An Efficient One-Shot Federated Medical Imaging via Variational Inference Parametric Feature Transfer

Omid Orang, Pedro H Barros, Giulia Zanon de Castro, Frederico Gadelha Guimarães

Keywords: Medical imaging, One-shot federated learning, Variational Inference, Parametric feature-transfer

TL;DR: This paper presents a one-shot federated learning method where clients share only variational posterior parameters, enabling the server to synthesize features and train a global cosine classifier efficiently under IID and heterogeneous settings.

Abstract: This study introduces a one-shot federated technique for medical imaging called FBPFT-VI, a Variational Inference parametric feature-transfer approach. Each client freezes an Attention-MobileNetV2 encoder to extract features, then fits a variational posterior over its class-conditional feature statistics and transmits only the posterior parameters. The server samples synthetic features from these posteriors and trains a cosine classifier head, using Variational Inference to combine client contributions in a single aggregation round. Across multiple medical imaging benchmarks under IID and heterogeneous settings, FBPFT-VI improves the communication–accuracy trade-off.

Submission Number: 16