Federated Learning for Lung Ultrasound Classification Across Age-diverse Patient Populations

Deep Learning (DL) has rapidly advanced Lung Ultrasound (LUS) image classification. Traditional DL uses Centralized Learning (CL) to train models, which requires large datasets for effective model training. While combining data across hospitals could provide sufficient samples, patient privacy may complicate direct data sharing. To solve this problem, Federated Learning (FL) can be employed. FL enables multiple clients to collaborate without exposing patient data, allowing each client to train models locally and share only weights for central aggregation. While previous works investigated the use of FL for LUS data analysis, the present study investigates, for the first time, whether FL maintains robustness across hospitals (compared with CL), specifically for the case of LUS data with diverse age distribution. Our experiments consider three configurations: (1) in Configuration 1, both FL and CL models are trained with adult LUS data from 4 hospitals; (2) in Configuration 2, a fifth client node with neonatal data is added; (3) in Configuration 3, the neonatal dataset is doubled. After training, all models are tested on two external test datasets. The results show that FL provides comparable and robust performance while keeping data private across demographically diverse institutions.