CFL-SparseMed: Communication-Efficient Federated Learning for Medical Imaging with Top-k Sparse Updates

Modern imaging modalities such as MRI, CT, ultrasound, and digital pathology have greatly expanded the volume of medical data [1, 2]. Concurrently, advances in artificial intelligence (AI), especially deep learning, have improved data processing efficiency and clinical decision-making [3]. However, applying AI in medical imaging remains challenging due to data privacy concerns, legal constraints, and the limited availability of diverse, high-quality datasets across healthcare institutions[4]. Individual organizations often possess insufficient data, leading to model overfitting and reduced generalizability. Moreover, centralizing healthcare data may violate privacy regulations such as Health Insurance Portability and Accountability Act (HIP AA) and General Data Protection Regulation (GDPR)[5], restricting inter-institutional collaboration [6]. To address these challenges, researchers and medical institutions have explored Federated Learning (FL) [6], which allows collaborative model training without sharing raw data.