🧠 Problem Statement Glaucoma is a chronic and irreversible eye disease that progressively damages the optic nerve, potentially leading to permanent vision loss if not detected early. Traditional diagnostic procedures require specialized expertise, are often time-intensive, and can be costly, limiting accessibility—especially in resource-constrained regions. With glaucoma being one of the leading causes of blindness globally, there is a pressing need for scalable, reliable, and automated diagnostic tools.

💡 Project Overview This project focuses on developing a deep learning-based system for glaucoma detection using fundus (retinal) images. By leveraging powerful computer vision techniques and transformer-based neural architectures, the solution aims to classify images as either Glaucoma Positive or Glaucoma Negative, providing a valuable tool for early and efficient screening.

🔧 Key Contributions Model Architecture: Utilized a ViT-B16 (Vision Transformer) model implemented with ViT-Keras for binary classification, allowing the model to capture global patterns in fundus images more effectively than traditional CNNs.

Image Preprocessing: Applied Contrast Limited Adaptive Histogram Equalization (CLAHE) and normalization techniques to enhance optic disc visibility and improve feature extraction.

Performance Evaluation: Achieved over 90% accuracy and validated the model using ROC-AUC, precision, recall, and F1 score to ensure robustness and clinical relevance.

🎯 Project Objectives High Accuracy: Maximize the model’s ability to correctly classify both positive and negative cases.

Low False Positives: Minimize incorrect identification of healthy eyes as glaucoma-affected.

Low False Negatives: Reduce the risk of missing actual glaucoma cases, which is critical for timely treatment.

🌍 Impact Once fully optimized and validated, this system can be integrated into clinical workflows or deployed in mobile/remote screening settings. It has the potential to support ophthalmologists, increase early diagnosis rates, and reduce the global burden of preventable blindness through fast, affordable, and accurate screening.

• numpy: Array manipulation and numerical operations
• pandas: Data manipulation and analysis
• scikit-learn: Machine learning utilities like metrics and preprocessing

• tensorflow: Deep learning framework; includes Keras for model building
• keras: Interface for the TensorFlow deep learning framework (optional if using TensorFlow 2.x)
• opencv-python: Image processing library (optional, for image transformations)
• Pillow: Image processing library, used with Keras for loading images

• matplotlib: Plotting and visualization of data and model metrics
• seaborn: Statistical data visualization (optional, enhances matplotlib)

You can install the modules individually with pip install, or add them to a requirements.txt file:

• pandas

• scikit-learn

• tensorflow

• opencv-python

• matplotlib

• scikit-learn

• seaborn

• numpy

  • Start by opening Google Colab.

• Ensure that you have access to a TPU by selecting Runtime > Change runtime type > Hardware accelerator > TPU.

• Clone the repository using the following command:

  !git clone https://github.com/CODEBRAKERBOYY/GlaucomaDetection.git

• Change into the cloned directory:

  %cd GlaucomaDetection

• Install the required modules by running:

  !pip install -r requirements.txt

• To train your model or make predictions, run the following command:

  !python main.py [arguments]

  Arguments:

  • train_model - Specify this if you want to train the model before inference.
  • existing - Use this after the train_model argument if you want to retrain an existing model.
  • make_predictions / None - This loads the existing model for inference.

Make sure you have the following modules for optimal performance:

• CUDA (if using a GPU): Ensure your system supports CUDA for faster training if you have a compatible NVIDIA GPU.
• Jupyter Notebook (optional): For interactive experimentation and visualization.

The dataset should be structured as follows:

DATASET
├── train/
│ ├── Glaucoma_Positive/
│ └── Glaucoma_Negative/
└── val/
├── Glaucoma_Positive/
└── Glaucoma_Negative/
└── test/
├── Glaucoma_Positive/
└── Glaucoma_Negative/

* While giving path as input always remember it consists only tow class one is Negative and other is Positive.

• Perform image preprocesssing if you think it's necessary otherwise skip it.

• Path for data images if you wants to retrain model:

• https://drive.google.com/drive/folders/1M89d5jKBInbhvmEC95zn51zD6A25HKbF?usp=share_link

  Accuracy plot