Radiomics for preoperative pancreatic ductal adenocarcinoma risk stratification: Cross-population validation, multidimensional integration, challenges, and future directions

Abstract

This editorial critically evaluated Liu et al's recent retrospective analysis of 283 Chinese patients with resectable pancreatic ductal adenocarcinoma (PDAC) that validated a preoperative computed tomography-based risk scoring system originally developed in South Korea. The scoring system incorporated five parameters: (1) Tumor size; (2) Portal venous phase density; (3) Necrosis; (4) Peripancreatic infiltration; and (5) Suspected metastatic lymph nodes. While demonstrating satisfactory recurrence prediction capability without requiring complex technologies, thereby supporting clinical utility in Chinese populations, the study exhibited notable limitations. Most analyzed patients lacked neoadjuvant chemotherapy exposure, resulting in underrepresentation of low-risk subgroups. Additionally, the short follow-up duration potentially compromised long-term prognostic assessment. Contemporary advances in radiomics coupled with machine learning have enhanced multimodal data integration for PDAC management. However, clinical implementation continues to confront challenges including variability in imaging parameters, incomplete understanding of molecular underpinnings, and confounding treatment effects. Future investigations should prioritize developing multidimensional predictive frameworks that synergize radiographic, molecular, and clinical data. Prospective multicenter validation and artificial intelligence-powered real-time risk stratification systems represent essential steps to overcome current barriers in precision medicine translation, ultimately advancing personalized therapeutic strategies for PDAC.

Key Words: Pancreatic ductal adenocarcinoma; Postoperative recurrence; Risk assessment system; Preoperative assessment; Radiomics

Core Tip: This editorial critically evaluated the retrospective study by Liu et al in validating a preoperative computed tomography-based risk scoring system, while high-lighting the potential of radiomics integrated with machine learning for multimodal data analysis. However, clinical translation continues to face significant challenges, including heterogeneity in imaging parameters, incomplete elucidation of molecular mechanisms, and confounding effects of therapeutic interventions. Future research should prioritize the development of multidimensional predictive frameworks that synergize radiological, molecular, and clinical data. Prospective clinical trials to validate model efficacy and artificial intelligence-driven real-time risk stratification systems represent critical initiatives to overcome current limitations in precision medicine implementation, thereby advancing personalized management strategies for pancreatic ductal adenocarcinoma.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) presents a major therapeutic challenge in global oncology due to its aggressive biological behavior and high postoperative recurrence rates. Although radical resection remains the only potentially curative intervention, the 5-year survival rate persists below 10%, with a postoperative recurrence risk reaching as high as 70% within 3 years[1,2]. Emerging as a key challenge in overcoming current prognostic limitations is the identification of high-risk recurrence populations through precision preoperative evaluation to guide therapeutic optimization. Recent advances integrating radiomics, machine learning, and multimodal clinical data have revolutionized PDAC preoperative risk stratification, with particularly notable progress in cross-population validation and multimodal integration studies of computed tomography (CT)-based risk scoring systems[3-5].

PDAC

Conventional imaging assessment demonstrates inherent limitations in quantifying tumor heterogeneity, as it primarily relies on radiologists' subjective evaluation of morphological characteristics, tissue density patterns, and invasion boundaries[6,7]. The preoperative CT-based risk stratification system developed by South Korean researchers—incorporating five imaging parameters (tumor size, portal venous phase density, tumor necrosis, peripancreatic infiltration, and suspected metastatic lymph nodes)—has demonstrated validated predictive value in indigenous populations[2], though its clinical generalizability across ethnic cohorts requires critical validation. A Chinese cohort study of 283 resectable PDAC patients has provided inaugural evidence demonstrating the model's robust risk stratification capability in Chinese populations (Table 1)[8]. Interobserver agreement analysis between two radiologists revealed a Kappa value of 0.78, confirming its clinical applicability[9]. These findings align with the Dutch nationwide study outcomes where asymptomatic recurrence patients demonstrated significantly higher treatment rates and improved survival prognosis, suggesting the potential generalizability of radiomics-based models across heterogeneous healthcare environments[10].

Table 1 Recurrence rates and median recurrence-free survival by risk stratification.

Risk groups (risk score)	Recurrence rates (reader 1)	Recurrence rates (reader 2)	Median recurrence-free survival
Low risk (< 2 points)	39.0%	50.0%	Significantly longer
Intermediate risk (2-4 points)	82.1%	79.5%	Shorter
High risk (≥ 5 points)	84.5%	88.9%	The shortest

Risk scores were calculated based on five computed tomography-derived variables: (1) Tumor size: 0 points for ≤ 2 cm, 1 point for 2-4 cm, 2 points for ≥ 4 cm; (2) Portal venous phase density: 0 points for isodense/hyperdense, 1 point for hypodense; (3) Tumor necrosis: 0 points for absent, 2 points for present; (4) Peripancreatic infiltration: 0 points for absent, 1 point for present; and (5) Suspicious metastatic lymph nodes: 0 points for absent, 2 points for present. Total scores were stratified into low risk (< 2), intermediate risk (2-4), and high risk (≥ 5).

Notably, this model's evaluation metrics are exclusively derived from routine dual-phase contrast-enhanced CT scans, requiring neither specialized techniques nor invasive procedures, thereby exhibiting favorable clinical applicability[9]. The composite feature of portal venous phase tumor hypodensity (suggestive of aberrant angiogenesis) combined with tumor necrosis (reflecting metabolic activity) carries a risk weight of 2-4 points, demonstrating strong positive correlation with postoperative recurrence risk. This integration of multidimensional biomarkers enables more comprehensive characterization of tumor biological behavior compared to conventional solitary pathological parameters such as lymph node metastasis.

