Radiomics for preoperative assessment of peritoneal metastasis in gastric cancer

Abstract

Peritoneal metastasis is a major determinant of prognosis and treatment strategy in gastric cancer, yet its accurate preoperative detection remains challenging with conventional imaging. Computed tomography often underestimates occult peritoneal disease, leading to delayed diagnosis or unnecessary exploratory procedures. Radiomics, enabling high-dimensional quantitative analysis of medical images, has emerged as a promising non-invasive approach for capturing tumor heterogeneity and improving metastatic risk assessment. Recent evidence, including the study by Mu et al, published in the World Journal of Gastrointestinal Oncology, highlights the value of multiphase contrast-enhanced computed tomography radiomics, where integration of arterial, venous, and delayed-phase features improves predictive performance, with reported area under the curves often exceeding 0.80. Despite these advances, significant challenges remain, including heterogeneity in imaging protocols, region-of-interest delineation, feature extraction pipelines, and the lack of prospective multicenter validation. Furthermore, the incremental value of radiomics over optimized clinical models and its impact on clinical decision-making require further clarification. This editorial critically appraises current advances and limitations of radiomics for peritoneal metastasis prediction, emphasizing methodological standardization and future directions toward clinically applicable, reproducible imaging biomarkers for gastric cancer staging.

Key Words: Gastric cancer; Peritoneal metastasis; Radiomics; Multiphase computed tomography; Machine learning; Preoperative staging

Core Tip: Radiomics based on multiphase contrast-enhanced computed tomography offers a promising non-invasive approach for the preoperative assessment of peritoneal metastasis in gastric cancer. By quantitatively capturing tumor heterogeneity beyond visual interpretation, multiphase radiomics models – particularly when integrated with clinical variables – demonstrate improved predictive performance compared with conventional imaging alone. However, heterogeneity in imaging protocols, feature extraction strategies, and model validation remains a major barrier to clinical translation. Standardization and prospective validation are essential before routine clinical implementation.

This editorial refers to "Application value of multiphase contrast-enhanced computed tomography radiomics in preoperative evaluation of peritoneal metastasis in gastric cancer" by Mu et al, 2026; https://dx.doi.org/10.4251/wjgo.v18.i2.115404.

---

INTRODUCTION

Peritoneal metastasis (PM) represents the most frequent and lethal dissemination pattern in gastric cancer and remains a pivotal determinant of treatment strategy and prognosis[1,2]. Patients with occult peritoneal disease often experience limited survival benefit from radical surgery, underscoring the importance of accurate preoperative staging[3-5]. Despite advances in imaging technology, conventional contrast-enhanced computed tomography (CT) still demonstrates suboptimal sensitivity for detecting small-volume or microscopic PM, frequently resulting in understaging and inappropriate surgical exploration[6].

In recent years, radiomics has emerged as a novel imaging paradigm capable of extracting high-dimensional quantitative features from routine medical images. These features reflect tumor shape, texture, intensity distribution, and spatial heterogeneity, providing surrogate information on underlying tumor biology that is not appreciable by visual assessment alone. Radiomics has shown increasing value in gastric cancer for predicting lymph node involvement, treatment response, molecular characteristics, and survival outcomes[7,8]. Among recent contributions, the study by Mu et al[9] published in this recent issue of World Journal of Gastrointestinal Oncology represents a notable advance by demonstrating the clinical utility of multiphase contrast-enhanced CT radiomics for predicting PM.

Multiphase contrast-enhanced CT, incorporating arterial, venous, and delayed phases, offers complementary temporal information on tumor perfusion, vascular permeability, and stromal interaction. In a landmark European Radiology study, Liu et al[10] demonstrated that venous-phase and multiphase CT radiomics achieved an area under the curve (AUC) of 0.87 for predicting occult PM in advanced gastric cancer, significantly outperforming conventional CT evaluation. Similarly, dual-energy CT radiomics models developed by Chen et al[7] achieved robust discrimination of PM by capturing iodine-related perfusion heterogeneity. Moreover, machine learning techniques, including random forest and deep learning frameworks, have further enhanced predictive accuracy by effectively handling high-dimensional radiomic data[11]. Notably, Mu et al[9] recently reported in World Journal of Gastrointestinal Oncology that a multiphase arterial-venous-delayed CT radiomics model achieved an AUC of 0.876 in an independent validation cohort, significantly outperforming all single-phase models for preoperative prediction of gastric cancer PM.

