Knowledge-Constrained Federated Learning for Medical Image Classification with SE-ResNeXt-50

¹ Department of Electronics and Communication Engineering, KG Reddy College of Engineering & Technology, Hyderabad, India
² Department of CSI, College of Science, Majmaah University, Al Majmaah, Saudi Arabia
³ Department of Electronics and Communication Engineering, Nitte Meenakshi Institute of Technology, (Nitte Deemed to be University), Bengaluru, India
⁴ Independent researcher
⁵ Department of Information Technology, College of Computing and Information Sciences, University of Technology and Applied Sciences, Shinas, Oman

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Recently, the classification of medical images plays an important role in healthcare, especially in diagnosing diseases and planning treatments. Advancements in intelligent systems operate in external or centralized environments concerning patient confidentiality, which is progressively essential. Traditional Machine Learning (ML) methods depend on data collection, which represents sensitive medical information and reduces the performance. To address this issue, Federated Learning (FL) is a suitable method that allows multiple institutions to collaboratively train models by incorporating knowledge constraints using their current data, without sharing confidential records. FL not only preserves privacy, but also supports robust and safe model development across various data sources. The proposed FL includes SE-ResNeXt-50, which combines the multibranch ResNeXt-50 design with a squeeze-and-excitation (SE) attention mechanism, allowing more discriminative feature extraction and adaptive channel weighting. The SE-ResNeXt-50 improves the models ability to handle heterogeneous and non-ID (non-identically distributed) medical imaging data across institutions. The proposed FL model integrating SE-ResNeXt-50 is evaluated using two different medical imaging datasets, BR35H (Brain Tumor Detection 2020) based on MRI images and SARS-CoV-2 CT on COVID-19, a lung CT scan image. The proposed FL model integrating SE-ResNeXt-50 achieves an accuracy of 98.8% for MRI images and 99.3% for CT scans, which is superior to the existing convolutional neural network (CNN) with the Gray-Level Co-occurrence Matrix (GLCM) model.