Manuscript received March 12, 2026; revised April 7, 2026; accepted June 10, 2026; published October , 2026.

Abstract—Early diagnosis of ocular pathologies is critical to avoiding the loss of vision; nevertheless, existing diagnostic imaging modalities are often based on expensive imaging devices, including fundus cameras or optical coherence tomography scanners, which are often unavailable in rural and low-resource environments. In this study, we introduce MobiShuffNetV2, a powerful hybrid deep-learning model that incorporates depth wise separable convolutions of MobileNetV2 with channel-shuffle components typical of ShuffleNetV2 and is capable of classifying five ocular conditions, including normal, cataract, conjunctivitis, uveitis, and eyelid disorders, without any extra information other than just simple ocular photography. The dataset used in this study was collected from Mendeley and was preprocessed with a domain-specific pipeline that included resizing, grayscale conversion, region-of-interest detection, contrast enhancement, and denoising, followed by targeted augmentation to balance underrepresented classes, resulting in 9861 images. The model achieved a test accuracy of 99.70% (95% confidence interval: [0.9911, 0.9990]) with perfect precision and recall on the categories of Normal and Uveitis, and a Cohen’s kappa of 0.9962, which reflects near-perfect interrater agreement. These findings were supported by cross-validation that gave a result of 99.54% ± 0.84% accuracy. The proposed architecture exhibited an inference latency of 0.0016 seconds per sample, which equates to an overall latency of 1.61 seconds, thus allowing for real-time usage with end-to-end predictions achieved within a three-second window. Also, Grad-CAM (Class Activation Mapping) visualization supported the accurate localization of pathological areas and thus improved the interpretability of the model. This study proposes a highly effective and scalable pre-screening tool in early diagnosis of ocular disorders by making use of simple ocular photographs and not advanced imaging tools, especially in underserved populations.

Keywords—eye disease classification, medical image analysis, eye disease detection, computer-aided diagnosis

Cite: Mahe Zabin, "MobiShuffNetV2: An Efficient Hybrid Deep Learning Framework for Accessible Eye Disease Screening Using Simple Ocular Photography," Journal of Image and Graphics, Vol. 14, No. 5, pp. 910-928, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).