Surgical removal of the tumor is the only potentially curative treatment for pancreatic ductal adenocarcinoma (PDAC). Yet even patients who undergo successful operations frequently see their cancer return within a year—a phenomenon called early recurrence. This happens because microscopic cancer cells, too small to see on scans, have already spread before surgery.
Early recurrence after surgery is a major driver of pancreatic cancer mortality. Patients who recur within 12 months have dramatically worse survival than those whose cancer stays controlled for longer. Knowing in advance who is likely to recur early would allow surgeons and oncologists to make better decisions about surgery timing, post-operative chemotherapy, and surveillance intensity.
This study set out to build an AI model that predicts early recurrence risk by combining two innovative data sources: radiomics (mathematical features extracted from CT scan images) from inside and around the tumor, and body composition measurements (quantifications of muscle and fat from the same CT scans).
The study enrolled 589 patients across four hospitals who underwent pancreatic cancer surgery between 2014 and 2023. Pre-operative CT scans were used to extract two types of information: radiomics features (mathematical texture and shape measurements from inside the tumor and the 5mm zone surrounding it) and body composition measurements (amounts of skeletal muscle, subcutaneous fat, and visceral fat).
Radiomic features capture information invisible to the human eye—such as tumor heterogeneity, internal texture patterns, and border irregularity—that may reflect underlying biology. Features from the surrounding area (peritumoral region) capture how the cancer interacts with neighboring normal tissue. Body composition is increasingly recognized as a prognostic factor, as low muscle mass (sarcopenia) predicts worse outcomes in many cancers.
Six machine learning algorithms were tested. The final model incorporated intratumoral radiomics, peritumoral radiomics, and body composition data together. SHAP (SHapley Additive exPlanations) analysis was used to visualize exactly which features drove predictions for individual patients, making the model transparent to clinicians.
The model that combined both intratumoral and peritumoral radiomics (the intra-peri-radiomics model) using a random forest algorithm achieved the best performance among pure imaging models: AUCs of 0.865, 0.849, and 0.839 in the training, internal validation, and external validation cohorts respectively.
Adding clinicopathological factors (information from the surgical pathology report, such as lymph node status and resection margin) to the radiomics model further improved performance, with AUCs of 0.936, 0.899, and 0.884 across the three cohorts. This combined model significantly outperformed models using radiomics or clinical data alone.
Notably, the peritumoral features (from the tissue surrounding the tumor) added substantial value beyond the intratumoral features alone. This suggests that the tissue around the tumor—which may reflect inflammatory infiltration, vascular involvement, or early invasion—contains independent prognostic information not captured by analyzing the tumor mass itself.
SHAP analysis identified specific features that most strongly drove early recurrence predictions. Peritumoral texture features—reflecting changes in the tissue immediately surrounding the tumor—were consistently among the top predictors, highlighting how cancer already begins reshaping its environment before spreading.
Body composition also emerged as a meaningful contributor, particularly measures of skeletal muscle quantity and quality. Patients with low muscle mass (sarcopenia) had higher predicted recurrence risk, consistent with research showing that sarcopenic patients have impaired immune function and worse treatment tolerance.
Importantly, the model was able to stratify patients within the same clinical stage into high and low recurrence risk groups—meaning it adds information beyond what pathology reports alone convey. Two patients with identical TNM staging might receive very different risk scores based on their CT imaging features.
Patients predicted to be at high recurrence risk by this model could be prioritized for more aggressive post-operative strategies: earlier initiation of chemotherapy, more frequent surveillance imaging (every 2-3 months rather than every 6 months), or enrollment in clinical trials of adjuvant therapies. Low-risk patients could be managed with standard protocols.
The model's use of pre-operative CT scans is a practical advantage—these scans are already obtained as standard of care before pancreatic cancer surgery. No additional procedures or tests are required. The computational analysis can be performed as an add-on to the existing radiology workflow.
Body composition analysis from routine CT scans also opens opportunities for intervention: patients identified as sarcopenic before surgery might benefit from prehabilitation programs (exercise and nutrition optimization before the operation) that could improve their functional reserve and reduce recurrence risk.
This study demonstrates that combining intratumoral radiomics, peritumoral radiomics, and body composition measurements into a machine learning framework provides significantly better prediction of early recurrence after pancreatic cancer surgery than any single data source alone.
The four-hospital validation design, encompassing 589 patients, provides stronger evidence of generalizability than most single-center studies in this field. The model maintained robust performance across geographically and demographically distinct patient cohorts.
Future work should focus on prospective clinical validation and, eventually, randomized trials testing whether risk-stratified post-operative management guided by this model improves survival. Integrating molecular biomarkers from surgical specimens with the imaging data could further refine risk predictions.