Distributed learning effectiveness in medical image analysis when trained with limited dataset

Raissa Souza, Pauline Mouches, Matthias Wilms, Anup Tuladhar, Sonke Langner, Nils Forkert

Keywords: Federated Learning, Travelling Model, Distributed Learning

TL;DR: This study aims to compare two distributed learning approaches, federated learning and travelling model, for cases where medical centers can only provide very small samples, such rare and pediatric diseases or small hospitals.

Abstract: Federated learning (FL) is a cutting-edge method for distributed learning used in many fields, including healthcare. However, medical centers need sufficient local data to train local models and participate in an FL network, which is often not feasible for rare and pediatric diseases or small hospitals with limited patient data. As a result, these centers cannot directly contribute to FL model development. To address this issue, this work explores the effectiveness of a different approach called the travelling model (TM). Specifically, this work evaluates the performances of FL and TM when only very small sample sizes are available at each center. Brain age prediction was used as an example case for comparison in this work. Our results indicate that the TM outperforms FL across all sample sizes tested, particularly when each center has only one sample.