AI-powered Early Detection for Pancreatic Cancer Via Non-contrast CT in Opportunistic Screening Cohort

Overview

Pancreatic ductal adenocarcinoma (PDAC) remains a therapeutic challenge with 5-year survival rates of 13%, primarily attributable to advanced-stage diagnosis (AJCC Stage III/IV in \>80% of cases). This prospective, observational, multi-center study will evaluate the performance of an AI-powered opportunistic screening system utilizing non-contrast computed tomography (NCCT) acquired during routine clinical encounters or health check-ups. The proposed AI model will perform automated detection of pancreatic parenchymal abnormalities, including PDAC and precursor lesions (intraductal papillary mucinous neoplasms \[IPMN\], mucinous cystic neoplasms \[MCN\]). Algorithm-positive cases will be independently reviewed by two radiologists. Highly suspected individuals will undergo further diagnostic verification, including serological tests and multimodal imaging confirmation. Patients with confirmed positive diagnosis will receive multidisciplinary consultation and specialized treatment, whereas those with negative results will undergo at least one-year clinical follow-up. This study will quantitatively evaluate the AI system's performance, and aims to advance PDAC early detection, improve patient outcomes, and make it accessible in underserved populations.

PDAC is projected to become the second-leading cause of cancer mortality by 2030, with stage-specific survival disparities reaching 83.7% for stage IA versus 2.9% for stage IV disease. This dramatic survival gradient highlights the transformative potential of stage migration through early detection. Screening-based early detection has demonstrated improved prognosis for PDAC patients; however, implementation faces dual challenges. he low incidence of PDAC renders population-wide screening cost-ineffective, while current screening methods are hampered by high false-positive rates and overdiagnosis risks. In this context, opportunistic screening has garnered attention for its unique implementation advantages. By leveraging existing imaging resources from routine clinical encounters or health check-ups, this approach obviates the need for additional screening infrastructure, potentially reducing healthcare resource consumption while effectively increasing screening coverage among high-risk populations. Non-contrast computed tomography (NCCT), despite its widespread clinical application and operational convenience, is limited by suboptimal soft tissue resolution, resulting in insufficient sensitivity for early pancreatic lesions (≤2 cm), thus significantly constraining its utility in opportunistic screening. Recent advancements in AI technology have significantly impacted the field of medical image analysis. These techniques have enabled the automation of the detection of subtle pancreatic lesion features in large-scale imaging data, with the potential to enhance the accuracy and efficiency of early pancreatic cancer detection. In preliminary research, a deep learning-based model for pancreatic cancer detection was developed by our team. This model demonstrated the ability to accurately detect and classify pancreatic lesions on NCCT images, with excellent performance in multicenter validation studies. The model also exhibited strong generalizability when applied to chest CT scans. Therefore, AI-powered NCCT shows significant potential for application in hospital-based opportunistic screening programs and may become an effective tool for early pancreatic cancer detection. However, further research is required to fully explore and realize this potential. This prospective, observational, multi-center study will evaluate the performance of an AI-powered opportunistic screening system utilizing NCCT acquired during routine clinical encounters or health check-ups. The deep learning-based detection system will perform automated identification of pancreatic lesions, including PDAC and precursor entities (intraductal papillary mucinous neoplasms \[IPMN\], mucinous cystic neoplasms \[MCN\]). Algorithm-positive cases will be independently reviewed by two radiologists. Individuals with high suspicion after radiologists review will undergo further validation via serological tests (e.g., CA19-9, CEA) and imaging studies (e.g., contrast-enhanced CT, contrast-enhanced MRI, EUS-FNA). Participants with a confirmed positive diagnosis will undergo multidisciplinary consultation and specialized treatment, while those with a negative diagnosis will be followed clinically for at least one year. The AI system's performance will be evaluated through three primary metrics: (1) Detection rate of PDAC and high-risk precursor lesions, defined as the proportion of histologically confirmed PDAC and precursor lesions (IPMN/MCN) meeting Sendai criteria among all participants undergoing CT screening. (2) Recall rate, defined as the proportion of individuals recalled for confirmatory testing after AI-positive screening and radiologist review among all participants undergoing CT screening. (3) Positive predictive value (PPV) defined as the proportion of histologically confirmed PDAC and high-risk precursor lesions among all AI-positive screening cases. Institutional Collaboration: Led by Shanghai Changhai Hospital (PI: Gang Jin, MD) with five regional centers (Yinzhou Hospital, Jiaxing University Hospital, Lishui Central Hospital, Jingning County Hospital) and Alibaba DAMO Academy (technical support).