Despite these encouraging findings, the clinical translation of radiomics for preoperative assessment of PM remains limited. Existing studies are predominantly retrospective, single-center in design, and characterized by substantial heterogeneity in imaging protocols, region-of-interest delineation, feature selection pipelines, and modeling strategies[12]. In addition, the incremental value of radiomics beyond optimized clinical models, as well as its role in guiding staging laparoscopy and individualized treatment planning, has not been systematically synthesized.

Therefore, the editorial aims to summarize current advances in radiomics for the preoperative assessment of PM in gastric cancer, with particular emphasis on multiphase CT methodologies, feature robustness, and model validation strategies. Methodological challenges, comparative performance, and future directions toward clinical implementation are critically discussed.

EDITORIAL PERSPECTIVE ON RECENT ADVANCES

The study by Mu et al[9] provides a timely and clinically relevant contribution by systematically evaluating multiphase CT radiomics for preoperative prediction of PM in gastric cancer. Their findings, demonstrating an AUC of 0.876 in the validation cohort, highlight the added value of integrating arterial, venous, and delayed-phase imaging features over conventional single-phase approaches. Importantly, the reported reduction in unnecessary exploratory surgeries underscores the potential clinical utility of radiomics-guided decision-making.

RATIONALE FOR RADIOMICS IN PREOPERATIVE ASSESSMENT OF PM

Accurate preoperative detection of PM remains one of the most challenging aspects of gastric cancer staging. Conventional contrast-enhanced CT relies primarily on morphological criteria, such as nodular peritoneal thickening, ascites, or omental caking, which are typically absent in early or occult disease. Consequently, CT frequently underestimates PM burden, leading to delayed diagnosis or unnecessary exploratory surgery[6].

Radiomics addresses these limitations by converting routine imaging data into high-dimensional quantitative features that capture subtle spatial heterogeneity and textural patterns within tumors and surrounding tissues. These features are hypothesized to reflect underlying biological processes such as angiogenesis, stromal remodeling, and tumor-microenvironment interactions, which precede overt radiological manifestations of peritoneal spread[7,13]. As such, radiomics provides a theoretical framework for identifying patients at high risk of PM even when conventional imaging appears negative.

MULTIPHASE CT RADIOMICS: METHODOLOGICAL CONSIDERATIONS

Imaging phase selection and biological implications

Multiphase contrast-enhanced CT, typically comprising arterial, portal venous, and delayed phases, offers complementary temporal information regarding tumor vascularity and contrast kinetics[14,15]. Arterial-phase images emphasize tumor perfusion and neovascularization, venous-phase images capture parenchymal enhancement and stromal interaction, while delayed-phase images may reflect extracellular contrast retention associated with fibrosis or infiltrative growth patterns. The overall analytical workflow of multiphase CT-based radiomics, from image acquisition and feature extraction to model construction and clinical decision-making, is schematically illustrated in Figure 1.

Figure 1 Schematic workflow of multiphase computed tomography-based radiomics for preoperative assessment of peritoneal metastasis in gastric cancer. CT: Computed tomography; DL: Deep learning; LASSO: Least absolute shrinkage and selection operator; mRMR: Maximum relevance and minimum redundancy; RF: Random forest; SVM: Support vector machine.

