Peritoneal metastases occur when pancreatic cancer seeds spread to the lining of the abdominal cavity. These are often invisible on standard CT scans—called occult peritoneal metastases (OPM)—because the deposits are too small to see. When surgeons open the abdomen expecting to remove a tumor, they sometimes discover the cancer has already spread in ways that preclude surgery.
Identifying OPM before surgery is critical because it could spare patients the risks and recovery of unnecessary operations. Current CT imaging misses OPM in a significant proportion of patients. A more sensitive detection tool could help route patients directly to chemotherapy or targeted therapy rather than futile surgery.
This retrospective study included 302 PDAC patients from two medical centers, split into training (167 patients), internal test (72 patients), and external test (63 patients) cohorts. Researchers extracted two types of features from CT images: traditional handcrafted radiomics (HCR) features based on pre-defined mathematical measurements, and deep learning radiomics (DLR) features learned automatically by a neural network.
Features were extracted from both the tumor and the peritoneum (the abdominal lining). Feature selection used mutual information and LASSO algorithms to identify the most informative measurements. A combined model integrating clinical and radiological characteristics—including CA 19-9 levels and CT-based T and N staging—with both HCR and DLR features was built using logistic regression and validated across all three cohorts.
The combined model achieved AUC values of 0.853 in training, 0.845 in internal testing, and 0.852 in external testing. This consistency across all three cohorts—including an entirely separate medical center—confirms the model generalizes beyond the data it was trained on. The high and stable AUCs indicate reliable discrimination between patients with and without occult peritoneal metastases.
Critically, the combined model significantly outperformed the clinical-radiological model alone. In the training cohort, the AUC jumped from 0.612 with clinical features only to 0.853 with the combined model. In the total test cohort, the improvement was from 0.638 to 0.842. This demonstrates that AI-extracted image features add substantial value beyond what radiologists and clinical data already capture.
Traditional handcrafted radiomics extract pre-defined features from images—measurements of texture, shape, and intensity that researchers specify in advance. Deep learning radiomics, by contrast, learns to identify patterns directly from pixel data without human-defined rules. The combination captures both the structured measurements clinicians value and the subtle patterns neural networks can detect.
In this study, feature selection retained 9 HCR tumor features, 14 DLR tumor features, and 3 HCR peritoneal features after filtering thousands of candidates. The fact that peritoneal features—extracted from the abdominal lining itself rather than just the tumor—contributed to the model underscores the importance of looking beyond the primary tumor mass when assessing metastatic risk.
The most direct clinical benefit of this model is reducing futile laparotomies—surgeries that open the abdomen only to find inoperable disease. Such operations cause significant morbidity and delay the start of systemic treatment. Identifying OPM-positive patients before surgery would allow them to start chemotherapy sooner.
Decision curve analysis showed the combined model had greater clinical benefit across a wide range of risk thresholds compared to clinical features alone. This type of analysis demonstrates that the model is not just statistically significant but clinically useful—it would improve decision-making for real patients across different risk tolerance scenarios.
This study demonstrates that combining deep learning radiomics with clinical and traditional imaging features can substantially improve pre-operative detection of occult peritoneal metastases in PDAC. The model's excellent performance across internal and external validation cohorts suggests it is robust and generalizable.
Integrating this type of AI model into pre-operative CT review workflows could change treatment pathways for a meaningful proportion of pancreatic cancer patients, directing them away from futile surgery and toward systemic therapies sooner. Prospective validation studies are the necessary next step before clinical adoption.