This systematic review and meta-analysis adhered to the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Diagnostic Test Accuracy) guidelines [22], CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies) checklists, PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist, and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) checklist, and was prospectively registered on PROSPERO (CRD420251167155). These 4 checklists were provided in Checklists 1-4.

A comprehensive literature search was conducted across 4 major databases—PubMed (via NCBI), Embase (via Ovid), Web of Science (via Clarivate), and the Cochrane Library (via Wiley). The original search strategy was completed on July 11, 2025. During revision, the strategy was reconstructed to improve reproducibility and database-specific syntax; it was rerun on March 17, 2026, and the finalized strategy was last updated on April 17, 2026. Thus, July 11, 2025, refers to the initial search cutoff, whereas April 17, 2026, refers to the most recent rerun of the reconstructed search strategy. The reconstruction did not change the review question or eligibility criteria; it made the previously used AI-related terms, VETC-specific terms, and HCC-specific terms explicit for each database. To ensure comprehensiveness, no limits were placed on language or publication year. Search results from all databases were exported to EndNote X20 (Clarivate Analytics) for systematic deduplication, followed by manual verification. Moreover, 2 independent reviewers (XH and RF) performed initial screening of titles and abstracts, followed by full-text assessment.

Studies were included based on the PITROS framework:

Studies were systematically excluded based on the following comprehensive criteria: (1) titles and abstracts irrelevant to the review question; (2) nonoriginal literature types, including reviews, case reports, conference abstracts, meta-analyses, and editorial correspondence; (3) studies without AI algorithms; (4) research not evaluating VETC; (5) studies lacking essential diagnostic data; or (6) studies with overlapping patient data. For publications reporting more than 1 cohort, such as training, internal-validation, and external-validation cohorts, each cohort was treated as a separate dataset only when the patients were mutually exclusive. Cohorts with overlapping participants were not entered separately, in accordance with the principle that double counting occurs when the same participants or evidence are included more than once in a synthesis. The screening process was independently conducted by 2 reviewers (XH and RF), who first assessed title and abstract relevance, then evaluated full-text articles against inclusion and exclusion criteria. Any disagreements between reviewers were resolved through discussion, with a third reviewer (ZY) consulted if consensus could not be reached.

We used the latest PROBAST+AI quality assessment tool [23], which supersedes the 2019 PROBAST version. This comprehensive evaluation framework comprises 2 distinct phases—model development and model assessment, with each phase encompassing 7 critical domains addressing participants, data sources, predictive factors, outcome assessment, and analytical methodologies. The assessment categorizes each domain’s evaluation results into 3 levels—low (L), high (H), and unclear (U), determined through specific signaling questions. These questions were graded using a nuanced scale: “Yes” (Y), “Probably Yes” (PY), “Probably No” (PN), “No” (N), “No Information” (NI), and occasionally “Not Applicable” (NA).

To ensure methodological rigor and minimize potential bias, 2 independent reviewers (XH and RF) conducted a comprehensive risk of bias assessment using the PROBAST+AI tool. Any disagreements between reviewers were systematically resolved through in-depth discussion and meticulous analysis, ensuring high reliability and consistency in the final assessment. Detailed signaling questions and comprehensive evaluation tables are provided in Tables S2 and S3 in Multimedia Appendix 1.

We used the GRADE framework to assess the certainty of the evidence for sensitivity and specificity. This method evaluates 5 key domains—risk of bias, indirectness, inconsistency, imprecision, and small-study effects—to systematically rate the evidence quality. The detailed assessment criteria and protocols are fully documented in Table S4 in Multimedia Appendix 1 to ensure transparency.

