A Collaborative Privacy Preserved Federated Learning Framework for Pneumonia Detection using Diverse Chest X-ray Data Silos

Pneumonia detection from chest X-rays remains one of the most challenging tasks in the traditional centralized framework due to the requirement of data consolidation at the central location raising data privacy and security concerns. The amalgamation of healthcare data at the centralized storage leads to regulatory concerns passed by the governments of various countries. To address these challenges, a decentralized, federated learning framework has been proposed for early pneumonia detection in chest X-ray images with a 5-client architecture. This model enhances data privacy while performing collaborative learning with diverse data silos and resulting in improved predictions. The proposed federated learning framework has been trained with a pre-trained EfficientNetB3 model in the Independent and Identically Distributed (IID) and non-IID data distributions, while the model updation has been performed using federated proximal aggregation. The configuration of the proximal term has been kept as 0.05, achieving an accuracy of 99.32% on IID data and 96.14% on non-IID data. In addition, the proximal term has also been configured to 0.5, resulting the accuracy levels of 92.05% and 96.98% in IID data and non-IID data distributions, respectively. The results of the proposed model demonstrate the effectiveness of the federated learning model in pneumonia detection, highlighting its potential for real-world applications in decentralized healthcare configurations.