After surgery for pancreatic ductal adenocarcinoma (PDAC), patients face widely varying outcomes — some survive for years while others experience rapid recurrence. Doctors currently rely on clinical staging, tumor size, and lymph node status to estimate prognosis, but these factors alone are often insufficient to reliably predict individual patient outcomes.
Digital pathology, which involves scanning tumor tissue slides at high resolution, produces images that contain vastly more information than the human eye can process. Deep learning models can analyze these whole-slide images to extract microscopic features that predict survival, potentially capturing biological information that standard pathological grading misses.
The study included 142 PDAC patients who underwent surgery, with 114 used for training and 28 for testing. Three pre-trained deep learning models — MobileNet, ResNet18, and DenseNet121 — were used to extract features from whole-slide tumor images. The DenseNet121 model performed best and was used to build the pathological risk model.
For the clinical risk model, researchers used LASSO regression and Cox survival analysis to identify independent predictors from blood test results and clinical-pathological data including tumor stage, lymph node involvement, and cancer biomarkers CA19-9 and CA242. The two models were then combined into a single unified model.
The pathological risk model alone achieved a C-index (a measure of prediction accuracy) of 0.82 in training and 0.73 in testing. The clinical risk model achieved 0.76 in training and 0.75 in testing. When combined into the unified model, performance improved to 0.86 in training and 0.77 in testing — demonstrating clear additive value.
Kaplan-Meier survival curves showed that patients classified as high-risk by the combined model had significantly shorter survival than low-risk patients, in both the training and testing cohorts. The model also showed good calibration at 1-, 3-, and 5-year survival predictions, meaning its predicted probabilities closely matched actual observed survival rates.
The clinical model identified several independent predictors of worse outcomes, including advanced tumor stage (T3), lymph node involvement (N1 stage), and elevated CA242 levels. Elevated platelet distribution width was also associated with worse prognosis — a finding consistent with its role in systemic inflammation and immune response.
The study confirmed that CA242, a less commonly used tumor marker, had strong prognostic value in this PDAC cohort. Patients with high preoperative CA242 had significantly worse survival, suggesting that this marker deserves more routine clinical measurement alongside the more commonly used CA19-9.
After pancreatic cancer surgery, decisions about whether to offer adjuvant chemotherapy, radiation, or intensive follow-up surveillance are partly guided by prognosis. A more accurate survival prediction model could help oncologists have more informed conversations with patients about their individual risk, and could help identify those who most need intensive postoperative treatment.
The model could eventually be used as a decision-support tool: high-risk patients identified by the combined model might be prioritized for clinical trials of novel adjuvant therapies, while lower-risk patients might be managed with standard protocols. The ability to stratify patients at the time of surgery, before any recurrence is visible, is clinically meaningful.
This study shows that deep learning analysis of routine pathology slides, when combined with standard clinical data, produces meaningfully better survival predictions than clinical data alone. The approach does not require additional tests or procedures beyond what is already part of standard post-surgical pathology workup.
Limitations include the relatively small cohort size and the retrospective design. Larger prospective studies are needed to validate these findings before clinical adoption. Nevertheless, the results provide a compelling foundation for integrating AI-driven pathology analysis into routine PDAC prognostication.