The primary outcome measures encompassed sensitivity, specificity, and AUC for AI-based approaches across different imaging modalities (CEMRI, CECT, CEUS, and [¹⁸F] FDG PET/CT), alongside their early recurrence HR values. Sensitivity, calculated as true positive (TP)/(TP+false negative [FN]), quantifies the AI model’s ability to accurately identify true positive cases, while specificity, calculated as true negative (TN)/(TN+false positive [FP]), evaluates the model’s proficiency in correctly recognizing negative cases. AUC represents the comprehensive discriminative performance of the receiver operating characteristic (ROC) curve, providing a holistic assessment of the model’s diagnostic capabilities. For early recurrence HR values, we directly extracted reported HRs with their 95% CIs from the original studies. In cases where direct reporting was absent, we planned to indirectly calculate HR values using published survival curves through precise data extraction and reconstruction techniques. An aggregated HR>1 with statistical significance would indicate the AI model’s significant prognostic stratification capability, demonstrating that patients categorized as “high-risk” by the AI model have a multiplicative early recurrence risk compared with the “low-risk” group. Early recurrence was defined according to the original studies; definitions varied across studies (eg,≤1 y or ≤2 y after surgery). For studies presenting multiple nonoverlapping patient datasets, we considered the contingency tables independent and extracted them comprehensively. When multiple AI algorithms were reported, we selected the highest-performing model based on AUC for analysis.

Two independent reviewers (XH and RF) systematically extracted data from full-text articles to assess their eligibility, with any discrepancies resolved through consultation with a third reviewer (ZY). The comprehensive data extraction process encompassed multiple critical domains, including (1) patient and study baseline information: authors, publication year, country, study design, imaging modalities (CEMRI, CECT, CEUS, and [¹⁸F] FDG PET/CT), reference standard, analysis methodology, total patient numbers (training and internal and external validation sets), and patient counts with positive VETC; (2) technical parameters: research center, data source, AI methodology, AI model, optimal AI algorithms, data splitting approach, and TP, FP, FN, and TN across different validation sets; and (3) imaging modality information: manufacturer, scanner modality, evaluation time, regions of interest, sequence parameters, enhancement phases, contrast agents, and radiotracer details. Given that most studies did not provide complete diagnostic contingency tables, we reconstructed TP, FP, FN, and TN values by systematically analyzing sensitivity, specificity, reference standard positive VETC cases, and total patient populations.

For the diagnostic meta-analysis, we addressed the observed between-study heterogeneity by using a bivariate random-effects model using the restricted maximum likelihood (REML) method. To enhance the robustness of our pooled estimates and account for uncertainty in variance, the Hartung-Knapp-Sidik-Jonkman (HKSJ) adjustment was applied to calculate the pooled sensitivity and specificity along with their corresponding 95% CIs. We calculated 95% prediction intervals (PIs) for the main diagnostic syntheses and for modality-specific subgroup syntheses only when random-effects pooling was performed with sufficient independent datasets. These PIs were reported as descriptive estimates of how performance might vary in a future clinical setting and were not used alone to justify additional subgroup analyses. Because subgroup analyses were prespecified by imaging modality, subgroup PIs were intended to contextualize heterogeneity within clinically defined modality groups rather than to generate new post hoc subgroup analyses [24]. The overall diagnostic accuracy was summarized by the AUC derived from the hierarchical summary receiver operating characteristic model, and the interpretation of these values was adjusted to acknowledge the impact of the substantial heterogeneity identified. Pooled HRs for early recurrence were estimated using a single prespecified REML+HKSJ random-effects model.

For CEMRI-based studies with more than 10 studies, meta-regression and bivariate boxplots were performed to investigate potential sources of heterogeneity. The prespecified covariates for meta-regression included center (single vs multi), validation type (internal vs external), data splitting method (hold-out vs independent validation), AI model type (radiomics vs radiomics and clinical), AI algorithms (deep learning vs machine learning), and MRI field strength (1.5T&3.0T vs 3.0T). The impact of different algorithms on diagnostic performance was also assessed. A bubble plot visualized the diagnostic odds ratio (DOR) over time. The clinical utility of CEMRI-based models was appraised using Fagan’s nomogram, and small-study effects were assessed using Deeks’ funnel plot asymmetry test (P value<.05 for the slope coefficient, indicating potential bias). A chord diagram illustrated the frequency of different imaging modalities and algorithms. All analyses were conducted using Stata (version 15.1; StataCorp LLC; with the Midas and Metadta packages) and R (version 4.3.1; R Core Team; using ggplot2 and tidyverse). Several deviations from the original registered protocol should be noted; the HKSJ adjustment and the reporting of 95% PIs were not prespecified but were incorporated post hoc to strengthen methodological rigor. Additionally, certain meta-regression covariates, such as MRI field strength and AI algorithm type, were added based on clinical relevance identified during the review process. No changes were made to the original inclusion or exclusion criteria.

