The following is a summary of “Large-scale pancreatic cancer detection via non-contrast CT and deep learning,” published in the November 2023 issue of Cardiology by Cao et al.

Pancreatic ductal adenocarcinoma (PDAC), known for late-stage diagnosis and low operability rates, poses a significant challenge in early detection due to its asymptomatic nature. Screening strategies for PDAC remain limited, often hindered by low prevalence and the potential for false positives. Non-contrast computed tomography (CT), a routinely conducted clinical procedure, offers a viable option for large-scale screening. However, identifying PDAC through non-contrast CT has historically been deemed unattainable. In this study, the researchers introduce a deep learning framework, termed pancreatic cancer detection with artificial intelligence (PANDA), designed to detect and accurately classify pancreatic lesions using non-contrast CT.

PANDA was trained using a dataset encompassing 3,208 patients from a single center. Upon validation across multiple centers involving 6,239 patients, PANDA demonstrated high accuracy, achieving an area under the receiver operating characteristic curve (AUC) ranging from 0.986 to 0.996 for lesion detection. Notably, compared to radiologist performance, PANDA outperformed 34.1% in sensitivity and 6.3% in specificity for PDAC identification. Furthermore, in a real-world multi-scenario validation comprising 20,530 consecutive patients, PANDA showcased a sensitivity of 92.9% and specificity of 99.9% for lesion detection.

Critically, PANDA’s application with non-contrast CT exhibited non-inferiority to radiology reports employing contrast-enhanced CT in distinguishing between common pancreatic lesion subtypes. These results underscore the potential of PANDA as a groundbreaking tool for large-scale pancreatic cancer screening, offering a promising avenue for early detection and improved clinical outcomes.

Source: nature.com/articles/s41591-023-02640-w