FL-PancreasNet: Federated Transformer Approach for Secure Multi-Center Pancreatic Lesion Analysis in CT Imaging

Computed Tomography (CT) is a mainstay for assessing pancreatic lesions. Clear delineation of the pancreas and tumour is essential for diagnosis and treatment planning. Training on data from multiple centres can improve robustness. In practice, however, clinical scans are usually kept within individual hospitals. Privacy rules and institutional policies often prevent central pooling. Differences in data distributions across sites can also undermine model stability. Most existing work relies on centralised three-dimensional U-Net variants or Transformer-based segmentation networks trained on pooled datasets. Such setups are difficult to realise in routine deployments. Federated Learning (FL) offers a practical alternative by enabling collaborative training without sharing raw scans. Yet, with heterogeneous client data, standard aggregation can become unreliable. To address this issue, a Federated Transformer-UNet hybrid, FL-PancreasNet, is introduced to improve cross-site generalisation. Secure training is supported through privacy-aware aggregation and robust handling of client updates. Experiments use the public Medical Segmentation Decathlon (MSD) Task07 Pancreas dataset. It contains 420 portal venous phase CT volumes with pancreas and tumour labels. A total of 282 cases are used for training and 139 for testing. On the MSD test set, FL-PancreasNet achieves a pancreas Dice of 0.892 and a tumour Dice of 0.748. The tumour Hausdorff Distance at the 95th percentile (HD95) is reduced to 15.6 mm. This exceeds the performance of Federated Averaging (FedAvg), which reports a tumour Dice of 0.721 and HD95 of 17.4 mm. With secure aggregation, performance is preserved at a tumour Dice of 0.747 with modest overhead. These results support privacy-preserving collaboration while maintaining strong segmentation accuracy in multi-centre settings.