EURO 2025 Leeds
Abstract Submission

2561. PFL-MDoE: A Personalized Federated Learning Framework with Mixture of Disease Experts for Robust Medical Image Classification

Invited abstract in session TA-28: AI and Machine Learning for Decision Support, stream Decision Support Systems.

Tuesday, 8:30-10:00
Room: Maurice Keyworth 1.03

Authors (first author is the speaker)

1. Tianrun Cai
Alliance Manchester Business School, University of Manchester
2. Qing Yin
Alliance Manchester Business School, University of Manchester
3. Yu-Wang Chen
Manchester Business School
4. Xian Yang
Alliance Manchester Business School, University of Manchester

Abstract

AI-driven decision methods have demonstrated superior performance for medical image classification tasks compared to traditional rule-based or statistical approaches. However, these methods need large-scale centralized data for model training which is impractical in healthcare due to patient data sensitivity and strict privacy regulations. Federated Learning (FL) offers a privacy-preserving solution by enabling local model training on private data for global model aggregation on a central server. However, FL faces significant statistical heterogeneity during model training, such as feature and label heterogeneity across medical institutions. Such heterogeneity often leads to aggregated global models making biased decisions in diverse institutions. Therefore, this paper introduces a novel framework called PFL-MDoE, an acronym for Personalized Federated Learning (PFL) with Mixture of Disease Experts (MDoE). The PFL component incorporates personalized modules that adapt to local data distributions, mitigating feature heterogeneity. In addition, the MDoE architecture incorporates dedicated expert modules designed for diseases with different frequencies. The gating mechanism dynamically assigns weights to the corresponding modules based on the disease category. By integrating specialized knowledge in this manner, MDoE addresses label heterogeneity. Consequently, our framework offers a personalized and secure AI training solution for diverse medical image classification tasks.

Keywords

Status: accepted

Back to the list of papers