FedFAT: Frequency adpative interpolation for federated domain generalization on heterogeneous medical images

Abstract

Multiple distributed medical institutions can leverage federated learning (FL) to collaboratively build a shared prediction model with privacy protection. However, the presence of non-independent and identically distributed (non-IID) data in medical imaging leads to data drift in practical learning scenarios, detrimentally affecting both convergence and generalization to the unseen domain. In this paper, we propose a novel framework named Federated Frequency Adaptive Interpolation(FedFAT), which leverages a frequency space adaptive interpolation mechanism to mitigate data drift in federated domain generalization. FedFAT enables clients to adaptively exchange partial amplitude information, leveraging multi-source data distributions to enhance generalization. Crucially, local phase information is retained to preserve privacy. To mitigate data drift, FedFAT employs cross-client feature alignment via amplitude normalization, which effectively batch-normalizes images from diverse source distributions. Furthermore, we introduce a client-specific weight perturbation mechanism designed to guide local models toward a consistent low-loss region. We have theoretically analyzed the proposed method and empirically conducted extensive experiments on two medical image classification and segmentation tasks, showing that FedFAT outperforms a set of recent state-of-the-art methods with average Dice improvement of 2.42 % and 10.61 % on the prostate MRI segmentation datasets (PROMISE12, PROSTATEx, and NCI-ISBI) and breast cancer classification datasets(CAMELYON17), respectively. FedFAT also has an improvement of 4.15 % on generalization performance. These results demonstrate the superiority of FedFAT in handling data drift and improving generalization performance.