This meta-analysis was prospectively registered with the PROSPERO (CRD42024613733). This systematic review and meta-analysis of diagnostic test accuracy was reported in accordance with the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Diagnostic Test Accuracy studies) [18] guidelines in Checklist 1. Due to the absence of subject information collection and its lack of impact on clinical diagnosis and treatment, ethical approval and informed consent were waived.

We conducted a systematic literature search in accordance with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for literature searches; completed checklist is available in Checklist 2) [19]. Relevant English-language publications were retrieved from PubMed, Web of Science, the Cochrane Library, and Embase, with the search covering all records up to October 16, 2025. The search strategy used both MeSH (Medical Subject Headings) and free-text keywords, including MeSH terms such as “Carcinoma, Hepatocellular,” “liver cell carcinoma,” and “deep learning.” Boolean operators were used to integrate MeSH terms and free-text terms, constructing tailored search queries for each database. Furthermore, the reference lists of identified review articles were manually screened to locate any additional eligible studies. No prospective study registries were searched, and no attempts were made to obtain unpublished data or to contact study authors for further information. The complete search strings for each database are provided in Table S1 in Multimedia Appendix 1.

Retrieved articles were imported into EndNote. After removing duplicates, the remaining articles were reviewed by title and abstract to identify initially eligible articles. Then, the full texts were downloaded and screened to determine the final eligible articles. Two researchers (WF and BQ, with 6 and 4 years of meta-analysis experience, respectively) performed the review independently. Interresearcher agreement during the literature screening was assessed using the κ coefficient (κ=0.93). Any disagreements were resolved in a consensus meeting with a third researcher (SH, with 10 years of experience in meta-analysis).

Before data extraction, a standardized spreadsheet was generated. The content to be extracted included the following: publication year, patient source, author, image source, manual segmentation, number of patients with MVI, total number of patients with HCC, number of patients with MVI in the training set, number of patients with HCC in the training set, validation set generation method, number of patients with MVI in the validation set, number of patients with HCC in the validation set, confusion matrix, sensitivity, specificity, precision, and accuracy.

Two researchers (WF and BQ, with 14 and 16 years of medical experience, respectively) carried out the data extraction independently. They then cross-checked their results. Any discrepancies were resolved through consultation with a third researcher (SH, with 13 years of medical experience).

The Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool was used to evaluate the overall risk of bias (ROB) and applicability of the eligible studies. The QUADAS-2 instrument encompasses 4 domains: index test, patient selection, flow and timing, and reference standard. Each domain includes specific questions. Answers to these questions are categorized as “yes,” “no,” or “unclear.” These answers correspond to ROB ratings of “unclear,” “high,” or “low.” If all key questions within a domain received a “yes” answer, the domain was rated as having a low ROB. If any key question received a “no” answer, a potential ROB was indicated, and the researcher judged the ROB according to established guidelines. An “unclear” rating was assigned when the literature did not provide sufficient detail for the researcher to make a judgment.

Two researchers completed the QUADAS-2 assessment independently. Any discrepancies were resolved through discussion with a third researcher.

A meta-analysis was performed using a bivariate mixed-effects model based on diagnostic 2×2 tables [20]. For studies that did not report these tables directly, we derived them from the available specificity, sensitivity, positive predictive value, accuracy, F₁-score, and case numbers. Throughout the analysis, the explicitly defined independent validation cohort from each study served as the unit of analysis. Each data point corresponded to distinct and nonoverlapping patient samples, which ensured the independence of the pooled results and mitigated potential bias from data reuse. Using the bivariate mixed-effects model, we computed the pooled estimates for specificity, sensitivity, negative likelihood ratio (LR–), positive likelihood ratio (LR+), diagnostic odds ratio (DOR), and area under the summary receiver operating characteristic (SROC) curve, along with their 95% CIs [20]. The Spearman correlation coefficient was used to evaluate the threshold effect and its contribution to between-study heterogeneity. Small-study effects were assessed using Deeks funnel plot asymmetry test. For subgroups with fewer than 10 studies, a Doi plot was used to informally assess publication bias. The degree of bias was determined based on the absolute value of the Luis Furuya-Kanamori (LFK) index. A value less than 1 suggests minor publication bias, a value between 1 and 2 indicates moderate publication bias, and a value exceeding 2 suggests substantial publication bias. During the meta-analysis, the validation set was used, and subgroup analyses were performed according to the validation set generation method and image type. A P value less than .05 was considered statistically significant.

