A Hybrid CNN-GNN Framework for Enhanced Glaucoma Detection

Using Retinal Fundus Images

S Hemalatha, S R Khrisha, K S Praneetheaswarr and M Sibi Logesh

Department of CSD, Kongu Engineering College, Perundurai, Tamilnadu, India

Keywords: Glaucoma Detection, Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), Medical

Image Classification, ORIGA Dataset, Feature Extraction, Graph Construction, Ensemble Learning, Precision

and Recall, Deep Learning Framework.

Abstract: Convolutional neural networks are very powerful tools used for the analysis of structured Euclidean space

data. However, it is on applications like image classification and audio analysis, language processing among

others. These models, therefore, can effectively retrieve critical features necessary for deciding. However,

many real-life problems involve data organized along non-Euclidean geometries, including social networks

and medical imaging, genetic studies among others. In these contexts, relationships of data points become

important. Particularly in medical image classification, using semantic relationships of features within images

improves detection accuracy of complex diseases considerably. Graph Neural Networks are proficient at

modeling relational data in such cases using graph structures. This paper presents a novel deep learning

framework that combines CNNs and GNNs, combining their complementary strengths. It utilizes CNNs for

feature extraction and GNNs for modeling relationships between features to offer a robust approach to

glaucoma detection from retinal fundus images. Using the ORIGA dataset, we designed a three-component

architecture: Feature Extractor, Graph Constructor, and Graph Classifier. Our experiments explored several

techniques of graph construction and similarity-measuring techniques, which lead to better classification

performance. Our proposed CNN-GNN ensemble was able to reach precision at 0.79, recall at 0.76, and F1

score at 0.77, outperforming previous approaches.

1 INTRODUCTION

Glaucoma is a progressive optic neuropathy this is

characterised by way of degeneration of retinal

ganglion cells and irreversible loss of vision if left

undetected and untreated. Being one of the important

reasons of blindness worldwide, the need for well

timed and accurate prognosis cannot be overstated

which will have right control and prevention.

Traditional techniques for diagnosing the situation

involve measurements of intraocular pressure and

sight view tests that can be subjective in nature and

regularly do now not indicate glaucoma till vast

damage has took place. (Kipf, and, Welling, 2017),

(He, Zhang, et al. , 2016)

Recent advances in system getting to know and

deep learning have revolutionized the landscape of

scientific photo evaluation, especially in

ophthalmology. Convolutional Neural Networks

(CNNs) have emerged as effective equipment for

extracting capabilities from medical pics, especially

in structured statistics domain names along with

image class. However, because of the inherent

boundaries of CNNs in shooting complex

relationships among functions, their overall

performance is frequently hindered while the

underlying data geometry isn't Euclidean, including

inside the case of scientific diagnostics. But GNNs

offer a viable alternative which promises to represent

the relational records amongst information points in

graph systems extra informatively. Indeed, it is rather

pertinent for applications of clinical picture category

responsibilities considering that semantic institutions

of functions are very crucial insights in the pathology.

Thus, incorporating GNNs with CNNs exploits the

strengths of both methodologies. The use of CNNs

extracts sturdy functions at the same time as that of

GNNs is about shooting complex relationships

between capabilities. This paper proposes a novel

deep learning architecture combining synergistic

598

Hemalatha, S., Khrisha, S. R., Praneetheaswarr, K. S. and Sibi Logesh, M.

A Hybrid CNN-GNN Framework for Enhanced Glaucoma Detection Using Retinal Fundus Images.