The initial database search identified 549 potentially relevant publications. Following the removal of 281 duplicate records, 268 articles proceeded to the preliminary screening phase. During this stage, 243 records were excluded based on title and abstract review, primarily due to clear irrelevance or nonconforming publication types. This resulted in 25 articles sought for full-text retrieval. Out of 25 articles, 1 was excluded as the full text was inaccessible. The remaining 24 articles underwent a detailed, full-text eligibility assessment. Of these, 9 studies were excluded due to insufficient or incomplete diagnostic data, with reasons including (1) without using AI algorithms, (2) nonevaluated VETC, (3) data (TP, FP, TN, and FN) not available, or (4) patients overlap. Consequently, 15 studies [6-8,12,24-34] met the predefined inclusion criteria and were incorporated into the final meta-analysis. The entire study selection process adhered to the PRISMA guidelines and is detailed in Figure 1.

A total of 15 eligible studies were included in this meta-analysis (Tables 1 and 2). The validation cohorts consisted of 12 studies [6-8,12,26-28,31-35] providing internal validation data (encompassing 729 patients) and 7 studies [6,7,27,29,30,32,36] providing external validation data (encompassing 613 patients). The distribution by primary imaging modality was as follows: 3 used CECT [6,27,36], 1 used [¹⁸F]FDG PET/CT [8], 1 used CEUS [12], and the remaining 10 studies used CEMRI [7,26,28-35]. All 15 studies, published between 2021 and 2025, were retrospective in design. The reference standard for confirming VETC status was pathological biopsy. With the exception of 1 study from Japan [35], all others originated from China. All analyses were conducted on a per-patient basis. When a publication reported training, internal-validation, and/or external-validation cohorts, these cohorts were extracted as separate datasets only if the original article indicated that the patients did not overlap.

The methodological quality of the included studies was appraised using the PROBAST+AI tool, with the detailed signaling-question tables provided in Tables S2 and S3 in Multimedia Appendix 1. For model development, the overall quality judgment was high-risk in 6.7% (1/15) of studies and low-risk in 93.3% (14/15). Applicability concerns for model development were high in 13.3% (2/15) of studies and low in 86.7% (13/15). For model validation, the overall risk of bias was high in 20% (3/15) of studies, unclear in 20% (3/15), and low in 60% (9/15). Applicability concerns for validation cohorts were low in all included studies. The main high-risk judgments were driven by analysis-related issues, whereas unclear judgments mainly reflected insufficient reporting of outcome assessment. These risk-of-bias findings are presented in the main manuscript together with the PRISMA flow diagram (Figure 1) and the prognostic forest plot (Figure 2), with full details retained in Multimedia Appendix 1.

The certainty of evidence was evaluated according to the GRADE framework and is now summarized in a main-text diagnostic Summary of Findings table. The table presents expected TP, FN, TN, and FP results per 1000 patients tested at the median VETC prevalence of 38%, together with the number of participants and studies, certainty ratings, and concise comments. Detailed domain judgments remain in Table S4 in Multimedia Appendix 1. Across outcomes, certainty ranged from low to moderate, with downgrading mainly driven by validation-related risk of bias, imprecision, and heterogeneity.

As illustrated in Figure 3B, the multilayer perceptron (MLP) algorithm demonstrated the highest sensitivity and specificity values at 0.89 (95% CI 0.26‐1.00) and 0.86 (95% CI 0.71‐0.97). Furthermore, Figure 3C indicates a slight decline in the DOR from 2021 to 2025, with random forest (RF) achieving relatively high DOR values while other algorithms maintained lower levels.