Database retrieval yielded 11,266 articles, of which 7539 remained after duplicate removal. Following title and abstract screening, 62 articles were selected for a full-text review. The full-text review subsequently excluded 10 records: 2 non-DL studies, 3 studies with insufficient data to construct diagnostic 2×2 tables, and 5 studies that used DL solely for medical image segmentation without establishing HCC MVI prediction models. Finally, 52 articles [14,16,21-70] met the eligibility criteria. The detailed process is illustrated in Figure 1.

This meta-analysis included 52 studies published between 2019 and 2025. All of the studies used histopathological diagnosis as the gold standard for MVI. All 52 studies were case-control studies, encompassing 19,531 individuals with HCC, of whom 8161 were MVI cases. Regarding population source, 32 (61.5%) of the studies were single-center, 19 (36.5%) studies were multicenter, and 1 (1.9%) study was based on a registry database.

These studies primarily used single-modality imaging techniques for image modeling methods: contrast-enhanced computed tomography (CECT, n=16), contrast-enhanced magnetic resonance imaging (CEMRI, n=19), MRI (n=5), contrast-enhanced ultrasound (CEUS, n=5), and pathological sections (n=2). Five additional studies integrated multimodal imaging (computed tomography + positron emission tomography: 1, CECT +CEMRI: 4) for modeling. Regarding image segmentation methods, 38 of the 52 (73.1%) studies used manual segmentation, while the remaining studies used fully automatic segmentation (n=5) or semiautomatic segmentation (n=9).

Concerning model validation strategies, 23 studies used randomly sampled internal validation, 15 used performance through cross-validation, and 14 used external validation with independent cohorts (Table 1).

Regarding patient selection, all eligible studies included consecutive or random cases. According to the QUADAS-2 assessment criteria, this study design carries an inherent high ROB in the “Patient Selection” domain. Consequently, all included studies received a “high” rating for ROB in this domain.

Due to the adoption of supervised DL, model training was based on clear pathological outcomes. However, since DL models predict by extracting inherent image features, rather than directly relying on clinical covariates, their training process’s dependence on known outcomes did not result in diagnostic information leakage. This led to a low ROB. Manual segmentation was used in 38 studies, which could have introduced operator subjectivity and led to a high ROB.

Regarding the implementation of the gold standard, all studies used histopathological diagnosis as the gold standard for MVI, ensuring the objectivity and consistency of disease classification. This indicated a low ROB in the implementation of the gold standard.

Regarding the item “the match between the conduct and interpretation of the index test and the review question” in the QUADAS-2 scale, 5 (9.6%) out of the 52 studies did not directly report specificity values. Since specificity is a key indicator for verifying the match between a DL model and a clinical question, the absence of such data may lead to an incomplete assessment of a model’s diagnostic efficacy and weaken the reliability of a study’s conclusions. Therefore, these 5 studies were determined to have a high ROB. The remaining 47 (90.4%) studies fully reported diagnostic performance indicators, ensuring the transparency in test conduct and interpretation, and indicating a low ROB.

There was a reasonable and appropriate time interval between imaging examinations and pathological diagnoses in all eligible studies. Therefore, there did not appear to be a significant impact on the cases’ process (Figure 2).

The current meta-analysis revealed that the modeling methods used for DL-based prediction of HCC MVI primarily used CECT, CEUS, CEMRI, MRI, multimodal imaging, and pathological image techniques. DL models based on medical imaging showed favorable overall diagnostic performance in predicting HCC MVI, with a pooled sensitivity of 0.80 (95% CI 0.78‐0.83) and specificity of 0.82 (95% CI 0.80‐0.85). Further analysis across imaging modalities revealed that CECT-based models achieved the highest diagnostic efficacy, showing a sensitivity of 0.84 (95% CI 0.79‐0.88) and specificity of 0.83 (95% CI 0.77‐0.88). Models based on CEUS exhibited particularly high specificity (0.88, 95% CI 0.82‐0.92). Furthermore, models using pathological slides, considered the diagnostic reference standard, attained the highest overall performance, with a sensitivity of 0.91 (95% CI 0.87‐0.94) and specificity of 0.90 (95% CI 0.68‐0.97). Therefore, these models appear promising as a diagnostic approach for MVI in HCC.

