Abstract:
Diabetic Retinopathy, a common diabetes
complication causes damages to the blood vessels of light
sensitive tissues in the human retina. Due to the limitations in
the manual screening process, there exists a compelling
requirement of an automated approach for the Diabetic
Retinopathy screening which can be applied regularly and in
abundance in any kind of a healthcare environment. This paper
suggests a Deep Learning based automated approach to classify
retinal fundus images into five major severity levels while
focusing on achieving the optimal accuracy-efficiency balance in
performance. In the classification task, a lightweight
Convolutional Neural Network (CNN) model with only 6
convolutional layers was suggested to classify retinal fundus
images to five major severity levels. CNN refinements such as
Hyperparameter Tuning, Regularization and Data
Augmentation were applied to increase the model accuracy. The
suggested model achieved an Accuracy of 72.28%, a Sensitivity
of 71.12% and a Specificity of 93.1% for a testing dataset of 267
retinal fundus images from Kaggle and Messidor-2 datasets. By
comparing with four pre-trained CNN models VGG16,
ResNet50, InceptionV3 and Xception, it was observed that the
accuracy of the suggested model is slightly lesser than that of
VGG16 and ResNet50 models. However, the number of FLOPs
in the suggested model is 23 times lesser than VGG16 and 6
times lesser than ResNet50, indicating that the suggested model
is more efficient than the mentioned pre-trained models. The
accuracy of the suggested model can be further improved
without increasing the number of FLOPs by increasing the
number of training data samples
