PDAC-CTFM — Scientific Report Paper

Overview

This work introduces a privacy-preserving federated deep learning framework for preoperative lymph node metastasis (LNM) detection in pancreatic ductal adenocarcinoma (PDAC). We integrate large-scale CT foundation model pre-training with a heterogeneity-aware federated aggregation strategy to enable collaborative training across three German institutions without sharing raw patient data.

Key contributions:

• Fine-tuning of the CT Vision Foundation Model (CT-FM) on multi-center PDAC CT data for LNM classification.
• A heterogeneity-aware federated aggregation strategy (PDAC-CTFMOurs) that jointly accounts for label-distribution imbalance and representation-level divergence across clients.
• Multi-center evaluation on 546 patients from three institutions (UKB Bonn, Charité Berlin, UMG Göttingen).

Main results (aggregated test set):

Our federated method outperforms all FL baselines (FedAvg, FedDisco, FedBN, FedALA) by +12.6% balanced accuracy over FedAvg while preserving data privacy.

2. Federated Learning — Proposed Method (Heterogeneity-Aware FedDisco)

Runs PDAC-CTFMOurs with default hyperparameters (a=0.1, b=0.1, γ=10). To sweep hyperparameters, edit the a_values, b_values, gamma_values lists at the bottom of the script:

### Using the code

The repository README covers installation, data, demo data (synthetic), using real data, federated learning setup. See [the repository on GitHub](https://github.com/albarqounilab/PDAC-CTFM#readme) for step-by-step instructions.