Akademik Veri Yönetim Sistemi
8th International Conference on Data Science and Machine Learning Applications, CDMA 2025, Riyadh, Suudi Arabistan, 16 - 17 Şubat 2025, ss.91-96, (Tam Metin Bildiri)
Traditional centralized machine learning faces chal-lenges with 'data label scarcity' and 'data privacy', especially in medical data classification. Semi-supervised learning leverages unlabeled data, while federated learning protects data privacy by decentralizing model training. In this study, we introduce an improved inter-client relation matching algorithm for semi-supervised federated medical image classification (iFedIRM). The FedIRM algorithm defines a loss function to transfer the knowledge of disease relationships from labeled clients to unlabeled clients to extract discriminative information from unlabeled data using per-category mean feature vectors. Our approach enhances the FedIRM algorithm by incorporating per-category feature covariance matrices for better feature representation and using group-based averaging for model aggregation. We utilize the ConvNeXt architecture as the backbone, and apply a confidence threshold to filter unreliable pseudo-labels. Experiments on three medical datasets show improved accuracy, F1, and AUC metrics. Code will be available at https://github.com/msoz7/iFedIRM.