Adaptive homomorphic federated learning framework for multi-institutional medical imaging with optimized diagnostic accuracy.
Authors
Affiliations (2)
Affiliations (2)
- Department of Computer Science and Engineering, School of Computing, SRM Institute of Science and Technology, Tiruchirappalli, Tamil Nadu, India. [email protected].
- Department of Computer Science and Engineering, School of Computing, SRM Institute of Science and Technology, Tiruchirappalli, Tamil Nadu, India.
Abstract
The growing use of artificial intelligence (AI) in medicine has highlighted the imperative for privacy-preserving and high-accuracy diagnostic systems. Traditional federated learning (FL) solutions allow multi-institutional collaborative training of a model without sharing raw patient data, but they tend to suffer from heterogeneous, multi-modal data, limited resilience to noisy or imbalanced data, and high communication overhead, which limits their practical clinical deployment. In order to overcome these constraints, we introduce the Next-Generation Adaptive Secure Federated Learning (NASFL) framework that aims at ultra-accurate, scalable, and secure multi-institutional medical AI. NASFL combines multi-level homomorphic encryption (MLHE) and stochastic differential privacy to provide patient confidentiality while using a transformer-guided ResNet backbone for adaptive multi-modal feature fusion between X-ray and CT imaging data. Institution-specific focus and trust-based aggregation dynamically scale model contributions to improve noise and low-quality dataset robustness. Communication efficiency is obtained from top-k gradient compression and adaptive learning rates, minimizing bandwidth consumption and speeding up convergence. The system was tested on publicly available multi-institutional datasets, namely NIH Chest X-ray14 and LIDC-IDRI CT scans, covering more than 112,000 images from over 30,000 patients, spanning 14 thoracic disease categories and lung nodules. NASFL exhibits exemplary performance, with 99.6% diagnostic accuracy, low convergence time (~ 65 communication rounds), good robustness in heterogeneous settings, and robust privacy assurance. These outcomes signify not only that NASFL surpasses traditional FL and state-of-the-art privacy-preserving protocols but also that it establishes a clinically sound platform for secure multi-institutional medical AI. Through its integration of adaptive multi-modal fusion, attention-guided aggregation, and strict privacy controls, NASFL sets a new benchmark for scalable, high-accuracy, and robust federated medical diagnostics, enabling wide-scale deployment in actual healthcare settings.