Sadia Alam Kotha, Abdullah Al Alam Shanto, Mumtahina Esha, Siam Tahsin Bhuiyan, Halima Khatun, Rashedur Rahman, Ashraful Islam, Saadia Binte Alam
2025 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE)
In: 2025 IEEE International Women in Engineering (WIE) Conference on Electrical and Computer Engineering (WIECON-ECE)
IEEE, pp. 480-485
Automated detection of retinal diseases from retinal color fundus images is crucial for early diagnosis and effective intervention. While deep learning approaches have achieved high accuracy, most existing methods analyze each eye independently and inconsistently apply preprocessing, limiting their ability to capture inter-eye asymmetries and subtle pathological features. In this study, we propose a bilateral ensemble framework for multi-class retinal disease classification using ConvNeXt-XLarge as the backbone. Fundus images were preprocessed with color normalization, a color correction technique to enhance feature consistency. Experiments were conducted on the ODIR-5K dataset, focusing on five clinically relevant categories: Normal, Diabetic Retinopathy, Cataract, Age-Related Macular Degeneration, and Myopia. We evaluated both single-eye and bilateral ensemble settings, with and without color correction, using Accuracy, Precision, Recall, F1 Score, Specificity. The bilateral ensemble with color-corrected images achieved the highest overall accuracy of 74.5 %, demonstrating that bilateral feature fusion and systematic preprocessing improve the model's ability to capture inter-eye asymmetries and subtle retinal patterns. These results underscore the potential of ensemble-based approaches to provide more reliable and clinically meaningful automated retinal disease detection.
Assistant Professor
Department of Computer Science and Engineering
Independent University, Bangladesh (IUB)Human-Computer Interaction, AI for Social Good, AI for Public Health, AI for Impact