Integrated Deep Learning Framework for glaucoma detection, Optic Disk/Cup Segmentation, and CDR Calculation

Abstract

The proposed system is a twostage deep learning framework for automated glaucoma detection and analysis of fundus images in this project. Once the image is processed, the system hands it off to a ResNet model to predict whether the image is positive for glaucoma or not. When it is a positive case, it segments the optic disc and cup using a YOLOv5 model and can then attempts to precisely calculate the cup to disc ratio (CDR), a key parameter in glaucoma risk determination. With combined classification and segmentation integration the proposed streamlines workflow greatly reducing manual examination-dependency and creates a more objective and accurate diagnostic support for ophthalmology. Clinicians can upload images, view segmentation results, and accessCDR values in a user friendly system to help make appropriate clinical decisions. The proposed system is an efficient and accurate framework with the potential for implementation in real time for early glaucoma diagnosis in high demand or resource limited locations or situations or settings

Introduction

Overview

Glaucoma is a progressive and asymptomatic eye disease leading to irreversible blindness. Early detection is crucial but challenging due to its silent progression. Traditional diagnostic methods rely on manual assessment of optic disc and cup regions in fundus images, which are time-consuming, subjective, and inconsistent.

Proposed Solution

This project introduces an automated, two-stage deep learning system for glaucoma detection:

• Stage 1 – Classification (ResNet50):

  • Uses ResNet50 to classify fundus images as glaucoma-positive or negative.

  • Employs transfer learning for better performance with limited data.

  • Achieves 92.5% accuracy, 93.8% sensitivity, and 92.6% F1-score.

• Stage 2 – Segmentation (YOLOv5):

  • Applied only to glaucoma-positive cases to segment optic disc and cup.

  • Calculates the Cup-to-Disc Ratio (CDR), a key indicator of glaucoma risk.

  • Achieves 89.9% Dice coefficient for optic disc and 86.5% for optic cup segmentation.

  • Average CDR deviation from expert annotations is just 2.3%.

User Interface (Flask App)

• Clinicians can upload fundus images and receive:

  • Classification results

  • Segmentation overlays

  • CDR values

  • Downloadable diagnostic reports

• Interface supports clinician feedback for future model refinement.

Methodology

• Data Collection: Annotated fundus images labeled for classification and segmentation.

• Preprocessing: Image resizing, normalization, and augmentation.

• CDR Calculation: Both area-based and diameter-based methods used, with a threshold (e.g., 0.6) for glaucoma risk assessment.

• Deployment Optimization: Includes model compression and GPU acceleration for real-time processing in resource-constrained settings.

Literature Survey Highlights

• Studies show benefits of combining classification and segmentation.

• Various deep learning models (e.g., Swin-Unet, MobileNet, RD-Net) have been explored.

• Self-attention, efficient CNNs, and hybrid approaches improve accuracy but face trade-offs in computation or generalizability.

Results

• Classification: High accuracy and sensitivity validate early glaucoma detection.

• Segmentation: Precise optic structure identification ensures reliable CDR calculations.

• System Reliability: Low deviation in CDR from expert ground truth highlights clinical readiness.

Conclusion

The two stage deep learning framework designed for automated glaucoma detection using a combination of ResNet and YOLOv5 for the classification stage and optic disc and cup segmentation stage is presented in this study. This system aims to make the diagnosis of glaucoma more accessible and also enable early and fast detection by classification, segmentation and cup to disc ratio (CDR) calculation. High sensitivity and specificity in classifying were shown in the ResNet model for screening purposes, and the YOLOv5 model accurately segmented the optic disc and cup regions, critically needed for accurate CDR measurement. Subsequent analysis demonstrates that the system can provide an objective diagnostic metric for risk assessment in glaucoma with an average deviation from expert CDR values of only 2.3 percent, mitigating both subjectivity and variability of manual assessment. The clinical applicability of the framework is further enhanced by the integration of a user-friendly interface that will allow ophthalmologists to upload images, view results, and make speedy decisions. This high accuracy, low cost and low logistics framework offers a valuable tool for the glaucoma diagnosis and particularly in high demand in less resource rich settings.

