Multi-Disease Detection Using Retinal Fundus Scans (MDDRFS)

Overview

This project focuses on multi-disease detection using retinal fundus scans and nerve segmentation through image processing and deep learning techniques. The main tasks include binary classification, multi-class classification, and nerve segmentation using different state-of-the-art deep learning architectures.

Key Features

• Binary Classification: Classifying retinal fundus images into disease or no disease.
• Multi-Class Classification: Detecting multiple diseases from retinal fundus images.
• Nerve Segmentation: Segmenting nerve structures from fundus images for further analysis.

Technologies Used

• TensorFlow, Keras, PyTorch for model building and experimentation.
• Python for scripting and data handling.
• Deep Learning Models like ResNet, EfficientNet, MobileNet, UNet, and more for classification and segmentation.

---

Directory Structure

├── Augmentations                   # Augmentation scripts for data preprocessing
├── Data Segregation and Loading    # Scripts for segregating and loading data
├── Disease Detection               # Notebooks and scripts for binary and multi-class classification
├── Experiments                     # Notebooks for model experimentation
├── Frontend                        # Frontend code for the web interface
├── Nerve Segmentation              # Nerve segmentation training notebooks
├── backend                         # Backend code for the web interface
├── .gitignore                      # Git ignore file for ignoring unnecessary files
├── LICENSE                         # License file (Apache 2.0)
├── requirements.txt                # requirements file     
└── README.md                       # This readme file

---

Data Reference

The reference public data has been uploaded to the Final Year drive. The Retinal Fundus Images Dataset contains images for Multi Disease Classification and Nerve Segementation folder contains images, masks and lables for nerve segmentation. Additional proprietery data was used for training the model in addition to the publicly available datasets listed below.

Data Sources: Segmentation:

1. High-Resolution Fundus (HRF) Image Database
2. Drive Grand Challenege
3. PAPILA: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment

Classification:

1. Retinal Image Analysis for multi-Disease Detection Challenge
2. FIVES: A Fundus Image Dataset for Artificial Intelligence based Vessel Segmentation
3. Retinal fundus images for glaucoma analysis: the RIGA dataset

---

Models Used

We use a variety of state-of-the-art deep learning models for the classification and segmentation tasks. These include:

Pre-trained Models (Transfer Learning)

• ResNet152V2
• EfficientNetB7
• MobileNetV2
• DenseNet201
• XceptionNet
• EfficientNetV2 (S, M, L)

Segmentation Models

• UNet

---

Installation and Setup

1. Clone the repository:

   git clone https://github.com/yasho191/MDDRFS.git
   cd MDDRFS

2. Create a Python virtual environment and install the dependencies:

   python3 -m venv env
   source env/bin/activate
   pip install -r requirements.txt

3. Download and set up the datasets. Follow the instructions for dataset references mentioned above.

4. Train the models:

   • For Binary Classification, refer to the Disease Detection folder for training scripts.
   • For Multi-Class Classification, follow the same procedure in the Disease Detection folder.
   • For Nerve Segmentation, the training notebooks are in the Nerve Segmentation folder.

---

How to Use

Augmentation Tasks:

We have implemented a set of augmentations in the Augmentations folder. These include:

• Rotate
• Flip Horizontal/Vertical
• Center Crop
• Contrast/Brightness adjustments

Classification Tasks

• Binary Classification: You can perform either transfer learning or training from scratch for disease detection.
• Multi-Class Classification: Similarly, multi-class disease detection supports transfer learning and scratch training.

Nerve Segmentation

• UNet architectures has been employed for nerve segmentation tasks. The models can be trained using the notebooks provided in the Nerve Segmentation folder.

---

Results

• The best-performing models from classification tasks will be ensemble for better accuracy.
• Segmentation results can be visualized using built-in functions from the notebooks.
• Model weights will not be made public. You can finetune your models using the provided training scripts.

---

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.