Esophageal cancer (EC), a prevalent and deadly neoplasm, has been identified as the seventh most commonly diagnosed cancer globally in 2020, with 604,000 new cases reported and the sixth leading cause of death, with 544,000 fatalities (1). Patients with EC are often diagnosed at advanced stages, making them unsuitable for surgery and leading to a poor prognosis with low 5-year survival rates of only 20%-30% (2–4). Lymph node metastasis (LNM) has been identified as a vital prognostic determinant for patient survival (5–7). The eighth edition of the AJCC’s International Staging Standard for Esophageal Cancer introduced a clinical staging system based on preoperative imaging; it incorporated the count of lymph node metastases in postoperative staging (8, 9). While pathological findings continue to be the gold standard for diagnosing LNM, lymph node biopsy represents an invasive procedure with a non-trivial incidence of complications (10). Therefore, accurate evaluation of lymph node status using non-invasive imaging methods is imperative for making informed treatment decisions and essential for precise prognostication.

The computed tomography (CT) scans is a widely used non-invasive imaging technique for acquiring preoperative and postoperative tumor-related data to assess lymph node status in EC patients (11). However, relying solely on morphological standards, including short diameter and shape measured by physicians with varying levels of clinical diagnostic expertise, conventional CT scans are inadequate for accurately identifying LNM. The accuracy of conventional CT scans for precise detection is suboptimal, with sensitivity ranging from 37.3% to 67.2% and specificity ranging from 63.9% to 96.4% (12). Furthermore, these criteria have limitations culminating in a markedly low diagnostic accuracy for normal-sized lymph nodes. Consequently, conventional evaluation of lymph node status through CT scans remains challenging.

Radiomics is an innovative technique that swiftly extracts numerous quantitative features from conventional medical images using high-throughput computation, yielding invaluable information for diagnostic and prognostic purposes (13–15). In recent years, radiomics has been extensively employed in detecting, grading, assessing the therapeutic response, and prognostic evaluation for patients with EC (16–20). CT-based radiomics have been utilized to predict LNM in patients with EC (11, 21–23). These findings suggest that this approach has great potential as an accurate and reproducible tool for non-invasive preoperative evaluation of LNM. Radiomics techniques offer a promising solution that overcomes some of the limitations of conventional CT imaging. However, owing to variations in imaging protocols, study design, sample size, modeling techniques, and software used for radiomics analysis across various studies on EC, the reported diagnostic efficacy of radiomics in preoperative identification of LNM has demonstrated significant variability. These inconsistencies have led to uncertainty regarding the effectiveness of using radiomics for this purpose in clinical practice. Therefore, we aimed to conduct a comprehensive meta-analysis that thoroughly assesses the diagnostic accuracy of CT-based radiomics in predicting LNM in individuals diagnosed with EC.

This study followed the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (24). The protocol for this review has been registered on the Open Science Framework (OSF) platform, with the registration link available at https://osf.io/5zcnd.

According to our current awareness, this study represented the first systematic review and meta-analysis investigating the efficacy of CT-based radiomics in assessing LNM among patients diagnosed with EC. Our findings demonstrated that the pooled sensitivity, specificity, and AUC for CT-based radiomics were 0.73 (95% CI, 0.67-0.79), 0.76 (95% CI, 0.69-0.83), and 0.78 (95% CI, 0.74-0.81), respectively, showcasing its significantly effective performance. The good sensitivity was crucial for accurately identifying most patients with LNM, thus reducing the chances of missed diagnoses. Furthermore, the high specificity was key in lowering false-positive rates, offering a reliable way to rule out LNM in patients and helping clinicians avoid unnecessary treatments and their potential complications. The inclusion of positive and negative likelihood ratios (PLR and NLR) of 3.1 and 0.35, respectively, further improved the diagnostic utility of CT-based radiomics. A PLR of 3.1 meant that patients with LNM were over three times more likely to have a positive test result than those without, greatly increasing the accuracy of identifying affected individuals. On the other hand, an NLR of 0.35 indicated that a negative result significantly reduced the chance of LNM by 65%, lowering the risk of false negatives. These ratios adjusted the post-test probability based on the initial probability, with a positive test increasing the probability from 20% to 44%, and a negative test decreasing it to about 8%. Such adjustments significantly changed how test results were interpreted clinically, boosting the predictive accuracy. Therefore, our study validated the precision and effectiveness of CT-based radiomics in predicting LNM in EC, facilitating personalized treatment plans. By precisely detecting LNM and improving cancer staging, it facilitated more personalized care, optimizing treatment for high-risk patients while avoiding unnecessary interventions for others.

