AI Model Detects Early Pancreatic Cancer on CT Scans

Medscape

13 hours ago

TOPLINE:
An artificial intelligence (AI) model achieved a sensitivity of 91.8% in detecting pancreatic cancer on CT scans at diagnosis, with a 53.9% sensitivity for scans taken 1 year or more before diagnosis. The model demonstrated a sensitivity of 82.9% for stage I pancreatic cancer detection, suggesting potential for earlier diagnosis.
METHODOLOGY:
This analysis included 1083 patients (mean age, 68.9 years; 575 men) with biopsy-confirmed pancreatic cancer from Danish medical registries between 2006 and 2016.
Researchers evaluated 1220 CT scans, including 1022 concurrent diagnosis scans acquired within 2 months of histopathologic diagnosis and 198 prediagnosis scans obtained before diagnosis (median, 7 months prior).
The PANCANAI model, previously trained on 2134 portal venous CT scans, was tested for pancreatic cancer detection through lesion identification and main pancreatic duct dilation assessment.
TAKEAWAY:
The AI model demonstrated a high sensitivity of 91.8% (95% CI, 89.9%-93.5%) for concurrent diagnosis scans and 68.7% (95% CI, 62.1%-75.3%) for prediagnosis scans.
The performance varied on the basis of the contrast phase, with sensitivities of 92.1% (95% CI, 90.3%-93.6%) for portal, 90.9% (95% CI, 83.6%-96.4%) for arterial, and 83.5% (95% CI, 70%-96.7%) for delayed phases.
The model maintained effectiveness across different cancer stages, achieving sensitivities of 83.1% for stage I, 85.5% for stage II, 94.9% for stage III, and 93.0% for stage IV cases.
For smaller subgroups, a sensitivity of 53.9% (95% CI, 41.8%-65.7%) was observed for CT scans acquired more than 1 year before diagnosis and 24.5% (95% CI, 6.3%-43.8%) for scans acquired more than 2 and a half years before diagnosis.
IN PRACTICE:
"This study showed that PANCANAI was able to detect pancreatic cancer in approximately half of the CT scans acquired more than a year before histopathologic diagnosis. This result suggests that the algorithm may enable timely diagnosis, which could drastically improve patients' survival," the authors of the study wrote.
SOURCE:
This study was led by Laura Degand, MSc, Novo Nordisk Foundation Center for Protein Research, Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark. It was published online on June 24, 2025, in Investigative Radiology.
LIMITATIONS:
The study cohort consisted of only patients with pancreatic cancer, limiting the evaluation to sensitivity without the initial specificity assessment. Additionally, most CT scans were from patients diagnosed at stage IV or with undetermined staging, constraining the evaluation of the algorithm on early-stage cases. The researchers also noted technical limitations that prevented a proper evaluation of the model's segmentation accuracy through radiologist verification.
DISCLOSURES:
No funding information was provided for this study. Several authors reported receiving funding from and having other ties with various sources.
This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.