Deep Red Lesion Classification for Early Screening of Diabetic Retinopathy

Diabetic retinopathy (DR) is an asymptotic and vision-threatening complication among working-age adults. To prevent blindness, a deep convolutional neural network (CNN) based diagnosis can help to classify less-discriminative small-sized red lesions in early screening of DR patients. However, training models with minimal data challenging task. Fine-tuning through transfer learning useful alternative, but performance degradation, overfitting, domain adaptation issues further demand architectural amendments effectively train models. Various pre-trained CNNs are fine-tuned on augmented set image patches. The best-performing ResNet50 model modified by introducing reinforced skip connections, global max-pooling layer, the sum-of-squared-error loss function. (DR-ResNet50) five public datasets found be better than state-of-the-art methods terms well-known metrics. highest scores (0.9851, 0.991, 0.9939, 0.0029, 0.9879, 0.9879) for sensitivity, specificity, AUC, accuracy, precision, F1-score, false-positive rate, Matthews’s correlation coefficient, kappa coefficient obtained within 95% confidence interval unseen test instances from e-Ophtha_MA. This high sensitivity low rate demonstrate worth proposed framework. It suitable due its performance, simplicity, robustness.