42: Machine-learning based investigation of prognostic indicators for oncological outcome of pancreatic ductal adenocarcinoma

Thursday, March 23, 2023

11:06 AM – 11:16 AM

Location: Ballroom B

Abstract Presenter(s)

Jeremy Chang, MD MS

Resident
University of Iowa Hospitals and Clinics
Iowa City, Iowa, United States

Disclosure(s): No financial relationships to disclose

First Author(s)

Jeremy Chang, MD MS

Resident
University of Iowa Hospitals and Clinics
Iowa City, Iowa, United States

Disclosure(s): No financial relationships to disclose

Author(s)

YL

Yanan Liu, PhD

Post Doc
University of Iowa Hospitals and Clinics - 200 Hawkins DriveC52-P General HospitalIowa City, IA 52242UNITED STATES - Iowa City, IA, United States

Disclosure information not submitted.
SS

Stephanie Saey, n/a

Medical Student
University of Iowa Hospitals, United States

Disclosure information not submitted.
KC

Kevin Chang, n/a

Medical Student
University of Iowa Hospitals and Clinics, United States

Disclosure information not submitted.
HS

Hannah Shrader, n/a

Research Assistant
University of Iowa Hospitals and Clinics, United States

Disclosure information not submitted.
KS

Kelsey Steckly, n/a

Research Assistant
University of Iowa Hospitals and Clinics, United States

Disclosure information not submitted.
MR

Maheen Rajput, MD

Clinical Professor
University of Iowa Hospitals and Clinics, United States

Disclosure information not submitted.
MS

Milan Sonka, PhD

Professor
University of Iowa College of Engineering, United States

Disclosure information not submitted.
Carlos HF Chan, MD, PhD

Associate Professor of Surgery
University of Iowa Hospitals and Clinics
Iowa City, Iowa, United States

Disclosure information not submitted.

Disclosure(s):

Jeremy Chang, MD MS: No financial relationships to disclose

Introduction: The ability of computerized tomography (CT) scans to assess nodal status and resectability in pancreatic ductal adenocarcinoma (PDAC) remains suboptimal and depends heavily on physician experience. Improved preoperative radiographic tumor staging with prediction of postoperative margin and lymph node status could have important implications in treatment sequencing. This paper proposes a novel machine learning model, utilizing a 3-dimensional convoluted neural network (3D-CNN), to reliably predict the presence of lymph node metastasis and postoperative positive margin status based on preoperative CT scans.

Methods: 110 patients presenting to the University of Iowa Holden Cancer Center from 2015-2021 were identified who had the diagnosis of PDAC, underwent surgery and had preoperative CT imaging avilable. Patient imaging data, tumor pathology and operative outcomes were abstracted. Patients were separated into training and validation groups. A 17-layer 3D-CNN was created to characterize CT imaging data. Algorithm prediction accuracy was assessed using per scan and majority voting per patient methods. Kaplan-Meier analysis and Cox proportional hazard ratio analysis was performed to assess overall survival (OS) and recurrence free survival (RFS).

Results: A total of 110 patients receiving 881 CT scans were included in study. In prediction of lymph node positivity, the 3D-CNN model achieved accuracy of 90% for per-patient analysis and 75% for per-scan analysis. In prediction of postoperative margin positivity, the 3D-CNN model achieved accuracy of 81% for per-patient analysis and 76% for per-scan analysis. Survival analysis performed demonstrated decreased median OS for patients with positive postoperative margin (14 months vs 27 months, p=0.015) and positive lymph node metastasis (22 months vs 29 months, p=0.04). In univariate cox analysis, positive margin was associated with worse OS (HR=2.10, p=0.016), local (HR=3.28, p=0.030) and distant RFS (HR=2.64, p=0.010), while positive lymph node was associated with worse OS (HR=2.68, p=0.004).

Conclusions: 3D-CNN based algorithms can predict lymph node metastasis and postoperative margin status with clinically relevant levels of accuracy. Additional studies should be performed with larger datasets to improve training and learning of the pilot model.

Learning Objectives:

Define a 3-dimensional convoluted neural network (3DCNN).
Discuss significance of lymph node metastasis and positive postoperative margin on oncologic outcomes of pancreatic cancer
Debate additional uses of Machine Learning Algorithms in surgical planning.