A clinically validated 3D deep learning approach for quantifying vascular invasion in pancreatic cancer.

Abstract

Vascular invasion assessment is critical for surgical planning in pancreatic ductal adenocarcinoma (PDAC). Current CT-based assessments often rely on radiologists' subjective 2D interpretations, which may not capture the continuous, three-dimensional tumor-vessel interactions and multiple vessel involvement, both essential for accurate preoperative evaluation. PAN-VIQ (Pancreatic Vascular Invasion Quantifier) is an automated deep learning framework to quantify tumor-vessel interactions from contrast-enhanced CT scans. It enables segmentation of pancreatic tumors and five critical vessels: celiac artery (CA), common hepatic artery (CHA), superior mesenteric artery (SMA), superior mesenteric vein (SMV), and portal vein (PV), quantifying vascular involvement through 3D encasement angles. PAN-VIQ was trained and internally validated on 2130 cases, and subsequently prospectively tested in 202 patients. External validation showed accuracies exceeding 90%. In prospective evaluation, the model outperformed junior radiologists and matched senior radiologists in accuracy and recall. These results underscore potential of PAN-VIQ to standardize vascular invasion assessment and reduce interobserver variability.