Deep Learning Detection of Direct and Indirect Imaging Findings Associated with Pancreatic Cancer at Contrast-enhanced and Noncontrast CT
Menée à partir de données portant sur 2 251 patients (âge moyen : 66 ans ; 850 hommes), cette étude évalue la performance de modèles d'apprentissage profonds pour diagnostiquer un cancer du pancréas à partir de signaux directs ou indirects provenant de tomographies numériques avec ou sans rehaussement de contraste
Background : Deep learning (DL) models have shown promise in diagnosing pancreatic cancer (PC); however, models that simultaneously detect both direct and indirect imaging findings associated with PC are lacking.
Purpose : To develop and evaluate DL models that detect direct and indirect imaging findings on noncontrast CT (NCCT) and contrast-enhanced CT (CECT) images for PC diagnosis.
Materials and Methods : This retrospective study from August 2007 to December 2022 included patients with PC and control patients. Two DL models were developed using NCCT and CECT to detect direct (pancreatic mass) and indirect (parenchymal atrophy, main pancreatic duct [MPD] dilatation, and MPD stenosis) imaging findings and diagnose PC based on these findings. For training and validation, CT scans from multiple institutions were manually annotated. Model evaluation was performed using two external test sets (CECT and NCCT sets). Receiver operating characteristic curve analysis was used to assess diagnostic performance. Model performance in detecting imaging findings and PC was compared with the performance of six physicians. The reference standard for PC diagnosis was histopathologic confirmation.
Results : This study included 2251 patients (mean age, 66 years ± 13.3 [SD]; age range, 20–96 years; 850 men). DL models demonstrated area under the receiver operating characteristic curve (AUC) values of 0.94, 0.90, 0.94, and 0.94 in the CECT set and 0.88, 0.88, 0.95, and 0.93 in the NCCT set for detecting pancreatic masses, parenchymal atrophy, MPD dilatation, and MPD stenosis, respectively. For PC diagnosis, DL models performed similarly to or better than the mean of six readers in the CECT (AUC, 0.99 vs 0.99; P = .84) and NCCT (AUC, 0.93 vs 0.91; P = .03) sets. For PCs that were 20 mm or smaller, the DL models demonstrated higher sensitivity than the reader mean in both the CECT (98% vs 82.6%; P < .001) and NCCT (86% vs 41.1%; P < .001) sets.
Conclusion : DL models detected direct and indirect imaging findings on CT images and diagnosed PC with performance comparable to or better than that of physicians, particularly for small PCs.
Radiology , article en libre accès, 2026