Optimized federated learning framework with RegNetZ and Swin-Transformer for multimodal pancreatic cancer detection1.

Pancreatic cancer is among the most lethal malignancies, marked by aggressive progression, late diagnosis, and limited screening methods, resulting in a five-year survival rate of less than 10%. Early-stage tumors are especially challenging to detect with standard CT and MRI imaging, leading to treatment delays and poor outcomes. While deep learning offers promise, centralized training in healthcare raises serious privacy and data-sharing concerns. This study introduces a federated learning framework that integrates RegNetZ and the Swin-Transformer for automated detection, subtype classification, and prognosis prediction from multimodal inputs, including CT, MRI, histology, genomic, and clinical records. The Swin-Transformer models long-range dependencies, whereas the lightweight RegNetZ backbone ensures efficient local feature extraction. A Hybrid Aquila-Grey Wolf Optimizer (HA-GWO) is incorporated to balance exploration and exploitation during hyperparameter tuning, providing faster convergence and reduced computational cost compared to conventional search strategies. The proposed framework, evaluated across 5-7 simulated client institutions, achieves 99.2% accuracy, 98.9% sensitivity, 99.0% precision, and 99.4% AUC, outperforming both CNN-only and transformer-only baselines. It further minimizes false positives and false negatives, improving both subtype classification (adenocarcinoma, neuroendocrine, cystic neoplasms) and prognosis risk prediction (high vs. low risk). Hyperparameter sensitivity analysis identifies a learning rate of 0.003 with a batch size of 64 as optimal. By enabling decentralized model training without raw data exchange, the system enhances diagnostic accuracy while preserving privacy, offering a practical solution for real-time pancreatic cancer detection in federated healthcare environments. The framework is scalable across medical institutions and supports precision oncology by enabling early and reliable diagnosis at low computational cost.