. Follow-up was performed using CT or MRI. Five studies reported on post-treatment follow-up duration, with a minimum of 24 months[14,16-18,21].

Results of selected studies

All ten studies (100%) reported on the performance of AI-based models, either DL or ML, for predicting HCC recurrence following thermal ablation. The reporting of specific outcome metrics was as follows: 8/10 (80%) studies reported the models’ AUC, 5/10 their sensitivity (50%) and 6/10 their specificity (60%), 2/10 their C-index (20%), 5/10 their PPV (50%), 4/10 their NPV (40%), and 5/10 their accuracy (50%).

The overall median and mean AUC were 0.865 (range 0.74-0.99) and 0.869, respectively. The overall median and mean sensitivity were 82% (range 71%-96%) and 81.4%, respectively. The overall median and mean specificity were 85.5% (range 67%-97%) and 82.83%, respectively. The overall median and mean C-index were 0.735 (range 0.67-0.8) and 0.735, respectively. The overall median and mean PPV were 0.79 (range 0.39-0.94) and 0.744, respectively. The overall median and mean NPV were 0.885 (range 0.80-0.91) and 0.87, respectively. The overall median and mean accuracy were 88% (range 71%-97.6%) and 85.24%, respectively. When considering only studies on DL models, the median and mean AUC, sensitivity, specificity, PPV, NPV, and accuracy were 0.87 and 0.87, 79% and 81.3%, 88.5% and 86.3%, 0.85 and 0.76, 0.90 and 0.89, and 88% and 85.2%, respectively[12,18-21]. None of the studies investigated the C-index.

When considering only studies on ML models, the median and mean AUC, specificity, and C-index were 0.86 and 0.86, 76% and 76%, and 0.70 and 0.70, respectively[13-17]. Only one study reported sensitivity, PPV, and NPV[15]. None of the studies investigated the accuracy of their models[13-17].

Risk of bias and certainty of evidence

As all studies had a retrospective design and the majority were monocentric, quality assessment using the GradePRO GDT revealed a high degree of clinical and methodological heterogeneity among the included studies, such that quantitative synthesis was not useful. Therefore, meta-analysis was not performed. Additional concerns include the heterogeneity of AI models applied, variability in input data, and the inconsistency in reported outcome metrics. Given these limitations, the overall quality of evidence was considered low/very low (Table 2).

Table 2 Quality assessment of the included studies using the GradePRO tool.

Outcome	Number of studies	Study design	Risk of bias	Inconsistency	Indirectness	Imprecision	Certainty of evidence
AUC	8	Retro	Moderate	Low	Low	Moderate	Low
Accuracy	5	Retro	Moderate	High	Moderate	High	Very low
C-Index	2	Retro	Moderate	Moderate	Low	Moderate	Low
Sens/Spec	5	Retro	Moderate	Moderate	Low	Moderate	Low
PPV/NPV	4	Retro	Moderate-High	High	Moderate	High	Very low

AUC: Area under the curve; Sens: Sensitivity; Spec: Specificity; PPV: Positive predictive value; NPV: Negative predictive value; Retro: Retrospective.

DISCUSSION

A key objective in oncologic care is to accurately predict treatment outcomes based on pre-treatment data or imaging. This is particularly important for HCC, where early recurrence significantly compromises prognosis. Being able to stratify patients based on recurrence risk is essential for tailoring the most effective and individualized therapeutic strategies, as well as for optimizing follow-up protocols and resource allocation. Recent advances in AI have introduced the possibility of enhancing predictive accuracy beyond traditional clinical judgement. For example, Iseke et al[22] trained an ML model on predicting tumor recurrence after thermal ablation, surgical resection, and liver transplantation from pretreatment clinical characteristics, laboratory data, and imaging features from MRI examinations, obtaining fair to good performances[22].

This systematic review highlights the growing interest in and application of AI-based models for predicting early recurrence of HCC following RFA or MWA. The systematic review focuses on clinical relevance of reported AI performance metrics, and the included studies demonstrated a median sensitivity of 82%, 85.5% specificity, and 88% accuracy[12-21]. The median AUC of 0.88 indicates good discriminatory performance, with some models, such as those by Li et al[21]and Lim et al[12], exceeding an AUC of 0.9. When confronting DL and ML models, we found similar median and mean AUCs (0.87 and 0.87, and 0.86 and 0.86 for DL and ML respectively), while we found higher median and mean specificity values in the DL subgroup (88.5% and 86.3% vs 76% and 76%, respectively); however, this result is limited by the scarce number of studies included.

Sato et al[13] reported a C-index of 0.67 using a gradient boosting decision tree ML model, which is indicative of a poor performance, while Peng et al[16] reported a C-index of 0.72 using a random survival forest ML model[23].

In their retrospective monocentric study, Yin et al[24] further extended the potential utility of AI models by demonstrating high predictive performance (AUC of 0.97, sensitivity of 95%, specificity of 91%, and an accuracy of 92.7%), even when applied to a mixed population of HCC and secondary liver lesions. However, these results must be interpreted cautiously as this study involved a relatively small cohort.

Despite these promising findings, several limitations affect the overall strength of evidence, as all studies were retrospective, with only 30% being multicentric. Moreover, sample size and characteristics varied significantly between studies, and there was heterogeneity in AI model types, as well as in the proposed treatment (RFA, MWA, or both), and in the pre-procedural diagnostic imaging techniques used (MRI, CT, or both). In addition, the time gap between pre-treatment imaging and ablation procedure was not specified, although 60% of studies specified a pre-treatment imaging window ranging from 2 weeks to 3 months; however, this time-gap is too broad and could be a source of bias. This interval should ideally be minimized to avoid tumor growth or disease progression in the interval, which could affect model accuracy. Moreover, only 5/10 studies (50%) reported a minimum follow-up of 24 months post-treatment, which also limits the comparability of the studies.

The heterogeneity of reported outcome measures limits the clinical application of these AI-based approaches; in particular, the absence of clearly defined performance thresholds that are clinically meaningful hinders clinical adoption: While many studies report high AUC values, these do not directly translate into patient-centered outcomes unless accompanied by high sensitivity and specificity levels that could be sufficient to guide decision-making in real-world practice. In a diagnostic setting, a sensitivity threshold above 90% may be necessary to minimize false negatives, whereas in prognostic models, specificity and predictive values may carry greater weight. This variability in model performance across various studies, together with incomplete reporting of key outcome metrics, underscores the need for uniform reporting guidelines in AI-related clinical research and for standardized benchmarks to evaluate AI tools.

Moreover, some AI-based models could be better suited for specific clinical contexts. These limitations underline the need for standardized, prospective, multicenter validation trials with harmonized methodology, defined endpoints, and external validation to confirm the generalizability and utility of AI-based prediction models in this setting.

CONCLUSION

This systematic review suggests that AI-based models show promising performance in predicting early tumor recurrence following heat-based thermal ablation treatments with RFA or MWA in patients with HCC. These models have the potential to inform clinical decision-making by identifying non-responders before treatment allocation, supporting personalized and more effective therapeutic strategies.