A novel convolutional neural network architecture for diabetic retinopathy screening

Diabetic retinopathy (DR) is a diabetes complication prevailing worldwide that may affect eye vision. Manual DR diagnosis is a tedious and complex task; thus, an automated approach is needed to cater to the need of people across the world. A lot of automated approaches have been developed in the recent past for DR classification. Most of these approaches are computationally expensive and involve a lot of pre-processing techniques or transfer learning-based algorithms. The present work focuses on developing a computationally-effective approach with high performance for DR classification. The present work proposes a novel lightweight convolutional neural network (CNN) based DR binary classification model using shortcut connections to reuse the features of previous convolution layers. The proposed model is computationally effective and requires 1.137M (Million) parameters, which is much less than the existing approaches. The experimental study is executed on the APTOS dataset. The parametric evaluation of the proposed model is also analysed with several transfer learning-based approaches. The model obtained 0.9754 (classification accuracy and sensitivity), 0.9666 (specificity), 0.9755 (precision), 0.9747 (F-1 score), 0.97 (AUC), and 0.9509 (kappa score) with APTOS. The cross-dataset validation has been performed on the IDRiD dataset. The proposed model achieved good and consistent performance on both APTOS and IDRiD datasets. (Abstract)