This study also noted that some researchers have discussed the use of machine learning for MVI in HCC. Xiao et al [71] and Liang et al [72], who focused on MRI and ultrasound radiomics, respectively, validated the predictive potential of single modalities, with pooled area under the curves (AUCs) of 0.87 and 0.81, respectively. However, their analyses included a limited number of studies and focused only on a single imaging modality, limiting the generalizability of their conclusions. Li et al [73] integrated multiple imaging modalities across 22 studies (involving 4129 participants), reporting a pooled AUC of 0.90 for radiomic models. However, their analysis primarily incorporated traditional machine learning models and did not stratify performance based on the validation set generation method. This limitation may lead to an overly optimistic assessment of model generalizability.

Unlike previous systematic reviews that focused on traditional machine learning models, our meta-analysis focuses on the value of DL algorithms for diagnosing HCC MVI. Methodologically, this study provides a deeper exploration of the sources of performance heterogeneity through subgroup analyses based on the validation set generation method and image source. A particular methodological strength is the inclusion of pathological sections, the diagnostic gold standard, as a benchmark for performance optimization. This meta-analysis synthesized data from 52 studies involving 19,531 patients and provided more robust and reliable conclusions than those from analyses with smaller sample sizes.

The imaging modalities used to construct DL models can be categorized into 2 main types: noninvasive and invasive. This study showed that the properties of different modalities and their ability to extract biological features directly affected the diagnostic efficacy of corresponding models.

Among noninvasive imaging modalities, CECT can effectively capture tumor heterogeneity, enhancement patterns, and peritumoral microenvironment changes due to its high spatial resolution and multi-phase dynamic imaging capabilities. These features support the superior diagnostic performance of DL models (AUC=0.90) [74,75]. Models developed from CEUS achieved commendable performance (AUC=0.89) and exhibited a high pooled specificity (0.88). This finding suggests that CEUS’s real-time hemodynamic properties may be valuable in ruling out MVI-negative cases. However, its comparatively lower sensitivity (0.70) concurrently suggests ongoing challenges in consistently identifying MVI-positive features. Integrating Sonazoid-based functional imaging with conventional ultrasound characteristics and serum markers in the future may be a promising way to improve performance [76]. CEMRI has unique advantages in depicting tumor boundaries and detecting subtle changes in the peritumoral liver parenchyma due to its excellent soft tissue contrast [77]. The model’s performance (sensitivity: 0.78, specificity: 0.81, AUC=0.86) surpasses the assessment that relies on traditional visual features, as reported by Wu et al [78] (sensitivity: 0.55, specificity: 0.87, AUC=0.80). However, its widespread adoption is limited by long examination times, high costs, and dependence on equipment. Nonenhanced MRI models also demonstrate predictive potential (sensitivity: 0.76, specificity: 0.80, AUC=0.84). Notably, multimodal fusion models, which are designed to integrate complementary information, have not demonstrated significant advantages in the studies included (AUC=0.82). This may be due to the simplified fusion strategies or data heterogeneity.

Among invasive imaging modalities, DL models based on pathological sections demonstrated the highest diagnostic efficacy in this study (AUC=0.92). This highlights the gold-standard status of pathology in evaluating MVI. However, the inherent invasiveness of pathological examination precludes its use in preoperative decision-making. Therefore, a central challenge for future research is to effectively transfer and integrate the gold-standard-level diagnostic insights from pathological sections into preoperative, noninvasive imaging. This objective requires leveraging advanced methodologies, such as transfer learning, cross-modal fusion, and generative models. Augmenting existing advantageous modalities, such as CECT, with these advancements could ultimately pave the way for a clinically viable preoperative “virtual biopsy.”