The results indicated that no significant threshold effect was observed for AI models based on CEMRI (Spearman ρ=−0.03, P=.93). The primary pooled analysis combined both internal and external validation datasets. For CEMRI-based AI models, the average pooled sensitivity was 0.84 (95% CI 0.73‐0.93; 95% PI 0.45‐1.00; τ²=0.03) with low certainty, and the average pooled specificity was 0.79 (95% CI 0.70‐0.86; 95% PI 0.50‐0.97; τ²=0.02) with low certainty. The 95% CIs describe uncertainty around the average pooled estimates, whereas the 95% PIs describe how accuracy may be distributed in a future setting; therefore, the wide CEMRI PIs indicate clinically important between-setting variability despite favorable average accuracy. For CECT-based AI models, sensitivity was 0.75 (95% CI 0.61‐0.86; 95% PI 0.62‐0.85; τ²<0.001) with moderate certainty, and specificity was 0.78 (95% CI 0.71‐0.83; 95% PI 0.69‐0.85; τ²=0) with moderate certainty. For [¹⁸F]FDG PET/CT and CEUS, only 1 study was available for each modality; therefore, their average estimates were not used to infer a distribution of effects across settings. The [¹⁸F]FDG PET/CT-based AI model had sensitivity and specificity of 0.74 (95% CI 0.52‐0.90), both with low certainty. The CEUS-based AI model had sensitivity of 0.81 (95% CI 0.54‐0.96) and specificity of 0.81 (95% CI 0.63‐0.93), both with low certainty.

The GRADE summary of findings, including the diagnostic summary of findings estimates and expected results per 1000 patients tested, is presented in Table 3.

As illustrated in Figure 2, regarding the prognostic value for predicting early recurrence, the pooled analysis of 10 nonoverlapping datasets from 7 studies [6,7,12,25-27,31] revealed an average HR of 2.34 (95% CI 1.93‐2.84) with the REML+HKSJ method. The CI reflects uncertainty around the average prognostic effect. Because between-study heterogeneity was very low (τ²<0.0001), the PI was similar to the CI, suggesting limited observed dispersion for the pooled HR; however, modality-specific HRs from single studies should not be interpreted as distributions across settings. In this survival-HR forest plot, Che et al [7] contributed 3 mutually exclusive patient cohorts (training, internal validation, and external validation), and Yang et al [27] contributed 2 mutually exclusive cohorts (training and external validation). Because these patient groups did not overlap, they were treated as separate datasets and did not constitute double-counting. When stratified by imaging modality, the HR was 2.32 (95% CI 2.04‐2.63; τ²=0) for CEMRI-based AI (8 datasets from 5 studies [7,25-27,31]), 8.76 (95% CI 3.10‐24.74) for CECT-based AI (1 study [6]), and 1.99 (95% CI 1.29‐3.06) for CEUS-based AI (1 study [12]). The extremely high HR observed in the CECT subgroup should be interpreted cautiously, as it was derived from only a single study with a wide CI. Similarly, the CEUS-based estimate was based on a single study.

The results of the meta-regression are illustrated in Table 4 and indicate that the heterogeneity in CEMRI-based AI models was primarily attributable to variations in center (specificity: P=.01); validation (specificity: P=.01); AI algorithms (specificity: P<.001); MRI field strength (specificity: P=.05). Specificity was significantly higher in single-center studies compared with multicenter studies (0.81 vs 0.75, P=.01). Similarly, internal validation cohorts demonstrated significantly higher specificity than external validation cohorts (0.81 vs 0.75, P=.01). Deep learning models achieved significantly higher specificity than machine learning models (0.83 vs 0.73, P<.001). Furthermore, studies using a 3.0T MRI field strength showed significantly higher specificity than those using a combination of 1.5T and 3.0T scanners (0.79 vs 0.77, P=.05).

it The bivariate box plot analysis identified the studies by Yu et al [26], Zhang et al [25], Yang et al [27], and Chu et al [32] as outliers, suggesting they may be potential sources of heterogeneity (Figure 4A).

