Аннотация
Medical image analysis (MIA) holds significant promise for advancing healthcare diagnostics, but privacy concerns often impede data sharing between institutions. Traditional methods require extensive labeled datasets, but annotating medical images is time-consuming and costly, leading to insufficient data for robust deep learning (DL) model training. Moreover, conventional federated learning (FL) approaches struggle with the heterogeneity of data across institutions and adapting to unseen labels, compromising model performance. This work addresses these challenges by integrating few-shot learning (FSL) with FL, leveraging the strengths of both methodologies. The proposed approach enables effective learning from limited and diverse data distributions while preserving patient privacy. By in-corporating meta-learning algorithms, our models quickly adapt to new, unseen data, enhancing robustness and generalizability in non-IID settings. This innovative framework promises significant improvements in MIA tasks, paving the way for more effective and privacy-preserving healthcare solutions.