Towards good practice for convolution and attention with PANs in federated medical image classification
Makhanov N. Nhan H.D. Wong K.-S. Anh Tu N.
January 2025Springer
Journal of Supercomputing
2025#81Issue 1
In the current healthcare landscape, accurately diagnosing patients with respiratory conditions while preserving data privacy has become a critical global concern. To address this issue, federated learning (FL) presents an innovative solution that allows multiple medical institutions to collaboratively train a deep learning (DL) model without sharing private data. Although there is a keen interest in exploring FL methods for medical image classification, the literature needs a comprehensive comparison and analysis of these approaches. In this paper, we present an extensive benchmarking framework for FL specifically tailored for medical image classification. Our framework systematically evaluates various deep learning architectures, including CNN-based, transformer-based, and hybrid models, to address the challenges posed by non-independent and identically distributed (non-IID) and imbalanced data scenarios. Among the architectures, hybrid models like CoAtNet, which combine the strengths of CNNs and transformers, have demonstrated superior performance in capturing both local and global features in image classification tasks. Building on this, we introduce a novel model, CoAtPENet, which integrates position-aware neurons (PANs) into the CoAtNet architecture, enhancing feature alignment across clients during the aggregation process. Experiments are conducted on three publicly available chest radiography datasets to assess the performance of various models. We mainly discuss two typical problems of image classification (multi-class classification and multi-label classification) and employ the strategies wherein clients are randomly allocated a specific number of images in each round to manage the settings of these problems in different FL environments. The results show that we can achieve promising classification performance when integrating DL models with PAN. Notably, CoAtPENet gains performance comparable to centralized training. This study not only advances the application of FL in medical image analysis but also lays the groundwork for future research in this promising area.
Deep learning , Federated learning , Medical image classification
Text of the article Перейти на текст статьи
Department of Computer Science, Nazarbayev University, Astana, Kazakhstan
College of Engineering and Computer Science, VinUniversity, Hanoi, Viet Nam
Department of Computer Science
College of Engineering and Computer Science
10 лет помогаем публиковать статьи Международный издатель
Книга Публикация научной статьи Волощук 2026 Book Publication of a scientific article 2026