Pancreatic cancer is the 12th most common cancer worldwide, but one of the deadliest, with an overall survival rate of only about 5%. Most patients are diagnosed at an advanced stage when treatment options are limited and curative surgery is no longer possible.
Early diagnosis is the single most important factor in improving survival, yet current screening methods — including ultrasound, blood markers, and CT scans — miss early-stage tumors too often. Identifying high-risk individuals before symptoms appear is a major challenge that AI may help address.
This scoping review mapped what AI models currently exist for predicting and diagnosing pancreatic cancer, covering 30 published studies from 2015 to 2022 across 14 countries.
The authors followed PRISMA-ScR guidelines and searched five major databases — PubMed, ScienceDirect, Google Scholar, BioRxiv, and MedRxiv — identifying 18,285 articles initially. After removing duplicates and irrelevant papers, 30 studies were included for analysis.
Two independent reviewers extracted data from each study, achieving 89% agreement. They recorded the type of AI used, the data inputs, the number of patients, and the performance outcomes reported.
The included studies covered a wide range of settings: 27% from the United States, 17% from China, and 13% from India, with participants averaging around 56 years of age.
Of the 30 studies reviewed, 37% used deep learning alone, 27% used traditional machine learning, and 23% combined both. The most common algorithm was the convolutional neural network (CNN), used in 60% of studies — mostly for analyzing medical images.
Radiology images were the most common data type (47% of studies), followed by clinical records (40%) and laboratory test results (30%). Dataset sizes varied enormously, from as few as 15 patients to over 13 million data points.
Most studies trained models on closed, non-public datasets, which limits how broadly findings can be applied. Only 13% of studies used publicly available open datasets.
AI models generally showed strong performance in distinguishing pancreatic cancer from healthy tissue when trained on imaging data. Convolutional neural networks in particular achieved high accuracy on internal test datasets.
Performance varied widely depending on dataset size and quality, the specific AI method, and whether models were validated on new patient cohorts. Many studies reported accuracy above 85%, though very few had been tested in real clinical settings.
The review found that combining multiple data types — imaging plus clinical data or lab results — tended to improve diagnostic performance over using any single data source alone.
AI has clear potential to improve early detection of pancreatic cancer, but most current models remain at the research stage. Very few have been externally validated or tested in prospective clinical trials with real patients.
Key gaps include the lack of large diverse public datasets, the dominance of single-institution studies, and the absence of standardized ways to compare model performance across studies.
The review calls for larger multi-center studies, better integration of AI into clinical workflows, and prospective validation trials before AI tools can be used to screen patients at scale.