Sensitivity analysis was performed by excluding the 4 outlier studies [25-27,32] to assess their impact on the pooled estimates for CEMRI-based AI models. After exclusion, the pooled sensitivity was 0.80 (95% CI 0.73‐0.87), specificity was 0.77 (95% CI 0.70‐0.84), and AUC was 0.85 (95% CI 0.82‐0.88), showing no substantial changes compared with the primary results. Additionally, a subgroup analysis was conducted by including only pure radiomics studies. In this analysis, the pooled sensitivity, specificity, and AUC for CEMRI-based models were 0.81 (95% CI 0.68‐0.92), 0.68 (95% CI 0.50‐0.84), and 0.75 (95% CI 0.71‐0.79), indicating a slight decrease in diagnostic performance. We also performed an analysis restricted to studies with external validation. In this analysis, the pooled sensitivity, specificity, and AUC of CEMRI-based models were 0.80 (95% CI 0.67‐0.90), 0.77 (95% CI 0.61‐0.90), and 0.84 (95% CI 0.80‐0.87), respectively.

Deeks’ funnel plot asymmetry test indicated no statistically significant small-study effects for CEMRI-based AI models (P=.08; Figure 4B). Furthermore, using the prespecified median prevalence of 38% as the pretest probability, the Fagan nomogram demonstrated a positive posttest probability of 71% and a negative posttest probability of 11% (Figure 4C).

Our meta-analysis demonstrated that AI models based on CEMRI yielded the highest sensitivity (0.84) and AUC (0.87) for predicting VETC in HCC, while models based on CEUS achieved the highest specificity (0.81). The superior sensitivity and AUC of CEMRI-based AI can be attributed to the ability of CEMRI to provide high soft-tissue contrast and detailed visualization of intratumoral heterogeneity and peritumoral vasculature [26]. This allows AI models to extract a large number of high-dimensional radiomic features, potentially maximizing the capture of subtle image phenotypes associated with the VETC pattern [35]. Conversely, the high specificity of CEUS-based AI likely stems from its capacity for real-time assessment of tumor hemodynamics and microvascular perfusion [36]. AI models can leverage this dynamic information to accurately identify the characteristic perfusion patterns of VETC-positive HCCs, thereby reducing false-positive diagnoses [37]. Furthermore, the typical enhancement patterns of non-VETC HCCs on CEUS may allow AI models to more reliably exclude VETC-negative cases [12]. However, as this finding was based on a single study, it should be interpreted with caution and requires validation in larger, independent cohorts.

Regarding the impact of different algorithms, our subgroup analysis revealed that the RF algorithm demonstrated the highest sensitivity (1.00), while both MLP and RF achieved the highest specificity values (0.86). The robust performance of RF, characterized by high sensitivity and specificity, can be explained by its ensemble nature. By constructing a multitude of decision trees and aggregating their predictions, RF effectively mitigates overfitting, often seen in single models, thereby enhancing its ability to correctly identify both positive and negative cases [38]. Furthermore, RF’s inherent capability to automatically assess feature importance allows it to select the most discriminative radiomics features for classification, improving its adaptability and discriminatory power when dealing with complex, high-dimensional medical imaging data [39]. However, the perfect sensitivity of 1.00 reported for the RF model should be interpreted with caution, as this finding may reflect study-specific characteristics, such as small sample sizes, overfitting, or the use of internal validation only. The high specificity attained by MLP, an artificial neural network, likely stems from its capacity to model complex, nonlinear relationships within the data. When the feature distribution of negative (non-VETC) samples is distinct, MLP can leverage its multiple hidden layers to learn highly specific decision boundaries, effectively separating them from positive cases and resulting in a low false-positive rate [40]. However, conclusions regarding MLP performance remain preliminary due to the limited number of studies.