Several studies demonstrate that radiomic features extracted from multiphase CT outperform single-phase models in predicting PM and other adverse pathological features[10,16]. For example, Liu et al[10] reported that a contrast-enhanced multiphase CT radiomics model achieved significantly higher accuracy than single-phase venous models in identifying occult PM in gastric cancer patients (AUC > 0.85). Similarly, Wang et al[17] showed that combining arterial and venous phase radiomic features improved discrimination of aggressive gastric cancer phenotypes associated with peritoneal spread. Dual-energy and spectral CT further enrich feature space by enabling iodine mapping and material decomposition, which may enhance sensitivity to subtle peritoneal involvement[7,17].

However, heterogeneity in phase selection across studies – some relying exclusively on venous-phase imaging, others incorporating all three phases – limits cross-study comparability and underscores the need for standardized acquisition protocols[12].

Region-of-interest delineation and feature robustness

Accurate segmentation remains a cornerstone of radiomics analysis. Most studies focus on primary tumor regions, while others extend segmentation to peritumoral fat or omental regions, aiming to capture early tumor-peritoneal interactions[18,19]. Both two-dimensional and three-dimensional segmentation strategies have been employed.

Comparative analyses suggest that two-dimensional radiomics may achieve performance comparable to three-dimensional approaches while offering greater feasibility and reproducibility in clinical practice[20]. Nonetheless, manual or semi-automatic segmentation remains labor-intensive and prone to interobserver variability, motivating increasing interest in deep learning-based automatic segmentation pipelines[21,22].

Feature stability is further influenced by CT reconstruction parameters and preprocessing steps. Studies employing intraclass correlation coefficient filtering and rigorous feature selection pipelines report improved robustness, but consensus standards are lacking[12].

MACHINE LEARNING MODELS FOR PREDICTING PM

Conventional machine learning approaches

Most radiomics studies adopt traditional machine learning algorithms such as logistic regression, support vector machines, and random forest classifiers. These algorithms are particularly well suited to radiomics because they can handle high-dimensional feature spaces with relatively limited sample sizes, especially when combined with feature selection methods such as least absolute shrinkage and selection operator or maximum relevance and minimum redundancy. Logistic regression provides transparent and clinically interpretable coefficients, support vector machines are effective in modeling complex nonlinear decision boundaries, and random forest classifiers offer robustness to noise and feature collinearity while capturing higher-order interactions among radiomic features. Feature selection methods, particularly least absolute shrinkage and selection operator and maximum relevance and minimum redundancy, are widely used to reduce dimensionality and mitigate overfitting[7,23].

Radiomics-based nomograms integrating imaging features with clinical variables – such as tumor stage and serum biomarkers [e.g., cancer antigen 125 (CA125)] – consistently demonstrate superior diagnostic performance compared with clinical models alone, with multiple representative studies reporting AUC values exceeding 0.85 and in some cases approaching 0.90 for individualized prediction of PM[24,25]. CA125 is a clinically established surrogate of peritoneal involvement and tumor burden in gastric cancer, and its incorporation into radiomics models provides complementary biological information that is not captured by imaging features alone. In a large Frontiers in oncology study by Huang et al[11], a radiomics nomogram integrating CT features and CA125 achieved an AUC of 0.88 for predicting PM, significantly outperforming clinical-only models. This finding illustrates that combining serum biomarkers reflecting tumor dissemination with radiomic features capturing spatial heterogeneity can synergistically enhance predictive accuracy beyond either modality alone. Similarly, positron emission tomography/CT-based radiomics nomograms developed by Xie et al[25] demonstrated AUC values exceeding 0.85 for individualized risk stratification of peritoneal dissemination.

However, the superiority of radiomics is not universal. Comparative analyses have shown that optimized clinical-CT models incorporating high-quality morphological imaging features and serum biomarkers can achieve performance comparable to, or in some cases exceeding, that of radiomics-only models, particularly in the prediction of occult PM[11]. These findings highlight that radiomics should be viewed as a complementary, rather than replacement, tool.

Deep learning-based radiomics

Deep learning has recently emerged as a powerful alternative to handcrafted radiomics. By automatically learning hierarchical imaging representations, deep convolutional neural networks can capture complex spatial patterns associated with PM risk. A pivotal multicenter study demonstrated that deep learning models achieved high sensitivity and specificity for noninvasive prediction of occult PM, with robust external validation[13].

