Automated Grading in Diabetic Retinopathy Using Image Processing and Modified EfficientNet.

We present our approach in achieving the Quadratic Weighted Kappa (QWK) score of 0.90 on the retinal image dataset from the APTOS 2019 Blindness Detection Kaggle challenge. We analysed various image preprocessing techniques then classified the images with a modified EfficientNet deep learning model. Our image preprocessing techniques helped to bring out the cell loss to the retina, highlight blood vessels, and centering the retina. We found that subtracting the average local color using a Gaussian mask was the most effective preprocessing technique, improving the QWK score by 0.03. We modified the EfficientNet-B5 network with the Batch Normalization layers replaced with Group Normalization and trained the network using the Rectified Adam (RAdam) optimizer. Group Normalization was found to do better for the batch size of 4 and RAdam trained the network better than Adam. This led to an increment of the QWK score by 0.02.