Fundus Image Disease Classification

This repository contains a Jupyter Notebook that implements a deep learning pipeline to classify retinal fundus images into four categories: Cataract, Diabetic Retinopathy, Glaucoma, and Normal.

Repository Structure

fundus-disease-classification/
├── notebook.ipynb         # Main Jupyter Notebook
├── requirements.txt       # Python dependencies
├── .gitignore             # Files to ignore in Git
└── split/                 # Dataset folder (download manually)
    ├── train/             # Training images
    ├── val/               # Validation images
    └── test/              # Test images

Dataset

1. Download the dataset from:

   https://www.kaggle.com/datasets/drskprabhakar/cataract-dr-normal-glaucoma-fundus-images-dataset

2. Unzip or extract so that the top-level folder is named split, containing train/, val/, and test/ subfolders.

3. Place the entire split/ folder at the root of this repository, alongside notebook.ipynb.

Setup

1. Clone this repository

   git clone https://github.com/Amaan-developpeur/fundus-disease-classification.git
   
   cd fundus-disease-classification

2. Create a virtual environment (optional but recommended)

   python -m venv venv
   source venv/bin/activate       # macOS/Linux
   venv\Scripts\activate        # Windows

3. Install dependencies

   pip install -r requirements.txt

4. Verify dataset placement Ensure the split/ folder (with train/, val/, test/) lives here:

   ls
   # you should see:
   # notebook.ipynb  requirements.txt  split/

Running the Notebook

• Launch Jupyter Lab or Notebook:

  jupyter notebook

• Open notebook.ipynb and run cells from top to bottom.

Notebook Outline

1. Step 1: Data Setup — Verify folder structure and list classes.
2. Step 2: Data Inspection — Image counts, sample viewing, blur & quality checks.
3. Step 3: Data Generators — Define ImageDataGenerator for train/val/test.
4. Step 4: Baseline CNN — Build & train a custom CNN from scratch.
5. Step 5: Transfer Learning — Feature extraction and fine-tuning with MobileNetV2.
6. Step 6: Logistic Regression Head — Frozen backbone + logistic classifier.
7. Step 7: Evaluation & Interpretation — Confusion matrices, classification reports, Grad-CAM.
8. Step 8: Conclusion & Future Work

Requirements

Contents of requirements.txt:

tensorflow>=2.10
keras
numpy
pandas
matplotlib
scikit-learn
opencv-python
Pillow

(Adjust versions as needed.)

Disclaimer

This notebook and the resulting model are for academic and research purposes only.This is a prototype model to assist the domain experts but not diagnostic tool.

---