pancreatic-cancer

Privacy-preserving AI detects pancreatic cancer lymph node spread across multiple hospitals.

Predicting lymph node metastasis (LNM) in pancreatic ductal adenocarcinoma (PDAC) before surgery from CT scans is difficult due to low sensitivity and inter-institutional heterogeneity. This retrospective multi-center study of 546 patients (3 institutions) developed a federated deep learning…

Dan Manning

Dan Manning

— 1 min read
Share
Privacy-preserving AI detects pancreatic cancer lymph node spread across multiple hospitals.

Privacy-preserving AI detects pancreatic cancer lymph node spread across multiple hospitals.

Predicting lymph node metastasis (LNM) in pancreatic ductal adenocarcinoma (PDAC) before surgery from CT scans is difficult due to low sensitivity and inter-institutional heterogeneity. This retrospective multi-center study of 546 patients (3 institutions) developed a federated deep learning framework that keeps patient data private while enabling collaborative model training.

A CT foundation model (pre-trained on 148,000 CT scans) was fine-tuned for LNM classification. Under centralized training, the model achieved balanced accuracy of 0.601 and DOR of 3.45. The novel heterogeneity-aware federated strategy outperformed standard FedAvg by 12.6% in balanced accuracy while maintaining strict data privacy.

This demonstrates a scalable, privacy-preserving pathway for multi-institutional AI model deployment in oncology.

Key Findings

  • CT foundation model pre-trained on 148,000 scans fine-tuned for PDAC LNM detection
  • Centralized model: balanced accuracy 0.601, DOR 3.45
  • Heterogeneity-aware federated strategy outperformed FedAvg by 12.6% balanced accuracy
  • 546 patients from 3 institutions analyzed while preserving data privacy
  • Framework scalable to distributed healthcare systems without sharing raw data

Implications

Federated learning enables multi-center AI model development without sharing patient data—critical for regulatory compliance and clinical adoption. Improved LNM detection could enhance surgical planning and treatment decision-making for PDAC patients.

Caveats

Retrospective study with limited performance (balanced accuracy 0.601—only modest improvement over chance); abstract-only. LNM prediction from CT remains inherently difficult. Clinical deployment requires further validation and regulatory approval.

Source: Scientific reports — 2026-04-09

Read more