Estratégia de seleção de clientes no aprendizado federado para a classificação de imagens médicas com otimização por enxame de partículas

Abstract

This work introduces a Federated Learning approach to assess the quality of biomedical images while protecting sensitive patient data. Unlike traditional methods that centralize data on a single server, Federated Learning trains models in a decentralized way, keeping data on the local devices of participating institutions. While Federated Learning offers privacy benefits, it struggles with data heterogeneity and inefficient client participation. To address this and im- prove the efficiency of collaborative training under heterogeneous data, the system integrates a client selection mechanism based on Particle Swarm Optimization (PSO), using factors such as data quality, sample diversity, and computational capacity. This strategy aims to boost overall model performance, reduce communication overhead, and avoid wasting resources. In addi- tion, the proposed method overcomes limitations of random client selection as used in FedAvg, which can become combinatorially intractable when searching for the best subset of clients. Experiments conducted on public disease datasets evaluate model robustness in heterogeneous environments. Results indicate that combining Federated Learning with PSO enables adaptive client selection across varying data heterogeneity levels while maintaining data security and complying with inherent medical domain privacy constraints, achieving up to 99.84% accuracy as one of the main performance gains of this work.