EUVIMEDEUVIMED entdecken
Crossref

An Empirical Evaluation and Comparative Study of Metaheuristic-Optimized Deep Learning for Four-Class Retinal Disease Classification

Audrey Huong, Wan Mahani Hafizah Wan Mahmud, Ser Lee Loh, Kim Gaik Tay, Xavier Ngu

Journal of Innovative Image Processing · 12.05.2026 · 8 (2) · 665-683

DOI: 10.36548/jiip.2026.2.012 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 12.05.2026. Verfasst von Audrey Huong und 4 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Traditional retinal disease diagnosis using colour fundus imaging is subject to inter- and intra-observer variability and subjective bias. Meanwhile, manual or heuristic hyperparameter selection for neural network training is vulnerable to suboptimal convergence. There are also common challenges, including the lack of standardized evaluation protocols, non-uniform experimental designs, and class imbalance, which further hinder the development of reliable classifiers. This research addresses these limitations by employing metaheuristic algorithms to optimize the training of AlexNet, GoogleNet, and EfficientNet-B0. Particle swarm optimization (PSO), grey wolf optimization (GWO), and wild horse optimization (WHO) are evaluated for their effectiveness in identifying optimal training hyperparameter configurations for a four-class fundus image classification task. Statistical analysis revealed no significant association between the choice of optimization algorithm and the prediction performance (ρ = 0.341). The difference is more pronounced when testing the relationship between the network architecture and the prediction outcomes (ρ = 0.015). Among the evaluated optimizer-architecture pairs, GWO-EfficientNet demonstrates superior performance, achieving an accuracy of 96.52%, precision of 96.50%, recall of 96.47%, specificity of 98.84%, F1 score of 96.48%, and a Matthew’s correlation coefficient (MCC) of 95.33%, outperforming other optimizer-architecture combinations. Repeated experiments show strong performance consistency across the best-performing pairs, with a standard deviation of below 2%. The high accuracy and low inference time highlight the potential of the proposed approach for real-time ophthalmology applications, supporting improved clinical decision-making and more efficient eye care delivery.

Zitieren

APA

Huong, A., Mahmud, W. M. H. W., Loh, S. L., Tay, K. G., & Ngu, X. (2026). An Empirical Evaluation and Comparative Study of Metaheuristic-Optimized Deep Learning for Four-Class Retinal Disease Classification. Journal of Innovative Image Processing, 8(2), 665-683. https://doi.org/10.36548/jiip.2026.2.012

Vancouver

Huong A, Mahmud WMHW, Loh SL, Tay KG, Ngu X. An Empirical Evaluation and Comparative Study of Metaheuristic-Optimized Deep Learning for Four-Class Retinal Disease Classification. Journal of Innovative Image Processing. 2026;8(2):665-683. doi:10.36548/jiip.2026.2.012

BibTeX
@article{Huong2026,
  author = {Huong, Audrey and Mahmud, Wan Mahani Hafizah Wan and Loh, Ser Lee and Tay, Kim Gaik and Ngu, Xavier},
  title = {An Empirical Evaluation and Comparative Study of Metaheuristic-Optimized Deep Learning for Four-Class Retinal Disease Classification},
  journal = {Journal of Innovative Image Processing},
  year = {2026},
  volume = {8},
  number = {2},
  pages = {665--683},
  doi = {10.36548/jiip.2026.2.012}
}
Kostenlos starten →