Despite these promising results, the “black-box” nature of deep learning remains a major barrier to clinical adoption and regulatory approval[26]. Unlike conventional radiomics models that rely on predefined, human-interpretable features, deep neural networks generate latent representations that are difficult to interpret, limiting clinicians’ ability to understand how predictions are derived and to trust model outputs in high-stakes decision-making[27].

Explainable artificial intelligence techniques, such as saliency maps, feature attribution methods, and layer-wise relevance propagation, have therefore gained increasing attention as tools to visualize and interpret deep learning–based radiomics models[28,29]. These approaches can help identify image regions and patterns that drive model predictions, thereby improving transparency, clinical confidence, and error detection[30].

Furthermore, hybrid frameworks that integrate deep learning–derived features with handcrafted radiomic features and clinical variables represent a promising strategy to balance predictive performance with interpretability[31,32]. By combining automatically learned representations with biologically and physically meaningful imaging features, such models may achieve both high accuracy and greater clinical trustworthiness[33].

CLINICAL INTEGRATION AND DECISION-MAKING IMPLICATIONS

Radiomics-based prediction of PM has direct implications for clinical management. Accurate preoperative identification of high-risk patients may optimize selection for staging laparoscopy, reduce non-therapeutic laparotomy, and facilitate personalized treatment strategies, including neoadjuvant therapy or palliative approaches[18,34].

Nomograms and risk scores derived from radiomics models provide intuitive clinical tools, enabling individualized probability estimation rather than binary classification. Moreover, integration of radiomics with clinical and pathological factors aligns with the paradigm of precision oncology, supporting tailored decision-making rather than uniform treatment algorithms. In particular, Mu et al[9] demonstrated that their multiphase radiomics model could avoid 33.7% of unnecessary exploratory surgeries at a clinically relevant threshold probability, underscoring the real-world utility of radiomics-guided triage for staging laparoscopy. An integrated overview of imaging strategies, radiomics workflows, machine learning models, validation approaches, and clinical application scenarios for preoperative assessment of PM in gastric cancer is summarized in Table 1.

Table 1 Integrated summary of radiomics-based approaches for preoperative assessment of peritoneal metastasis in gastric cancer.

