Contrastive prototype federated learning against noisy labels in fetal standard plane detection.

Purpose: This study aims to improve federated learning (FL) for ultrasound fetal standard plane detection by addressing noisy labels and data size variability across decentralized clients. We propose a federated denoising framework leveraging prototypes from the largest dataset in the federation to refine noisy labels and enhance predictions in all clients while preserving privacy.

Methods: The proposed framework consists of two main steps. First, contrastive learning (SimCLR) is applied to the images of the largest client, generating robust embeddings. These embeddings are used to refine noisy labels in the same client by leveraging the latent space structure using a threshold-based k-nearest neighbors re-labeling strategy. As a second step, image prototypes, computed from the embeddings with noise-free labels, along with SimCLR trained backbone, are shared with the smallest client to guide the FL process effectively, without requiring the use of labels from the smallest client. To address possible image distribution shifts, an ensemble strategy is introduced, which uses a majority voting scheme to optimize label refinement in the smallest dataset while minimizing image discard.

Results: Our framework showed improved performance compared to traditional FL approaches in standard plane detection, achieving the highest mean F1-score across planes.

Conclusions: The proposed strategy effectively improves fetal standard plane detection by leveraging high-quality prototypes, enabling robust performance even with noisy and heterogeneous data size across clients, while preserving privacy.