Diabetic-Retinopathy-Detection

Diabetic retinopathy (DR) is an eye disease that seriously impairs the vision of diabetic patients. In this project, we mainly developed a detection model based on deep convolutional neural networks. This model is based on IDRID and Kaggle EyePACS datasets, which are also augmented through Sample Pairing. After analysis using deep visualization, we selected various architectures, and the model reached an binary accuracy of nearly 90% through ensemble learning. In the end, we concluded that the size of the dataset is currently the most critical factor that determines the detection ability.

Content

• Input pipeline with TFRecord of dataset IDRID und EyePACS;
• Various CNN models (VGG, Inception, SEResNeXt, RepVGG) built by ourselves;
• Different kinds of metrics (binary/multi-accuracy, binary/multi-confusion-matrix, precision and recall) and corresponding evaluations;
• Image preprocessing (cutting black edges, resizing, normalization) and different kinds of data augmentation (random rotation, random shear, random crop, random flip, random change of contrast/saturation/hue);
• Deep visualization (Grad-CAM, Guided Backpropagation, Guided Grad-CAM, Integrated Gradients);
• Transfer learning models (Densenet and Efficientnet);
• Ensemble learning models;
• Freely selectable detection models (binary classification, multi-classification, regression)
• Sample Pairing (Data augmentation based on image mixing)

How to run the code

• Download the original data file and modify the parameter load.data_dir in config.gin and tuning_config.gin to your corresponding data directory
  • IDRID dataset: https://ieee-dataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid
  • Kaggle Challenge Dataset provided by EyePACS: https://www.kaggle.com/c/diabetic-retinopathy-detection/data
• If you want to start training a new model, run main.py directly (To evaluate the model, just change the parameter train in main.py to False)
• If you want to start fine-tuning of the model, run tune.py directly
• If you want to use ensemble learning, run directly ensemble_learning.py, but make sure that there are already well trained models to be fused (you need to give names and types of these models to the parameter model_list in ensemble_learning.py)

Note: If you want to change the type and parameters of the model, you need to modify the parameters in config.gin, tuning_config.gin, main.py and tune.py

Note: The default parameters are: VGG16 regression model for main.py; EfficientNet regression model for tune.py

Results

Single model results:

Model Name	Model Type	Binary Accuracy	5-Class Accuracy
VGG16	Regression	85.44%	53.40%
Inception	Regression	83.50%	49.51%
SEResNeXt	Regression	82.52%	51.46%
DenseNet	Regression	85.44%	57.28%
DenseNet	5-Class Classification	85.44%	62.14%
EfficientNet	Regression	87.38%	55.34%
EfficientNet	5-Class Classification	86.41%	54.39%
RepVGG	Regression	82.52%	45.63%

Ensemble learning results:

Metrics	Value
Binary Accuracy	88.35%
Binary Balanced Accuracy Score	89.62%
Precision	96.43%
Recall	84.38%
F1 Score	89.90%
5-class Accuracy	55.40%
5-class Balanced Accuracy Score	44.67%

EfficientNet Regression Model achieves the best binary classification performance, when DenseNet 5-Class Classification Model has the highest 5-Class accuracy.

Ensemble learning can effectively improve the overall performance of diabetic retinopathy detection.