DOI: 10.5220/0013632900004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futur istic Technology (INCOFT 2025) - Volume 3, pages 598-602

ISBN: 978-989-758-763-4

CNNs and GNNs to enhance the diagnosis of

glaucoma in images of the retinal fundus. Based on

the publicly available ORIGA dataset, we devise a

three-module architecture involving Feature

Extractor, Graph Constructor, and Graph Classifier

modules, then perform extensive experiments

involving several techniques for graph construction,

together with various similarity measures in order to

achieve optimum classification performance. Our

results show that the ensemble approach proposed by

us delivers promising results and competitive

precision, recall, and F1 scores. (Wang, Li, et al. ,

2021), (Wu, Pan, et al. , 2020), (Zhang, Ning, et al. ,

2020)

2 DESIGN AND PRINCIPLE OF

OPERATION

2.1 Designing the CNN-GNN

Framework for Glaucoma

Detection

The proposed CNN-GNN framework was designed to

leverage the complementary strengths of

Convolutional Neural Networks (CNNs) and Graph

Neural Networks (GNNs) for glaucoma detection

from retinal fundus images. The architecture

comprises three primary modules: the Feature

Extractor, Graph Constructor, and Graph Classifier.

The Feature Extractor employs a fine-tuned ResNet-

18 model to extract robust feature embeddings from

retinal fundus images, effectively capturing spatial

patterns such as the optic disc shape and retinal layer

thickness. These embeddings are then used to

construct graph representations in the Graph

Constructor module. This module supports two types

of graph structures: sparse graphs, which retain

significant relationships between features by utilizing

similarity measures like cosine similarity or

correlation, and complete graphs, which include all

possible node connections. The graph construction

process ensures that relational information is

preserved while maintaining computational

efficiency through sparse adjacency matrices.

(Kingma, Ba, et al. , 2014), (Paszke, Gross, et al. ,

2019)

2.2 Operation of Graph Neural

Networks (GNNs) for Classification

Once the graphs are constructed, the Graph Classifier

employs a Graph Convolutional Network (GCN) to

process these graphs. The GCN aggregates

information from neighboring nodes through iterative

convolutional operations, effectively capturing local

and global relational patterns within the data. The

final graph-level embeddings are passed through fully

connected layers and a softmax function to classify

the retinal images into glaucoma or non-glaucoma

categories. To further enhance performance, an

ensemble learning approach combines predictions

from multiple models: a baseline CNN, a sparse

graph-based GNN classifier, and a complete graph-

based GNN classifier. The ensemble aggregates

predictions through majority voting or weighted

averaging, allowing the framework to capture both

local feature information and inter-feature

relationships comprehensively. This design enables

precise and robust glaucoma detection,

outperforming conventional approaches.

The simulation results validate the effectiveness

of the proposed hybrid CNN-GNN framework for

glaucoma detection. Key performance metrics, such

as accuracy, precision, recall, F1 score, and

specificity, were analyzed to assess the model’s

performance. (Deng, Dong, et al. , 2009), (Zhou, Hao,

et al. , 2020)

3 SIMULATION RESULTS AND

DISCUSSION

The simulation results validate the effectiveness of

the proposed hybrid CNN-GNN framework for

glaucoma detection. Key performance metrics, such

as accuracy, precision, recall, F1 score, and

specificity, were analyzed to assess the model’s

performance.

3.1 CNN Accuracy Analysis

The training accuracy improved steadily, eventually

reaching near-perfect values. However, the testing

accuracy stabilized at approximately 0.9 with minor

fluctuations, indicating potential overfitting. These

fluctuations highlight the need for regularization

techniques such as dropout or early stopping to

enhance the model's generalization to unseen data.

A Hybrid CNN-GNN Framework for Enhanced Glaucoma Detection Using Retinal Fundus Images

599

Figure 1: CNN Accuracy Analysis

3.2 GNN Accuracy Analysis

During the initial 50 epochs, the training loss

consistently decreased, while the testing loss

plateaued with minor oscillations. Although the

training accuracy reached 0.9, the testing accuracy

lagged behind at approximately 0.8, indicating

overfitting. Regularization techniques and early

stopping could reduce these fluctuations, thereby

improving generalization.

Figure 2: GNN Accuracy Analysis.

3.3 Ensemble Model Performance

The ensemble approach combined the strengths of

CNNs and GNNs to achieve improved performance.

