Benchmarking CNN Architectures for Multi-Class Retinal Disease Classification

This repository contains the implementation and benchmarking of six deep learning architectures for automated classification of Normal, Diabetic Retinopathy, Cataract, and Glaucoma from retinal fundus images.

The project performs a detailed comparison of multiple CNN models under identical preprocessing, augmentation, and training pipelines to identify the most efficient and accurate architecture suitable for real-world clinical deployment.

Project Structure

• .ipynb_checkpoints/
• dataset/
  • dataset.zip # Original compressed dataset (~753 MB)
• eye-diseases-classification-DenseNet201 Final.ipynb
• eye-diseases-classification-EffNetb3 Final.ipynb
• eye-diseases-classification-IncResNetV2 Final.ipynb
• eye-diseases-classification-MobileNetv3L Final.ipynb
• eye-diseases-classification-ResNet50v2 Final.ipynb
• eye-diseases-classification-Xception Final.ipynb
• README.md

Notebook Descriptions

Notebook	Model
MobileNetv3L Final.ipynb	MobileNetV3-Large
EffNetb3 Final.ipynb	EfficientNetB3
DenseNet201 Final.ipynb	DenseNet201
Xception Final.ipynb	Xception
ResNet50v2 Final.ipynb	ResNet50V2
IncResNetV2 Final.ipynb	InceptionResNetV2

Each notebook includes the full pipeline:

• Importing dataset
• Preprocessing
• Data augmentation
• Transfer learning
• Model training
• Confusion matrix & metrics
• Grad-CAM heatmaps
• Performance summary

Dataset

• Total images: 4,217
• Classes:
  • Normal
  • Diabetic Retinopathy
  • Cataract
  • Glaucoma
• Format: .jpg / .png

Models Compared

This project benchmarks:

• MobileNetV3-Large
• EfficientNetB3
• DenseNet201
• Xception
• ResNet50V2
• InceptionResNetV2

Common configuration:

• Transfer learning (ImageNet)
• Image size: 256 × 256
• Optimizer: Adam (lr=0.0001)
• Loss: Categorical Cross-Entropy + L2 regularization
• Epochs: 50 (with Early Stopping)
• Batch size: 32
• Validation: 10%

Results

Overall Accuracy Comparison

Model	Accuracy
MobileNetV3-Large	0.8483
EfficientNetB3	0.7986
DenseNet201	0.7536
Xception	0.7109
ResNet50V2	0.6991
InceptionResNetV2	0.2607

Key Insight

Lightweight architectures (MobileNet) outperform heavy ones (InceptionResNetV2) in medical imaging due to reduced overfitting and better generalization.

Model Explainability

Each model includes:

Grad-CAM Visualizations

Saliency Maps

These highlight clinically relevant regions (optic disc, hemorrhages, microaneurysms), improving transparency for medical use.

Conclusion

MobileNetV3-Large emerges as the most practical architecture for real-world retinal disease screening:

Highest accuracy

Lowest computational cost

Fast inference

Suitable for edge devices & low-resource clinics

This demonstrates that efficient model design > model size for medical imaging applications.

Authors – Team Oculus (Group 2)

Aashman Sarkar

Abhay Rathore

Arpan Mitra

Varun Mohanta

Jayasree Chakraborty

Sanjana Biswas

Supervisor: Dr. Ranjita Kumari Dash School of Computer Engineering, KIIT University