Application of dual-energy computed tomography combined with radiomics in the clinical diagnosis of lung cancer: a systematic review and meta-analysis.

Introduction
Lung cancer is one of the most common types of cancer globally and the leading cause of cancer-related death. According to global cancer statistics by region in 2022, lung cancer ranks as the most prevalent cancer, with nearly 2.5 million new cases (accounting for 12.4% of all global cancers) and 1.82 million deaths (accounting for 18.7% of cancer deaths) (1). In China, the incidence rate [age-standardized incidence rate (ASIR): 39.38 per 100,000] and mortality rate [age-standardized mortality rate (ASMR): 31.70 per 100,000] of lung cancer rank first globally. It is projected that the number of global lung cancer cases will increase by 86.2% by 2050, with new cases in China accounting for 30% of the global increase (2), indicating an especially severe situation. Therefore, there is a significant correlation between early diagnosis of lung cancer and improved prognosis. However, current methods for clinical diagnosis still face challenges: conventional computed tomography (CT) has insufficient sensitivity in detecting early-stage lung cancer, and pathological biopsy is invasive with sampling limitations.
Dual-energy computed tomography (DECT) has introduced multiple innovative applications into clinical practice, such as virtual monochromatic imaging (VMI) and iodine concentration (IC) maps (3), and has shown potential in the preoperative diagnosis of metastatic lymph nodes in patients with non-small cell lung cancer (NSCLC) (4). As a noninvasive and rapid imaging technique, DECT can identify potential malignant nodules at an early stage, providing robust evidence for timely intervention and treatment decision-making (5). DECT enables low-dose scanning to obtain diagnostic images while acquiring multiple quantitative parameters that conventional CT cannot provide (6); however, in the clinical assessment of complex lesions, the accuracy of its feature extraction and comprehensive analytical capability remain insufficient. Radiomics is a field of study, it processes medical images through automated high-throughput analysis methods to acquire quantitative and reproducible tumor information indiscernible to the human eye, providing an important basis for radiologists in the differentiation and diagnosis of lung cancer (7). Its core value lies in transforming the visual information that traditional imaging cannot quantify into objective data, thereby achieving precise characterization of tumor heterogeneity. Radiomics algorithms include two categories: machine learning and deep learning. Machine learning is mainly used for the selection of traditional features and the construction of clinical models, while deep learning shows advantages in automatic segmentation, high-order feature learning, and multimodal fusion. The combination of the two has promoted the precision of lung cancer diagnosis, prognosis, and treatment response assessment (8,9).
In this context, the integrated application of DECT and radiomics has increasingly become a research focus in the precision imaging diagnosis of lung cancer. The multidimensional physical parameters provided by DECT offer richer quantitative data for lung cancer diagnosis (10), whereas the high-dimensional analytical capability of Radiomics facilitates further exploration of potential diagnostic information under DECT multiparametric settings. Their combination is expected to overcome the limitations of traditional imaging diagnosis, such as restricted resolution (11) and challenges in distinguishing complex lesion structures. Nevertheless, existing relevant studies are characterized by small sample sizes, significant variability in conclusions, and a lack of systematic evidence synthesis. Therefore, this meta-analysis aimed to evaluate the incremental diagnostic value of the combined diagnostic model integrating DECT and radiomics in lung cancer. We present this article in accordance with the PRISMA reporting checklist (12) (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2449/rc).

Methods

Literature search strategy
The study protocol was registered with PROSPERO (registration No. CRD420261296133). Two investigators systematically searched the PubMed, Embase, and Cochrane Library databases, with the search conducted up to May 29, 2025. Boolean logic was employed for keyword searches, and the search strategy was as follows: (‘dual energy computed tomography’ OR ‘energy CT’ OR ‘spectral CT’ OR ‘dual-energy CT’ OR ‘DECT’ OR ‘Dual-energy CT’ OR ‘Dual-Energy Computed Tomography’) AND (‘lung cancer’ OR ‘pulmonary cancer’ OR ‘pulmonary nodules’ OR ‘Lung Cancer’) AND (‘radiomics’ OR ‘radiomic analysis’). Any disagreements between the investigators were resolved through discussion.

Inclusion and exclusion criteria
Two researchers strictly adhered to the population, intervention, comparison, outcome (PICO) framework to formulate specific inclusion and exclusion criteria. The inclusion criteria were as follows: (I) prospective or retrospective diagnostic trials; (II) studies enrolling patients with pathologically confirmed lung cancer or benign lesions; and (III) studies on the diagnosis of lung cancer using DECT or in combination with radiomics. The exclusion criteria were as follows: (I) reviews, conference abstracts, animal studies, letters, and case reports; (II) studies with low relevance to the current research topic; and (III) studies with limited or incomplete data availability despite attempts to contact the original authors for additional data.

