A Meta-Analysis of Radiomics-Based Models for Preoperative Assessment of Ki-67 Status in Hepatocellular Carcinoma

Summary

This meta-analysis of 17 studies (1,708 patients with hepatocellular carcinoma) found pooled sensitivity, specificity, and area under the curve values of 0.87, 0.79, and 0.90, respectively, for radiomic prediction of high Ki-67 expression; substantial heterogeneity indicates a need for standardized research.

Abstract

This study aimed to evaluate the diagnostic value of radiomic features in predicting Ki-67 expression levels in hepatocellular carcinoma (HCC) through a meta-analysis. Electronic databases, including PubMed, Web of Science, the Cochrane Library, and Embase, were systematically searched for relevant clinical studies published through 20 August 2025. Studies using radiomic features to predict Ki-67 expression levels in patients with HCC were included. Sensitivity, specificity, and summary receiver operating characteristic curves were evaluated, and the area under the curve (AUC) was calculated. A total of 17 studies involving 1,708 patients with HCC were included. The pooled sensitivity was 0.87 (95% confidence interval [CI], 0.81–0.91), the pooled specificity was 0.79 (95% CI, 0.71–0.85), and the overall AUC was 0.90 (95% CI, 0.87–0.93). The pooled AUC values for Ki-67 cutoff values of 10% and >10% were 0.89 (95% CI, 0.86–0.92) and 0.90 (95% CI, 0.87–0.92), respectively. The AUC values for magnetic resonance imaging- and ultrasound-derived radiomic features were 0.88 (95% CI, 0.85–0.91) and 0.92 (95% CI, 0.89–0.94), respectively. The AUC for prediction models based on logistic regression was 0.89 (95% CI, 0.86–0.92). Radiomic features showed promising pooled diagnostic performance for predicting high Ki-67 expression in HCC. However, substantial heterogeneity was present among the included studies. Further standardized research is needed to validate these findings.

Introduction

Over the past decade, the incidence and prevalence of liver cancer have progressively increased, making it a significant and growing disease burden worldwide. Specifically, hepatocellular carcinoma (HCC) accounts for 90% of all primary liver cancer diagnoses and represents the fourth most common cause of cancer-related deaths globally, with 840,000 new cases reported in 2018^1,2. Despite recent advances in surgical techniques, local interventions, and systemic therapies, the long-term prognosis of patients with HCC remains suboptimal, with 5-year overall survival rates ranging between 13% and 36%^3,4,5. This is largely attributable to its insidious onset, difficulties in early diagnosis, and high tumor biological heterogeneity. These challenges underscore a critical limitation of current diagnostic strategies: traditional histopathological assessments are invasive and unable to provide real-time, comprehensive evaluation of dynamic tumor biology. Therefore, the development of novel, noninvasive tools that can accurately quantify the intrinsic malignant potential of tumors is important for advancing personalized medicine in HCC.

Among various prognostic molecular biomarkers, Ki-67 is a well-established indicator of cellular proliferation and is closely associated with prognosis in multiple malignancies. Ki-67 reflects proliferative activity and is widely expressed in tumor cells; its high expression is strongly correlated with increased invasiveness and poor differentiation, serving as a predictor of tumor aggressiveness⁶. In HCC, multiple studies have confirmed that high Ki-67 expression is significantly associated with larger tumor size, vascular invasion, higher Edmondson–Steiner grade, and advanced tumor node metastasis stage. Moreover, it serves as an independent predictor of postoperative recurrence and worse overall survival^7,8,9,10. As Ki-67 expression levels mirror the proliferative state of tumor cells—with elevated levels indicating greater invasiveness and more rapid disease progression—this biomarker holds considerable value in prognostic stratification, treatment decision-making, and therapeutic monitoring in HCC. However, current clinical assessment of Ki-67 relies on invasive tissue sampling. This approach not only carries procedural risks and patient discomfort but, more importantly, given the substantial spatial and temporal heterogeneity of HCC, a single biopsy may fail to represent the overall proliferative status of the tumor, potentially leading to clinical misinterpretation¹¹. Thus, there is an urgent clinical need for a noninvasive, reproducible method that can comprehensively capture the spatial heterogeneity of Ki-67 expression.

The emergence of radiomics offers a potential approach to address this limitation. Radiomics involves the high-throughput extraction of numerous quantitative features from standard medical images, such as computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET), and the construction of predictive models using machine learning or deep learning algorithms to decode underlying tumor biology¹². In HCC research, radiomics has been successfully applied in tumor grading, prediction of microvascular invasion, assessment of gene expression, and treatment response monitoring. Although several studies have attempted to develop radiomic models for predicting Ki-67 expression^13,14, there remains considerable variability across studies in sample size, imaging modalities, feature extraction methods, and model performance. Therefore, a systematic meta-analysis is required to quantitatively synthesize these heterogeneous findings, establish robust pooled diagnostic estimates, and identify key sources of clinical and methodological variation, thereby providing an evidence-based foundation for future standardization and clinical translation. In practice, the clinical translation of radiomics is challenged by issues with feature reproducibility and modality-specific limitations, such as MRI’s dependence on scanning parameters and ultrasound’s dependence on the operator. Consequently, the clinical utility of these models must be critically appraised based on their methodological standardization and external validation status. The clinical applicability and robustness of these models have yet to be systematically evaluated. Therefore, this study comprehensively assesses the diagnostic performance of radiomic models in predicting Ki-67 expression status in HCC, thereby providing an evidence-based foundation for future standardization and clinical translation.

