Pancreatic cysts are increasingly detected on imaging performed for unrelated reasons. The critical clinical decision is whether a cyst is malignant or has high malignant potential, requiring surgery, or is benign, permitting watchful waiting. Making this distinction before surgery is challenging because imaging alone often cannot characterize cyst contents accurately.
Endoscopic ultrasound (EUS)-guided fine needle aspiration (FNA) allows direct sampling of cyst fluid, which can be analyzed for tumor markers and cells. Standard markers include CEA (carcinoembryonic antigen), which is elevated in mucinous cysts, and CA19-9, CA125, and amylase. However, these markers individually have modest sensitivity and specificity, leaving many cases indeterminate.
This study investigated whether a deep learning model trained on a panel of cyst fluid markers combined with clinical data could outperform individual biomarkers and cytology in distinguishing malignant from benign pancreatic cystic lesions.
The study included 85 patients with pancreatic cystic lesions who underwent EUS-FNA: 23 malignant and 62 benign, classified based on pathological examination of surgical specimens or clinical follow-up. This malignant-to-benign ratio reflects the typical prevalence in clinical practice.
Input features for the model included the four cyst fluid markers (CEA, CA19-9, CA125, amylase) plus clinical variables such as patient age, sex, cyst size, and imaging characteristics. The model was implemented using TensorFlow, Google's open-source deep learning framework, with a neural network architecture suited to the tabular (non-image) data.
Performance was evaluated by area under the ROC curve (AUC), sensitivity, specificity, and accuracy. The AI model was directly compared to CEA alone and to cytology alone -- the two most commonly used diagnostic methods in clinical practice -- using the same patient cohort, allowing a fair head-to-head comparison.
Because the dataset was small, performance metrics were computed with bootstrapping to provide confidence intervals and assess the stability of the results. The small sample size is a recognized limitation but also reflects the practical challenge of assembling cohorts with complete surgical pathology confirmation for all cyst types.
The deep learning model achieved an AUC of 0.966, substantially better than CEA alone (AUC=0.719) and cytology alone (AUC=0.739). These differences represent a clinically meaningful improvement in discriminatory ability -- the AI model is far more likely to correctly order malignant and benign cases than either individual test.
For sensitivity -- the ability to correctly identify malignant cases -- the AI achieved 95.7%, compared to only 60.9% for CEA and 47.8% for cytology. Missing a malignancy (false negative) carries severe consequences for patients who proceed without surgery, making high sensitivity the most critical performance metric in this diagnostic context.
Overall accuracy of the AI model was 92.9%, compared to lower accuracy for both CEA and cytology. The model achieved this high accuracy while maintaining specificity sufficient to avoid unnecessary surgery in most benign cases, suggesting an improvement in the overall diagnostic decision rather than just a trade-off toward higher sensitivity at the cost of specificity.
CEA and cytology each provide partial and sometimes complementary information about cyst malignancy potential. CEA is elevated in mucinous cysts regardless of malignancy, while cytology detects malignant cells but is limited by the sparse and often nondiagnostic cell yields from FNA. Neither marker alone captures the full picture.
A deep learning model trained on multiple inputs simultaneously can learn complex, non-linear relationships between the markers and the outcome that no single threshold or linear combination could capture. For example, a cyst with borderline CEA but elevated CA19-9 and specific clinical features might have a very different malignancy probability than one with borderline CEA and normal other markers -- a nuance the AI can learn but a clinician applying a CEA cutoff cannot.
The substantially higher sensitivity compared to individual markers suggests the AI is rescuing cases that would be missed by applying any single marker in isolation. This is the core value of the multi-input approach: each additional informative variable increases the opportunity to correctly classify challenging borderline cases that are diagnostically ambiguous on a single test.
The study's most significant limitation is its small size: 85 patients, including only 23 malignant cases. With so few malignant examples, the neural network has limited training data for the outcome of primary interest, and performance estimates carry wide confidence intervals.
All patients were from a single institution, and the EUS-FNA technique, specimen handling, and biomarker assay methods were standardized locally. Performance may differ at other centers with different assay platforms or specimen processing protocols, as cyst fluid CEA and cytology results are known to vary considerably across laboratories.
The model's internal architecture and the specific decision boundaries it learned are difficult to interpret directly -- it functions as a black box that takes inputs and produces a probability score without explaining which features drove the prediction in any individual case. Explainability tools similar to those used in imaging studies could help build clinician trust if applied to this type of tabular data model.
This study demonstrates that deep learning applied to a panel of cyst fluid biomarkers and clinical data can dramatically improve the identification of malignant pancreatic cysts compared to the individual markers currently used in clinical practice. An AUC of 0.966 and sensitivity of 95.7% represent a meaningful advancement in diagnostic capability.
If validated in larger prospective cohorts, an AI-assisted cyst fluid analysis tool could be integrated directly into EUS-FNA reporting workflows, providing a malignancy probability score based on the biomarker panel results. This would help clinicians triage patients more accurately toward surgery or surveillance without requiring additional procedures.
The success of combining CEA, CA19-9, CA125, and amylase with clinical variables suggests that further improvements may be possible by adding additional markers -- such as molecular markers from cyst fluid DNA sequencing -- to the input panel. Multimarker AI approaches represent the future of pancreatic cyst risk assessment, moving beyond the current reliance on single markers with mediocre performance.