Literature screening and data extraction
The retrieved articles were imported into EndNote (v20.0), and duplicate entries were removed. The initial screening of titles and abstracts was subsequently performed, and articles meeting the study requirements were retained. On the basis of the predetermined inclusion and exclusion criteria, full-text articles were carefully reviewed, and a total of 21 articles were ultimately included in the study. The keywords of these studies (Figure 1) show and confirm their relevance to this meta-analysis.
To facilitate the extraction of relevant data, two researchers designed a table of basic characteristics of the included studies, which specifically included basic study information (first author, year of publication, country), study design, manufacturer of DECT equipment, tube voltage, key DECT parameters, tumor segmentation, feature selection, and the radiomics algorithm, among others. Counts of true positive (TP), false positive (FP), true negative (TN), and false negative (FN) were extracted from the included articles for the synthesis of diagnostic effect size point estimates. All TP/TN/FP/FN data included in the study were independently extracted by two researchers. The sample sizes for the training set and testing set (or validation set) derived from the included studies were randomly allocated in a 7:3 ratio to ensure balanced data distribution and maintain methodological rigor. Consistency was verified (P<0.001), and discrepancies were resolved through third-party arbitration. When multiple models were present in a study, the model with the best overall performance [highest sensitivity, specificity, and area under the curve (AUC) value] was selected for meta-analysis.

Risk of bias assessment
To evaluate the overall risk of bias among the included studies, two researchers independently employed the QUADAS-2 tool (13), which encompasses four domains: ‘Patient Selection’, ‘Index Test’, ‘Reference Standard’, and ‘Flow and Timing’. Each domain is further divided into two sections: ‘Risk of Bias’ and ‘Concerns Regarding Applicability’. All 21 included studies were assessed individually against the criteria of the selected tool.

Statistical analysis
Two researchers independently conducted MetaDiSc 1.4 and Stata/MP 15.1 to calculate effect size indices. The sensitivity, specificity, AUC with 95% confidence interval (CI), positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR) were selected as the primary effect sizes. Heterogeneity was assessed via the I2 statistic (with I2<25% indicating low heterogeneity, 25–50% indicating moderate heterogeneity, and >50% indicating high heterogeneity) and the Q test. If heterogeneity existed, the random-effects model was prioritized for pooling; otherwise, the fixed-effects model was used. Subgroup analyses were performed via Stata/MP 15.1 with clear subgrouping dimensions, and forest plots were generated to present indices of each subgroup to explore the sources of heterogeneity, where P<0.05 was considered statistically significant. Moreover, a sensitivity analysis was conducted using Stata/MP 15.1.
Diagnostic test performance metrics (sensitivity, specificity) are confounded by sample size, study design, and technical parameters. Small sample sizes often lead to non-normal effect size distributions, violating the assumptions of parametric tests (e.g., Pearson correlation). Thus, Spearman rank correlation was used to assess variable associations, as it relies on ranked values and requires no normality assumption—making it ideal for non-normal or ordinal data. Specifically, this test quantified the relationship between sensitivity and 1 − specificity to evaluate threshold effects.

Results

Literature search and selection results
Two researchers independently conducted comprehensive searches in three public databases: Embase, PubMed, and the Cochrane Library. The initial search yielded 1,407 articles. Duplicate articles (n=184) were removed via EndNote. Following careful evaluation of titles and abstracts, 1,193 articles, including reviews, conference reports, animal studies, and other publications irrelevant to the topic of this study, were excluded. During the secondary screening, the full texts of the remaining 30 articles were carefully reviewed on the basis of predetermined inclusion and exclusion criteria; of these, nine articles were excluded (one due to failure to meet the inclusion criteria in terms of quality, and eight due to incomplete data precluding the extraction of effect sizes). Ultimately, a total of 21 articles meeting the inclusion criteria of this study were included in the meta-analysis. The detailed literature selection process is illustrated in a PRISMA flow diagram (14) (Figure 2).

Baseline characteristics of the included studies
This meta-analysis included 21 studies published between 2016 and 2025 (4-6,10,15-31), most of which were retrospective diagnostic studies. A total of 2,899 lesions were included across the 21 studies (1,215 malignant and 1,684 benign), among which 1,818 lesions were diagnosed via DECT combined with radiomics. The general characteristics of the included studies are summarized in Table 1.