Dimension	Integrated findings from current evidence	Clinical implications
Imaging modality	Contrast-enhanced CT is the most widely used modality for radiomics analysis in gastric cancer PM. Positron emission tomography/CT and magnetic resonance imaging are explored as complementary modalities, while dual-energy and spectral CT provide additional quantitative information such as iodine distribution	CT-based radiomics is currently the most feasible approach for clinical translation, while multimodal imaging may further improve predictive accuracy in selected settings
Imaging phase strategy	Multiphase CT (arterial, venous, delayed) consistently outperforms single-phase imaging by capturing complementary information related to tumor perfusion, stromal infiltration, and contrast retention. Venous phase remains the most commonly used baseline phase	Multiphase acquisition enhances sensitivity for occult PM and should be preferred when image quality and workflow allow
Region-of-interest definition	Most models focus on the primary tumor, with increasing inclusion of peritumoral fat, omentum, or adjacent structures to reflect tumor-microenvironment interactions. Both 2D and three-dimensional segmentation strategies are used	Incorporating peritumoral regions may improve detection of early peritoneal spread; 2D segmentation offers a pragmatic balance between performance and feasibility
Radiomic feature categories	First-order intensity features, texture features (GLCM, GLRLM, GLSZM), and wavelet-transformed features are most frequently retained after feature selection. Shape features play a secondary role	Texture-based features appear particularly sensitive to subtle heterogeneity associated with early metastatic potential
Feature selection methods	Least absolute shrinkage and selection operator and maximum relevance and minimum redundancy are the dominant feature selection approaches to reduce dimensionality and improve model stability. Intraclass correlation filtering is commonly applied to ensure robustness	Standardized feature selection pipelines are critical to minimize overfitting and improve reproducibility across centers
Modeling strategies	Conventional machine learning models (logistic regression, random forest, support vector machine) are widely applied. Deep learning approaches enable automated feature learning and segmentation, often achieving higher sensitivity	Machine learning models provide interpretable risk estimates, while deep learning offers automation but requires careful validation and explainability
Integration with clinical variables	Combined radiomics-clinical models consistently outperform radiomics-only or clinical-only models. Common clinical variables include tumor stage, cancer antigen 125 (a marker of peritoneal tumor burden), and nodal status, which provide complementary biological and anatomical information to radiomic features	Multimodal integration aligns with precision oncology and improves individualized risk stratification
Predictive performance	Most integrated models achieve good-to-excellent discrimination for PM, with reported area under the curves frequently exceeding 0.80 and, in some cases, approaching 0.90	Radiomics demonstrates meaningful incremental value over conventional imaging, particularly for occult disease
Validation strategy	Internal validation is common; external and multicenter validation is less frequent but increasingly reported, particularly for deep learning models. Prospective validation remains scarce	Robust external validation is essential before routine clinical implementation
Clinical application scenarios	Preoperative risk stratification, selection for staging laparoscopy, avoidance of non-therapeutic surgery, and individualized treatment planning	Radiomics may function as a decision-support tool rather than a standalone diagnostic test
Key methodological limitations	Retrospective design, heterogeneous imaging protocols, variable segmentation strategies, lack of standardization, and limited interpretability of complex models	These limitations currently restrict widespread clinical adoption and regulatory approval
Future research priorities	Standardized imaging and radiomics pipelines, prospective multicenter studies, explainable artificial intelligence, and integration with molecular and pathological biomarkers	Addressing these gaps is essential for clinical translation and guideline incorporation

CT: Computed tomography; PM: Peritoneal metastasis; 2D: Two-dimensional.

Despite these advantages, real-world implementation remains limited. Few studies have prospectively evaluated whether radiomics-guided decision-making improves patient outcomes, resource utilization, or cost-effectiveness. Workflow integration, computational demands, and reproducibility across institutions continue to pose significant barriers[12].

METHODOLOGICAL LIMITATIONS AND RESEARCH GAPS

Current evidence is constrained by several limitations. Most studies are retrospective and single-center in design, raising concerns regarding selection bias and generalizability. Imaging protocols, feature extraction pipelines, and machine learning strategies vary substantially, impeding reproducibility and meta-analytic synthesis[12].

Importantly, this methodological heterogeneity partly explains the conflicting results reported in the literature. While many radiomics models demonstrate superior discrimination over conventional CT or clinical imaging, other studies show that carefully optimized clinical-CT models can achieve similar predictive performance, especially for occult PM[11,35]. Differences in patient selection, imaging quality, reference standards, and feature engineering likely contribute to these discrepancies.

Furthermore, demographic heterogeneity and treatment-related confounders – such as neoadjuvant chemotherapy – are insufficiently explored, despite their potential impact on radiomic signatures[13]. Standardized reporting guidelines, prospective multicenter validation, and harmonization of imaging and analysis workflows are urgently needed to advance clinical translation.

CONCLUSION

Radiomics based on multiphase contrast-enhanced CT represents a promising non-invasive strategy for improving the preoperative assessment of PM in gastric cancer. By quantitatively capturing tumor heterogeneity and microenvironmental characteristics, multiphase radiomics models – particularly when combined with clinical variables and advanced machine learning techniques – demonstrate superior predictive performance compared with conventional imaging alone. However, current evidence is largely derived from retrospective studies with heterogeneous methodologies. Future research should prioritize standardized radiomics pipelines, prospective multicenter validation, and integration of explainable artificial intelligence to facilitate reliable clinical implementation. With these advances, radiomics may evolve into a clinically applicable imaging biomarker that meaningfully improves preoperative staging and personalized management of gastric cancer.