Pancreatic cystic lesions (PCLs) are fluid-filled sacs within the pancreas. They are found in approximately 15-20% of people who undergo abdominal imaging for unrelated reasons, and their incidence increases with age. While most PCLs are entirely benign, a subset can develop into pancreatic cancer if left untreated.
The challenge is that PCLs look deceptively similar on imaging, yet they have vastly different malignant potentials. Aggressive management of all cysts would subject many people to unnecessary surgery; overly conservative management would miss cancers at a preventable stage. The ideal strategy is to accurately identify which cysts need intervention and which can be safely observed.
Pancreatic cystic lesions are therefore a critical target for early cancer prevention. If high-risk cysts can be reliably identified and removed before they become invasive cancer, pancreatic cancer deaths can potentially be prevented entirely in those patients - a rare opportunity for true primary prevention of this deadly disease.
Serous Cystadenomas (SCN) are typically composed of clusters of small cysts (microcystic) and are almost always benign. They rarely require surgical removal and can usually be safely monitored. On imaging, they often show a characteristic honeycomb or cluster-of-grapes appearance, sometimes with a central scar.
Mucinous Cystic Neoplasms (MCN) occur almost exclusively in middle-aged women, typically in the body or tail of the pancreas. They produce mucin and have malignant potential, particularly when they contain solid components, calcifications, or grow large. Surgical resection is generally recommended for MCN in suitable surgical candidates.
Intraductal Papillary Mucinous Neoplasms (IPMN) arise within the pancreatic duct system and are subdivided by which duct is involved - main duct, branch duct, or mixed-type. Main duct IPMNs carry the highest cancer risk (up to 70%). Solid Pseudopapillary Epithelial Neoplasms (SPEN) are rare, low-grade malignant tumors most common in young women. Understanding which type a cyst is drives management decisions entirely.
The most widely used international guidelines for managing IPMNs are the Fukuoka Criteria (from the International Association of Pancreatology). These criteria classify patients into low-risk (observation), intermediate-risk (further investigation), and high-risk (surgery recommended) groups based on imaging features such as cyst size, duct dilation, presence of solid components, and mural nodules.
While the Fukuoka Criteria provide a useful framework, their accuracy is imperfect. Studies have shown sensitivity of approximately 60-80% and specificity of 50-75% for predicting malignancy, meaning a substantial proportion of patients are still misclassified. Some patients receive unnecessary surgery, while others with high-grade lesions are incorrectly placed in observation.
The limitations of existing guidelines create a compelling rationale for radiomics and machine learning approaches. Where human guidelines based on a handful of imaging features fall short, machine learning can analyze hundreds of quantitative image features and potentially uncover subtle patterns that more accurately distinguish benign from malignant cysts.
Radiomics is the process of converting medical images into high-dimensional quantitative data through the extraction of mathematical features. For pancreatic cysts, this typically involves first segmenting the cyst boundary from CT or MRI images, then calculating features describing the cyst's intensity distribution, shape, and internal texture patterns.
Texture features are particularly important for distinguishing cyst types. Methods like the Gray Level Co-occurrence Matrix (GLCM) quantify how pixel intensities relate to their neighbors, capturing information about roughness, homogeneity, and pattern regularity inside the cyst wall and fluid. These microscopic texture patterns, invisible to the human eye in routine interpretation, may encode information about the biological nature of the cyst lining.
After feature extraction, machine learning classifiers are trained to distinguish different cyst types or to predict malignancy risk. The resulting radiomic signatures can potentially outperform visual assessment because they are quantitative, objective, and capture far more information per image than any human observer can process.
The review summarizes several published radiomics studies specifically targeting IPMN risk stratification - distinguishing low-grade dysplasia (benign) from high-grade dysplasia or invasive cancer. Most studies used CT or MRI features combined with machine learning classifiers including support vector machines, random forests, and logistic regression.
Across these studies, radiomic models generally achieved AUC values between 0.75 and 0.90 for predicting malignant or high-grade IPMNs, with several studies demonstrating that radiomic models outperformed established clinical guidelines when evaluated on the same patient cohort.
Certain radiomic features appeared consistently predictive across multiple studies, particularly those describing internal heterogeneity and texture irregularity - consistent with the biological concept that more irregular, heterogeneous cysts are more likely to harbor dysplastic changes. However, the review notes that cross-study comparison is complicated by differences in imaging protocols, segmentation approaches, and feature extraction software.
The review identifies several major challenges that must be overcome before radiomics can be routinely used to stratify pancreatic cysts in clinical practice. The first is reproducibility: radiomic features can vary substantially depending on image acquisition parameters, scanner type, CT slice thickness, and how the cyst boundary is drawn. Small changes in segmentation can produce meaningfully different feature values.
A second challenge is small sample sizes. Pancreatic cysts with confirmed histopathological diagnosis are relatively rare in any single institution, meaning most published studies involve only dozens to a few hundred cases. Machine learning models trained on such small datasets are at high risk of overfitting - performing well on training data but failing when applied to new patients from different institutions.
The review stresses that external validation - testing a model on a completely independent patient cohort, ideally from a different institution - is absolutely necessary before claiming clinical utility. Many published radiomic studies for PCLs have only been internally validated, making it impossible to know whether their performance will generalize to real-world clinical populations.
Radiomics works best not in isolation but as one component of a multi-parametric risk model. Clinical variables such as patient age, symptoms (jaundice, pancreatitis), family history of pancreatic cancer, and laboratory values (CA19-9, CEA in cyst fluid) provide complementary information that imaging alone cannot capture.
Cyst fluid analysis is also highly informative. Endoscopic ultrasound-guided sampling can provide cyst fluid for measurement of carcinoembryonic antigen (CEA) - elevated CEA in cyst fluid is associated with mucinous cysts - as well as next-generation sequencing of cyst fluid DNA for KRAS and GNAS mutations, which are strongly associated with mucinous and IPMN histology respectively.
Integrating radiomic imaging features with these clinical and molecular data sources using machine learning creates a richer, more informative model than any single data type alone. Several studies reviewed here demonstrated significant performance improvements when radiomics was combined with cyst fluid molecular markers, suggesting a synergistic relationship between imaging and biological characterization.
The review concludes optimistically about the potential of AI-based radiomics to transform pancreatic cyst management, while acknowledging the significant work still required. The theoretical advantage is compelling: if machine learning can reliably distinguish benign from premalignant cysts, unnecessary surgeries can be avoided while ensuring high-risk cysts are removed before they become invasive cancer.
Key priorities for the field include standardizing imaging protocols across institutions, developing harmonized approaches to cyst segmentation, and building large multi-center datasets that provide sufficient statistical power to train generalizable models. International collaboration through groups like the International Study Group on Pancreatic Surgery (ISGPS) will be important in this effort.
The authors envision a future clinical workflow in which every pancreatic cyst discovered incidentally on CT or MRI is automatically analyzed by an AI system that provides a risk score and management recommendation. This system would incorporate imaging radiomics, cyst fluid analysis, and clinical variables into a single decision-support tool, helping clinicians make more consistent, evidence-based management decisions.