This project presents an AI-powered deep learning pipeline that detects and classifies diabetic lesions (microaneurysms, hemorrhages, and exudates) in retinal fundus images to assist in early diagnosis of Diabetic Retinopathy (DR). Leveraging Vision Transformers and UNet, it provides high accuracy for both classification and lesion-level segmentation.
Manual diagnosis of DR via fundus images is time-consuming, expertise-dependent, and often unavailable in rural settings. We aim to develop a robust, scalable, and interpretable deep learning model that aids ophthalmologists in clinical and remote environments.
- Languages: Python, Dart
- Frameworks: PyTorch, TensorFlow, Keras, Flask, Flutter
- Models: Vision Transformer (ViT), UNet, EfficientNet, ResNet, YOLOv8
- Deployment Tools: Flask API, Flutter frontend
- APTOS 2019
- IDRiD
Augmentation expanded dataset size to 50,000+ samples for training.
- Classification using ViT (Severity Levels: 0–4)
- Segmentation using UNet (Lesion Types: MA, HEM, SE, EX)
- Backend: Python & Hugging face API for inference
- Frontend: Flask app for user upload & result display
| Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| ViT | 91% | 0.85 | 0.82 | 0.83 |
| ResNet50 | 84% | 0.82 | 0.77 | 0.79 |
| UNet (Lesions) | 74% IoU | Varies | Varies | Varies |
-
Due to the large size of the deep learning models used in this project (ViT, UNet, EfficientNet), they are not included in this repository directly.
-
👉 All pre-trained models, weights, and inference configurations are available on Hugging Face:
- Also Checkout the Generated report from "Images&Report/IDRiD_01_Report" which gives all the necessary data to the user, through which the user can consult the doctor
- Upload fundus image
- View DR severity and lesion map
- Lightweight and easy-to-use interface
- Real-time mobile deployment
- Offline inference capabilities
- Larger, diverse datasets for better generalization
- Kaif Nasim Tokare – [211755]
- Mohammad Aqeel Memon – [211721]
- Mohammed Irfan Siddiqui – [211751]
Supervisor: Prof. Faiz Rangari
College: M.H. Saboo Siddik College of Engineering
This project is licensed under the MIT License - see the LICENSE file for details.