Risk of bias assessment via the QUADAS-2
Figure 3 summarizes the risk of bias and concerns regarding applicability for each study. Most studies presented a low risk of bias and minimal concerns regarding applicability.

Diagnostic performance of DECT for lung cancer
This meta-analysis included 21 studies, involving a total of 2,899 lesions. The forest plot (Figure 4) visually illustrates the differences in results among the included studies. Pooled results, as shown in the summary receiver operating characteristic (SROC) curve (Figure 5), revealed a pooled sensitivity of 0.85 (95% CI: 0.82–0.88), a specificity of 0.83 (95% CI: 0.79–0.87), a PLR of 5.1 (95% CI: 4.1–6.4), an NLR of 0.18 (95% CI: 0.14–0.22), a DOR of 29.0 (95% CI: 21–42), and an AUC of 0.91 (95% CI: 0.88–0.93) based on the SROC curve, indicating that DECT exhibits good diagnostic accuracy and clinical utility in lung cancer diagnosis. Meanwhile, Spearman’s correlation test showed no significant correlation between sensitivity and “1 − specificity” (ρ=0.038, P=0.87), suggesting the absence of threshold effect among the included studies.

Heterogeneity assessment and subgroup analysis
In the study design subgroup, “Yes” indicated the use of prospective studies, and “No” indicated the use of retrospective studies; in the race subgroup, “Yes” indicated the inclusion of Asian populations, and “No” indicated the inclusion of non-Asian populations; in the tumor segmentation subgroup, “Yes” indicated the use of automatic tumor segmentation, and “No” indicated the use of manual tumor segmentation; in the feature selection subgroup, “Yes” indicated the use of least absolute shrinkage and selection operator (LASSO), and “No” indicated the nonuse of LASSO; in the radiomic algorithm subgroup, “Yes” indicated the use of deep learning, machine learning, and other radiomic algorithms, and “No” indicated the nonuse of these radiomic algorithms.
To further explore heterogeneity, we performed meta-regression analysis across 12 subgroups categorized by study design, race, tumor segmentation, feature selection, and radiomics algorithm. The results revealed that, with the exception of race, study design, tumor segmentation, feature selection, and radiomics algorithms were sources of heterogeneity (P<0.05). Among these factors, feature selection and the radiomics algorithm exerted a particularly prominent influence, contributing significantly more to the heterogeneity in the meta-analysis of diagnostic tests than other factors did. In the subgroup analysis stratified by feature selection, the pooled sensitivity of the LASSO group was significantly greater than that of the non-LASSO group (difference =0.04, 95% CI: 0.83–0.91 vs. 0.79–0.87, P<0.001). The magnitude of this difference was greater than the sensitivity difference induced by the tumor segmentation method (difference =−0.02, 95% CI: 0.79–0.88 vs. 0.81–0.89, P<0.01). Additionally, the radiomic group had a significantly greater pooled sensitivity than the non-radiomic group did (difference =0.06, 95% CI: 0.80–0.96 vs. 0.75–0.88, P<0.001). These results further confirmed that the feature selection method and radiomics algorithm are key factors influencing study heterogeneity (Figure 6; Table 2).

Sensitivity analysis
To verify the impact of the 11 studies that did not report radiomics algorithms on the pooled results, we conducted a sensitivity analysis: after excluding these 11 studies, we re-pooled the data from the 10 studies that clearly reported the algorithms. The results showed that after excluding the 11 studies that did not report the algorithms, there were no significant differences in sensitivity (0.85 vs. 0.87), specificity (0.83 vs. 0.82), and AUC (0.91 vs. 0.90) for the DECT combined with radiomics model in lung cancer diagnosis (P>0.05). This indicates that the core conclusions of this study are robust and not significantly affected by the studies that did not report the algorithms.

