[Paper Review] Deep Learning based Framework for Automatic Diagnosis of Glaucoma based on analysis of Focal Notching in the Optic Nerve Head

Sneha Dasgupta, Rishav Mukherjee+2 more|arXiv (Cornell University)|Dec 10, 2021

Retinal Imaging and Analysis4 citations

TL;DR

This paper proposes a deep learning framework for automatic glaucoma diagnosis using retinal fundus images, combining multi-class U-Net for optic disc and cup segmentation with SVM-based classification using cup-to-disc ratio and focal notching features. The method achieved 93.33% accuracy on the DRISHTI-GS dataset, improving detection by integrating ISNT rule-based rim thickness analysis to address limitations of CDR alone.

ABSTRACT

Automatic evaluation of the retinal fundus image is emerging as one of the most important tools for early detection and treatment of progressive eye diseases like Glaucoma. Glaucoma results to a progressive degeneration of vision and is characterized by the deformation of the shape of optic cup and the degeneration of the blood vessels resulting in the formation of a notch along the neuroretinal rim. In this paper, we propose a deep learning-based pipeline for automatic segmentation of optic disc (OD) and optic cup (OC) regions from Digital Fundus Images (DFIs), thereby extracting distinct features necessary for prediction of Glaucoma. This methodology has utilized focal notch analysis of neuroretinal rim along with cup-to-disc ratio values as classifying parameters to enhance the accuracy of Computer-aided design (CAD) systems in analyzing glaucoma. Support Vector-based Machine Learning algorithm is used for classification, which classifies DFIs as Glaucomatous or Normal based on the extracted features. The proposed pipeline was evaluated on the freely available DRISHTI-GS dataset with a resultant accuracy of 93.33% for detecting Glaucoma from DFIs.

Motivation & Objective

• To improve early detection of glaucoma by automating the analysis of retinal fundus images using deep learning.
• To address the limitation of cup-to-disc ratio (CDR) in cases of naturally large optic cups by incorporating focal notching and ISNT rule-based neuroretinal rim thickness.
• To develop a joint segmentation pipeline for optic disc and optic cup to extract clinically relevant features for glaucoma classification.
• To enhance the accuracy of computer-aided diagnosis (CAD) systems by combining morphological features with machine learning classification.
• To validate the framework on a publicly available dataset (DRISHTI-GS) for reproducibility and benchmarking.

Proposed method

• A multi-class U-Net architecture with encoder-decoder structure is used for semantic segmentation of optic disc (OD) and optic cup (OC) regions in retinal fundus images.
• The network is trained using the Adam optimizer with a binary cross-entropy loss function on resized 256×256 pixel images.
• Focal notching features are extracted by analyzing the neuroretinal rim thickness in the four ISNT sectors (Inferior, Superior, Nasal, Temporal) to detect asymmetry indicative of glaucoma.
• Eight quantitative features are extracted: area CDR, diameter CDR, cup diameter, S-distance, I-distance, disc diameter, cup area, and disc area.
• Support Vector Machine (SVM) is employed as the classifier, using the extracted features to distinguish between glaucomatous and normal cases.
• Data augmentation and validation set monitoring are applied to prevent overfitting during training.

Experimental results

Research questions

• RQ1 Can a deep learning-based segmentation model achieve high accuracy in simultaneously segmenting optic disc and optic cup from retinal fundus images?
• RQ2 Does incorporating focal notching features based on the ISNT rule improve glaucoma classification accuracy beyond cup-to-disc ratio alone?
• RQ3 Can a hybrid pipeline of U-Net segmentation and SVM classification outperform traditional CAD systems in glaucoma detection?
• RQ4 How does the proposed method perform on the DRISHTI-GS dataset in terms of segmentation and classification metrics?
• RQ5 To what extent do morphological features like cup diameter and rim thickness contribute to distinguishing glaucomatous from normal eyes?

Key findings

• The proposed multi-class U-Net achieved 99.697% accuracy, 98.103% recall, and 92.654% precision for optic disc segmentation on the DRISHTI-GS test set.
• For optic cup segmentation, the model achieved 99.513% accuracy, 91.459% recall, and 89.498% precision, demonstrating high segmentation fidelity.
• The SVM-based classification model achieved 93.33% overall accuracy, with 90.9% sensitivity and 94.73% specificity on the test dataset.
• The integration of ISNT rule-based notching features significantly improved classification robustness, especially in cases with high baseline CDR due to anatomical variation.
• The F1-score for optic cup segmentation was 0.89728, indicating a strong balance between precision and recall.
• The confusion matrix confirmed a low false positive rate (1/19) and low false negative rate (1/11), supporting reliable clinical decision support.