Patient specific Pancreatic Ductal Adenocarcinoma segmentation in multiphase CTs through a registration methodology

Sara Ghezzi, Edoardo Maria Polo, Cristian Drudi, Riccardo Levi, Rami Abou El Zahab, Silvia Carrara, Cristiana Bonifacio, Alessandro Zerbi, Riccardo Barbieri

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Pancreatic Ductal Adenocarcinoma, Deep Learning, Multi-phase CT, Image Registration, Tumor Segmentation, nnU-Net

TL;DR: Image registration methodology on multi-phase abdominal CTs to obtain precise PDAC segmentations of different contrast phases

Abstract:

Dual-phase computed tomography, comprising late arterial (LA) and portal venous (PV) phases, is essential for Pancreatic Ductal Adenocarcinoma (PDAC) assessment. However, automated PDAC detection remains challenging due to phase-dependent contrast variations. We present a patient-specific paradigm using clinically-validated LA phase segmentation and inter-phase registration to align PV phase segmentations with the corrected LA phase, enabling more accurate PV phase annotation. In 21 PDAC patients, we observed consistent automated LA segmentation of single-mass tumors, while 9.5% of automated PV phase segmentations exhibited disconnected regions. Our registration-based approach achieved a median Dice score of 0.85 for pancreas segmentation, significantly improving upon the unregistered PV phase performance of 0.79 and approaching the automated LA phase performance of 0.90. Furthermore, it demonstrated comparable centroid distance accuracy to automated LA segmentation (p>0.05). This approach enables efficient multi-phase PDAC analysis by only requiring manual correction in the optimal LA phase.
Our methodology addresses phase-dependent segmentation challenges while optimizing clinical workflow, potentially improving diagnostic efficiency and treatment planning.

Track: 3. Imaging Informatics

Registration Id: 3LN3KP9YCTM

Submission Number: 398