Diabetic Retinopathy Detection using Color Fundus Photos

---

This repository contains the implementation of a project Diabetic Retinopathy Detection using Color Fundus Photos. It is the final project demonstration of CSCE 566 Data Mining course.

Dataset

The project used ODIR color fundus image dataset, stored in .npz format, which is preprocessed for training, validation, and test.

Models

• Implemented a pipeline with multiple models:
  • VGG19
  • ResNet50
  • EfficientNetB0
• Pre-trained with ImageNet weights.
• Fine-tuned using various optimizers, loss functions, and hyperparameters.

DenseNet201 was initially included but removed due to poor performance.

Results

• VGG19: Achieved overall AUC of 0.98; 0.97 for males and 0.98 for females.
• ResNet50: Achieved overall AUC of 0.98, and 0.98 for both groups.
• EfficientNetB0: Achieved the highest overall AUC of 0.99 and 0.99 for both groups.

Run

You can run this project either cloning (not updated yet) or run the Detection Task.ipynb file.