Diabetic Retinopathy Detection using Transfer Learning from Pre-trained Convolutional Neural Network Models

1. We developed a novel predictive model based on Efficient Net b3 pre-trained Convolutional Neural Network for reliable classification of Diabetic Retinopathy into 5 Stages, namely: 0 – No DR, 1 – Mild DR, 2 – Moderate DR, 3 – Severe DR, 4 – Proliferate DR. These 5 stages are based on the severity of diabetic retinopathy. Since treatment varies based on the severity of diabetic retinopathy. So, if 5 stages are predicted appropriately then it helps doctor to diagnose the patient based on the stage.

2. We used three different pre-trained CNN models: VGG Net, Res Net (which are scaled by their depth to get higher versions), and Efficient Net (which uses an efficient compound scaling method that scales width, height, and resolution to get higher versions). We observed Efficient Net could classify the images into more than one class, unlike Res Net and VGG Net. The reason could be models like Res Net and VGG Net is not capable of extracting complex features from all the stages of diabetic retinopathy.

3. We considered quadratic weighted kappa as the primary evaluation metric because it considers the level of disagreement between the actual label and predicted label, unlike accuracy, F1-score, etc. When one class is detected with high probability and other classes aren’t detected metrics like accuracy can be higher because they do not consider the level of disagreement whereas the quadratic weighted kappa considers it and gives the low score or zero itself. Hence, quadratic weighted kappa is a reliable metric for this project.

4. We achieved a quadratic weighted kappa of 0.85 using Efficient Net b3 trained for 60 epochs on the resized version of a dataset containing digital fundus images of retina taken from Kaggle.