Artificial Intelligence-powered Low-Dose Computed Tomography for Screening of Pancreatic Cancer

High-grade Pancreatic Intraepithelial Neoplasia, Intraductal Papillary Mucinous Neoplasm, Mucinous Cystic Neoplasm, Pancreatic Cancer, PDAC - Pancreatic Ductal Adenocarcinoma

Screening, Early Diagnosis, Pancreatic Cancer, Artificial Intelligence, Computed Tomography

Pancreatic ductal adenocarcinoma (PDAC) has a poor prognosis, with early diagnosis crucial for improving survival. Due to the absence of effective screening methods, most patients are diagnosed at advanced stages. The population undergoing low-dose computed tomography (LDCT) screening significantly overlaps with those at high risk for PDAC; however, traditional imaging methods have limited sensitivity for detecting pancreatic lesions. This study utilizes the Pancreatic Cancer Detection with Artificial Intelligence (PANDA) system to enhance LDCT for pancreatic cancer screening in a prospective, multicenter, observational cohort. PANDA will analyze LDCT images, followed by a multidisciplinary team (MDT) reassessment of abnormal interpretations. Based on MDT evaluation, individuals will be recalled for further examination, placed under a personalized follow-up plan, or monitored for at least one year. The primary outcomes include pancreatic cancer detection rate, positive predictive value, consensus rate, and recall rate, while secondary outcomes focus on early-stage cancers, resectable tumors, and safety indicators such as false positive rates and unnecessary procedures. This study aims to assess the effectiveness and safety of AI-assisted LDCT for PDAC detection, providing a practical solution for improving public health and enhancing early diagnostic capabilities.

Pancreatic ductal adenocarcinoma (PDAC) is an extremely aggressive cancer with a dismal 5-year survival rate of just 13%. The key to improving outcomes lies in early detection, as patients diagnosed at an early stage (IA) can achieve an 80% 5-year survival. However, current screening methods are limited, focusing only on high-risk populations and lacking effectiveness for the general public due to the cancer's relatively low incidence and high false-positive risks. Contrast-enhanced CT (CE-CT), the primary imaging modality, faces barriers for widespread implementation due to its invasiveness, high costs, and need for contrast agents. In this context, low-dose CT (LDCT) emerges as a promising alternative, having demonstrated success in lung cancer screening by reducing radiation exposure. Retrospective analysis revealed that one-third of pancreatic abnormalities were missed during routine LDCT interpretations, suggesting the untapped potential of LDCT-based pancreatic lesion screening. Breakthroughs in AI have transformed medical imaging. Our PANDA (pancreatic cancer detection with artifcial intelligence) system excels at pancreatic cancer detection, utilizing innovative registration techniques and a cascaded deep learning framework (UNet+Max-Deeplab) for comprehensive lesion analysis. Validated across 10 centers (6,239 patients), PANDA outperformed radiologists. Real-world testing (20,530 cases) demonstrated remarkable accuracy: 92.9% sensitivity and 99.9% specificity, maintaining 92.2% sensitivity even for small T1 tumors. On LDCT, PANDA achieved 0.979 AUC without protocol modifications, confirming LDCT+AI as a viable screening approach. China's health check-up environment presents three key advantages: First, LDCT delivers just 1/4-1/5 the radiation of standard abdominal CT, staying within ICRP safety guidelines (\<3mSv). Second, LDCT offers superior cost-effectiveness compared to CE-CT by eliminating contrast agent expenses. Third, China's extensive annual health check-up infrastructure provides an unparalleled foundation for widespread implementation. In the study, we will conduct a prospective, multicenter, observational cohort design targeting a health check-up population, utilizing the PANDA system to enhance LDCT for pancreatic cancer screening. Initially, PANDA analyzes the LDCT images of participants and provides interpretation results. Subsequently, a multidisciplinary team (MDT) will re-evaluate the cases with positive AI findings (including PDAC, pancreatic precursor lesions and benign lesion) and determine whether to recall the individuals: (1) Suspected PDAC and pancreatic precursor lesions are referred for hospital examination with diagnostic results collected; (2) Benign lesion cases receive personalized monitoring until endpoint events or study end; (3) Cases with positive AI findings but MDT-confirmed normal pancreatic issues receive at least one year of follow-up. If any abnormal results arise, management will transition to either plan (1) or (2). The primary outcome measures include pancreatic cancer detection rate, positive predictive value, consensus rate, and recall rate. Secondary outcome measures include the proportion of early-stage pancreatic cancers and resectable tumors. Safety indicators include the false positive rate, the proportion of unnecessary invasive procedures, and the proportion of unnecessary surgeries. This study aims to evaluate the effectiveness and safety of AI-powered LDCT in detecting pancreatic cancer within a health check-up population, offering a practical solution to improve public health and early diagnosis for pancreatic cancer.

No related papers found yet.

China