DECT diagnosis vs. DECT combined with radiomics diagnosis
This study also systematically evaluated the diagnostic performance of DECT in three different clinical scenarios, including differentiating between benign and malignant lesions of lung cancer (18-20,25-28), predicting invasiveness (based on the comparison of invasiveness differences among pathological subtypes, i.e., differentiation of invasiveness among adenocarcinomas in situ, minimally invasive adenocarcinomas, and invasive adenocarcinomas) (17,29-31), and predicting lymph node metastasis (LNM) (4,15,16,21-24). Key diagnostic performance metrics, such as sensitivity, specificity, AUC, PLR, NLR, and DOR, were analyzed (Table 3). For the differentiation of benign and malignant lung lesions, DECT exhibited strong diagnostic capability—with an AUC of 0.93 (95% CI: 0.90–0.95), balanced sensitivity (0.87) and specificity (0.88), plus a high DOR of 48.00, indicating robust discriminative value. In predicting lung cancer invasiveness, it maintained moderate performance (AUC =0.85, 95% CI: 0.81–0.88), though specificity (0.79) was slightly lower than sensitivity (0.83). For LNM prediction, DECT achieved a sensitivity of 0.84, paired with a specificity of 0.83 and an AUC of 0.90, supporting its potential in staging assessment. Analysis of the diagnostic performance metrics indicated that DECT meets the clinical diagnostic performance criteria for the differentiation of benign and malignant lung lesions and the prediction of LNM, whereas its performance in predicting lung cancer invasiveness still has substantial potential for optimization.
The combined diagnostic technology of DECT and radiomics has shown significant gains in diagnostic efficacy in both the differentiation of benign and malignant lesions and the prediction of invasiveness in lung cancer (Table 4). For the differentiation of benign and malignant lesions, the training set (25,27,28) exhibited excellent discriminative performance (sensitivity =0.90, specificity =0.88, AUC =0.94, DOR =55.82). Notably, the validation set (10,25,26) maintained comparable diagnostic accuracy (AUC =0.94, 95% CI: 0.89–0.99), with only marginal decrements in sensitivity (0.87) and specificity (0.81)—a pattern indicative of good cross-cohort generalizability. Compared with DECT, the training set demonstrated a 3.4% increase in sensitivity and a 1.1% increase in AUC, whereas the AUC of the test set remained stable, further supporting its incremental diagnostic value. For the prediction of invasiveness, the training set (6,29-31) of the combined model achieved balanced diagnostic metrics (sensitivity =0.84, specificity =0.84, AUC =0.91, 95% CI: 0.88–0.93; DOR =28.01). Compared with DECT, there was a 1.2% increase in sensitivity, a 6.3% increase in specificity, and a 7.1% increase in AUC, which can effectively reduce the misclassification of low-invasive tumors as highly invasive ones.

Analysis of publication bias
Deeks’ funnel plot asymmetry test revealed that the distribution of included studies in the present study showed no significant publication bias (P=0.44), suggesting that studies with positive and negative results had a balanced probability of publication and that the pooled effect size was not significantly affected by publication bias (Figure 7).

