Intraductal papillary mucinous neoplasms (IPMNs) are fluid-filled cysts that form in the pancreas. They are increasingly found by accident during imaging done for other reasons, and while most are benign, some can progress to cancer over time.
Managing IPMN is a delicate balancing act: operating too soon exposes patients to the risks of major pancreatic surgery unnecessarily, but waiting too long risks missing a dangerous transition to cancer. Three major international guidelines exist to guide decisions, but they often disagree — and none are perfectly accurate.
Artificial intelligence tools that can analyze the imaging features of these cysts may be able to better predict which ones are truly dangerous, helping doctors make safer and more accurate decisions about surgery versus surveillance.
This scoping review systematically searched the medical literature up to December 2021 for studies on AI models that use CT or MRI imaging to predict whether an IPMN is likely to be malignant. Twelve studies met the eligibility criteria and were included.
The review categorized models into two types: radiomics-based models (which extract numerical features from imaging data) and deep learning models (which learn patterns directly from images). Both types were evaluated for performance and methodological quality.
Two independent reviewers assessed each study's quality using a validated scoring tool — the modified Radiomics Quality Score — to identify strengths and weaknesses in how the research was conducted.
Of the 12 included studies, 10 used radiomics approaches and 2 used deep learning. Most models performed well in their own datasets, achieving AUC values ranging from 0.76 to 0.98 — all above the threshold of 0.75 considered sufficient for clinical research.
One deep learning model based on endoscopic ultrasound achieved an AUC of 0.98, while an MRI-based deep learning model reached 0.78 — similar to the performance of current clinical guidelines. A combined CT and MRI radiomics model significantly outperformed the Fukuoka guidelines (AUC 0.94 vs 0.77).
However, methodological quality was generally poor. The median score on the quality assessment tool was just 11.5 out of a possible 36 for radiomics models. Key weaknesses included small sample sizes, lack of external validation, and failure to compare models head-to-head against clinical guidelines.
Despite impressive reported accuracy figures, none of the 12 models have been validated well enough to be used in routine clinical practice. Most were tested on only a few dozen to a few hundred patients at a single institution — far too few to ensure they work across the diversity of real-world patients.
A critical gap is that only 2 of the 12 models were directly compared to existing clinical guidelines in terms of diagnostic accuracy. Without this comparison, it is impossible to know whether the AI provides genuine added value over what clinicians already do.
The high variability in study design, imaging types, and target patient populations also makes it difficult to draw generalizable conclusions or to combine results across studies.
The review identifies clear priorities for the next generation of AI research on IPMN: larger multi-center datasets, rigorous external validation, and direct comparison with established clinical guidelines as the reference standard.
Standardizing how AI models are built and reported — using frameworks like the TRIPOD guidelines for prediction models — would help ensure future studies are comparable and reliable enough to support clinical use.
If properly developed and validated, AI tools could one day help doctors identify which patients with IPMN truly need surgery, sparing many from unnecessary operations while ensuring those at genuine risk are treated in time.