[Paper Review] Eye Disease Classification Using Deep Learning Techniques

Tareq Babaqi, Manar Jaradat+3 more|arXiv (Cornell University)|Jul 19, 2023

Retinal Imaging and Analysis19 citations

TL;DR

The paper compares CNN and transfer learning using EfficientNet for multi-class eye disease classification (normal, cataract, diabetic retinopathy, glaucoma), achieving 94% accuracy with transfer learning vs 84% with CNN.

ABSTRACT

Eye is the essential sense organ for vision function. Due to the fact that certain eye disorders might result in vision loss, it is essential to diagnose and treat eye diseases early on. By identifying common eye illnesses and performing an eye check, eye care providers can safeguard patients against vision loss or blindness. Convolutional neural networks (CNN) and transfer learning were employed in this study to discriminate between a normal eye and one with diabetic retinopathy, cataract, or glaucoma disease. Using transfer learning for multi-class classification, high accuracy was achieved at 94% while the traditional CNN achieved 84% rate.

Motivation & Objective

• Motivate early detection of eye diseases due to vision loss risks and limited access to ophthalmologists.
• Investigate whether transfer learning improves CNN-based eye disease classification.
• Evaluate performance using standard metrics on a dataset of four classes (normal, cataract, diabetic retinopathy, glaucoma).
• Utilize EfficientNet as the pre-trained model for transfer learning to classify eye diseases.

Proposed method

• Use CNN and transfer learning with a pre-trained EfficientNet architecture.
• Rescale images from 512x512x3 to 224x224x3 to improve generalization.
• Split data into 70% training, 20% testing, 10% validation.
• Evaluate using precision, recall, F1-score, and accuracy.
• Compare training behavior via accuracy/loss plots and confusion matrices.

Experimental results

Research questions

• RQ1 Does transfer learning improve CNN accuracy for multi-class eye disease classification?
• RQ2 What are the class-wise performance differences between CNN and transfer learning on normal, cataract, diabetic retinopathy, and glaucoma?
• RQ3 Is EfficientNet-based transfer learning effective for automatic eye disease screening from color fundus images.

Key findings

• Transfer learning achieved higher accuracy (94%) than CNN (84%).
• Diabetic retinopathy shows very high performance with both methods (CNN: 100% precision/recall/F1; TL: 100% precision, 99% recall, 99% F1).
• For Normal class, TL achieved 86% precision, 96% recall, 91% F1; CNN achieved 86% precision, 56% recall, 68% F1.
• Cataract class: TL (97% precision, 96% recall, 96% F1); CNN (87% precision, 96% recall, 91% F1).
• Glaucoma class: TL (94% precision, 85% recall, 89% F1); CNN (67% precision, 85% recall, 75% F1).
• Transfer learning produced smoother training accuracy/loss curves and clearer improvement in confusion matrices compared to CNN.