Research Article

Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification

Volume: 21 Number: 2 December 21, 2025
EN TR

Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification

Abstract

Glaucoma is a critical ophthalmological disease affecting millions of people worldwide, leading to irreversible optic nerve damage and permanent vision loss if diagnosed late. The disease often presents with no obvious clinical findings in its early stages, coupled with a high reliance on expert judgment, leading to time-consuming and misclassification risks in current diagnostic processes. This makes the integration of artificial intelligence (AI)-based automated systems into clinical decision support processes a crucial requirement for early and accurate glaucoma detection. This study proposes a hybrid approach based on the integrated use of lightweight deep learning (DL) architectures and machine learning (ML)-based classifiers for the automatic classification of glaucoma from fundus images. The proposed hybrid architecture utilizes MobileNetv1, MobileNetv2, and MobileNetv3 (small and large) architectures as the core components of the model. The MobileNet family was chosen due to its low parameter count, high computational efficiency, suitability for real-time operation in mobile and embedded systems, and its ability to effectively extract meaningful deep features from fundus images. Furthermore, in this study, high-dimensional deep feature vectors were extracted from fundus images using these pre-trained models. These features were then processed with different ML classifiers, such as Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Lightweight Gradient Boosting Machine (LGBM), Categorical Gradient Boosting (CatBoost), and k-Nearest Neighbor (kNN), creating a comprehensive hybrid classification framework that combines the strengths of both DL and traditional ML methods. Classification metrics such as accuracy, precision, recall, F-score, and area under curve (AUC) were used for performance evaluation. Experimental findings indicate that the MobileNetv2+SVM hybrid model, in particular, exhibits significant superiority. This hybrid model achieved the highest performance levels in the study, with 0.9409 accuracy, 0.9229 recall, 0.9221 F-score, and 0.9229 AUC. However, the highest precision value (0.9264) was obtained with the MobileNetv3(small)+LGBM hybrid model, demonstrating that different MobileNet variants can provide strong discrimination performance when effectively integrated with various classifiers. The results demonstrate that MobileNet-based deep feature extraction offers high discrimination in glaucoma classification and that the proposed hybrid approach is a reliable, fast, and computationally efficient solution suitable for use in clinical decision support systems. This study provides an important foundation for the development of low-cost and real-time early glaucoma detection systems.

Keywords

References

  1. [1] M. Ashtari-Majlan, M. M. Dehshibi, and D. Masip, “Glaucoma diagnosis in the era of deep learning: A survey,” Expert Systems With Applications, vol. 256, Art. no. 124888, Aug. 2024, doi: 10.1016/j.eswa.2024.124888.
  2. [2] S. K. Sharma, D. Muduli, R. Priyadarshini, R. R. Kumar, A. Kumar, and J. Pradhan, “An evolutionary supply chain management service model based on deep learning features for automated glaucoma detection using fundus images,” Engineering Applications of Artificial Intelligence, vol. 128, Art. no. 107449, Nov. 2023, doi: 10.1016/j.engappai.2023.107449.
  3. [3] A. Karimi, A. Stanik, C. Kozitza, and A. Chen, “Integrating deep learning with electronic health records for early glaucoma detection: A multi-dimensional machine learning approach,” Bioengineering, vol. 11, no. 6, Art. no. 577, Jun. 2024, doi: 10.3390/bioengineering11060577.
  4. [4] X. C. Ling et al., “Deep learning in glaucoma detection and progression prediction: A systematic review and meta-analysis,” Biomedicines, vol. 13, no. 2, Art. no. 420, Feb. 2025, doi: 10.3390/biomedicines13020420.
  5. [5] H. Elmannai et al., “An improved deep learning framework for automated optic disc localization and glaucoma detection,” Computer Modeling in Engineering & Sciences, vol. 140, no. 2, pp. 1429–1458, May 2024, doi: 10.32604/cmes.2024.048557.
  6. [6] S. Islam et al., “Novel deep learning model for glaucoma detection using fusion of fundus and optical coherence tomography images,” Sensors, vol. 25, no. 14, Art. no. 4337, Jul. 2025, doi: 10.3390/s25144337.
  7. [7] N. A. Alkhaldi and R. E. Alabdulathim, “Optimizing glaucoma diagnosis with deep learning-based segmentation and classification of retinal images,” Applied Sciences, vol. 14, no. 17, Art. no. 7795, Sep. 2024, doi: 10.3390/app14177795.
  8. [8] V. K. Velpula et al., “Glaucoma detection with explainable AI using convolutional neural networks based feature extraction and machine learning classifiers,” IET Image Processing, vol. 18, pp. 3827–3853, 2024, doi: 10.1049/ipr2.13211.

Details

Primary Language

English

Subjects

Image and Video Coding

Journal Section

Research Article

Early Pub Date

December 21, 2025

Publication Date

December 21, 2025

Submission Date

December 13, 2025

Acceptance Date

December 21, 2025

Published in Issue

Year 2025 Volume: 21 Number: 2

APA
Baydogan, C. (2025). Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification. Electronic Letters on Science and Engineering, 21(2), 123-135. https://izlik.org/JA23MC39GD
AMA
1.Baydogan C. Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification. Electronic Letters on Science and Engineering. 2025;21(2):123-135. https://izlik.org/JA23MC39GD
Chicago
Baydogan, Cem. 2025. “Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification”. Electronic Letters on Science and Engineering 21 (2): 123-35. https://izlik.org/JA23MC39GD.
EndNote
Baydogan C (December 1, 2025) Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification. Electronic Letters on Science and Engineering 21 2 123–135.
IEEE
[1]C. Baydogan, “Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification”, Electronic Letters on Science and Engineering, vol. 21, no. 2, pp. 123–135, Dec. 2025, [Online]. Available: https://izlik.org/JA23MC39GD
ISNAD
Baydogan, Cem. “Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification”. Electronic Letters on Science and Engineering 21/2 (December 1, 2025): 123-135. https://izlik.org/JA23MC39GD.
JAMA
1.Baydogan C. Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification. Electronic Letters on Science and Engineering. 2025;21:123–135.
MLA
Baydogan, Cem. “Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification”. Electronic Letters on Science and Engineering, vol. 21, no. 2, Dec. 2025, pp. 123-35, https://izlik.org/JA23MC39GD.
Vancouver
1.Cem Baydogan. Lightweight Deep Learning Architectures for Ophthalmic Disease Detection: MobileNet Variants Applied to Glaucoma Classification. Electronic Letters on Science and Engineering [Internet]. 2025 Dec. 1;21(2):123-35. Available from: https://izlik.org/JA23MC39GD