Privacy-preserving federated learning with MobileViT for chest X-ray classification in Nigerian hospitals

Abstract

Nigeria faces challenges in the accurate diagnosis of respiratory diseases such as tuberculosis (TB), pneumonia, and COVID-19 because of limited radiology resources and stringent patient-privacy requirements. Deep learning models can provide strong diagnostic performance but often rely on centralised data, raising ethical and security concerns. This study presents an application-oriented privacy-preserving federated learning framework integrating Mobile Vision Transformer (MobileViT) for chest X-ray classification across simulated Nigerian healthcare environments. Using the Nigeria Chest X-ray Dataset (2,600 radiologist-labelled images across four classes), experiments were conducted in a simulated federated environment on a single compute instance with five clients as a proof of concept for multi-hospital collaboration. Differential privacy was implemented through differentially private stochastic gradient descent (DP-SGD), and performance was evaluated using accuracy, area under the curve (AUC), precision, recall, F1-score, and class-wise sensitivity. In this simulated setting, the differentially private federated learning (DP-FL) MobileViT model achieved test accuracy above 92% and a macro F1-score above 89% at moderate privacy budgets (varepsilon  approx 5--6), while maintaining TB sensitivity above 85% and pneumonia sensitivity above 88%. These proof-of-concept results suggest that privacy-preserving federated learning could support reliable and ethical AI diagnostics in resource-limited Nigerian hospitals, pending real-world deployment validation.