Background Accurate preoperative grading of pancreatic neuroendocrine tumors (PNETs) is essential for optimal treatment selection, yet endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) yields inadequate tissue in up to 40% of cases and carries procedural risks, necessitating reliable noninvasive alternatives. Materials and methods This multicenter retrospective study included 407 surgically confirmed PNET patients across training (n = 244), validation (n = 106), and external test (n = 57) cohorts. We developed a pancreatic radiomics integrated scoring model for PNET (PNET-PRISM), integrating multidimensional CT radiomics features from intratumoral, peritumoral, habitat, and deep learning domains using automated segmentation. A multidimensional deep learning radiomics score (M-DLR Score) was constructed from 13,542 features and combined with clinical variables for preoperative grade prediction. Results PNET-PRISM demonstrated robust performance with AUCs of 0.92, 0.89, and 0.87 in training, validation, and external test sets, respectively, significantly outperforming clinical-only models (ΔAUC = 0.15–0.22, all p < 0.001). The model achieved perfect sensitivity (100%) in external validation and provided accurate grading in 13 of 25 patients (52%) where EUS-FNA yielded insufficient tissue. Net Reclassification Improvement analysis demonstrated significant improvement over clinical models across all datasets (NRI = 0.318–0.406, p ≤ 0.070). M-DLR Score stratification showed a significant association with progression-free survival (HR = 2.050, 95% CI: 1.484–2.833, p < 0.001). Conclusions This validated radiomics-based nomogram serves as a powerful noninvasive decision-support tool for PNET risk stratification, effectively complementing EUS-FNA limitations and enabling optimized treatment pathways, particularly when biopsy is contraindicated or nondiagnostic. Critical relevance statement This CT-based radiomics nomogram reliably grades pancreatic neuroendocrine tumors (PNETs) and predicts prognosis. This study addresses endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) limitations and advances clinical radiology by enabling safer triage and personalized management when tissue diagnosis is uncertain or unavailable. Key Points <UnorderedList Mark="Bullet"> <ItemContent> A CT-based pancreatic radiomics integrated scoring model for PNET (PNET-PRISM) helps when endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) fails. </ItemContent> <ItemContent> In 407 patients, PRISMPNET-PRISM achieved a high area under the curve (AUC) and 100% external sensitivity for triage. </ItemContent> <ItemContent> The multidimensional deep learning radiomics (M-DLR) score stratified progression-free survival (hazard ratio (HR) ≈ 2.05) and rescued nondiagnostic biopsies. </ItemContent> </UnorderedList> Graphical Abstract

PNET-PRISM: a multicenter-validated radiomics nomogram for noninvasive grading of pancreatic neuroendocrine tumors

期刊论文
出版模式：

开放获取
发表日期：

2026年3月24日

查看全文

Ying Li,
Chengwei Chen,
Mingzhi Lu,
Jiajun Liu,
Jieyu Yu,
Danqun Zheng,
Yilun Zheng,
Yixuan Shen,
Fang Liu,
Tiegong Wang,
Xu Fang,
Jing Li,
Jianping Lu,
Chengwei Shao,
Yun Bian

摘要

Background

Accurate preoperative grading of pancreatic neuroendocrine tumors (PNETs) is essential for optimal treatment selection, yet endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) yields inadequate tissue in up to 40% of cases and carries procedural risks, necessitating reliable noninvasive alternatives.

Materials and methods

This multicenter retrospective study included 407 surgically confirmed PNET patients across training (n = 244), validation (n = 106), and external test (n = 57) cohorts. We developed a pancreatic radiomics integrated scoring model for PNET (PNET-PRISM), integrating multidimensional CT radiomics features from intratumoral, peritumoral, habitat, and deep learning domains using automated segmentation. A multidimensional deep learning radiomics score (M-DLR Score) was constructed from 13,542 features and combined with clinical variables for preoperative grade prediction.

Results

PNET-PRISM demonstrated robust performance with AUCs of 0.92, 0.89, and 0.87 in training, validation, and external test sets, respectively, significantly outperforming clinical-only models (ΔAUC = 0.15–0.22, all p < 0.001). The model achieved perfect sensitivity (100%) in external validation and provided accurate grading in 13 of 25 patients (52%) where EUS-FNA yielded insufficient tissue. Net Reclassification Improvement analysis demonstrated significant improvement over clinical models across all datasets (NRI = 0.318–0.406, p ≤ 0.070). M-DLR Score stratification showed a significant association with progression-free survival (HR = 2.050, 95% CI: 1.484–2.833, p < 0.001).

Conclusions

This validated radiomics-based nomogram serves as a powerful noninvasive decision-support tool for PNET risk stratification, effectively complementing EUS-FNA limitations and enabling optimized treatment pathways, particularly when biopsy is contraindicated or nondiagnostic.

Critical relevance statement

This CT-based radiomics nomogram reliably grades pancreatic neuroendocrine tumors (PNETs) and predicts prognosis. This study addresses endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) limitations and advances clinical radiology by enabling safer triage and personalized management when tissue diagnosis is uncertain or unavailable.

Key Points

A CT-based pancreatic radiomics integrated scoring model for PNET (PNET-PRISM) helps when endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) fails.

In 407 patients, PRISMPNET-PRISM achieved a high area under the curve (AUC) and 100% external sensitivity for triage.

The multidimensional deep learning radiomics (M-DLR) score stratified progression-free survival (hazard ratio (HR) ≈ 2.05) and rescued nondiagnostic biopsies.

Graphical Abstract