Soft attention with Convolutional Neural Network for Grading Diabetic Retinopathy

Diabetes is one of the most common illnesses, and diabetes causes Diabetic Retinopathy (DR), which can lead to vision problems and blindness if not detected early. Early detection of DR is possible with the help of a computer-aided diagnostic (CAD) system, thus preventing many people from losing their eyesight. In order to classify fundus images into five categories, this article makes use of soft attention as feature refinement in convolutional neural networks (CNN). The proposed framework initially utilizes Contrast Limited Adaptive Histogram Equalization (CLAHE) as a pre-processing technique for improving the contrast level of the fundus images and then applied concatenation based CNN architecture coupled with soft attention for feature extraction and classification. The data has been balanced using oversampling so that each DR grade contains equal number of samples during the training process. The experimental results reveal that, unlike current methods, the proposed model detects all the five stages of DR.