Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers, with surgery remaining the only potentially curative treatment. However, cancer recurs in up to 80% of patients within 2 years after surgery, which severely limits long-term survival.
A critical contributor to early recurrence is extrapancreatic perineural invasion (EPNI), a process where tumor cells invade the nerve fibers that extend beyond the pancreas. EPNI is found in 52 to 76% of surgical specimens and is closely linked to poor prognosis.
The peripancreatic region is anatomically complex, with major blood vessels like the celiac artery, superior mesenteric artery, and portal vein running alongside a dense network of nerve fibers. This makes accurate pre-surgical assessment of EPNI especially challenging.
Traditional CT imaging methods for evaluating EPNI rely on subjective radiologist interpretation of microstructural features, resulting in moderate specificity of only 64.7%. This limitation drives the need for more objective and automated approaches.
Researchers developed a fully automated artificial intelligence (AI) model based on CT imaging to detect EPNI before surgery, enabling more precise treatment planning for patients with PDAC.
The key innovation was identifying the perivascular space as a critical anatomical landmark. Since pancreatic nerves run alongside blood vessels, the space immediately surrounding vessels serves as a proxy for where nerve invasion is likely to occur.
The AI model uses deep learning segmentation via the nnUNet framework to automatically outline the tumor and surrounding blood vessels in CT scans. Vessel boundaries are then mathematically expanded by 1 pixel to simulate the perivascular nerve plexus region.
EPNI is classified as positive when the tumor overlaps with this expanded perivascular zone, and negative when no overlap exists. This transforms a complex pathological concept into a precise, computable geometric measurement.
The study enrolled 1,065 patients with PDAC from two Shanghai hospitals between June 2014 and April 2023. Patients were divided into training (n=497), internal validation (n=212), internal test for neoadjuvant therapy (n=180), and external test (n=176) sets.
All CT images were analyzed using the nnUNet framework, which automatically selects training parameters and supports both 2D and 3D segmentation tasks. The model was trained using cross-entropy and Dice loss functions over 1,000 epochs with five-fold cross-validation.
Image preprocessing included normalization using global foreground mean and standard deviation, and voxel-spacing adjustment via third-order spline interpolation. Data augmentation with rotation, flipping, and other transformations improved model robustness.
Six models were built using vessel boundary expansions of 0 to 5 pixels. The 1-pixel expansion model achieved the highest AUC in the validation set and was selected as the final diagnostic model, capturing the nerve plexus space while avoiding false positives from larger expansions.
The 1-pixel expansion AI model achieved strong diagnostic performance across all four patient groups. Area under the curve (AUC) values were 0.87 in training, 0.88 in validation, 0.82 in internal test, and 0.83 in external test sets.
The model demonstrated high sensitivity of 97.3% in the training set, 98.1% in the validation set, and 96.4% in the internal test set, meaning it correctly identified nearly all EPNI-positive cases. Specificity ranged from 68 to 78% across sets.
Overall accuracy exceeded 85% in all test sets, with positive predictive values above 82%, confirming reliable performance across different hospitals and patient populations. The model code and inference examples are publicly available on GitHub.
These results exceeded prior conventional imaging methods that relied on subjective radiologist evaluation, establishing the AI approach as a more objective and reproducible tool for EPNI assessment.
Patients identified as EPNI-positive by the AI model had significantly worse overall survival than EPNI-negative patients. In the training set, median survival was 52.9 months for EPNI-negative versus 29.5 months for EPNI-positive patients.
Multivariate Cox regression confirmed that AI-predicted EPNI was an independent preoperative predictor of worse overall survival, with a hazard ratio of 1.51 (95% CI: 1.05 to 2.15; P=0.02), even after adjusting for clinical and pathological factors.
Subgroup analyses showed that EPNI-positive patients had poorer survival regardless of tumor size (T1-T2), lymph node involvement (N0, N1-2), lymphovascular space invasion (LVSI) status, differentiation grade, or clinical stage, demonstrating broad prognostic applicability.
Notably, even in early-stage disease (stage IA and IB), AI-identified EPNI-positive patients had significantly poorer survival than EPNI-negative patients (P=0.004), highlighting the value of AI detection beyond what standard staging captures.
Neoadjuvant therapy (NAT), which involves chemotherapy or radiation given before surgery, is increasingly used for borderline resectable PDAC. However, assessing its effectiveness using standard criteria like RECIST 1.1 or tumor regression grade (TRG) alone is often insufficient.
In the 180 patients who underwent NAT, RECIST 1.1 classification and TRG response alone did not significantly differ in overall survival (P=0.239 and P=0.291, respectively), indicating these conventional measures miss important biological information.
When the AI-predicted EPNI was combined with TRG, the survival difference became significant (P=0.026). Similarly, combining AI-EPNI with RECIST 1.1 improved prognostic prediction (P=0.016), demonstrating synergistic value of the AI model.
These findings suggest that integrating AI-detected EPNI status into standard response evaluation criteria can more precisely identify which NAT patients are truly benefiting from treatment, enabling better personalized therapy decisions.
This study establishes a fully automated, CT-based AI model that reliably detects extrapancreatic perineural invasion before surgery, providing an objective and reproducible tool that outperforms traditional imaging approaches.
The model's ability to identify EPNI as an independent prognostic factor adds meaningful information beyond TNM staging, helping clinicians better stratify patients and identify those who may require more aggressive or targeted treatment strategies.
By enabling non-invasive, preoperative EPNI detection, the AI model supports precision medicine by identifying high-risk patients who may benefit from extended resection or intensified neoadjuvant regimens before surgery.
Future work could integrate this model into broader clinical decision-support systems, potentially combining EPNI status with other imaging biomarkers to further refine pre-surgical planning and improve long-term outcomes for patients with PDAC.