Our pooled prognostic analysis, based on 10 nonoverlapping datasets from 7 studies [6,7,12,25-27,31], indicated that patients with HCC classified as VETC-positive by AI models had a significantly higher risk of early recurrence, with an HR of 2.34 (95% CI 1.93‐2.84). This suggests that AI-based VETC prediction offers prognostic capability for early recurrence risk. Notably, the AI model based on CECT yielded the highest HR of 8.76, although its wide 95% CI (3.10‐24.74) warrants caution in interpretation, likely attributable to the limited number of studies available for this modality. Given the small number of contributing studies, this estimate may be unstable. Previous research findings indicated that in the pathological classification, the VETC-positive group had an HR (95% CI 1.44‐2.07) for early recurrence [21], and the results of this study suggest that the AI model’s predictive capability for early recurrence risk is not inferior to that of traditional pathological risk classification.

A recent meta-analysis by Wang et al [41] specifically evaluated the performance of MRI-based radiomics for predicting VETC in HCC, reporting a sensitivity of 0.88, specificity of 0.86, and an AUC of 0.93. This meta-analysis, which included a broader scope of imaging modalities, found a comparatively lower pooled predictive performance for CEMRI-based AI. This discrepancy may be explained by our inclusion of a larger number of studies investigating MRI-based AI models, which provides a more generalizable, albeit slightly less optimistic, estimate of real-world performance. Our study extends the findings of Wang et al [41] by providing a comprehensive comparative analysis of AI performance across multiple imaging modalities, including CEUS, CECT, and [¹⁸F] FDG PET/CT, in addition to CEMRI. We further investigated the influence of different AI algorithms on predictive performance. Most significantly, our work provides a novel contribution by quantitatively synthesizing the prognostic value of AI-predicted VETC status, demonstrating that the AI model exhibits good prognostic capability for predicting early recurrence.

Significant heterogeneity was observed among the included studies, which is an acknowledged challenge in radiomics meta-analyses. We accounted for this inherent variability a priori by using a bivariate random-effects hierarchical summary ROC model with HKSJ adjustment for diagnostic metrics, and a prespecified REML+HKSJ random-effects model for the prognostic HR synthesis, with DerSimonian-Laird and uncorrected REML reported only as sensitivity analyses. Meta-regression and bivariate boxplot analysis identified several potential sources of this heterogeneity. The significantly higher specificity in single-center versus multicenter studies (0.81 vs 0.75, P=.01) may reflect the greater standardization in imaging protocols and patient populations within a single institution, whereas multicenter studies encompass variations in scanners, acquisition parameters, and patient demographics that can challenge model generalizability [42]. The higher specificity in internal validation cohorts compared with external validation sets (0.81 vs 0.75, P=.01) underscores the performance drop commonly encountered when models are applied to independent, external data with different distributions [43]. The superior specificity of deep learning models over machine learning models (0.83 vs 0.73, P<.001) may be attributed to their ability to automatically learn complex, hierarchical feature representations directly from images, potentially capturing more discriminative patterns to rule out negative cases [44]. Furthermore, studies using 3.0T MRI scanners showed higher specificity than those using a mix of 1.5T/3.0T devices (0.79 vs 0.77, P=.05), likely due to the superior signal-to-noise ratio and spatial resolution of 3.0T, leading to more consistent and detailed image data [45]. The bivariate boxplot also pinpointed specific studies as potential outliers contributing to heterogeneity [24,27,28,33].