A moderate degree of heterogeneity in terms of specificity was indicated by the meta-analysis of the studies included, which is worth noting. To determine the possible sources of heterogeneity, a univariate meta-regression analysis and subgroup analysis were conducted, as the Spearman’s correlation coefficient test revealed that heterogeneity was not associated with threshold effects. The results suggested that the heterogeneity in specificity could be attributed to the utilization of different CT scan phases and variations in sample size. However, upon conducting a subgroup analysis, it was observed that the sensitivity and specificity were not significantly affected by these factors. It is important to acknowledge that due to the variations in methodologies employed among the included studies, it was challenging to identify all the factors contributing to the observed heterogeneity. Interestingly, while there was no heterogeneity in the pooled sensitivity across all studies, several subgroups showed significant results (P < 0.05) in the univariate meta-regression analysis. Therefore, further research with methodological standardization is necessary to improve accuracy, eliminate heterogeneity, and provide more robust evidence for using CT-based radiomics in predicting LNM in patients with EC.

LASSO regression is widely used for feature selection and dimensionality reduction, aiming to reduce the number of features and eliminate irrelevant ones (36, 37). In the subgroup analysis of this study, using LASSO alone for dimensionality reduction resulted in slightly higher specificity than combining it with other methods. This observation may be attributed to the distribution of data features and the division of subgroups. Most studies on radiomics diagnostic models used LR due to the binary nature of LNM status. Nevertheless, studies utilizing other algorithms, including SVM and RF, showed higher sensitivity rates than those using LR. Additionally, previous studies have shown that neural network models or RF based on clinical features could more effectively predict LNM than traditional LR, exhibiting higher AUC, specificity, positive predictive value, and accuracy (38, 39). Regrettably, owing to limited available literature, only one article each for SVM and RF was retrieved, making it difficult to draw reliable conclusions regarding the comparison between SVM or RF and LR.

Image segmentation is a pivotal element of radiomic analysis, incorporating manual delineation using 2D or 3D images, as well as semi-automatic and fully automatic techniques. However, a universally accepted standard for tumor segmentation remains elusive (40). Although manual segmentation offers high precision, it is labor-intensive, subjective, and lacks standardization, leading to limited reproducibility and elevated time and labor expenses. Semi-automatic segmentation necessitates manual refinement, whereas automatic segmentation employs sophisticated computer algorithms for efficient and reproducible lesion boundary identification (41, 42). However, it’s crucial to mention that the studies incorporated in this research exclusively used manual delineation for image segmentation. Moreover, only one study in this meta-analysis employed the 2D method, and the subgroup analysis did not reveal a significant difference in sensitivity and specificity between the 2D and 3D methods. Nonetheless, the majority of previous studies have recognized that radiomics-based 3D imaging traits offer a wider and more diverse range of specific information, covering the entire tumor volume and providing a more comprehensive and accurate representation of its shape, size, and texture. Furthermore, 3D segmentation enhances reproducibility by reducing interobserver variability and offers a standardized approach to tumor delineation (43, 44). Hence, future research could concentrate on investigating the advantages and limitations of manual, semi-automatic, and fully automatic delineation in both 2D and 3D methods in radiomics-based imaging analysis, with the goal of determining the most suitable imaging technique for specific clinical situations and enhancing the accuracy and reproducibility of radiomics-based tumor characterization.

