PURPOSE
In this work, we explore to use very deep convolutional neural network (CNN) for the automatic classification of diabetic retinopathy using color fundus image.
METHODS
We apply translation, stretching, rotation and flipping to the labeled dataset. The original number of labeled-frames is 3000, while after augmentations, 6000 frames with labels are used for the CNN training task. Several different CNN architectures have been proposed and tested. The architecture of our network contains 18 layers with parameters, consists of 12 convolutional layers, some of which followed by max-pooling layers, and two fully connected layers. For a given input, the networks output two probabilities sum to 1, one for each class (as our problem is a two class classification problem).
RESULTS
To make a comparison, the performance of CNN-based are compared to previous automatic classification attempts using the hand-crafted features, such as: hard exudates, red lesions, micro-aneurysms and blood vessel detection. We obtained an accuracy of 94.54% on our dataset using CNN, which rank as the highest with the comparison with previous handcrafted features-based classifiers.
CONCLUSION
With limited number of medical staff, an automated system can significantly decrease the tedious manual labor involved in diagnosing large quantities of retinal images. In this work, we explore the potential usage of CNN in retinal image classification. The results are encouraging and a clinical evaluation will be undertaken in order to be able to integrate the presented algorithm in a tool to diagnosis of diabetic retinopathy.