An AI-Based Framework for Glaucoma Diagnosis Using Deep Learning and Machine Learning Classifiers

Abstract

Glaucoma is an ocular neuropathy that progresses over time and causes blindness worldwide. Early and accurate identification is crucial to preventing vision loss, but manually evaluating retinal fundus images takes a lot of time and is unreliable. This work uses retinal fundus images to automatically identify and assess the risk of glaucoma using deep learning and machine learning techniques. The suggested approach incorporates a hybrid convolutional neural network that combines DenseNet121 and EfficientNetB3 for reliable feature extraction and preliminary classification, followed by LightGBM for precise and final classification. The optic disc and optic cup are segmented using a U-Net-based segmentation algorithm to improve clinical relevance. This allows for the calculation of the Cup-to-Disc Ratio (CDR), which is essential for glaucoma diagnosis. To create an enhanced feature representation, the deep learning features are combined with the calculated CDR value. A LightGBM classifier is then used to process the representation in order to make the final choice. Standard performance indicators such as Accuracy, Precision, Recall, F1-Score, AUC-ROC, Specificity, and Cohen\'s Kappa coefficient are used to assess the model. According to experimental findings, the suggested hybrid architecture improves classification robustness and diagnostic reliability compared to standalone approaches. The created system offers an automated glaucoma screening and decision support solution that is both scalable and clinically interpretable.

Introduction

The text presents a research study on an AI-based glaucoma detection system using retinal fundus images to enable early, accurate, and interpretable diagnosis of glaucoma, a disease that can cause irreversible blindness if undetected early.

Because manual diagnosis using optic disc and cup analysis is subjective and requires specialists, the study proposes an automated system that combines deep learning with clinically meaningful measurements, especially the Cup-to-Disc Ratio (CDR).

The proposed approach is a multi-stage hybrid framework:

1. Hybrid CNN Classification (EfficientNetB3 + DenseNet121):
   Extracts deep image features and provides initial glaucoma classification.
2. U-Net Segmentation:
   Identifies optic disc and optic cup regions and computes the CDR, improving clinical interpretability.
3. Feature Fusion + LightGBM Classifier:
   Combines deep learning features with CDR values and uses LightGBM for final, refined classification.

The system is designed to overcome limitations of existing methods, which often either focus only on image classification or ignore anatomical indicators. By integrating segmentation and clinical biomarkers, the model aims to be both accurate and explainable.

Evaluation uses standard metrics such as accuracy, precision, recall, F1-score, and ROC-AUC, supported by confusion matrix and ROC curve analysis. The system is also deployed using a user-friendly interface (Streamlit) for real-world screening.

Conclusion

To improve diagnostic precision and clinical interpretability, a multi-stage hybrid architecture combining deep feature extraction, anatomical segmentation, and machine learning-based classification was devel- oped for automated glaucoma detection using retinal fundus pictures. In order to precisely calculate the Cup-to-Disc Ratio (CDR), the system integrates hybrid convolutional neural networks for semantic feature learning with U-Net-based optic disc and cup segmentation. This captures both structural biomarkers and pathological visual patterns linked to glaucoma. In comparison to solo methods, a Gradient Boosting classifier applied to the fused feature representation increases robustness and decreases misclassification, resulting in an overall classification accuracy of 97.14%.

References

[1] Karanbir Singh Grover ,Nitika Kapoor ,\" Artificial Intelligence for Glaucoma and Diabetic Retinopathy:A Review” , Lecture Notes in Networks and Systems ((LNNS,volume1181),pp:217–227,February2025,https://doi.org/10.1007/978-981-97-8861-3_19. [2] Thukkaram Umashree and S. Narendran,\" Convolutional Neural Network Analysis for Glaucoma Detection in Retinal Image Analysis Compared with the K-Nearest Neighbors Algorithm’s Accuracy\", Lecture Notes in Networks and Systems ((LNNS,volume 1038), pp 651–658 ,September 2024,https://doi.org/10.1007/978-981-97-4149-6_45. [3] Mandicou Ba(B), Amadou Mbaye, Idy Diop, and Alassane Bah, \"Comparative Study of Machine Learning Models for Glaucoma Detection Based on Retinal Images \", Lecture Notes in Networks and Systems ((LNNS,volume 1421), pp 372–385, June 2025, https://doi.org/10.1007/978-3-031-92545-0_34. [4] Wisal Hashim Abdulsalam, Rasha H. Ali, Sawsan H. Jadooa, Samera Shams Hussein, “Automated Glaucoma Detection Techniques: A Literature Review”, Pages:19891-19897,Volume:15,Issue:1,February2025,doi: https://doi.org/10.48084./etasr.9316 [5] Geetha, M. Carmel Sobia, D. Santhi, A. Ahilan, “DEEP GD: Deep learning based snapshot ensemble CNN with EfficientNet for glaucoma detection”, Biomedical Signal Processing and Control, Elsevier, Volume 100,Part B, February 2025,doi: https://doi.org/10.1016/j.bspc.2024.106989 [6] Jie Xu, Erkang Jing & Yidong Chai, “A deep learning model integrating domain-specific features for enhanced glaucoma diagnosis”, BMC Medical Informatics and Decision Making, Volume :25, Article no:195, May 2025, https://doi.org/10.1186/s12911-025-02925-9 [7] Hubert Swierczynski, Juliusz Pukacki, Szymon Szczesny,Cezary Mazurek, Robert Wasilewicz, “Application of machine learning techniques in GlaucomAI system for glaucoma diagnosis and collaborative research support”, Scientific Reports, Article No:7940, March 2025, https://doi.org/10.1038/s41598-025-89893-2

Copyright

Copyright © 2026 Pedapalli Sobhitha, K. Santhi Sree. 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.