Adaptive Federated Learning With Intelligent Client Selection For Automated Chest X-Ray Pneumonia Detection

Abstract

While federated learning is a promising approach for privacy-preserving medical image analysis, optimal participation of the clients is essential for achieving good global model convergence, classification accuracy and communication efficiency, but it is not always observed. In this study, an adaptive Federated Learning with intelligent client selection for automatic Chest X-ray pneumonia detection is proposed. The goal is to design the best selection of clients for a distributed healthcare system in terms of maximized diagnostic accuracy, and minimised communication overhead. The proposed framework combines three intelligent client selection strategies: Random, Performance-Based, and Diversity-Based, Federated Averaging aggregation, privacy-preserving collaborative learning and the ImageNet-pretrained MobileNetV2. The Chest X-Ray Pneumonia images are split into 10 simulated hospitals and local models are trained independently with each hospital without sharing patient information and are then iteratively aggregated. The experimental results show that Performance Based selection round with 3 clients gives the best accuracy of the test of 83.36% which is more than the Random selection round (80.32%) and Diversity Based selection round (76.80%). In the 2-client scenario, the Random selection is the best with a score of 80.64%, surpassing the Performance-based (77.28%) and Diversity-based (80.16%). The higher the round participation of the clients, the better the test accuracy becomes with a 2.72% increase when the participation is increased from two to three clients per round, but the communication cost rises from 16 to 24 client-rounds. This novelty of the work is that for a given participation budget, the intelligent selection of clients has been studied under the same architecture, thus allowing for a thorough communication–accuracy trade off analysis. The study provides practical suggestions for adaptive client scheduling in privacy-preserving medical imaging systems and proves that the intelligent participation of patients with quality-aware substantially improves federated pneumonia detection, with retaining efficiency of communication and privacy of patient's data.