Pancreatic cystic lesions (PCLs) - fluid-filled pockets in the pancreas - are being found more and more often as CT and MRI scanning has become routine. They affect a significant portion of the population, appearing in about 2.7% of CT scans and 24.8% of MRI scans. The key clinical question is: which ones are dangerous?
PCLs range from completely benign serous cysts to pre-malignant mucinous cysts like intraductal papillary mucinous neoplasms (IPMN) and mucinous cystic neoplasms (MCN), which can progress to full pancreatic ductal adenocarcinoma (PDAC). The problem is that standard CT imaging can only correctly classify a cyst as a specific type in 40-81% of cases, leaving major uncertainty about which patients need surgery and which need surveillance only.
Surgery for PCLs - a major pancreatic operation - carries 2-4% mortality even at high-volume centers, meaning unnecessary operations harm patients. This study aimed to build machine learning models using radiomic features from CT images to predict which PCLs have high malignant potential, avoiding both unnecessary surgeries and missed dangerous cysts. Uniquely, the study also analyzed actual cyst fluid to understand the biology behind the imaging patterns.
The study included 362 retrospective patients from two hospitals, divided into training (n=262), internal test (n=50), and external test (n=50) sets, plus a separate prospective cohort of 34 patients. All patients had undergone surgical resection of their PCLs, providing definitive pathological diagnosis as the ground truth. The training set had 39% malignant PCLs, the internal test set 40%, and the external test set 54%.
From each preoperative CT scan, 321 radiomic features were extracted covering seven categories: 3D shape features, first-order intensity statistics, gray-level co-occurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). Features were selected using LASSO regression, retaining 19 final radiomic features. These were combined with clinical variables (CEA, CA199, lesion size, wall characteristics, solid component) to build the final models.
Eleven different machine learning classifiers were trained and compared: adaptive boosting (AB), bagging (BAG), gradient boosting (GB), Gaussian naive Bayes (GNB), k-nearest neighbors (KNN), linear discriminant analysis (LDA), logistic regression (LR), neural network (NN), quadratic discriminant analysis (QDA), random forest (RF), and support vector machine (SVM). Additionally, cyst fluid from the prospective cohort was analyzed by proteomics and lipidomics to investigate the biological basis of the radiomic scores.
The combined clinical-radiomic models dramatically outperformed radiomic-only models. All 11 clinical-radiomic classifiers achieved AUC greater than 0.93 across the training, internal test, and external test sets. The random forest model achieved the highest overall performance with AUC 1.00 on training, 0.95 on internal test, and 0.96 on external test. Even in the prospective cohort, all 11 models achieved AUC ranging from 0.92 to 0.96, confirming robust real-world performance.
The adaptive boosting (AB) clinical-radiomic model showed a particularly strong improvement over the radiomic-only version: AUC improved from 0.88 to 0.96 on the internal test set and from 0.83 to 0.96 on the external test set. Across all models, the solid component (the presence of solid tissue within the cyst) and the radiomic score were consistently the most significant predictors in multivariate analysis.
The proteomic analysis of cyst fluid from 25 prospective patients identified 347 differentially expressed proteins between high-risk and low-risk radiomic score groups. The top upregulated proteins included MIF (immune checkpoint protein) and KRT19 (epithelial marker associated with cancer). Downregulated proteins were largely digestive enzymes like CTRB2 and CELA3A. Pathway analysis showed upregulated proteins enriched in Th17 cell differentiation and mucin biosynthesis, while downregulated proteins were linked to pancreatic secretion loss - a hallmark of malignant transformation.
The lipidomic analysis of cyst fluid detected 468 lipid molecules across 28 lipid classes and identified 33 significantly different lipid molecules between high and low radiomic score groups. The most striking finding was that ceramide (Cer) was the predominant differentially expressed lipid class. Ceramide is a bioactive sphingolipid that plays critical roles in cell death pathways and inflammation.
The top 5 most important lipid molecules by variable importance in projection (VIP) scores included two ceramide species upregulated in high-risk cysts, alongside phosphatidylethanolamine and two phosphatidic acid species that were downregulated. This suggests that elevated ceramide signaling in high-malignancy-risk PCLs may reflect ongoing apoptosis resistance or inflammatory remodeling within the cyst wall - providing a molecular explanation for why these cysts look different on CT imaging.
Together, the proteomic and lipidomic data provide the first biological validation of what CT-based radiomic scores are actually measuring in PCL malignancy risk. This connection between imaging features and molecular biology makes the clinical-radiomic models more scientifically credible and could guide development of minimally invasive cyst fluid tests as alternatives or complements to imaging alone.
The clinical-radiomic models developed in this study are visualized as nomograms - scoring tools that clinicians can use with routine preoperative CT data to predict malignancy risk for any given PCL. Because the inputs are all standard clinical and imaging variables (size, wall features, CEA, CA199, and radiomic score), no additional tests or procedures are required beyond what is already collected in standard care.
The prospective validation confirms that the models maintain their accuracy when applied to new patients sequentially, not just retrospectively. This is an important milestone because many radiomics studies show good performance in retrospective data but fail prospectively due to changes in patient characteristics or imaging protocols over time. The consistent AUC of 0.92-0.96 in the prospective cohort demonstrates real-world robustness.
The multi-omics underpinning of the models - connecting CT radiomic features to proteomic and lipidomic changes in cyst fluid - opens a path toward combined imaging and biofluid testing for PCL management. Future clinical trials could evaluate whether using these models to select patients for surgery (versus surveillance) reduces unnecessary operations while maintaining cancer detection rates, ultimately improving outcomes for the growing number of patients with incidentally discovered pancreatic cysts.