Click here to

Session: Imaging Data Science for Treatment Assessment [Return to Session]

Delta Radiomic Features Predict Failure and Survival Outcomes for Surgically Resected Pancreatic Cancer Patients Treated with Neoadjuvant Therapy

K Wang*, A Elamir, J Karalis, S Enrico, P Polanco, T Aguilera, M Ligorio, J Wang, UT Southwestern Medical Center, Dallas, TX


TH-C-202-3 (Thursday, 7/14/2022) 10:00 AM - 11:00 AM [Eastern Time (GMT-4)]

Room 202

Purpose: Neoadjuvant therapy (NAT) has emerged as one of the standards of care for pancreatic ductal adenocarcinoma (PDAC) patients undergoing surgical resection. However, the operations to resect pancreatic cancer are morbid, and some patients do not receive a meaningful clinical benefit. As such, improved tools to predict patient survival before surgery are critically needed. In this study, we built radiomics-based models to predict overall survival (OS), disease-free survival (DFS), and distant-failure-free survival (DFFS), using the baseline radiomic features and NAT-induced change in radiomic features.

Methods: We retrospectively queried our institutional database to identify PDAC patients treated with NAT followed by surgical resection who had pre-NAT and post-NAT contrast-enhanced CT scans available. We extracted 257 radiomic features from the segmented GTVs in each scan. The net changes between pre-and post-NAT radiomic features were defined as NAT-induced delta-radiomic features (DRF). After removing similar features, we used Cox regression with a step forward feature selection method to construct the models. The average risk scores of each patient in repeated 5-fold cross-validation were recorded and used to differentiate high- and low-risk groups for the three outcomes. We also built survival prediction models using pre-NAT radiomic features, post-NAT radiomic features, and post-operative clinical features separately for comparison. C-index and log-rank tests were used to assess the models.

Results: Our study cohort was comprised of 58 patients. The DRF-based models performed better than other models built with radiomic features or post-operative clinical features, with C-indices of 0.74, 0.74, and 0.73 for OS, DFS, and DFFS, respectively. For all three outcomes, log-rank tests showed the low-risk groups have significant better survival than the high-risk groups.

Conclusion: The DRF method is a promising tool to predict the survival of patients undergoing surgical resection of PDAC following NAT. These findings should be prospectively validated in an independent patient cohort.


Not Applicable / None Entered.


IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email