Emerging research suggests that deep learning may provide equivalent detection to that of radiologists for pancreatic cancer (PC) on non-contrast and contrast-enhanced CT (CECT) and significantly better detection for smaller pancreatic cancers.
For the retrospective study, recently published in Radiology, researchers evaluated deep learning models for non-contrast CT (NCCT) and CECT for detection of pancreatic masses as well as indirect findings such as parenchymal atrophy, main pancreatic duct (MPD) dilatation and MPD stenosis. The study authors compared detection with the deep learning models to that of six unassisted radiologists in a total cohort of 2,251 patients (mean age of 66).
In external multicenter validation testing, the researchers found that the CECT DL model provided an equivalent AUC to the mean AUC among radiologists for detecting PC (99 percent for both). The NCCT DI model offered a slightly higher AUC than radiologists (93 percent vs. 91 percent), according to the study authors.
The researchers also pointed out that the AI models for NCCT and CECT provided significantly higher sensitivity rates in comparison to unassisted radiologists. The study authors found that the AI CECT model had a greater than 10 percent higher sensitivity (99 percent vs. 88.9 percent) while the AI NCCT model offered a nearly 31 percent higher sensitivity (84 percent vs. 53.4 percent).
“In multicenter external test sets, both DL models demonstrated diagnostic performance comparable to or better than that of six physicians with varying levels of experience,” noted lead study author Takeru Yamaguchi, MD, who is affiliated with the Department of Radiology at the Kobe University Graduate School of Medicine in Kobe, Japan, and colleagues.
Specifically, the CECT DL model and NCCT DL model had 94 percent and 88 percent AUCs, respectively, for detection of pancreatic masses. For detection of parenchymal atrophy, MPD dilatation and MPD stenosis, the CECT DL model provided 90 percent, 94 percent and 94 percent AUCs, respectively, and the NCCT DL model had 88 percent, 95 percent and 93 percent, according to the study authors.
The researchers also noted significant differences between the DL models and unassisted radiologist interpretation for the detection of pancreatic cancers < 20 mm. The study authors found that the CECT DL model had over a 15 percent higher sensitivity (98 percent vs. 82.6 percent) and the NCCT DL model provided over a 44 percent higher sensitivity.
“ … DL models may be more sensitive to subtle grayscale and morphologic changes than the human visual system and may compensate for limited reader training and experience in NCCT-based PC diagnosis,” posited Yamaguchi and colleagues.
Three Key Takeaways
• Deep learning matches or outperforms radiologists for PC detection. On CECT, the DL model achieved equivalent diagnostic accuracy to radiologists (99 percent AUC for both), while the NCCT model slightly exceeded radiologist performance (93 percent vs. 91 percent AUC), suggesting AI can serve as a reliable diagnostic aid across both imaging modalities.
• AI demonstrates substantially higher sensitivity, especially on non-contrast CT. The NCCT DL model showed nearly 31 percent higher sensitivity than unassisted radiologists (84 percent vs. 53.4 percent), which is clinically significant given that non-contrast CT is more widely available and commonly used as a first-line study. This may potentially expand early detection opportunities in lower-resource settings.
• The advantage is most pronounced for small tumors. For pancreatic cancers under 20 mm — the stage at which surgical resection and survival outcomes are most favorable — the CECT DL model showed over 15 percent higher sensitivity and the NCCT model over 44 percent higher sensitivity compared to radiologists. This is the finding with the greatest potential mortality benefit, though prospective trials are still needed to confirm real-world clinical impact.
While emphasizing the need for subsequent prospective trials in an accompanying editorial, Rajesh Bhayana, MD, and Pranav Rajpurkar, PhD, noted the potential of these DL models in possibly facilitating earlier detection of PC.
“The clinical hypothesis behind every model discussed here is that AI-assisted detection at CT will shift cases toward an earlier stage at diagnosis and that this shift will produce measurable mortality benefit. The present study’s 98% sensitivity for tumors 20 mm or smaller is exactly the capability this hypothesis turns on,” maintained Dr. Bhayana, an abdominal radiologist and artificial intelligence and information technology lead for the Joint Department of Medical Imaging at the University of Toronto, and Dr. Rajpurkar, an associate professor of biomedical informatics at Harvard Medical School.
(Editor’s note: For related content, see “Can a Radiomic AI Model Facilitate Earlier CT Detection of Pancreatic Ductal Adenocarcinoma?,” “Predicting Diabetes on CT Scans: What New Research Reveals with Pancreatic Imaging Biomarkers”and “SNMMI: Preliminary Research Suggests Dual-Targeting Radiopharmaceutical May Have Impact in Multiple Cancers.”
In regard to study limitations, the authors acknowledged a reference standard derived by expert consensus, a relatively homogeneous cohort with respect to race and ethnicity, and a lack of assessment for the AI models in distinguishing pancreatic cancer from other pancreatic diseases.