The findings of this meta-analysis hold several implications for clinical practice. The identified strengths of specific imaging-modality and algorithm combinations—such as the high sensitivity of CEMRI-based AI and the high specificity of CEUS-based AI—could inform the development of noninvasive tools for preoperative VETC prediction and risk stratification [46]. The demonstrated prognostic value of AI-classified VETC status for early recurrence, which was comparable with pathological classification, suggests its potential role in guiding postoperative surveillance and adjuvant therapy decisions, thereby facilitating timely management of HCC [47]. It is crucial to emphasize that any implemented AI should function as a decision-support tool, augmenting rather than replacing clinician judgment [48]. A key consideration for clinical implementation is model generalizability; the fact that only 7 included studies [6,7,24,27,28,30,34] reported external validation performance underscores the need for future research to prioritize robust, multicenter external validation [42]. Current unimodal AI models also have inherent limitations, as they may not capture the comprehensive information available from a multiparametric imaging workup [43]. Future AI systems may need to integrate findings across multiple imaging modalities, correlate them with clinical context, and communicate synthesized insights [44]. Other barriers to widespread adoption include the scarcity of expert-annotated data, regulatory hurdles, and challenges regarding model interpretability and transparency [45]. The “black-box” nature of many complex models remains a significant obstacle to clinical trust and integration [49]. Advances in techniques, like few-shot learning, self-supervised models, and explainable AI, alongside collaborative platforms, are needed to build a robust and trustworthy AI ecosystem in radiology [50]. AI-assisted VETC prediction may be particularly beneficial for high-risk surgical candidates, such as patients with large tumors, borderline resectable disease, or elevated AFP levels [51].

Several limitations of this meta-analysis should be acknowledged when interpreting the results. First, the predominantly retrospective design and relatively small sample sizes of the included studies may introduce potential biases in patient selection and data collection. Therefore, large-scale, prospectively designed studies are warranted to validate these findings [52]. Second, to avoid potential patient overlap across multiple algorithms reported in a single study, we extracted data only for the best-performing algorithm from each publication. While this approach mitigates bias, it prevents a comprehensive evaluation of the performance spectrum across all developed algorithms and may present an overly optimistic view; future studies with independent test sets for all models are needed [53]. Third, the generalizability of our conclusions may be limited by the geographical distribution of the included literature. Specifically, 14 out of the 15 included studies (93.3%) were conducted in China, representing a significant geographic concentration within Asian populations. While these findings are robust for the studied cohorts, validation in more ethnically and geographically diverse populations outside of China is necessary to ensure the widespread applicability and diagnostic stability of these AI models across different health care systems and genetic backgrounds [54]. Fourth, the average pooled diagnostic effects should be distinguished from the distribution of effects across settings. The 95% CIs describe uncertainty around the average sensitivity or specificity, whereas the 95% PIs describe how performance may vary in future settings. For example, the CEMRI sensitivity CI was 0.73‐0.93, but the PI was 0.45‐1.00, indicating that substantial heterogeneity may remain despite favorable average performance, as recommended by Borenstein [55] for the interpretation of heterogeneity. Fifth, the GRADE certainty of evidence was low for most outcomes and moderate only for CECT sensitivity and specificity, mainly because of imprecision, validation-related risk of bias, and limited external validation. Given these limitations, the pooled results should be interpreted cautiously to avoid overstatement. Future studies with standardized imaging protocols, unified reporting guidelines, prospective designs, and geographically diverse external validation are needed to reduce heterogeneity and improve certainty. Finally, differences in the definition of early recurrence across studies may have influenced the pooled HR estimates [56], and other potential sources of heterogeneity—such as variations in patient demographics, tumor stage, and imaging acquisition protocols—may also contribute to the observed variability.

In conclusion, the available evidence indicates that AI models derived from contrast-enhanced imaging can identify VETC status with clinically meaningful average accuracy, particularly for CEMRI, and that model-predicted VETC positivity is associated with a higher risk of early recurrence. The main contribution of this synthesis is to define the current evidentiary boundary for preoperative VETC risk stratification; AI outputs may help clinicians recognize tumors with aggressive vascular architecture, select patients for intensified follow-up, and generate hypotheses for individualized perioperative planning. These applications remain conditional. Most included cohorts were retrospective and geographically concentrated, external validation was sparse, and uncertainty was amplified by heterogeneity, imprecision, and low-to-moderate GRADE certainty. Before routine clinical adoption, future multicenter prospective studies should use prespecified thresholds, standardized image acquisition and reporting, independent external validation, calibration assessment, and evaluations of whether AI-informed decisions improve patient outcomes. Until such evidence is available, AI-based VETC prediction should be used as investigational support within multidisciplinary assessment rather than as a standalone basis for management.