This study[9] demonstrates significant progress in validating the cross-population efficacy of CT-based risk stratification models, while several limitations necessitate further investigation. First, the majority of enrolled patients did not receive neoadjuvant chemotherapy, a therapeutic intervention proven to enhance tumor resectability and reduce postoperative recurrence risk, potentially constraining the model's applicability in modern multimodal treatment regimens. Second, the proportion of low-risk tumors (≤ 2 centimeters) in the Chinese cohort (23.3%) was lower than that reported in the original Korean study, potentially compromising the model's robustness in predictive performance. Smaller tumors typically demonstrate less aggressive biological behavior with greater heterogeneity, and their underrepresentation may introduce bias in early-stage disease risk stratification. Furthermore, insufficient incorporation of population-specific epidemiological characteristics—particularly genetic variations and comorbidity distributions prevalent in Chinese cohorts—necessitates future model calibration using indigenous biomarker data. Finally, the 24-month median follow-up duration precludes comprehensive evaluation of 5-year survival rates and late-phase recurrence patterns, thereby limiting prognostic precision for long-term clinical outcomes.

The evolution of precision medicine has revealed inherent limitations in unidimensional imaging-based models, positioning the integration of multidimensional data as a critical determinant for optimizing predictive accuracy. In radiomics research, an investigation incorporating 1409 CT textural features demonstrated model efficacy using the extreme gradient boosting (XGBoost) algorithm, achieving training and validation area under the receiver operating characteristic curve (AUC) values of 0.93 and 0.63, respectively. This framework successfully quantified tumor-stroma ratio (TSR), with TSR values demonstrating strong correlations with tumor microenvironment fibrotic status and immune evasion mechanisms in subsequent validation cohorts[5,11,12]. Elevated preoperative serum levels of cancer antigen 19-9 and C-reactive protein, along with decreased skeletal muscle index (indicative of sarcopenia), have been identified as significant predictors of adverse postoperative outcomes in pancreatic cancer patients. These findings suggest that systemic inflammatory status and metabolic dysregulation may adversely affect prognosis through their potential role in facilitating the development of micrometastases[13-15].

Significantly, machine learning exhibits distinct strengths in integrating multi-modal data. Research demonstrates that predictive models leveraging venous-phase radiomic characteristics attain 0.801 accuracy in validation cohorts, with inclusion of supplementary clinical parameters (including tumor marker concentrations and lesion morphology) augmenting the model's AUC to 0.904[16]. Multi-omics data synthesis has further proven capable of delineating cross-modality biological interactions, effectively addressing fundamental constraints of single-modality approaches[17]. These observations imply that coordinated incorporation of radiomic profiles with novel liquid biopsy markers—specifically circulating tumor DNA (ctDNA) mutational load and longitudinal ctDNA surveillance—may refine clinical prediction models, thereby advancing evidence-based decision architectures in precision oncology.

Despite substantial advancements in current research, clinical translation continues to confront significant obstacles. Primarily, in multicenter studies, heterogeneity in CT acquisition parameters (including reconstruction kernel, slice thickness, and tube voltage) significantly impacts model generalizability. Variations in reconstruction kernels demonstrate the most substantial influence on radiomic feature stability, while adjustments in slice thickness and field of view may reduce feature reproducibility. Differences in dose output and reconstruction algorithms across CT manufacturers lead to significant variations in radiomic feature stability. To address these challenges, standardized imaging protocols or transfer learning methods could harmonize radiomic feature disparities between European and Asian medical centers[18-21]. Secondarily, the underlying mechanisms linking imaging biomarkers to molecular pathology remain incompletely characterized. The spatial mapping between TSR-related radiomic patterns and corresponding histopathological stromal proportions with T-cell infiltration densities requires elucidation through integrated spatial transcriptomic analyses. Furthermore, contemporary models predominantly derive from treatment-naïve cohorts, whereas neoadjuvant chemotherapy-induced radiographic alterations (particularly tumor regression grade variations) may introduce prediction bias, mandating prospective validation studies across diverse therapeutic contexts[22-24].

Future research should focus on three key directions. First, multidimensional predictive networks should be established through the development of integrated imaging-molecular-clinical models incorporating genomic alterations, metabolic biomarkers, and gut microbiome profiles. Second, systematic validation through prospective clinical trials should assess the clinical utility of these models in guiding neoadjuvant therapeutic decision-making. Finally, implementation of artificial intelligence (AI)-enabled electronic health record (EHR) systems with real-time data integration could enable adaptive risk stratification and establish predictive-therapeutic-monitoring frameworks to support precision medicine initiatives. A nationwide study in the Netherlands demonstrated that early detection, personalized treatment, and standardized monitoring significantly influence pancreatic cancer prognosis. AI-enabled EHR systems can facilitate the translation of these findings into precision medicine tools through data aggregation, intelligent analysis, and adaptive decision-making.

CONCLUSION

The study conducted by Liu et al[8] validated the cross-population application of radiomics models, marking a significant step toward standardization in preoperative prediction of PDAC. Three key priorities for further research include: (1) Developing multidimensional predictive models that synthesize imaging, molecular, and clinical data through incorporation of genomic profiles, metabolic features, and gut microbiota characteristics; (2) Validating the clinical efficacy of these models in guiding neoadjuvant treatment strategies via prospective trials; and (3) Establishing EHR-integrated AI systems to enable dynamic risk stratification and implement predictive-therapeutic-monitoring closed-loop management frameworks. Interdisciplinary collaboration and standardized imaging protocols constitute the foundational requirements for mitigating multicenter data heterogeneity and advancing clinical translation, ultimately supporting personalized therapeutic approaches for PDAC patients.