Loading ...

TOPLINE:

An automated AI tool (REDMOD) detected hidden pancreatic ductal adenocarcinoma (PDA) on routine CT scans a median of 475 days before clinical diagnosis, achieved nearly double the sensitivity of expert radiologists, showed consistent performance on repeat scans, and maintained performance across external datasets.

METHODOLOGY:

• Routine CT scans often miss early, visually hidden PDA. Manual segmentation of the pancreas on CT images is labor-intensive and inconsistent, underscoring the need for automated AI tools.
• In a retrospective, multi‑institutional study, researchers trained an AI framework (REDMOD) to detect prediagnostic PDA signatures on routine contrast-enhanced abdominal CT scans obtained 3-36 months before diagnosis; control scans were obtained from patients who remained free of PDA for at least 3 years of the follow-up period.
• A fully automated 3D pancreas segmentation was applied. Hundreds of radiomic features, derived using multiscale filters, were extracted and reduced to a 40-feature signature.
• In the training dataset, researchers used synthetic oversampling to balance the small number of prediagnostic cases and trained three machine learning algorithms.
• The performance of the model was evaluated on an independent test set, with a clinically relevant ratio of control and case scans (about 6:1), representing the low prevalence of early pancreatic cancer.

TAKEAWAY:

• Researchers analyzed prediagnostic case scans from 219 patients and control scans from 1243 individuals (median age, 69 years and 64 years, respectively; male-to-female sex ratio, 1.3 and 1.4, respectively). They assigned 969 scans (156 case scans and 813 control scans) to train the model and 493 scans (63 case scans and 430 control scans) to independently test it.
• On the test subset, REDMOD achieved an area under the curve (AUC) of 0.82 for detecting stage 0 PDA, accurately identifying 73.0% of case scans (sensitivity) and ruling out 81.1% of control scans (specificity) at a prespecified threshold. The model identified PDA a median of 475 days before clinical diagnosis.
• REDMOD achieved nearly twofold higher sensitivity in detecting occult PDA than two radiologists (AUC, 0.82 vs 0.69; sensitivity, 73.0% vs 38.9%; P < .001 for both). Sensitivity was nearly threefold higher for scans taken more than 24 months before diagnosis (68.0% vs 23.0%).
• REDMOD showed agreement with earlier repeat scans in about 90%-92% of cases. The model correctly flagged normal pancreases in external datasets about 81.3%-87.5% of the time.

IN PRACTICE:

“This work overcomes key barriers in the field by providing a scalable objective tool that addresses a critical diagnostic gap. While prospective validation is paramount to confirm clinical utility, the REDMOD framework represents a significant advance towards shifting the paradigm for sporadic PDA from a late-stage symptomatic diagnosis to proactive preclinical interception, offering tangible hope for improving outcomes in this challenging disease,” the authors of the study wrote.

SOURCE:

This study was led by Sovanlal Mukherjee, PhD, Mayo Clinic in Rochester, Minnesota. It was published online in Gut.

LIMITATIONS:

This study was not designed to test performance across different racial or ethnic groups, limiting applicability to all populations. Specificity held up in external cohorts, but sensitivity could not be validated because no suitable public prediagnostic datasets were available. REDMOD only measured diagnostic accuracy, not survival.

DISCLOSURES:

This study received funding from multiple sources, including the National Institutes of Health, Mayo Clinic Comprehensive Cancer Center, and Champions for Hope Pancreatic Cancer Research Program of the Funk Zitiello Foundation. The authors did not declare any competing interests.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.