Retinal Disease Detection using Deep Neural Networks

EfficientNetB0 and ResNet50 used as baseline model with initial weights from ImageNet and will retrain all layers for this project.

Augmenting images, will have more data and make the training set become more regularize. In Keras 'ImageDataGenerator' has used to feed image to the model.

Instead of average pooling to try with max pooling.

Two callbacks were added to the training to save the best model on validation set and reduce learning rate during training.

Visualizing the heatmap/saliency/features to demonstrate what regions of interest contribute to Diabetic Retinopathy and Glaucoma, respectively.

Using the unlabeled data set in the 'test' folder augmenting the training data (Semi-Supervised learning) and reporting the variation in classification performance on test data set.

The training data set contains 3435 retinal images that represent multiple pathological disorders. The pathology classes and corresponding labels are included and each image can have more than one class category (multiple pathologies).

Initially the exploratory data analysis done and able to see the details about distribution on each label in training dataset. and also combination of labels in training dataset.

Since this Model is for specialist of diabetic retinopathy and glaucoma only, other disorders have put into label called 'other' and 'normal' label will remain as it is. So only four labels will contain in training dataset.

This is followed by distributing all these training dataset images into Train/Validation/Test datasets.

We augmented the split Train dataset images to have more data and make the training set become more regularize.

Defined two models with the baseline, both EfficientNetB0 and ResNet50 has used which is pretrained on ImageNet dataset. The classifier of model would be replaced with a new dense layer to make the output suit to this project.

The GradCAM heatmaps are generated to incorporate Explainable AI. After that the trained model has been saved into H5 file formats for the deployment purpose.