Scalable knowledge distillation for pancreatic cancer diagnosis: Real-time ResNet-based stage stratification with global deployment feasibility.

Abstract

e16412 Background: Pancreatic cancer remains one of the deadliest cancers, with a 5-year survival below 12%. Early-stage disease is often misdiagnosed as benign lesions or chronic pancreatitis, delaying treatment. Deep learning models show high diagnostic accuracy on cross-sectional imaging, but infrastructure and computational demands limit clinical use. Knowledge distillation can maintain performance while enabling real-time deployment. We evaluated whether a distilled ResNet18 could retain ResNet152-level accuracy for stage stratification and differential diagnosis while supporting global scalability. Methods: We analyzed 2,840 contrast-enhanced CT scans from multiple countries: pancreatic ductal adenocarcinoma (n = 1,400; 700 resectable/borderline-resectable, 700 locally advanced/metastatic), chronic pancreatitis (n = 900), and benign lesions (n = 540). Ground truth was established via multidisciplinary consensus using histopathology, surgical findings, and longitudinal follow-up. A ResNet152 mentor network (60.2M parameters; 28.6 GFLOPs) served as reference. A ResNet18 mentee (11.7M parameters; 1.8 GFLOPs) was trained via temperature-scaled knowledge distillation with regularized cross-entropy loss. Accuracy, sensitivity, specificity, F1-score, and AUROC were measured. Radiologists and oncologists across six continents evaluated clinical utility and deployment feasibility. Results: The distilled ResNet18 achieved 96.6% accuracy for stage stratification (sensitivity 97.4%, specificity 95.8%, AUROC 0.989). Differential diagnosis accuracy was 94.8% for pancreatic cancer vs chronic pancreatitis and 95.6% vs benign lesions. External validation ranged from 93% to 97% across institutions. Mean inference time was 41 ms per image vs 335 ms for ResNet152, enabling real-time integration into CT workflows. Overall, 93.9% of clinicians rated it clinically valuable for triage and staging support. Conclusions: Knowledge distillation allows a lightweight ResNet18 to achieve near–ResNet152 performance while reducing computational cost by over 90%. This approach addresses key translational barriers in pancreatic cancer imaging and supports real-time, globally scalable clinical decision support. Prospective studies will assess its impact on diagnostic delay, surgical referral, and treatment allocation.