Figure 3 depicts the accuracy variations with different

Figure 3: Ensemble Model Performance

numbers of estimators. As the number of estimators

increased from 2 to 6, the accuracy improved

significantly, reaching around 0.91. Beyond six

estimators, accuracy gains diminished, stabilizing

with minor fluctuations. The optimal performance

was observed with 12–14 estimators, maintaining an

accuracy of approximately 0.91.

3.4 Comparative Analysis

It compares the performance metrics of CNN, GNN,

and ensemble models. While CNNs provided a strong

foundation for feature extraction, the GNNs captured

relational information more effectively. The

ensemble model outperformed the individual models,

achieving the best overall performance.

Figure 4: Comparative Analysis

3.5 Summary of Key Metrics

The hybrid framework demonstrated superior

accuracy (95.48%), sensitivity (97.30%), specificity

(94.52%), and AUC (97%), with an F1 score of 97%.

These results highlight the robustness of the ensemble

approach in detecting glaucoma.

The simulation results confirm that combining

spatial and relational feature extraction is highly

effective. While the CNN model excels in extracting

localized features, the GNN enhances performance

by capturing feature interdependencies. This

complementary relationship allows the ensemble

model to outperform standalone CNN and GNN

implementations.

Further improvements, such as advanced

regularization, alternative graph architectures, and

dynamic graph construction techniques, can enhance

the model's generalization and adaptability across

diverse datasets.

INCOFT 2025 - International Conference on Futuristic Technology

600

4 CONCLUSIONS

We proposed a hybrid framework combining CNNs

and GNNs for glaucoma detection from retinal

fundus photographs. The ensemble technique, which

used CNNs for characteristic extraction and GNNs

for modeling relational information among features,

outperformed CNN-handiest fashions. In specific, the

ensemble received a precision of zero.79, a consider

of 0.76, and an F1 score of 0.77 on the ORIGA

dataset, surpassing contemporary methods. This

research opens up the opportunity of graph-based

totally techniques in medical image analysis with the

ability for massive improvement in glaucoma

detection. Future paintings may be in increasing this

method to different medical conditions and imaging

duties by exploring other graph structures.

ACKNOWLEDGEMENTS

The author extends sincere thanks to the mentor for

their valuable contributions in discussing the results

and providing feedback.

REFERENCES

Kipf, T. N., & Welling, M. (2017). Semi-Supervised

Classification with Graph Convolutional Networks. In

International Conference on Learning Representations

(ICLR). Retrieved from

https://arxiv.org/abs/1609.02907

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual

Learning for Image Recognition. In Proceedings of the

IEEE Conference on Computer Vision and Pattern

Recognition (CVPR) (pp. 770-778).

Wang, L., Li, Z., & Zhang, X. (2021). A Deep Learning-

Based Framework for Glaucoma Detection Using

Retinal Fundus Images. IEEE Transactions on Medical

Imaging, 40(6), 1541-1552.

https://doi.org/10.1109/TMI.2021.3053295

Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Philip, S.

Y. (2020). A Comprehensive Survey on Graph Neural

Networks. IEEE Transactions on Neural Networks and

Learning Systems, 32(1), 4-24.

https://doi.org/10.1109/TNNLS.2020.297838

Zhang, X., Ning, L., Yang, J., Chen, W., & Li, H. (2020).

Graph-Based Medical Diagnosis Using GCN:

Application in Glaucoma Detection. In Proceedings of

the AAAI Conference on Artificial Intelligence (Vol.

34, No. 07, pp. 7952-7960).

https://doi.org/10.1609/aaai.v34i07.6846

Kingma, D. P., & Ba, J. (2014). Adam: A Method for

Stochastic Optimization. In Proceedings of the

International Conference on Learning Representations

(ICLR). Retrieved from

https://arxiv.org/abs/1412.6980

Fauw, J. D., Ledsam, J. R., Romera-Paredes, B., Nikolov,

S., Tomasev, N., Blackwell, S., ... & Mohamed, S.

