Pancreatic ductal adenocarcinoma (PDAC) is expected to become the second leading cause of cancer death by 2030. Over 80% of diagnoses occur at an advanced stage, when the five-year survival rate is below 10%. Yet if caught early, survival rates can rise dramatically — up to 50% for very early-stage disease.
The core challenge is that PDAC is asymptomatic or causes only vague, non-specific symptoms until it reaches an advanced stage. This means there is no reliable way to prompt early screening in the general population. Risk prediction — identifying individuals likely to develop PDAC years in advance — offers a path forward: if high-risk individuals are found, they can enter surveillance programs and be diagnosed earlier.
This review article examines how AI combined with pancreatic imaging could transform risk prediction, discussing current limitations of existing risk indicators and building a case for an AI-powered, imaging-based prediction framework.
Several risk factors for PDAC are well established: family history of pancreatic cancer (familial pancreatic cancer, FPC), new-onset diabetes (particularly in individuals over 50), chronic pancreatitis, obesity, smoking, heavy alcohol use, and poor diet. Each of these individually raises risk, but none is specific enough alone to justify targeted PDAC screening.
New-onset diabetes is particularly important: patients diagnosed with diabetes have a two- to eight-fold higher risk of developing PDAC within three years compared to the non-diabetic population. About 3-10% of new-onset diabetics over 50 will be found to have PDAC. Smoking increases PDAC risk by 25% and obesity by 20% according to the American Cancer Society.
The fundamental limitation is that none of these risk factors is specific enough to PDAC. They are general indicators associated with many other diseases. A more refined approach requires combining multiple risk indicators — and AI is particularly well suited to integrate these diverse data streams into a coherent risk score.
Pancreatic imaging — CT, MRI, and endoscopic ultrasound — plays a central role in PDAC management. But this review makes the case that imaging holds an underutilized potential role in risk prediction, not just diagnosis. The pancreas undergoes biological adaptations as it progresses toward cancer, manifesting as subtle morphological and textural changes on imaging that appear before a visible tumor forms.
These changes include pancreatic duct dilation, pancreatic atrophy (shrinkage), textural heterogeneity (uneven tissue appearance), and the presence of precancerous lesions like intraductal papillary mucinous neoplasms (IPMNs) and pancreatic intraepithelial neoplasia (PanINs). While experienced radiologists can identify overt changes, subtle micro-level alterations typically go unnoticed in routine reads.
AI models trained on large imaging datasets can potentially detect these subtle precancerous changes with superhuman consistency, identifying patterns that predict future cancer development years in advance — a form of imaging-based early warning system.
The review builds its case for imaging-based PDAC risk prediction by pointing to successful applications in other cancers. In breast cancer, AI models analyzing mammograms can predict future cancer development years in advance by detecting texture changes invisible to human readers. In lung cancer, AI applied to low-dose CT scans identifies high-risk nodules requiring follow-up.
In prostate cancer, AI integration of MRI features with clinical variables has improved risk stratification beyond PSA testing alone. The common thread: AI finds patterns in imaging data that encode biological processes preceding overt cancer — processes that are systematic and learnable but too subtle for routine human detection.
The authors argue that PDAC is particularly suited to an imaging-based risk prediction approach because the precancerous stages (IPMNs, PanINs) are biologically progressive, last years, and produce measurable imaging changes. The challenge is building the large, well-annotated imaging datasets needed to train such models.
The authors propose a framework where AI models routinely analyze abdominal imaging — obtained for any clinical reason — to flag individuals whose pancreatic appearance suggests elevated cancer risk. These individuals would then enter structured surveillance programs, catching cancers at resectable, curable stages.
Key challenges for realizing this vision include the need for large longitudinal imaging databases with long follow-up periods, standardization of imaging protocols to ensure AI models generalize across institutions and scanner types, and development of AI tools that can explain their predictions in terms radiologists and oncologists can act on.
The review concludes with a call to action for collaborative imaging databases dedicated to pancreatic cancer risk prediction — emphasizing that building this resource requires coordinated effort across medical centers and will take years, but the potential payoff of catching PDAC at earlier, curable stages justifies the investment.