Automated Diabetic Retinopathy Detection Based on Binocular Siamese-Like Convolutional Neural Network
This paper aim to detect the diabetic disease identification using deep learning methods.
Besides a binocular model for the ?ve class DR detection task is also trained and evaluated to further prove the effectiveness of the binocular design. The result shows that, on a 10% validation set, the binocular model achieves a kappa score of 0.829 which is higher than that of existing non ensemble model. Finally the comparison between confusion matrices obtained through models with paired and unpaired inputs is performed and it demonstrates that the binocular architecture does improve the classi?cation performance.
For a deep learning model, the most important parts that should be focused on are data set, network architecture and training method. Before being used to train our model, fundus images data set obtained from public resources is preprocessed and augmented. The model accepts two fundus images corresponding to the left eye and right eye as inputs and then transmits them into the Siamese-like blocks. The information from two eyes is gathered into the fully-connected layer and ?nally the model will output the diagnosis result of each eye respectively.