Accuracy vs. Privacy: a Federated Learning Approach for Lung Ultrasound Pattern Classification

The exchange and flow of healthcare data are crucial and beneficial to develop automated methods for providing better care to patients. However, privacy is one of the major concerns when it comes to sharing medical data. Due to these concerns, limited data are available to develop such automated methods. To overcome this challenge, federated learning (FL), has gained significant interest within the research community as it allows to train deep learning (DL) models on different data sources without having to share data. In this study, our focus is on the evaluation of DL models trained to classify lung ultrasound (LUS) patterns in data acquired from multiple medical centers in FL setting. These patterns include horizontal artifacts, vertical artifacts, and small to large consolidations. To classify these patterns, we used ResNet-18 with spatial attention, trained and tested on 2104 LUS videos from 135 patients across 6 medical centers with a train-test split at the patient level. The classification performance of the patterns at frame and video levels is evaluated and the model's capability to perform patients' prognostic stratification is also assessed. We also compare it with an identical, centrally-trained model as the baseline method. Results show that the FL model achieved an overall accuracy of 57.5%, 48.6%, and 75% for frame, video, and prognostic levels, respectively, compared to 66.7%, 47.4%, and 80.8% achieved by the monolithic model. Despite a slight reduction in the classification performance compared to the monolithic model, FL demonstrates overall comparable results. These results underscore the efficacy of FL in enabling medical centers to learn collaboratively without the need to share raw data, thereby preserving privacy.