Protocol

Search strategy and information source
In accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement¹⁵, a systematic search was performed across PubMed, Web of Science, Cochrane Library, and Embase from database inception to 20 August 2025, without language restrictions (see Supplemental File 1). The search strategy was constructed by grouping keywords and subject headings (such as Medical Subject Headings [MeSH] in PubMed and Emtree in Embase) into three logical blocks using the Boolean operators ‘AND’ and ‘OR’. These blocks were applied to the title, abstract or topic fields in each database as follows: (1) radiomics, models and imaging: ‘radiomics’, ‘texture analysis’, ‘image features’, ‘magnetic resonance imaging’ (or ‘MRI’), ‘tomography, X-ray computed’ (or ‘CT’), ‘ultrasonography’ (or ‘ultrasound’), ‘deep learning’, ‘machine learning’, and ‘logistic’; (2) disease: ‘hepatocellular carcinoma’, ‘liver cancer’, ‘HCC’ and ‘hepatoma’; and (3) antigen: ‘Ki-67 Antigen’, ‘Ki-67’, ‘Ki67’, and ‘proliferation index’. The exact search syntax was adjusted to fit the specific rules of each database (e.g. using [Title/Abstract] and [MeSH] tags in PubMed, and ‘:ti,ab’ or ‘/exp’ tags in Embase). A manual search of the bibliographies of the retrieved studies, together with relevant review articles, was also performed to identify any additional eligible publications. The study protocol has been retrospectively registered in the INPLASY database under the registration number INPLASY202670061.

Eligibility criteria
To ensure a clinically relevant study selection, the eligibility of retrieved studies was determined using the PICOS (Population, Intervention, Comparison, Outcome, and Study design) framework.

The inclusion criteria were as follows: (1) study population consisting of patients diagnosed with HCC based on histopathological examination of resected/biopsied tissue, or according to established clinical and imaging guidelines; (2) use of radiomics, deep learning or texture features derived from medical imaging modalities (including CT, MRI or ultrasonography) to predict tumor Ki-67 expression levels; (3) histopathological evaluation of Ki-67 expression levels on tumor specimens, with a clearly defined cutoff value to classify patients into high-expression and low-expression groups; (4) availability of diagnostic outcomes in the form of a 2 × 2 contingency table enabling calculation of test performance; and (5) original, peer-reviewed clinical studies.

Studies were excluded if they (1) were conference abstracts, case reports, systematic reviews, or other non-original research; (2) contained insufficient outcome data for quantitative analysis; (3) were duplicate publications; and (4) were not available in full text.

Study screening and data extraction protocol
To execute the literature screening, all retrieved citations from the electronic databases were merged, and duplicate records were systematically identified and removed. Subsequently, two investigators independently screened the remaining literature based on the predefined inclusion and exclusion criteria. Initially, titles and abstracts were reviewed, and potentially eligible studies underwent full-text assessment. Any discrepancies between the two reviewers were resolved through discussion or by consulting a third researcher to reach a consensus.

After completing the screening process, the same two investigators independently extracted data using a standardized, predefined extraction form. The extracted information included basic study characteristics, demographic details of the study population, and diagnostic performance data, specifically, the number of true positives, false positives, true negatives, and false negatives. For studies with missing data or unresolved ambiguities, we contacted the corresponding authors via email to request the original data and excluded studies if the required data remained unavailable.

Quality assessment and risk-of-bias evaluation
The quality of the included diagnostic studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool, developed by Whiting et al. at the University of York¹⁶. This instrument assesses four domains: patient selection, index test, reference standard, and flow and timing. Each domain was rated for risk of bias as “high,” “low,” or “unclear.” Two investigators independently conducted the quality assessment. Any discrepancies were resolved through discussion or by consulting a third researcher to reach consensus.

Statistical analysis
Statistical analyses were conducted using Stata version 16.0 and Review Manager (RevMan) version 5.3. RevMan was used to evaluate methodological quality, assess the risk of bias, and generate the corresponding visual summaries. Diagnostic accuracy was analyzed in Stata 16.0 with the hierarchical midas module. Pooled sensitivity and specificity, together with their 95% confidence intervals (CIs), were estimated using a bivariate random-effects model. Overall diagnostic performance was summarized by constructing a hierarchical summary receiver operating characteristic (HSROC) curve and calculating the corresponding area under the curve (AUC). Between-study heterogeneity was examined with Cochran's Q test and the I² statistic, with I² values below 50% or Q test P values greater than 0.10 considered indicative of acceptable homogeneity. To investigate potential sources of heterogeneity, prespecified subgroup analyses were carried out according to Ki-67 cutoff values, imaging modality, and prediction model. The presence of a threshold effect was evaluated by determining the Spearman correlation between the logit-transformed sensitivity and the logit of (1 − specificity). Potential publication bias was assessed using Deeks' funnel plot asymmetry test. Unless stated otherwise, all statistical tests were two-tailed, and statistical significance was defined as P < 0.05.

Results

Characteristics of included studies
Through a systematic electronic database search, 3,923 studies were initially identified for screening. After excluding 2,032 duplicate records and 1,719 irrelevant studies, 172 articles underwent full-text review. Based on the predefined inclusion and exclusion criteria, 17 studies^{13,14,17,18,19,}

Discussion

This systematic meta-analysis evaluated the diagnostic performance of radiomic features for predicting Ki-67 expression in HCC. The pooled sensitivity was 0.87, the pooled specificity was 0.79, and the summary AUC was 0.90. Although estimates were broadly similar across the subgroups analyzed, substantial heterogeneity limits the generalizability of the pooled estimates. These findings suggest that radiomics may be a complement to tissue-based assessment, but its clinical utility requires further validation.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Review Manager	Nordic Cochrane Centre	Version 5.3	Risk-of-bias assessment and mapping
Stata	StataCorp	Version 16.0	Diagnostic meta-analysis using the midas module