Fig. 2.

Personalized federated learning based on feature transformation network (FedFTN). Each institution contains one FTN-modulated denoising network (Figure 1). At each global iteration, each FTN-modulated denoising network at each institution is first trained locally for a fixed number of epochs. During model aggregation, the FTNs are kept local and only the denoising networks’ parameters are uploaded to the central cloud server for parameter averaging. The algorithm is summarized in Algorithm 1.