Discussion
This study focused on evaluating the diagnostic efficacy of DECT and DECT combined with radiomics algorithms in diagnosing lung cancer. The conclusions focused on the prominent diagnostic value demonstrated by DECT combined with radiomics algorithms, and the findings were systematically elaborated by closely integrating the current status of the research field and the needs of clinical practice.
Analysis of the diagnostic performance of DECT in three distinct clinical scenarios revealed favorable performance in differentiating between benign and malignant nodules and predicting LNM, with AUC values of 0.93 (95% CI: 0.90–0.95) and 0.90 (95% CI: 0.87–0.93), respectively. In contrast, its performance was relatively inferior in predicting invasiveness, with an AUC value of 0.85 (95% CI: 0.81–0.88). The reasons underlying these differences in diagnostic performance may be associated with the adaptability of technical principles and pathological characteristics. DECT spectral imaging can effectively distinguish between benign and malignant nodules and metastatic lymph nodes on the basis of differences in tissue composition, whereas lung cancer invasiveness involves microscopic biological behaviors, which are difficult to fully reflect via macroscopic imaging features.
In terms of core diagnostic efficacy, the combined technique of DECT and radiomics yielded AUC values of 0.94 (95% CI: 0.85–1.00) and 0.94 (95% CI: 0.89–0.99) in the training and test sets, respectively, for differentiating between benign and malignant lung lesions, with an AUC of 0.91 (95% CI: 0.88–0.93) in the training set for predicting invasiveness. These results highlight that this combined technique is crucial for overcoming the bottleneck in precise lung cancer diagnosis, as its incremental value far exceeds that of DECT [AUC values of 0.93 (95% CI: 0.90–0.95) and 0.85 (95% CI: 0.81–0.88)] for benign-malignant differentiation and invasiveness prediction, respectively). Our findings indicate that although DECT has demonstrated favorable efficacy in clinical applications, it still has significant limitations in predicting lung cancer invasiveness. After combining it with radiomics algorithms, this technical shortcoming was effectively ameliorated and supplemented. In the differentiation of benign and malignant lesions, the AUC of the combined technique in the training set increased by 1.1% compared with that of DECT, and the performance in the validation set remained stable, confirming its excellent generalizability and feasibility for clinical application. The core advantage lies in its ability to predict invasiveness, with the AUC and specificity in the training set increasing by 7.1% and 6.3%, respectively, which can effectively reduce the misclassification of low-invasive lesions as high-invasive lesions, providing precise imaging evidence for formulating personalized clinical treatment strategies.
Subgroup analysis revealed that the heterogeneity stemmed from the study design, tumor segmentation, feature selection, and radiomics algorithms, with feature selection and radiomics algorithms being the primary sources (P<0.001). In the present study, prospective studies present advantages in the standardization of data collection and completeness of follow-up, which helps reduce bias. In contrast, retrospective studies, which are limited by the integrity of historical data, may compromise the evaluation of diagnostic performance. Manual segmentation outperformed automatic segmentation in both sensitivity and specificity. Through manual intervention on the basis of physicians’ experience, manual segmentation can obtain more accurate image information, whereas automatic segmentation may exhibit variations in boundary identification in complex scenarios. LASSO is a feature selection method commonly used in radiomic algorithms (32), and the present meta-analysis included a total of 6 LASSO-related studies. Compared with non-LASSO methods, LASSO exhibited greater sensitivity, whereas the specificity was comparable between the two groups. These findings suggest that LASSO can enhance diagnostic sensitivity by optimizing feature quality without compromising specificity (0.88 vs. 0.84), making it highly suitable for multiparametric analysis scenarios that combine DECT with radiomics. Additionally, we observed that the sensitivity of machine learning algorithms (0.89) outperformed that of deep learning (0.83). The underlying mechanisms of this difference can be explained from the following two aspects. First, machine learning algorithms exhibit superior adaptability to low-dimensional handcrafted features, enabling them to accurately capture subtle imaging patterns relevant to disease diagnosis, thereby reducing the false-negative diagnosis rate (33). Second, in this study, the machine learning group had a larger sample size and more balanced data distribution, which provided a data foundation for the model to establish stable decision boundaries, whereas deep learning models tend to exhibit performance fluctuations in scenarios with small sample sizes or imbalanced data distributions (34). However, both machine learning and deep learning exhibited higher sensitivity than non-radiomics approaches did (0.82), which paves the way for the optimization of algorithm selection in DECT combined with radiomic technology and for precise application in high-sensitivity diagnostic scenarios in the clinic.
Our study has certain limitations. First, among the 21 included studies, 16 were retrospective, and only five were prospective, which may lead to selection bias in case selection. Furthermore, there was only one non-Asian study in the race subgroup, making it difficult to validate the applicability of the combined technique in non-Asian populations, which limits its global generalizability. Second, we did not subclassify the lung cancer subtypes to evaluate their efficacy, so the advantages of the combined technique for specific subtypes cannot be clarified. Moreover, in the radiomic algorithm subgroup, owing to the small sample size, it was difficult to screen the optimal algorithm suitable for DECT data accurately. Finally, the number of eligible studies on the combined technique for the diagnosis of LNM in lung cancer patients was insufficient; therefore, a combined analysis in this context has not yet been conducted. Additionally, the AUC results of the invasiveness prediction validation set may have certain biases, so the results are not presented. Therefore, future studies should focus on optimizing DECT combined with radiomics technology. They should conduct more prospective, multicenter studies (35) to reduce bias, include greater representation of non-Asian populations, adjust parameters and features accordingly to enhance racial universality, deepen research on tumor subtypes and radiomics algorithms, expand more application scenarios, and optimize diagnostic efficacy through the integration of the two technologies.

Conclusions
DECT has favorable clinical application efficacy in differentiating between benign and malignant lung lesions, predicting invasiveness, and predicting LNM, thereby providing reliable imaging evidence for the formulation of clinical diagnosis and treatment decisions in patients with lung cancer. Further stratified analysis confirmed that the integration of DECT with radiomics algorithms not only significantly enhances the accuracy of differentiating between benign and malignant lung lesions but also effectively compensates for the limitations of DECT in predicting invasiveness, offering crucial technical support for optimizing precise imaging assessment strategies in lung cancer. Future studies could further validate the clinical universality and promotional value of this combined technique by expanding the sample size and incorporating more lung cancer cases with diverse pathological subtypes and clinical stages, thereby laying a more solid imaging foundation for the advancement of precision medicine in lung cancer.

Supplementary
The article’s supplementary files as