Pancreaticoduodenectomy, also known as the Whipple procedure, is the primary surgical treatment for pancreatic head tumors and periampullary cancers. While it offers the best chance of cure for resectable disease, it carries substantial risk of serious postoperative complications.
The most feared complication is clinically relevant postoperative pancreatic fistula (CR-POPF), which occurs when the surgical connection between the pancreatic remnant and the digestive tract leaks pancreatic fluid. This can cause severe infection, bleeding, and prolonged hospital stays, and it significantly increases mortality risk.
Accurately predicting which patients are at high risk for CR-POPF before surgery would allow surgeons to modify their operative strategy, choose the most appropriate drainage approach, or implement enhanced postoperative monitoring. However, current risk prediction tools have important limitations in discriminating intermediate-risk patients.
The standard clinical tool, the Fistula Risk Score (FRS), combines four intraoperative variables including pancreatic duct diameter and gland texture. While useful, FRS requires information only available during surgery and performs poorly for patients categorized as intermediate risk, a large and clinically problematic group.
This study developed a Deep Learning Score (DLS) using preoperative CT imaging data from 513 patients across four institutions. Preoperative CT was chosen because it is routinely obtained before pancreatic surgery and is available before any intraoperative assessment, making it ideal for truly prospective risk prediction.
A convolutional neural network (CNN) was trained to extract imaging features from CT slices through the pancreatic body and tail region. The model learned to identify subtle texture, density, and morphological patterns associated with soft, fatty pancreatic tissue, which is the key anatomical predictor of fistula risk.
The multi-institutional dataset spanning four centers was critical for ensuring the model would generalize across different CT scanner types, imaging protocols, and patient populations. Training on data from a single institution would risk overfitting to institution-specific imaging artifacts.
The DLS was then compared head-to-head against the established Fistula Risk Score using matched patient cohorts. Performance was assessed using area under the receiver operating characteristic curve (AUC) in both training and independent test cohorts, along with stratified analysis by FRS risk category.
In the training cohort, the DLS achieved an AUC of 0.85 compared to 0.78 for the FRS, demonstrating meaningfully superior discriminative ability even among the data used for model development. This gap widened in the independent test cohort, where DLS reached AUC 0.89 versus 0.73 for FRS.
The advantage of DLS was most pronounced among patients classified as intermediate risk by FRS. In this challenging subgroup, DLS correctly classified 92.1% of test cases compared to only 65.8% by FRS, a dramatic improvement with direct clinical implications for the most difficult risk category.
Analysis of the imaging features most influential in DLS predictions revealed correlations with pancreatic fibrosis, main pancreatic duct (MPD) diameter, and remnant pancreatic volume. These features correspond to the known pathophysiology of fistula formation, supporting the biological validity of the learned model.
Grad-CAM (Gradient-weighted Class Activation Mapping) visualizations were generated to show which regions of the CT scan most influenced model predictions. These heatmaps confirmed that the model focused on anatomically appropriate pancreatic tissue regions rather than irrelevant background structures.
Gradient-weighted Class Activation Mapping (Grad-CAM) is an interpretability technique that highlights which regions of an input image most strongly influenced a neural network's prediction. For medical imaging models, this type of visualization is essential for building clinician trust and verifying that the model is using clinically meaningful features.
In this study, Grad-CAM maps showed activation concentrated over the pancreatic parenchyma, particularly in regions corresponding to the soft, low-density fatty tissue characteristic of high fistula risk. This alignment with established pathophysiology is reassuring and supports the model's validity.
The ability to visualize model attention also enables prospective quality assurance: in clinical deployment, an operator could check that Grad-CAM activation maps are focusing on the correct anatomical regions before trusting a given prediction, providing a mechanism to flag potentially unreliable model outputs.
This level of interpretability distinguishes the DLS approach from purely black-box neural networks and makes it more suitable for clinical adoption in a high-stakes surgical planning context where unexplained predictions would be difficult to act upon.
A key advantage of DLS over FRS is that it relies entirely on preoperative CT data, making it available before the patient enters the operating room. This creates the opportunity to modify surgical strategy, such as choosing between different types of pancreatic anastomosis, before the incision is made.
Surgeons managing patients identified as high risk by DLS might elect to place additional drainage catheters, use fibrin glue sealants, or consider alternative reconstruction techniques specifically designed to reduce fistula risk. These decisions are more impactful when made proactively rather than reactively.
For the intermediate-risk FRS category, where the traditional score provides insufficient guidance, DLS offers a new level of discrimination that could allow clinicians to separate genuinely high-risk from genuinely low-risk patients within this ambiguous group.
Multi-institutional validation is a critical strength of this study because it demonstrates the tool can function reliably across different hospitals with varying patient populations, scanner hardware, and imaging protocols, a prerequisite for real-world deployment.
This study demonstrates that a deep learning model trained on preoperative CT imaging can predict CR-POPF after pancreaticoduodenectomy with greater accuracy than the current standard clinical scoring system, particularly for the difficult intermediate-risk group.
The multi-institutional validation across four centers provides strong evidence that DLS generalizes beyond the training environment and is a practical candidate for broader clinical testing. The model's interpretability via Grad-CAM further supports its suitability for clinical use.
Future work should include prospective clinical trials that test whether using DLS to guide preoperative decision-making actually reduces fistula rates or improves outcomes, rather than just improving risk classification accuracy.
More broadly, this study is an example of how AI applied to routine preoperative imaging can extract clinically valuable information that exceeds what trained human observers can extract through conventional reading, opening new possibilities for surgical risk stratification across many procedures.