A Cross-Disease Comparative Analysis of CNN, Transformer, Hybrid, and Self-Supervised Models for Ocular Disease Classification
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.01710106 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Hazem Y. Al-Sebaay und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Deep learning architecture comparisons in ophthalmic imaging are almost always conducted on a single disease and dataset, leaving open whether an architecture’s apparent advantage is a genuine, transferable property or an artifact of that one task. This study presents a controlled comparison of nine architectures spanning four families (CNN, Transformer, hybrid, and self-supervised DINOv2), trained under a harmonized protocol across three disease-specific fundus datasets: ORIGA (glaucoma), AMDNet23 (age-related macular degeneration), and ODIR-5K (cataract), with three-seed reproducibility. Bootstrap confidence intervals, Holm–Bonferroni-corrected pairwise De-Long tests, calibration and clinical operating points are reported, together with Spearman rank-correlation of performance across diseases as the central, exploratory analysis. Individual-model rank correlates only weakly and non-significantly between any pair of diseases (rho = 0.50 to 0.65, all p > 0.05), with DeiT-tiny obtaining the highest and EfficientNet-B0 the lowest mean AUROC on all three datasets, while the middle of the ranking reorders. At the family level, the two CNN configurations rank last on every disease, but which family takes first place changes with the disease, and the test is underpowered (n = 4). In addition, a lightweight, untrained heuristic for opticdisc localization, applied only where anatomically appropriate (glaucoma, not AMD or cataract), recovers detail lost to whole-image downsampling and yields the largest gain among the pipeline interventions tested (0.030 mean AUROC across all nine models). These findings suggest that disease-conditional evaluation, rather than a single “best” architecture, should inform model selection for ocular disease screening research.
Zitieren
APA
Al-Sebaay, H. Y., El-Saadawy, H., Al-Berry, M. N., & Amin, S. E. (2026). A Cross-Disease Comparative Analysis of CNN, Transformer, Hybrid, and Self-Supervised Models for Ocular Disease Classification. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.01710106
Vancouver
Al-Sebaay HY, El-Saadawy H, Al-Berry MN, Amin SE. A Cross-Disease Comparative Analysis of CNN, Transformer, Hybrid, and Self-Supervised Models for Ocular Disease Classification. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.01710106
BibTeX
@article{AlSebaay2026,
author = {Al-Sebaay, Hazem Y. and El-Saadawy, Hadeer and Al-Berry, Maryam N. and Amin, Safaa El-Sayed},
title = {A Cross-Disease Comparative Analysis of CNN, Transformer, Hybrid, and Self-Supervised Models for Ocular Disease Classification},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.01710106}
}
EUVIMED