Deep Learning Project for Diabetic Retinopathy Detection

Diabetic retinopathy is a serious eye condition associated with diabetes, leading to damage in the retina. This project aims to provide automatic methods for the detection of diabetic retinopathy using deep learning techniques.

Project Overview

This repository contains code for training and evaluating deep learning models to classify and detect diabetic retinopathy from retinal images. We use the IDRID (Indian Diabetic Retinopathy Image Dataset) and implement various models like VGG, ResNet, and a transfer learning approach for this purpose.

Getting Started

Prerequisites

• Python 3
• TensorFlow
• Other necessary libraries and dependencies (listed in requirements.txt)

Preparing the Dataset

Convert the IDRID dataset into TFRecords for efficient input pipeline processing:

• under the folder ./input_pipeline/, run the following command to convert IDRID dataset into TFRecords python3 create_smaller_tfrecord.py.

Diabetic Retinopathy

• Binary classification:

  Training:

  • train vgg model:

    python3 main.py --mode train --model_name vgg --num_classes 2

  • train resnet model:

    python3 main.py --mode train --model_name resnet --num_classes 2

  • train transfer model:

    python3 main.py --mode train --model_name transfer --num_classes 2

  Test:

  • evaluate vgg model:

    python3 main.py --mode evaluate --model_name vgg --num_classes 2

  • evaluate resnet model:

    python3 main.py --mode evaluate --model_name resnet --num_classes 2

  • evaluate transfer model:

    python3 main.py --mode evaluate --model_name transfer --num_classes 2

  • evaluate ensemble model:

    python3 ensemble.py --num_classes 2

  Deep visualization:

  • visualization for vgg model:

    python3 deep_visualization.py --model_name vgg --num_classes 2

    visualization images are located under the folder: diabetic_retinopathy/checkpoint/vgg/visualization/

  • visualization for resnet model:

    python3 deep_visualization.py --model_name resnet --num_classes 2

    visualization images are located under the folder: diabetic_retinopathy/checkpoint/resnet/visualization/

• Multi-classes classification:

  Training:

  python3 main.py --mode train --model_name transfer --num_classes 5

  Test:

  python3 main.py --mode evaluate --model_name transfer --num_classes 5

  Results and Visualizations

Our deep learning models provide insightful visualizations that help in understanding the decision-making process. Below are some of the results using techniques like Gradient-weighted Class Activation Mapping (Grad-CAM) and Guided Backpropagation.

Class-Discriminative Visualizations

• Grad-CAM: Highlights the regions in the image that are important for the model's decision.

• Guided Grad-CAM: Combines Grad-CAM with Guided Backpropagation to provide high-resolution class-discriminative visualizations.

High-Resolution Visualizations

• Guided Backpropagation: Visualizes the gradients flowing into the final convolutional layer of the model to understand which features activate the neurons the most.

The visualization techniques applied here serve to make the model's predictions interpretable to humans, by mapping the areas of interest that lead to a particular output. These insights can be invaluable for further research and improvements in the field of medical image analysis.

If you want to get a overview of results for this project please visit my Googld Drive: https://docs.google.com/presentation/d/17PHRd5ZhU06U6DkaT6uEixp0sByi4pTk/edit?usp=drive_link&ouid=102690024146339890642&rtpof=true&sd=true