(2018). Clinically Applicable Deep Learning for

Diagnosis and Referral in Retinal Disease. Nature

Medicine, 24(9), 1342-1350.

https://doi.org/10.1038/s41591-018-0107-6

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., ... & Chintala, S. (2019). PyTorch: An

Imperative Style, High-Performance Deep Learning

Library. In Advances in Neural Information Processing

Systems (pp. 8024-8035).

Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P.,

Danesh, J., ... & Collins, R. (2015). UK Biobank: An

Open Access Resource for Identifying the Causes of a

Wide Range of Complex Diseases of Middle and Old

Age. PLOS Medicine, 12(3), e1001779.

https://doi.org/10.1371/journal.pmed.1001779

Deng, J., Dong, W., Socher, R., Li, L. J., Li, K., & Fei-Fei,

L. (2009). ImageNet: A Large-Scale Hierarchical

Image Database. In Proceedings of the IEEE

Conference on Computer Vision and Pattern

Recognition (CVPR) (pp. 248-255)

Zhou, C., Li, H., Hao, X., & Xu, X. (2020). Graph

Convolutional Network-Based Glaucoma Detection

With Multi-View Features. IEEE Access, 8, 92390-

92398. ttps://doi.org/10.1109/ACCESS.2020.2994419

Xu, M., & Cheng, F. (2021). Medical Image Analysis by

Convolutional Neural Networks with Dynamic Graphs.

IEEE Journal of Biomedical and Health Informatics,

25(2), 501-509.

https://doi.org/10.1109/JBHI.2020.2988444

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., & Batra, D. (2020). Grad-CAM: Visual

Explanations from Deep Networks via Gradient-Based

Localization. International Journal of Computer Vision,

128(2), 336-359. https://doi.org/10.1007/s11263-019-

01228-7

Liu, X., Faes, L., Kale, A. U., Wagner, S. K., Fu, D. J.,

Bruynseels, A., ... & Topol, E. J. (2019). A Comparison

of Deep Learning Performance Against Health-Care

Professionals in Detecting Diseases From Medical

Imaging: A Systematic Review and Meta-Analysis.

The Lancet Digital Health, 1(6), e271-e297.

https://doi.org/10.1016/S2589-7500(19)30123-2

Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V.,

& Leskovec, J. (2020). Open Graph Benchmark:

Datasets for Machine Learning on Graphs. In Advances

in Neural Information Processing Systems (Vol. 33, pp.

22118-22133).

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net:

Convolutional Networks for Biomedical Image

Segmentation. In Medical Image Computing and

Computer-Assisted Intervention (MICCAI) (pp. 234-

241). https://doi.org/10.1007/978-3-319-24574-4_28

Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking

Model Scaling for Convolutional Neural Networks. In

Proceedings of the 36th International Conference on

A Hybrid CNN-GNN Framework for Enhanced Glaucoma Detection Using Retinal Fundus Images

601

Machine Learning (ICML) (pp. 6105-6114).

https://doi.org/10.48550/arXiv.1905.11946

Defferrard, M., Bresson, X., & Vandergheynst, P. (2016).

Convolutional Neural Networks on Graphs With Fast

Localized Spectral Filtering. In Advances in Neural

Information Processing Systems (pp. 3844-3852).

Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M.,

Blau, H. M., & Thrun, S. (2017). Dermatologist-Level

Classification of Skin Cancer With Deep Neural

Networks. Nature, 542(7639), 115-118.

https://doi.org/10.1038/nature21056

Zhang, Y., Lin, Q., Gan, Y., Zhou, Y., & Li, X. (2020).

Knowledge-Guided Convolutional Neural Network for

Glaucoma Diagnosis. IEEE Access, 8, 55675-55684.

https://doi.org/10.1109/ACCESS.2020.2978859

INCOFT 2025 - International Conference on Futuristic Technology

602