Seema Gulati
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India; & Department of Computer Science and Engineering, Graphic Era Hill University, Dehradun, Uttarakhand, India.

Kalpna Guleria
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India.

Nitin Goyal
Department of Computer Science and Engineering, School of Engineering and Technology, Central University of Haryana, Mahendergarh, Haryana, India.

Sushil Kumar
Department of Computer Science and Engineering, G. B. Pant DSEU Campus, Delhi, India.

Sunita Tiwari
Department of Computer Science and Engineering, G. B. Pant DSEU Campus, Delhi, India.

DOI https://doi.org/10.33889/IJMEMS.2026.11.5.092

  ;

Abstract

Diabetic Retinopathy (DR), Diabetic Macular Edema (DME), and glaucoma are significant ocular diseases that severely impair vision and require early automated detection for effective clinical management, but the traditionally adopted centralised deep learning based detection poses serious privacy problems for the participating organisations and also the data heterogeneity is a practical issue. The article proposes a privacy-preserving Federated Learning (FL) framework that enables collaborative model training across multiple decentralized datasets without sharing sensitive patient data for detecting DR, DME, and glaucoma from high-quality coloured fundus images, employing a Federated Deep Learning (FDL) approach coupled with the FedProx aggregation for effectively handling the non-IID data distributions in a multi-disease diagnostic setting. The framework evaluates four federated deep learning models based on EfficientNetB0, MobileNetV2, AlexNet, and InceptionV3 architectures across multiple communication rounds to analyze their effectiveness in a decentralized learning environment. A curated dataset of 2,700 high-resolution fundus images (balanced across four classes) is constructed from multiple public sources and augmentation was applied to the training set, increasing the effective training samples while keeping the test set unchanged. The dataset is distributed across three clients in a 30:35:35 ratio and trained over 20 communication rounds with consistent hyperparameters. The obtained results demonstrate that EfficientNetB0 and InceptionV3-based federated models consistently achieve more than 97% accuracy, with the EfficientNetB0-based model showing superior stability and emerging as the most suitable architecture with a prediction accuracy of 97.96% along with high precision, recall, and F1-score, indicating robust and stable multi-disease classification under heterogeneous federated settings.

Keywords- Federated learning, Automatic detection, Diabetic eye diseases, FedProx.

Citation

Gulati, S., Guleria, K., Goyal, N., Kumar, S., & Tiwari, S. (2026). FedEye: Detecting Diabetic Eye Diseases using Federated Deep Learning and Handling Data Heterogeneity with Fedprox Aggregation. International Journal of Mathematical, Engineering and Management Sciences, 11(5), 2292-2317. https://doi.org/10.33889/IJMEMS.2026.11.5.092.