INTEGRATING POSITION-AWARE NEURONS IN DEEP LEARNING MODELS FOR EFFECTIVE FEDERATED MEDICAL IMAGE CLASSIFICATION

Abstract

Privacy regulations create significant barriers to accessing distributed medical imaging datasets, impeding collaborative efforts to improve diagnostic accuracy for respiratory diseases. This thesis presents a novel Federated Learning (FL) framework that enables collaboration between healthcare institutions without compromising data privacy. The research makes three key contributions. First, it develops a comprehensive benchmarking framework to evaluate FL performance in medical image classification. This framework systematically compares various deep learning (DL) architectures, including Convolutional Neural Networks (CNNs), Transformers, and hybrid models, under challenging real-world conditions such as non-IID (non-independent and identically distributed) and imbalanced datasets. Second, the research introduces CoAtPENet, a hybrid model that enhances the CoAtNet architecture by incorporating Position-Aware Neurons (PANs). This integration specifically addresses the neuronal alignment issues that commonly occur during model aggregation in FL, improving model consistency across distributed training. Third, extensive empirical validation using three publicly available chest X-ray datasets demonstrates that the proposed approach achieves classification performance comparable to centralized training methods in both multi-class and multi-label classification scenarios. The results confirm that CoAtPENet effectively handles data heterogeneity while maintaining diagnostic accuracy. By enabling secure collaboration between healthcare facilities with protected datasets, this research advances both the theoretical understanding and practical application of FL in medical imaging, ultimately supporting a more accurate diagnosis of respiratory diseases without compromising patient privacy.