Convolutional Vision Transformer Based Automatic Grading of Diabetic Retinopathy Images

Diabetic Retinopathy (DR) is a condition which causes blindness among people who have diabetes. To preserve the patient's existing level of eyesight an early identification of DR is essentially required. This early DR identification has been aided by machine learning-based medical image analysis methodologies. In this paper, we have applied auto-cropping on raw images which removes the unwanted black background. After that, Ben Graham's method is applied with sigma=30 keeping the images in RGB format. Then, CLAHE technique is applied on those images to amplify their contrast followed by data augmentation. Thereafter, those pre-processed images fed to a Convolutional Vision Transformer (CvT), which surpassed Vision Transformer (ViT) in terms of performance and efficiency with fewer parameters and lower FLOPs. CvT combines the best aspects of Convolutional Neural Networks (CNNs) and ViT. Various experiments have been conducted using our methodology and it has been found that in all experiments ours has achieved better accuracy than existing ones on APTOS 2019 and Messidor-2 datasets. The highest accuracy achieved on the 20 % test set of APTOS is 86.9% and on Messidor-2 it is 75.43%.