Previous studies have highlighted the potential impact of variations in manufacturers and devices on the reproducibility of radiomics features, which could affect the precision of image diagnosis (13, 45). Similarly, the subgroup analysis results of this study revealed that radiomics features derived from distinct CT devices had an impact on the pooled specificity. Nonetheless, it is important to interpret these results cautiously, considering the limited number of studies included in the meta-analysis and the potential for bias due to the small sample size. Multicenter studies can validate the generalization ability of radiomics models by overcoming data differences across regions and devices, thereby improving the stability and reliability of the model (46). However, it is worth noting that all the studies included in this analysis were conducted in the same geographical region, China, which introduces a potential geographical bias. To gain a deeper understanding of the value of radiomics in diagnosing LNM in EC, further analysis is required through more prospective, multi-regional, and high-quality studies.

To assess the robustness of our study, we conducted a sensitivity analysis by sequentially removing one literature source at a time. The results showed no significant changes in the combined DOR after each exclusion, indicating that individual studies did not significantly influence our meta-analysis and that the conclusions were stable and reliable. Moreover, the lack of publication bias, as evidenced by Deeks’ funnel plot, further supports the credibility of our findings.

To bolster the robustness and reproducibility of radiomics methodologies, Lambin et al. introduced the RQS guidelines in 2017 (13), aiming to establish a benchmark for quality in radiomics research. However, the absence of standardized quality thresholds remained a notable gap. In response, Wesdorp et al. (47) suggested adopting a 30% cut-off score to enhance clarity and consistency across studies. Despite the RQS percentage of included studies in this meta-analysis ranging from 38.9% to 44.4%, surpassing the 30% threshold, and the pooled diagnostic efficacy demonstrating commendable performance in detecting LNM, the methodological quality of included studies remained a concern. This was because none of the studies utilized phantoms to assess robustness against inter-scansner discrepancies and vendor-specific characteristics. Additionally, comprehensive cost-effectiveness analysis, discussions on potential biological correlations, and a prospective study design were lacking in these studies. Therefore, caution is advised when interpreting the study outcomes.

Several constraints should be considered in the meta-analysis. Firstly, a constrained number of studies met our selection criteria. Secondly, the exclusively retrospective studies analyzed, all conducted in China and solely encompassing English-language publications, may have introduced selection biases and affected quality assessment, thereby potentially constraining the generalizability of our findings. Thirdly, despite conducting various analyses, heterogeneity persisted, emphasizing the need for cautious interpretation of the pooled quantitative results. During data extraction, the highest diagnostic performance model was chosen among multiple models, potentially leading to overestimating the radiomics diagnostic accuracy. Lastly, radiomics could be influenced by factors such as imaging equipment technology and protocols, contributing to heterogeneity. Therefore, establishing standardized presentation protocols in future radiomics research papers is necessary.

Our findings indicated that the CT-based radiomics demonstrated good diagnostic accuracy in predicting LNM in EC, with commendable sensitivity and specificity levels. However, considering the suboptimal RQS and observed heterogeneity among the included studies, it is essential to conduct additional high-quality, multicenter, and large-scale prospective trials to establish more robust and conclusive evidence for the findings presented in this research.

The original contributions presented in the study are included in the article/ Supplementary Materials , further inquiries can be directed to the corresponding author/s.

LL: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Writing – original draft, Writing – review & editing, Project administration. HL: Data curation, Formal analysis, Investigation, Methodology, Validation, Writing – original draft, Writing – review & editing. YZ: Data curation, Methodology, Software, Writing – original draft, Writing – review & editing. JY: Formal analysis, Writing – original draft, Writing – review & editing. CW: Methodology, Visualization, Writing – original draft, Writing – review & editing. LD: Methodology, Writing – original draft, Writing – review & editing. PX: Software, Writing – original draft, Writing – review & editing. WW: Supervision, Visualization, Writing – review & editing. MX: Supervision, Visualization, Writing – review & editing. DS: Conceptualization, Funding acquisition, Project administration, Supervision, Visualization, Writing – review & editing.