Retinal Fundus Classification

Transfer Learning for Diabetic Retinopathy Grading: Comparing CNN and Vision Transformer Approaches

Built by Riya Shet | MSc Health Data Science, University of Birmingham

---

Overview

This project compares three deep learning approaches for grading diabetic retinopathy (DR) severity from retinal fundus photographs, using the APTOS 2019 Blindness Detection dataset (3,662 images, 5 severity levels).

Model	Approach	Quadratic Weighted Kappa	Accuracy
Baseline CNN	3-layer CNN, trained from scratch	0.4697	58.6%
ResNet-50	ImageNet pre-trained, fine-tuned	0.8877	78.6%
ViT-B/16	ImageNet pre-trained, fine-tuned	0.8839	81.4%

Transfer learning nearly doubled the baseline's kappa score. ResNet-50 achieved the highest kappa (0.89), while ViT-B/16 achieved the highest accuracy (81.4%). Both pre-trained models exceeded the 0.85 QWK threshold for "strong agreement".

Key Features

• Ben Graham preprocessing -- crop, resize, and colour-normalise fundus images to enhance vascular detail
• Three architectures -- Baseline CNN (from scratch), ResNet-50 (transfer learning), ViT-B/16 (transfer learning)
• Mixed precision (AMP) training -- ~2x speedup on T4 GPUs via automatic mixed precision
• Class-imbalance handling -- inverse-frequency weighted cross-entropy loss
• Grad-CAM explainability -- heat map visualisations showing where each model focuses its attention, with ViT-compatible reshape transform
• Multi-platform support -- auto-detects Google Colab, Kaggle Notebooks, and local environments
• Stratified train/val/test split (70/15/15) with data augmentation via albumentations

Dataset

The APTOS 2019 Blindness Detection dataset contains 3,662 retinal fundus images labelled by severity:

Label	Severity	Description
0	No DR	Healthy retina
1	Mild	Microaneurysms
2	Moderate	Hemorrhages and exudates
3	Severe	Widespread vascular damage
4	Proliferative	Neovascularisation (most dangerous)

To download: Create a Kaggle account, accept the competition rules, then download via the Kaggle API:

kaggle competitions download -c aptos2019-blindness-detection

Place train.csv and the train_images/ folder in the same directory as the notebook.

Quick Start

# Clone the repository
git clone https://github.com/riyashet-hds/retinal-fundus-classification.git
cd retinal-fundus-classification

# Install dependencies
pip install -r requirements.txt

# Download the APTOS dataset (requires Kaggle API)
kaggle competitions download -c aptos2019-blindness-detection
unzip aptos2019-blindness-detection.zip

# Open the notebook
jupyter notebook retinal_classification.ipynb

The notebook auto-detects Google Colab, Kaggle Notebooks, and local environments and adjusts paths accordingly.

Project Structure

retinal-fundus-classification/
├── retinal_classification.ipynb   # Full pipeline: preprocessing, training, evaluation, Grad-CAM
├── requirements.txt               # Python dependencies
├── LICENSE                        # MIT License
└── README.md

Results Highlights

• Transfer learning gap: Pre-trained models improved QWK by ~0.42 over the from-scratch baseline
• ResNet-50 vs ViT-B/16: ResNet-50 achieved the highest kappa (0.89 vs 0.88), while ViT-B/16 achieved the highest accuracy (81.4% vs 78.6%) -- different error patterns explain the divergence
• Hardest classes: Severe DR (F1 = 0.44--0.46) and Proliferative DR (F1 = 0.48--0.55) were most challenging due to rarity and visual ambiguity with adjacent classes
• Fastest convergence: ViT peaked at epoch 8 (QWK = 0.91) and triggered early stopping at epoch 13 due to overfitting
• Grad-CAM insights: ResNet produces single concentrated activation regions; ViT produces multi-focal distributed attention across multiple anatomical landmarks

References

1. Zhou, Y. et al. (2023). A Foundation Model for Generalizable Disease Detection from Retinal Images. Nature.
2. Oquab, M. et al. (2024). DINOv2: Learning Robust Visual Features without Supervision. TMLR.
3. Graham, B. (2015). Kaggle Diabetic Retinopathy Detection Competition, 1st Place Solution.
4. Selvaraju, R.R. et al. (2017). Grad-CAM: Visual Explanations from Deep Networks. ICCV 2017.
5. APTOS 2019 Blindness Detection. Kaggle Competition.

License

This project is licensed under the MIT License.