Deep learning model automates pancreas lesion classification on CT across 18 institutions

A deep learning model classified pancreas lesions on portal-venous CT as normal, benign, or malignant, achieving an area under the curve (AUC) of 0.84 for all lesions and 0.97 for malignant lesions specifically in an external test set of 166 patients.

• 1,037 contrast-enhanced CT scans from 18 centers were split into training (n=732) and test sets for internal (n=139) and external (n=166) validation.
• The primary finding was consistent performance in internal and external validation (final score 0.72), with malignant lesion detection AUC reaching 0.97 externally.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.