Accurate image segmentation is essential for building reliable DL models. However, the literature incorporated in this meta-analysis primarily relies on manual segmentation. This method can introduce subjective bias (38/52, 73.1%). While advanced network architectures have achieved expert-level precision in segmenting normal liver tissue, as demonstrated by Dice coefficients ranging from 0.968 to 0.982 [79,80], comparative data from the Liver Tumor Segmentation challenge reveal significant discrepancies. Specifically, the accuracy of liver tumor segmentation (Dice: 0.739) is considerably lower than that of liver parenchyma segmentation (Dice: 0.963) [81]. This discrepancy primarily stems from the heterogeneity of HCC lesions, suboptimal image contrast, and a lack of high-quality annotated data. Concurrently, selecting the appropriate segmentation strategy is critical. While 3D segmentation comprehensively captures spatial heterogeneity, the more clinically feasible 2D approach sacrifices substantial volumetric information [82]. Given the demand for submillimeter-level precision in MVI prediction, there are 2 key technological pathways for enhancing model stability. First, exploring segmentation paradigms that require fewer annotations, such as weakly or self-supervised learning, can reduce dependency on annotations. Second, developing novel network architectures designed specifically to address HCC heterogeneity and boundary ambiguity is equally crucial.

The rigor of validation strategies is paramount for evaluating the real-world generalizability of DL models in predicting MVI of HCC [83,84]. The present analysis reveals that, despite exemplary performance during internal validation, consistent performance declines emerge in independent external validation cohorts. This finding clearly shows that relying too much on internal validation can lead to overestimating a model’s true efficacy [85]. Concern regarding validation strategies is not unique to DL research. A recent meta-analysis [86] that focused on MRI-based radiomics for predicting HCC recurrence and MVI similarly concluded that the current predominance of internal validation results in an overestimation of model generalizability as well. Consequently, any proclaimed superior performance may substantially diminish when confronting real-world heterogeneity, if the evaluation framework remains confined to internal validation, irrespective of the underlying algorithm—be it DL or radiomics. Therefore, promoting rigorous external validation and establishing standardized, cross-institutional imaging protocols are essential steps toward reliable clinical translation in this field.

There is substantial heterogeneity among the included studies. Notably, high levels of heterogeneity persist within subgroups, even after stratifying analyses by imaging modality and validation strategy. This observation objectively reflects the inherent complexity of artificial intelligence (AI)–based medical imaging research and is a common challenge for meta-analyses in this field. The heterogeneity primarily stems from 3 levels. Technically, variations in critical parameters, including imaging equipment, magnetic field strength, and slice thickness, directly influence image texture and quality. These variations are a significant technical source of variability in model performance. Methodologically, diversity in study design, validation strategies, segmentation techniques, and DL network architectures introduces additional variation in model construction and performance interpretation. Clinically, differences in patient populations regarding geographic distribution, underlying liver disease etiology, and disease stage may also affect model performance and generalizability. While this heterogeneity limits the direct interpretability of the pooled results to some extent, it accurately reflects the diversity in methodology and clinical practice within the field. Future investigations should adhere to the Findable, Accessible, Interoperable, and Reusable principles, providing detailed reporting of imaging acquisition parameters, model architectures, and training specifics. Such comprehensive reporting will facilitate the in-depth exploration of heterogeneity sources via methods such as meta-regression. This will promote the identification of key influencing factors and the standardization of methodologies.

This systematic review used a composite strategy to assess the ROB. First, all included studies were rigorously evaluated according to the QUADAS-2 guidelines. The results showed that all primary studies were rated as having a high ROB in the “Patient Selection” domain of QUADAS-2 due to the widespread use of the retrospective case-control design. While this outcome aligns with the QUADAS-2 assessment principles, it also reveals a limitation of the tool when evaluating machine learning-based diagnostic studies that use retrospective data. The tool struggles to differentiate nuances in data construction quality among studies.

To conduct a more granular assessment of data-level bias risk, a supplemental analysis was performed using items from the Quality Assessment of Diagnostic Accuracy Studies for Artificial Intelligence (QUADAS-AI) tool targeting “Study Participant Selection.” The QUADAS-AI tool provides more detailed criteria for this dimension, including an explicit description of data source, size, and quality characteristics, use of open-source datasets, a clear rationale for splitting data into training, validation, and test sets, performance of image preprocessing, and provision of scanner model information. The analysis using QUADAS-AI items showed that all 52 studies appropriately described the source, size, and quality of the input data and clearly defined the patient inclusion criteria. Among these studies, only one used an open-source dataset. All studies provided a rationale for the data split. Image preprocessing was performed in all studies. However, 11 studies did not report the scanner model used for image acquisition (Table S2 in Multimedia Appendix 15 [14,16,21-70]).

