Eye Disease Detection Model

Abstract

According to the World Health Organization (WHO), an estimated 253 million people worldwide are visually impaired, with 36 million of them being blind, predominantly due to eye diseases. In the pursuit of early diagnosis and intervention for eye diseases, we present an advanced eye disease detection system harnessing the power of the ResNet (Residual Networks) architecture. Our research focuses on accurate, efficient, and accessible detection of eye diseases, addressing a critical global health concern. The system developed promises to aid eye disease diagnostics, particularly benefiting low-income and middle-income countries with limited healthcare resources and high healthcare costs. We also emphasise its role as a preliminary diagnostic tool.

Methodology Overview

We explored multiple models, including Inception v3, ResNet, and EfficientNet, and found that ResNet-50 outperformed the others, which led us to select it as our final model.

Data Preprocessing and Model Architecture

1. Data Source: Kaggle Dataset Link
2. Eye Diseases Detected: Cataract, Glaucoma, Diabetic Retinopathy
3. Split Ratio : 80% Training, 10% Testing, 10% Validation
4. Preprocessing:
   • Resizing Images to 224x224 Pixels ( As per Resnet requirements)
   • Normalization using Mean and Standard Deviation of the Dataset
   • De-noising with ldm-super-resolution-4x-openimages
5. Transfer Learning with Resnet-50, with additional custom layers to enhance its performance.

Results & Performance Metrics

Performance Metric	Training Data	Validation Data	Test Data
Accuracy	0.92899	0.8975	0.8825
Precision	0.93245	0.89370	0.88347
Recall	0.92899	0.8975	0.8825
F1-Score	0.92783	0.89422	0.88218

Visualisations

Eye Disease Image Analysis

Original Image Vs Normalised Vs Grad-CAM

Conclusions

The ResNet-50 architecture was chosen and fine-tuned to achieve promising results. Below are the key findings and takeaways from our work.

• Model Performance: The trained model demonstrated competitive performance in classifying eye diseases, achieving a high accuracy rate on the test dataset.
• Streamlit Application: Model was successfully deployed using a user-friendly Streamlit web application.

Takeaways:

• High-quality fundus images are needed to accurately detect if eye diseases like cataracts and glaucoma are present.
• Deployment: Denoising the fundus images helps to improve the prediction score.