FedPA: Federated Learning with Performance-Based Averaging for Efficient Medical Image Classification

Atif Mahmood^1,*, Yasin Saleem¹, Usman Tariq², Yousef Ibrahim Daradkeh³, Adnan N. Qureshi⁴
1 Faculty of Data Science and Information Technology, INTI International Univeristy, Nilai, 71800, Malaysia
2 Prince Sattam bin Abdulaziz University, Al-Kharj, 16278, Saudi Arabia
3 Department of Computer Engineering and Information, College of Engineering in Wadi Alddawasir, Prince Sattam bin Abdulaziz University, Al-Kharj, 16273, Saudi Arabia
4 Faculty of Arts, Society and Professional Studies, Newman University, Birmingham, B32 3NU, UK
* Corresponding Author: Atif Mahmood. Email: email , email

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2025.073501

Received 19 September 2025; Accepted 12 December 2025; Published online 02 March 2026

Download PDF

Abstract

Federated learning is a decentralized model training paradigm with significant potential. However, the quality of Federated Network’s client updates can vary due to non-IID data distributions, leading to suboptimal global models. To address this issue, we propose a novel client selection strategy called FedPA (Performance-Based Federated Averaging). This proposed model selectively aggregates client updates based on a predefined performance threshold. Only clients whose local models achieve an F1 score of 70% or higher after training are included in the aggregation process. Clients below this threshold receive the updated global model but do not contribute their parameters. In this way, the low-performance clients are still in the process of learning and, after some rounds, will be able to contribute. If no client meets the performance threshold in a given round, the system falls back to standard FedAvg aggregation. This ensures the global model continues to improve even when most clients perform poorly. We evaluate FedPA on a subset of the MURA dataset for abnormality detection in radiographs of four bone types. Compared to baseline federated learning algorithms such as Federated Averaging (FedAvg), Federated Proximal (FedProx), Federated Stochastic Gradient Descent (FedSGD), and Federated Batch Normalization (FedBN), FedPA consistently ranks first or second across key performance metrics, particularly in accuracy, F1 score, and recall. Moreover, FedPA demonstrates notable efficiency, achieving the lowest average round time (≈2270 s) and minimal memory usage (≈645.58 MB), all without relying on GPU resources. These results highlight FedPA’s effectiveness in improving global model quality while reducing computational overhead, positioning it as a promising approach for real-world federated learning applications in resource-constrained environments.