Prevalence and malignant potential of pancreatic cysts: Pancreatic cystic lesions are increasingly detected incidentally on cross-sectional imaging, found in up to 20-40% of patients undergoing abdominal MRI or CT for unrelated reasons. While many cysts are benign, specific subtypes - particularly Intraductal Papillary Mucinous Neoplasms (IPMNs) - carry significant risk of malignant transformation and require careful surveillance or surgical resection.
Diagnostic challenge of cyst classification: The four main pancreatic cyst types - IPMN, Mucinous Cystic Neoplasm (MCN), Serous Cystadenoma (SCA), and Solid Pseudopapillary Neoplasm (SPN) - have overlapping imaging features on CT and MRI. Distinguishing between them is critical because management differs dramatically: SCA is almost always benign and can be observed, while MCN and branch-duct IPMN require risk-stratified management and main-duct IPMN typically requires resection.
Need for automated analysis: Radiological interpretation of pancreatic cysts requires specialized expertise, and the growing incidental detection rate is creating a capacity challenge for radiologists and gastroenterologists. An automated system that can segment, characterize, and classify cysts from CT images could standardize and accelerate the triage process, ensuring appropriate follow-up for high-risk lesions while avoiding unnecessary interventions for clearly benign ones.
End-to-end pipeline design: The 3D Virtual Pancreatography system was developed at Stony Brook University as a complete end-to-end pipeline. Starting from a standard abdominal CT scan, the system performs: (1) automated pancreas segmentation to localize the organ, (2) lesion detection and segmentation within the pancreatic region, (3) feature extraction from the lesion geometry and imaging characteristics, and (4) classification into one of the four cyst subtypes. The output is presented as an interactive 3D visualization of the pancreas with lesion overlays.
Segmentation methodology: Pancreas and cyst segmentation were performed using convolutional neural networks trained on annotated CT datasets. The team used a multi-scale approach that captures both fine-grained lesion details and global pancreatic context, addressing the challenge that pancreatic cysts can range from a few millimeters to several centimeters and occur anywhere along the gland from head to tail.
Feature extraction and classification: After segmentation, radiomic features were extracted from each lesion, including morphological features (shape, size, wall thickness), texture features from the lesion interior, and relationship features describing the lesion's location relative to the pancreatic duct. A machine learning classifier was then trained on these features to distinguish among IPMN, MCN, SCA, and SPN subtypes using labeled training cases with pathologically confirmed diagnoses.
3D visualization component: The virtual pancreatography visualization renders the segmented pancreas and lesions in three dimensions, allowing clinicians to interactively rotate and examine the pancreas, view lesion size and location, and assess the relationship between cysts and the main pancreatic duct - a critical factor in IPMN classification and management decisions.
Pancreas segmentation accuracy: The automated pancreas segmentation component achieved strong performance on the test dataset, providing a reliable region of interest for downstream lesion detection. Accurate organ segmentation is a prerequisite for the entire pipeline; errors at this stage would propagate through lesion detection and classification.
Cyst detection and segmentation: The lesion segmentation module successfully detected and delineated cystic lesions, with performance varying based on lesion size and complexity. Larger, unilocular cysts with well-defined margins were segmented most accurately, while complex multilocular cysts with thin septations posed greater difficulty - a pattern consistent with the general challenges of segmenting complex anatomical structures.
Classification results by subtype: The classification module demonstrated the ability to distinguish among the four cyst subtypes with accuracy exceeding random chance. Classification performance was highest for SCA (which has characteristic microcystic morphology) and SPN (which tends to occur in younger women and has distinctive features). IPMN and MCN classification was more challenging due to overlapping imaging appearances, consistent with the known difficulty radiologists face in distinguishing these entities.
Comparison to radiologist interpretation: System outputs were compared against expert radiologist classifications to assess clinical validity. In cases where the AI classification agreed with the radiologist, the confidence in the correct diagnosis was reinforced. Discrepant cases were analyzed to identify systematic error patterns - for example, whether the AI tended to misclassify MCN as IPMN in certain imaging contexts.
Workflow integration potential: The 3D visualization component was evaluated by clinicians for its utility in understanding cyst anatomy before treatment planning. The interactive 3D rendering was found to be particularly helpful for communicating lesion location and duct involvement to referring clinicians and during multidisciplinary tumor board discussions, where comprehensive visual presentations support complex management decisions.
Potential for surveillance standardization: One of the key clinical benefits highlighted was the potential to standardize surveillance reporting. Currently, radiologist descriptions of cyst size, morphology, and concerning features vary in completeness and terminology. Automated extraction and structured reporting of cyst characteristics could improve consistency and enable better longitudinal tracking of lesion changes across imaging time points.
Addressing the cyst management crisis: The increasing incidental detection of pancreatic cysts is creating significant burden on healthcare systems, with millions of patients potentially requiring long-term surveillance. The 3D Virtual Pancreatography system addresses this by automating the most labor-intensive steps of cyst characterization, potentially allowing radiologists to focus their expertise on edge cases where human judgment is most needed.
Integration with clinical guidelines: The Fukuoka and American College of Gastroenterology guidelines provide risk stratification criteria for pancreatic cyst management. A future development direction is to automate the application of these guidelines based on AI-extracted features, generating automated management recommendations (observe, surveillance imaging, endoscopic ultrasound, or surgical referral) that clinicians could review and accept or override.
Limitations and future work: The study was limited by dataset size and the availability of pathologically confirmed diagnoses as ground truth - many cysts managed conservatively never receive histological confirmation. Future work should focus on larger multi-institutional datasets, incorporation of MRI (which provides better soft tissue contrast for cyst characterization than CT), and prospective validation in clinical pancreatic cyst surveillance programs.