References

[1] B. S. Abraham, D. G. Bhanu, A. Thomas and J. Anitha, \"A Deep Dive into Optic Disc and Cup Segmentation for Glaucoma Diagnosis with Deep Learning,\" 2023 International Conference on Emerging Research in Computational Science (ICERCS), Coimbatore, India, 2023, pp. 1-6, doi: 10.1109/ICERCS57948.2023.10434051. [2] R. Krishnan, V. Sekhar, J. Sidharth, S. Gautham and G. Gopakumar, \"Glaucoma Detection from Retinal Fundus Images,\" 2020 International Conference on Communication and Signal Processing (ICCSP), Chennai, India, 2020, pp. 0628-0631, doi: 10.1109/ICCSP48568.2020.9182388. [3] F. Z. Zulfira and S. Suyanto, \"Detection of Multi-Class Glaucoma Using Active Contour Snakes and Support Vector Machine,\" 2020 3rd International Seminar on Research of Information Technology and Intelligent Systems (ISRITI), Yogyakarta, Indonesia, 2020, pp. 650-654, doi: 10.1109/ISRITI51436.2020.9315372. [4] J. K. Dewa, E. Rachmawati and G. Kosala, \"Investigating Self-Attention in Swin-Unet Model for Disc and Cup Segmentation,\" 2023 10th International Conference on Information Technology, Computer, and Electrical Engineering (ICITACEE), Semarang, Indonesia, 2023, pp. 401-406, doi: 10.1109/ICITACEE58587.2023.10276855. [5] Y. Yang, G. Yang, D. Ding and J. Zhao, \"Contour Offset Map: A New Component Designed for Smooth and Robust Optic Disc/Cup Contour Detection,\" 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Las Vegas, NV, USA, 2022, pp. 1773-1776, doi: 10.1109/BIBM55620.2022.9995691 [6] D. I. C. Wiguna, E. Rachmawati and G. Kosala, \"Glaucoma Detection Based on Joint Optic Disc and Cup Segmentation Using Dense Prediction Transformer,\" 2023 10th International Conference on Information Technology, Computer, and Electrical Engineering (ICITACEE), Semarang, Indonesia, 2023, pp. 407-412, doi: 10.1109/ICITACEE58587.2023.10277656. [7] G. -R. Huang and T. -R. Hsiang, \"A Simplified Deep Network Architecture on Optic Cup and Disc Segmentation,\" 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1-7, doi: 10.1109/IJCNN48605.2020.9206670. [8] I. D. S. Ubaidah, Y. Fu’Adah, S. Sa’Idah, R. Magdalena, A. B. Wiratama and R. B. J. Simanjuntak, \"Classification of Glaucoma in Fundus Images Using Convolutional Neural Network with MobileNet Architecture,\" 2022 1st International Conference on Information System & Information Technology (ICISIT), Yogyakarta, Indonesia, 2022, pp. 198-203, doi: 10.1109/ICISIT54091.2022.9872945. [9] A. Septiarini, H. Hamdani, E. Setyaningsih, E. Arisandy, S. Suyanto and E. Winarno, \"Automatic Segmentation of Optic Nerve Head by Median Filtering and Clustering Approach,\" 2021 13th International Conference on Information & Communication Technology and System (ICTS), Surabaya, Indonesia, 2021, pp. 118-122, doi: 10.1109/ICTS52701.2021.9608854. [10] Preity, A. K. Bhandari, A. Jha and S. Shahnawazuddin, \"RD-Net: Residual-Dense Network for Glaucoma Prediction Using Structural Features of Optic Nerve Head,\" in IEEE Transactions on Artificial Intelligence, doi: 10.1109/TAI.2024.3447578.

Copyright

Copyright © 2025 Sarthak Mudaliar, Adibhatla Srinivasa Chakravarthi , Deebalakshmi R . 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.