A primary strength of this research is that it is the first large-scale, systematic meta-analysis to evaluate medical imaging-based DL models for predicting MVI in HCC. The analysis included 52 studies with 19,531 patients, providing the field with comprehensive evidence. The analytical process strictly adhered to the PRISMA guidelines, and bias risk was evaluated using QUADAS-2. These measures ensured methodological rigor and transparency. However, several limitations warrant consideration. First, the training cohorts in most primary studies were small. Only 3 studies had a sample size greater than 1000 patients. The robustness of DL models depends heavily on large volumes of high-quality data. Therefore, restricted training sample sizes are a potential methodological limitation. This likely contributes to overfitting in some models, which is probably an internal reason for the observed performance degradation during external validation. This finding underscores the fundamental importance of acquiring large-scale, high-quality datasets to enhance model generalizability in the development of current DL models. Second, the 95% CIs for the pooled AUCs and for most subgroup analyses were exceptionally wide. This constrained the interpretation of the result precision to some extent. This primarily stems from the significant heterogeneity among the included studies. The limited number of studies in subgroup analyses exacerbated data sparsity. Third, Doi plot analyses for some subgroups indicated moderate to substantial publication bias. These analyses showed that the pooled results for these subgroups may be influenced by unpublished negative studies, which poses a risk of overestimating diagnostic performance. Fourth, the rigor of the validation strategies needs to be improved. Most studies relied on internal validation. Only a few conducted stringent external validation. This reliance may introduce an optimistic bias into the overall assessment of the models’ real-world generalizability. While the overall performance on external validation sets was discussed, the limited quantity of external validation data prevented a more in-depth subgroup analysis of validation strategies by different imaging modalities. Fifth, the vast majority of primary studies inadequately reported model calibration metrics or details about network complexity. This omission hinders a quantitative, systematic evaluation of predictive reliability and overfitting risk at the review level and impacts the comprehensive assessment of clinical applicability. This reflects a general deficiency in the transparency of methodological reporting within the current research landscape.

This study indicates that medical imaging-based DL models, particularly those using preoperative CECT, demonstrate promising diagnostic performance in predicting MVI of HCC. These models have the potential to assist in personalized surgical planning. However, translating them into clinical practice faces multiple challenges. One primary issue is establishing clinical decision thresholds. While this analysis quantified predictive probabilities via Fagan nomograms, there is currently a lack of evidence-based guidelines defining “at what predicted probability of MVI the surgical margin should be adjusted.” Future work must integrate clinical outcome data and empirically explore the net benefit of different thresholds using methods such as decision curve analysis. Second, these models’ generalizability and deployment feasibility need urgent enhancement. Performance degradation during external validation suggests susceptibility to variations in imaging protocols, equipment, and patient populations. Furthermore, model interpretability is crucial for gaining clinical trust. This necessitates developing transparent methods for presenting decision rationale.

To bridge the gap between “high performance” and “high utility,” future efforts must focus on 3 interconnected levels. At the research level, the focus must shift from model construction to rigorous, prospective, multicenter validation to unequivocally assess generalizability. At the algorithmic level, it is essential to explore cross-modal information fusion, especially using transfer learning, to bring “gold standard”–level diagnostic insights from histopathological sections to preoperative, noninvasive imaging. At the clinical level, establishing standardized, cross-institutional imaging protocols and developing decision support systems that integrate seamlessly into clinical workflows is imperative. This integrated approach is vital for reliably translating technology from innovation to tangible patient benefit.

This systematic review and meta-analysis demonstrate that medical imaging-based DL models, especially those leveraging preoperative CECT, hold significant promise for the noninvasive preoperative prediction of MVI in HCC. Unlike previous reviews that focused on radiomics or single imaging modalities, this study conducted a comprehensive comparison across multiple modalities. The study also emphasizes the critical role of external validation in the real-world generalizability of a model. However, substantial heterogeneity across studies and the performance degradation observed during independent external validation suggest that their generalizability to the real world must be confirmed through more rigorous study designs. Consequently, future research should prioritize establishing model robustness via prospective, multicenter external validation, coupled with efforts to standardize methodologies and improve reporting transparency. A critical step toward reliable clinical translation and achieving the ultimate goal of a “virtual biopsy” is developing algorithms that can translate pathology-grade diagnostic insights into preoperative, noninvasive imaging.