TO THE EDITOR
We read with great interest the study by Wang et al[1] published in the recent issue of the World Journal of Gastroenterology, which developed and validated a computed tomography (CT)-based nomogram for predicting the depth of portal-systemic venous invasion (PSVI) in borderline resectable pancreatic ductal adenocarcinoma (BR-PDAC). Their shift toward quantifying histologic invasion depth, rather than relying on simple binary assessment, is clinically meaningful. Previous large-scale surgical data have demonstrated a stepwise decline in survival from adventitial to intimal invasion[2]. A preoperative tool capable of anticipating this gradient may refine patient selection for vascular resection and inform decisions regarding neoadjuvant therapy.
Despite these strengths, several limitations restrict the generalizability of the proposed model. As acknowledged by Wang et al[1], the retrospective nature of the cohort and the variability in neoadjuvant strategies may introduce bias and limit transportability. Critical treatment-related variables including regimen type, dose intensity, radiotherapy parameters, and radiologic tumor response were not incorporated. Because neoadjuvant therapy is now strongly emphasized in contemporary BR-PDAC guidelines[3], a model based solely on baseline anatomical and serologic parameters may be insufficient. Pancreatic ductal adenocarcinoma is characterized by aggressive biology and variable chemosensitivity, and tumor characteristics including stromal composition, proliferative activity, and tumor-vessel interface behavior may evolve during neoadjuvant treatment. Consequently, static baseline models may underperform in predicting the true extent of PSVI at the time of surgical exploration. Incorporating dynamic treatment-response markers may therefore improve the robustness and external applicability of PSVI prediction models.
Reliance on carbohydrate antigen (CA) 19-9 as a primary predictor may further limit robustness. Although the authors performed bilirubin-normalized subgroup analyses, CA19-9 is highly susceptible to confounding by biliary obstruction, inflammation, and variability in tumor shedding, all of which are common in BR-PDAC. Notwithstanding these limitations, when confounding factors such as biliary obstruction are adequately accounted for, the dynamic kinetics of CA19-9 particularly its longitudinal change during neoadjuvant therapy remain a powerful and clinically meaningful indicator of tumor biology and treatment response. Additionally, the nomogram did not include radiomic or biological variables, despite growing evidence that vascular engagement is influenced by molecular subtype heterogeneity, stromal activation, and differential chemosensitivity[4]. Integrating radiomics-derived metrics that characterize interface complexity and incorporating biomarker signatures from endoscopic ultrasound-guided tissue acquisition (EUS-TA) may therefore enhance model accuracy and allow a more biologically grounded assessment of PSVI.
Accumulating evidence indicates that PSVI reflects underlying biological programs rather than purely anatomical interactions. Processes such as epithelial mesenchymal transition, stromal remodeling, and perivascular niche alteration mediated by pathways including transforming growth factor-β and Hedgehog signaling modulate tumor-vessel interactions and invasive behavior[5-7]. These mechanisms evolve during disease progression and treatment exposure and therefore cannot be adequately captured by single-time-point, anatomy-based imaging models alone. This biological complexity underscores a fundamental limitation of purely morphologic PSVI assessment and provides a strong rationale for integrating complementary biological and functional parameters into predictive frameworks.
EUS-TA offers an opportunity to incorporate biological information into preoperative assessment. Biomarkers measurable in EUS-TA specimens including human equilibrative nucleoside transporter 1, deoxycytidine kinase, Ki-67, and stromal activation markers have demonstrated associations with chemosensitivity and survival[8-11]. Although these biomarkers do not anatomically quantify invasion depth, they provide indirect yet biologically strong and clinically meaningful insight into tumor aggressiveness, stromal activation, and treatment responsiveness, all of which are closely linked to invasive behavior and PSVI risk. Thus, rather than representing a limitation, this complementary biological information enhances the interpretability of anatomy-based PSVI assessment when integrated with CT morphometrics. Incorporating such biological signatures with CT morphometrics may therefore complement anatomical assessment and enhance the biologic interpretability of PSVI risk models. However, the clinical application of EUS-TA based biomarker profiling is currently limited by incomplete assay standardization, variable institutional availability, and additional costs related to tissue processing and molecular analysis, which may restrict widespread implementation. Importantly, most evidence supporting these biomarkers is derived from single-center or region-specific cohorts, and robust external validation across diverse patient populations and treatment settings remains limited. This lack of generalizability represents a key barrier to their routine incorporation into PSVI prediction models.
Artificial intelligence (AI)-driven radiomics further expand the potential of PSVI modeling. Radiomic signatures including entropy-based texture metrics, gradient-derived attenuation features, and perivascular interface heterogeneity have shown correlations with vascular invasion and aggressive tumor phenotypes[12-14]. A recent systematic review underscored the diagnostic and prognostic potential of radiomics and deep learning across pancreatic cancer imaging[15]. Radiogenomic analyses also demonstrate that radiomic patterns correlate with molecular subtypes, stromal characteristics, and proliferation profiles[16]. Nonetheless, scanner variability, segmentation inconsistency, and domain shift pose substantial challenges, reinforcing the need for harmonized multicenter validation before widespread adoption. In addition, the requirement for specialized computational infrastructure and expertise may further increase operational complexity and limit accessibility in routine clinical settings. Moreover, many radiomic signatures have yet to undergo rigorous external validation across heterogeneous scanners, institutions, and ethnic populations, underscoring the necessity of standardized, multicenter prospective studies.
Taken together, these insights suggest that the next generation of PSVI predictive models should integrate CT morphometrics, EUS-TA biomarker profiling, AI-derived radiomic features, and treatment-response variables within a unified, biologically informed framework. Rather than aiming for a single, all-encompassing “super-nomogram”, we propose a stepwise decision algorithm in which CT-based morphologic assessment serves as the initial screening modality, with selective incorporation of EUS-TA derived biomarkers and AI-based radiomic features in patients with high-risk or indeterminate findings (Figure 1). Such an incremental approach may help balance potential gains in predictive accuracy against increased cost and complexity of care. Future studies systematically correlating EUS-TA derived biomarkers and radiomic features with histologically confirmed PSVI across diverse populations would be essential to establish reproducibility and clinical credibility. Because EUS-TA is routinely performed in BR-PDAC, the addition of biomarker analysis would not impose additional procedural burden.
Figure 1 Stepwise and clinically pragmatic decision algorithm for preoperative portal-systemic venous invasion risk stratification in borderline resectable pancreatic ductal adenocarcinoma.
BR-PDAC: Borderline resectable pancreatic ductal adenocarcinoma; CT: Computed tomography; CA19-9: Carbohydrate antigen 19-9; hENT1: Human equilibrative nucleoside transporter 1; dCK: Deoxycytidine kinase; PSVI: Portal-systemic venous invasion; AI: Artificial intelligence.
