First PANORAMA findings support AI-based opportunistic screening

Sunday, December 1 | 10:40 a.m.-10:50 a.m. | S2-SSGI01-2 | E451B

Transparently benchmarked, expert AI can enable opportunistic screening to start catching pancreatic ductal adenocarcinoma (PDAC) on contrast-enhanced computed tomography (CECT) at earlier stages, according to first results of the PANORAMA (Pancreatic cancer diagnosis: Radiologists meet AI) study.

Early-detected PDAC patients have improved survival, but 40% of lesions are missed in prediagnostic imaging, noted the team from Radboud University Medical Center, which includes John Hermans, MD, PhD, Megan Schuurmans, Henkjan Huisman from the Diagnostic Imaging Analysis Group at Radboud University, and others.

With 165 registered participants, the PANORAMA study aims to benchmark PDAC detection AI algorithms, developed in a Grand Challenge, against abdominal radiologists participating in the reader study -- to evaluate the clinical viability of modern pancreas-AI at PDAC detection and diagnosis in CECT, states the project website. The challenge designers hypothesized that state-of-the-art AI algorithms are noninferior to radiologists. See the Panorama timeline here.

PANORAMA's study protocol involved an international scientific advisory board of 13 experts in AI, radiology, and histopathology. According to the researchers, the study aims to ensure transparent development and validation of pancreas-AI as a benchmark for clinical translation.

Currently, 45 radiologists from 26 centers in 13 countries participate, according to the study abstract. In Chicago, the researchers will present findings from their retrospective study of 3,338 CECT abdominal scans of patients without a prior history of treatment or positive histopathology findings of PDAC.

Scans were acquired between 2006 and 2021 from five centers in the Netherlands, Norway, and Sweden. Among the findings to be explained Sunday, the baseline AI (nnU-Net with cross-entropy loss) achieved an area under the receiver operating characteristic curve of 0.98 in the tuning dataset. In addition, the researchers expect a significant improvement over the baseline in the testing set, in line with previous challenges' results.

Proving noninferiority of AI to expert radiologists could enable AI-based opportunistic screening. Don't miss this session as we get set for an announcement of the challenge winners!

Page 1 of 2
Next Page