Horizontal Federated Learning for Advancing Pneumonia Detection with Distributed Medical Imaging Data

Today pneumonia is still a global health problem and in many such situations diagnosis comes too late. PneumoNet, an innovative deep learning and federated learning approach to automate pneumonia detection in chest X-rays with improved performance in terms of diagnostic accuracy and patient data privacy, is presented. Decentralized data from multiple healthcare institutions is used to train a robust convolutional neural network (CNN) across multiple datasets, without encumbering sensitive patient information in pneumoNet. Using local model updates as input, the federated averaging system alternates between global model update and local model updates to improve the global model’s performance. PneumoNet is trained to $95.2 \%$ training accuracy, $93.8 \%$ validation and $92.5 \%$ test accuracy. The architecture of the federated learning ensures data privacy compliance while maintaining generalization insights across discrete institutions, which enables it to be a scoping and applicable solution to pneumonia detection. Included in the system is a user-friendly interface for intuitive prediction and upload of images, as well as for the delivery of downloadable medical reports to healthcare providers.