A Deep Neural Network Method for Diabetic Retinopathy Detection by introducing a Novel Loss Function

Diabetic retinopathy (DR) is the primary cause of vision loss and even blindness among individuals who have had diabetes for a long time. As such, detecting DR in its early stages is the only way to avoid serious complications to the vision of these individuals. However, the current detection methods available are not only inefficient but prone to human error. Therefore, the present study developed a new strategy that improves the diagnostic and classification capabilities of a model for DR detection. This was accomplished by combining focal (FL) and dice loss (DL) functions into a custom loss function. As class imbalances are common in datasets of medical images, a total of 35126 retinal images were first pre-processed using K-means clustering and data augmentation to ensure that there were a sufficient number of images to equally represent each of the five stages of DR. The proposed hybrid loss function was then used to further overcome the class imbalance issue. The proposed hybrid loss function significantly improved the precision of the ResNet-50 model, with an accuracy of 95 to 99%. Therefore, the FL function decreased the inherent bias of the model toward the majority class while the DL function increased the model’s precision at a pixel level. As such, the modified ResNet-50 model was better able to identify minor changes in retinal images, even in the early stages of DR.