Diabetic Retinopathy Classification Model CNN

This repository contains a classification model made by Matlab for Diabetic Retinopathy Detection.

Dataset

• For classification with 5 classes, I have installed a set of images from this Dataset (approximately 450MB).
• For classification with 3 classes, I used the same dataset as the previous one (with 5 classes). However, I combined "Mild nonproliferative diabetic retinopathy (NPDR)" and "Moderate NPDR" into 'Moderate', and "Severe NPDR" and "PDR" into 'Advanced'.

Architecture

This model is built using a pretrained model 'AlexNet', along with 3 additional layers.

The model uses the following optimization algorithm:

• Adam ('adam') for training.

The training settings are as follows:

• Mini-batch size (MiniBatchSize): 32
• Maximum number of epochs (MaxEpochs): 10
• Initial learning rate (InitialLearnRate): 0.0001
• Validation frequency (ValidationFrequency): Calculated based on the number of training images
• Displaying

Training Progress

For Model with 5 Classes

For Model with 3 Classes

Accuracy

• The accuracy for the model with 5 classes: 62,14% .
• The accuracy for the model with 3 classes: 73,79% .
• The accuracy is not perfect but it is good, especialy for the 3 classes.

Other Statistiques for Model with 5 Classes

Confusion Matrix

True Positive Rate - False Positive Rate

Other Statistiques for Model with 3 Classes

Confusion Matrix

True Positive Rate - False Positive Rate

License

This repository is licensed under the MIT License.