Medical imaging analysis and, specifically, Magnetic Resonance Imaging (MRI) have seen notable progress thanks to deep learning methods. However, the design of diagnostic models through Machine Learning (ML) encounters several problems related to data owner privacy and the heterogeneity of data involved in model training. Federated Learning (FL), one of the most innovative distributed learning techniques, thanks to its characteristics of data protection, represents a possible solution to enable ML-based diagnostic tools. It enables an approach in which a global model is trained in a collaborative fashion, ensuring data remains safely stored within each private device and infrastructure. Despite the benefits of FL, a critical aspect in the training of a collaborative model is related to the quality of the local data of each client and, in particular, the heterogeneous quality of each MRI acquisition. Traditional FL aggregation strategies, such as Federated Averaging (FedAvg), typically assign importance to participating clients only depending on the size of each local dataset, potentially allowing lower-quality imaging data to contribute to the aggregation of the global resulting model. In this paper, we introduce Federated Signal-to-Noise Ratio (FedSNR), a novel FL aggregation strategy specifically designed for the healthcare field and MRI diagnosis model training. FedSNR incorporates a new parameter in the weighted average performed during the aggregation process, correlating client contributions with their imaging quality metrics. The key insight is that MRI data with higher Signal-to-Noise Ratio (SNR) inherently contains more reliable …