Pancreatic cancer remains one of the most lethal malignancies, with most patients diagnosed only after the disease has spread. Traditional diagnostic tools like CT, MRI, and endoscopic ultrasound (EUS) are effective but invasive, expensive, or require specialist facilities.
Precursor lesions — including intraductal papillary mucinous neoplasms (IPMNs) and pancreatic cysts — can be identified before cancer develops, but distinguishing dangerous from harmless lesions is difficult even for expert radiologists.
This review focused on AI methods that could detect early pancreatic lesions and predict cancer noninvasively, covering 44 studies published over the past decade.
The authors searched PubMed using terms including 'artificial intelligence', 'deep learning', 'radiomics', and 'pancreatic cancer', narrowing from 97 candidate articles to 44 after removing duplicates and off-topic papers.
The 44 studies were divided into two groups: those evaluating AI diagnostic methods for pancreatic cancer broadly (24 studies), and those focusing on AI for early lesion detection and cancer prediction (10 studies).
Studies covered input data including CT images, MRI scans, EUS findings, radiomics features, and epidemiological variables.
CT-based AI models have shown the ability to flag suspicious features months before a clinical diagnosis, including irregular tumor margins, duct dilation, and vascular invasion. However, sensitivity for very early lesions remains low.
EUS achieves 75-97% detection rates for pancreatic cancer and is particularly effective at identifying early lesions and cystic tumors. AI tools trained on EUS images have improved detection of mural nodules indicating malignancy in pancreatic cysts.
Radiomics — extracting quantitative features from medical images — combined with machine learning has shown promise in distinguishing malignant from benign pancreatic lesions without requiring biopsy.
AI has demonstrated utility in helping clinicians decide which pancreatic cysts need surgery and which can be safely monitored — a major clinical challenge because cyst management guidelines vary and are imprecise.
Machine learning models combining imaging findings with patient history and biomarkers such as CA19-9 have outperformed individual criteria in predicting which patients will develop invasive cancer.
The review notes that AI's greatest near-term clinical value may be as a second reader or decision-support tool that flags high-risk findings for specialist review, rather than replacing human judgment entirely.
Despite encouraging results, no AI system for pancreatic cancer detection has yet been prospectively validated in a large, multi-center clinical trial. Most evidence comes from retrospective studies at single institutions.
Standardization of imaging protocols, data sharing, and agreed performance benchmarks are needed so that AI models developed at one hospital can be applied reliably at others.
The authors conclude that AI holds genuine promise for transforming noninvasive detection, but a coordinated international research effort is needed to move from proof-of-concept to clinical reality.