Diabetic Retinopathy Benchmark

Blindness detection from diabetic retinopathy fundus images using transfer learning CNN backbones and a classical machine-learning baseline.

Overview

This repository is the polished, GitHub-ready version of a final-year B.Tech Computer Science capstone project. It focuses on automated diabetic retinopathy grading from retinal fundus photographs and packages the work as:

• a configurable CNN training pipeline
• a classical ML baseline for comparison
• evaluation artifacts for reproducibility
• a GitHub Pages UI in docs/ for presentation

The project no longer depends on the old Kaggle-only directory layout and works from a clean local structure.

GitHub Pages UI

Live at - https://vermashaurya.github.io/CNN-Health

The repository includes a static presentation site in docs/ designed for GitHub Pages. It explains:

• project motivation and pipeline
• backbone comparison
• result surfaces and evaluation outputs
• implementation details
• future scope

Once the repo is on GitHub, enable Pages from the docs/ folder and the UI becomes the public-facing project page.

Pipeline

flowchart LR
    A["Fundus Retinography Images"] --> B["Preprocessing<br/>resize, rescale, augmentation"]
    B --> C["CNN Training<br/>EfficientNetB0 / ResNet152V2 / NASNetLarge"]
    B --> D["ML Baseline<br/>image statistics + Random Forest"]
    C --> E["Predicted DR Severity"]
    D --> F["Baseline Comparison"]
    E --> G["Evaluation<br/>accuracy, confusion matrix, report"]
    F --> G

Loading

Repository Structure

diabetic-retinopathy-benchmark/
├── docs/                     # GitHub Pages UI
├── data/
│   └── aptos2019-blindness-detection/
│       ├── train.csv
│       └── train_images/
├── main.py                   # Final CLI entrypoint
├── efficientnet.py           # Backbone shortcut
├── resnet152v2.py            # Backbone shortcut
├── nasnetlarge.py            # Backbone shortcut
├── diagrams.py               # Diagram generator
├── requirements.txt
└── README.md

Dataset Layout

Expected local structure:

data/
└── aptos2019-blindness-detection/
    ├── train.csv
    └── train_images/
        ├── 000c1434d8d7.png
        └── ...

If your dataset lives elsewhere, override the default paths with CLI flags.

Installation

pip install -r requirements.txt

Usage

Smoke test:

python3 main.py train --backbone efficientnetb0 --smoke-test

Final training run:

python3 main.py train --backbone efficientnetb0 --epochs 30

Classical ML baseline:

python3 main.py baseline

Backbone-specific shortcuts:

python3 efficientnet.py
python3 resnet152v2.py
python3 nasnetlarge.py

Model Comparison

Model	Input Size	Role in Repo	Characteristics
EfficientNetB0	256x256	Recommended default	Good balance of accuracy and efficiency
ResNet152V2	224x224	Deep residual benchmark	Strong feature extraction with higher depth
NASNetLarge	331x331	High-capacity benchmark	Larger model with heavier compute cost
RandomForestClassifier	Handcrafted features	Classical ML baseline	Useful non-deep-learning comparison

Result Artifacts

The training pipeline exports:

• best_model.keras
• final_model.keras
• training_log.csv
• history.csv
• evaluation.json
• run_summary.json

The baseline writes its own outputs to artifacts/baseline/.

This repository intentionally does not fabricate performance numbers. Once you run your final training locally, the generated artifacts become the source of truth for reported metrics.

Resume Summary

Built a diabetic retinopathy grading pipeline on retinal fundus images using transfer learning CNN backbones (EfficientNetB0, ResNet152V2, NASNetLarge) and a classical ML baseline, with configurable training, experiment logging, evaluation exports, and a GitHub Pages showcase UI.

License

This project is licensed under the MIT License

Copyright (c) 2026



Feel free to take inspiration.
Happy Coding !